the application of acoustic emission monitoring - the UWA

TTHHEE AAPPPPLLIICCAATTIIOONN OOFF AACCOOUUSSTTIICC EEMMIISSSSIIOONN MMOONNIITTOORRIINNGG

TTOO TTHHEE DDEETTEECCTTIIOONN OOFF FFLLOOWW CCOONNDDIITTIIOONNSS

IINN CCEENNTTRRIIFFUUGGAALL PPUUMMPPSS

JJooaannnnaa ZZooffiiaa SSiikkoorrsskkaa

B.E. (Hons)

This thesis is submitted for the degree of Doctor of Philosophy

at the University of Western Australia

School of Mechanical Engineering

April 2006

SSTTAATTEEMMEENNTT OOFF OORRIIGGIINNAALLIITTYY

To the best of my knowledge and belief, all the material presented in this thesis, except where

otherwise referenced, is my own original work and has not been published previously for an award

of any other degree or diploma at any other university.

Joanna Zofia Sikorska

August 2005

ii

SSTTAATTEEMMEENNTTSS FFRROOMM CCOO--AAUUTTHHOORRSS

I, Mr Paul Kelly as co-author of “Development of an AE data management and analysis system”, agree for

Joanna Sikorska to include this work as part of her PhD thesis. My contribution to this paper

included:

• Setting up and configuring the SQL server hardware and software;

• Providing advice on the overall database architecture;

• Writing advanced Visual Basic and TransactSQL routines used to populate data tables and

extract data for subsequent analysis routines supplied by Ms Sikorska;

• Troubleshooting hardware and software problems.

Sections of the publication that were written by me, or relate primarily to the aforementioned work

listed, have been clearly identified in Chapter 5 by my initials in the section’s header. Ms Sikorska

was primarily responsible for all remaining sections.

Paul John Kelly,

I, Dr Melinda Hodkiewicz, as co-author of “Comparison of acoustic emission, vibration and dynamic pressure

measurements for detecting change in flow conditions on a centrifugal pump”, agree for Joanna Sikorska to

include this work as part of her PhD thesis. In this publication I was responsible for collecting and

analysing vibration and dynamic pressure data, whilst Ms Sikorska completed all other work therein.

Melinda R. Hodkiewicz,

iii

AACCKKNNOOWWLLEEDDGGEEMMEENNTTSS

This journey would never have begun, or for that matter reached its end, if it was not for the

generosity and help of many individuals and organisations along the way.

Firstly, on a professional note, thankyou to Don, Mick, Brent and everyone else at Clyde for

teaching me what it really means to be an engineer; to Tadashi Kataoka, Mark Friesel, Bob Reuben

and Phil Cole for much needed early guidance and willingness to reply to numerous emails sent by

a stranger on the other side of the world; to SEDO, MERIWA, Blakers Pump Engineers, John

Crane, CSC and Imes Group for providing funding, equipment or access to facilities; to Dr Mike

Lowe at Imperial College for determining DISPERSE mode shapes used in the waveguide work; to

Jon Romaine at PAC for spending so many nights and weekends helping me unravel data-file

formats and commission new equipment; to Claire, Karl and Emma for your helpful thoughts and

insights; to the late Steve “Grumpy” Armitt for always being happy to see me, and for constructing

all things mechanical; to Rob and Peter for not yelling every time I blew something up; to Angus

for your invaluable stylistic and linguistic input; to the late Professor Michael Norton for always

believing in me and reminding me to do the same; to Professor Jie Pan for really caring; and to

Melfort, Kev, Derek, Sarah and Jim at Imes for giving me the time and space to finish this.

On a more personal note, thankyou to my parents for teaching me the meaning of hard work,

integrity and sheer pig-headedness, in which this project was so firmly grounded… you are my

inspiration; to my sister, for reminding me that I am my own worst critic and for an endless supply

of good karma; to my Grandmother for forgiving when I didn’t call; to Melinda for sharing this

journey with me, endless discussions about work and life, and for reading countless drafts; and to

Jenny and Gia for still speaking to me after so many years of neglect.

And finally, to my loving husband Paul… For your help in programming, electronics design,

database development and general IT support: my deepest gratitude, admiration and respect. As

for your love, generosity, support, patience, resilience, forgiveness and endless cups of tea, no

words will ever express the true extent of my appreciation. I owe this to you.

iv

TTHHEESSIISS AABBSSTTRRAACCTT

Centrifugal pumps are the most prevalent, electrically powered rotating machines used today. Each

pump is designed to deliver fluid of a given flow rate at a certain pressure. The point at which

electrical energy is converted most efficiently into increased pressure is known as the Best

Efficiency Point. For a variety of reasons, pumps often operate away from this point (intentionally

or otherwise), which not only reduces efficiency, but also increases the likelihood of premature

component failure.

Acoustic emissions (AE) are high frequency elastic waves, in the range of 20-2000kHz, released

when a material undergoes localised plastic deformation. Acoustic emission testing is the process

of measuring and analysing these stress waves in an attempt to diagnose the nature and severity of

the underlying fault. AE sensors mounted on the surface of a machine or structure also detect any

stress waves generated within the fluid being transmitted through to the structure.

Unfortunately, attempts to detect incipient component faults in centrifugal pumps using acoustic

emission analysis have been complicated by the sensitivity of AE to a pump’s operating state.

Therefore, the aim of this thesis was to determine how acoustic emission monitoring could be used

to identify the hydraulic conditions within a pump. Data was collected during performance tests

from a variety of small end-suction pumps and from one much larger double-suction pump.

A system was developed to collect, process and analyse any number of AE features (be they related

to discrete AE events, or due to the continuous background AE level) from continuously operating

equipment. Based on a relational database, this collated results and initiated processing routines,

independently from software and hardware used for data gathering. Processing methods included

traditional AE analysis techniques as well as others, such as octave band analysis and wavelet

decomposition, more commonly applied to vibration monitoring.

To help identify relationships between AE features and hydraulic conditions, resulting features were

averaged, trended and 90% confidence limits for these point estimates determined. This facilitated

identification of which features were most useful for quantifying changes in hydraulic conditions.

To reduce the effects and consequences of noise contamination, a variety of techniques, including

averaging, Swansong hit filtering and wavelet denoising were applied to identify infiltration at the

time of data gathering, and/or to reduce the effects thereafter. Although most unwanted signal

sources were appropriately managed, no technique was identified that could remove periodic

broadband impulsive bursts created by variable frequency drives.

v

Waveguides are often required to access remote pump components or hot surfaces. Consequently,

the effects of short cylindrical rods on the temporal and frequency characteristics of received

acoustic emission signals were analysed. A variety of different waveguide materials, dimensions and

face angles were tested, including the use of a pointed source end instead of adequate couplant.

Findings indicate that the use of short, solid, cylindrical rods should be avoided. Where absolutely

necessary, rods should be as narrow and short as possible, made from very high acoustic materials

with inherent damping (eg. ceramics) and have flat ends attached with a good couplant.

Results from pump performance tests showed that, particularly in large pumps where hydraulic

changes are significant, various AE energy features can be very effective for detecting changes in

flow and movement away from best efficiency point. By measuring AE signals from both suction

and discharge flanges, automating the identification process in large pumps should be readily

achievable. Unfortunately, results from smaller pumps were less conclusive, particularly at low

flows, probably due to the relatively small changes in hydraulic energy across the range of flows,

and consequent sensitivity to the testing process. However, even in these pumps consistent

patterns in hit energies were observed resulting in the conclusion that low to medium flows in

centrifugal pumps are typified by a very large number of very low energy (VLE) events. These

decrease in number and increase in energy as flow approaches BEP and/or is reduced to very low

flows. High flows above BEP are marked by an absence of these VLE events, with bursts having

significantly higher energies and spread over a much greater range. Unfortunately, these VLE

events are too small to affect averaged trends, indicating that further work on a suitable filter is

required.

vi

TTAABBLLEE OOFF CCOONNTTEENNTTSS

1. INTRODUCTION..................................................................... 1-1

1.1 Prelude ..................................................................................................................... 1-1

1.2 Background.............................................................................................................. 1-2

1.3 Thesis Structure ....................................................................................................... 1-6

1.3.1 Overall layout ................................................................................................................... 1-6

1.3.2 Chapter topics .................................................................................................................. 1-7

2. REVIEW OF AE TECHNOLOGY ........................................... 2-1

2.1 Introduction .............................................................................................................2-2

2.2 AE research ..............................................................................................................2-3

2.3 Existing Technology................................................................................................2-6

2.3.1 Shock pulse – SPM Instruments ................................................................................... 2-6

2.3.2 CSI PeakVue..................................................................................................................... 2-9

2.3.3 Spike Energy – Entek ................................................................................................... 2-11

2.3.4 Stress Wave Analysis – Swantech Ltd ........................................................................ 2-12

2.3.5 SEE – SKF ..................................................................................................................... 2-14

2.3.6 Acoustic Emission – Holroyd Instruments............................................................... 2-15

2.4 Comparisons and Discussion ................................................................................ 2-16

2.4.1 Source Characteristics ................................................................................................... 2-16

2.4.2 Transducers .................................................................................................................... 2-17

2.4.3 Gain and Threshold Level............................................................................................ 2-18

2.4.4 Differences in Functionality......................................................................................... 2-18

2.5 Conclusions............................................................................................................ 2-19

2.6 Postscript (NC) ...................................................................................................... 2-20

vii

3. AE SIGNAL PROCESSING ..................................................... 3-1

3.1 Introduction ............................................................................................................. 3-1

3.2 Hit Parameters.........................................................................................................3-3

3.3 Continuous Signal Parameters.................................................................................3-4

3.3.1 Traditional Time-dependant Parameters ..................................................................... 3-4

3.3.2 Calculating Statistical Parameters from Raw Waveforms ......................................... 3-7

3.3.3 Frequency Parameters..................................................................................................... 3-9

3.3.4 Joint-time-frequency analysis ....................................................................................... 3-11

3.4 Ensemble Statistics................................................................................................ 3-13

3.4.1 Confidence limits ........................................................................................................... 3-13

4. DENOISING AE DATA ............................................................ 4-1

4.1 Introduction ............................................................................................................. 4-1

4.2 Hit Filtering .............................................................................................................4-2

4.3 Frequency Filters .....................................................................................................4-5

4.4 Averaging .................................................................................................................4-6

4.5 Wavelet Denoising ...................................................................................................4-9

4.5.1 Continuous Wavelets ...................................................................................................... 4-9

4.5.2 Discrete Wavelets .......................................................................................................... 4-10

4.6 Conclusions............................................................................................................ 4-16

5. DATA MANAGEMENT........................................................... 5-1

5.1 Introduction ............................................................................................................. 5-1

5.2 Database Theory (PK) .............................................................................................5-4

5.3 Hardware Setup Used ..............................................................................................5-5

5.4 Implementation Details ...........................................................................................5-6

5.4.1 Metadata Input ................................................................................................................. 5-9

viii

5.4.2 Primary data input ......................................................................................................... 5-11

5.4.3 Secondary data input ..................................................................................................... 5-11

5.4.4 Defining session subsets............................................................................................... 5-13

5.4.5 Further analysis and graphing...................................................................................... 5-13

5.5 Examples of Results Extracted from AEData ....................................................... 5-14

5.6 Optimising Performance (PK)............................................................................... 5-15

5.6.1 Hardware......................................................................................................................... 5-15

5.6.2 Indexes ............................................................................................................................ 5-16

5.6.3 Stored Procedures.......................................................................................................... 5-16

5.7 Conclusions............................................................................................................ 5-16

5.8 Postscript ............................................................................................................... 5-17

6. EFFECT OF WAVEGUIDES - PART ONE............................. 6-1

6.1 Relevant Ultrasonic Theory .....................................................................................6-2

6.2 Experimental Method ..............................................................................................6-4

6.3 Results and Observations ........................................................................................6-9

6.3.1 Burst Profiles .................................................................................................................... 6-9

6.3.2 Amplitude and Energy.................................................................................................... 6-9

6.3.3 Rise time, Duration and Counts.................................................................................. 6-13

6.4 Discussion.............................................................................................................. 6-14

6.4.1 Effect of Material........................................................................................................... 6-14

6.4.2 Effect of Length, Diameter and Face angle .............................................................. 6-15

6.4.3 Effect of Pointed Source End ..................................................................................... 6-15

6.5 Conclusions............................................................................................................ 6-15

7. EFFECT OF WAVEGUIDES - PART TWO............................. 7-1

7.1 Relevant Theory .......................................................................................................7-2

7.1.1 Effect of Finite Ends ...................................................................................................... 7-2

ix

7.1.2 Frequency Signal Processing.......................................................................................... 7-2

7.1.3 Advanced Joint-Time-Frequency Techniques ............................................................ 7-3

7.1.4 Modal AE.......................................................................................................................... 7-3

7.2 Experimental Method ..............................................................................................7-4

7.2.1 Additional Post-Processing ............................................................................................ 7-4

7.2.2 Modal Analysis ................................................................................................................. 7-8

7.3 Frequency Results....................................................................................................7-8

7.3.1 Flat Waveguides ............................................................................................................... 7-8

7.3.2 Angled Waveguides ....................................................................................................... 7-10

7.3.3 Pointed Waveguides ...................................................................................................... 7-10

7.3.4 Modal data ...................................................................................................................... 7-10

7.4 Discussion.............................................................................................................. 7-12

7.4.1 Effect of material ........................................................................................................... 7-12

7.4.2 Effect of length and diameter...................................................................................... 7-17

7.4.3 Effect of face angle........................................................................................................ 7-17

7.4.4 Effect of a pointed source end.................................................................................... 7-17

7.5 Conclusions............................................................................................................ 7-17

8. FLOW MONITORING OF A DOUBLE-SUCTION PUMP..8-1

8.1 Introduction ............................................................................................................. 8-1

8.2 Detection of Adverse Hydraulic Conditions............................................................8-3

8.2.1 Stable Cavitation (low NPSHA).................................................................................... 8-3

8.2.2 Recirculation..................................................................................................................... 8-4

8.2.3 Identifying BEP ............................................................................................................... 8-5

8.3 Experimental Setup .................................................................................................8-5

8.3.1 Pump Rig .......................................................................................................................... 8-5

8.3.2 Vibration and Pressure (MH) ........................................................................................ 8-6

8.3.3 AE Measurement ............................................................................................................. 8-6

x

8.4 Results and discussion.............................................................................................8-7

8.4.1 Dynamic Pressure and Vibration Results (MH) ......................................................... 8-7

8.4.2 AE Results ........................................................................................................................ 8-8

8.5 Conclusions and Recommendations ..................................................................... 8-12

9. FLOW MONITORING OF END-SUCTION PUMPS ........... 9-1

9.1 Introduction ............................................................................................................. 9-1

9.2 Experimental Setup .................................................................................................9-2

9.2.1 DAQ Hardware and Software ....................................................................................... 9-2

9.2.2 Sensors............................................................................................................................... 9-3

9.3 Observations ............................................................................................................9-4

9.3.1 Test 2 ................................................................................................................................. 9-4

9.3.2 Test 4 ................................................................................................................................. 9-4

9.3.3 Test 5 ................................................................................................................................. 9-4

9.3.4 Test 6 ................................................................................................................................. 9-6

9.3.5 Test 7 ................................................................................................................................. 9-6

9.3.6 Test 8 ................................................................................................................................. 9-6

9.4 AE Results................................................................................................................9-9

9.4.1 Changes at High flow...................................................................................................... 9-9

9.4.2 Changes at Low Flow.................................................................................................... 9-12

9.4.3 Results at BEP................................................................................................................ 9-13

9.4.4 Other observations........................................................................................................ 9-13

9.5 Discussion & conclusions...................................................................................... 9-14

9.5.1 Detection of changes in hydraulic conditions........................................................... 9-14

9.5.2 Pump Testing Process .................................................................................................. 9-14

10. CONCLUSIONS & RECOMMENDATIONS ....................... 10-1

10.1 Discussion...............................................................................................................10-1

xi

10.2 Summary of Findings............................................................................................ 10-10

10.3 Recommendations for Future Work ..................................................................... 10-11

10.4 Final words............................................................................................................ 10-11

11. REFERENCES .........................................................................11-1

APPENDICES

APPENDIX A: Derivation of Statistics

APPENDIX B: In-house AE Hardware

APPENDIX C: Signal Processing Algorithms

APPENDIX D: AEData Tables

APPENDIX E: Pump Information

APPENDIX F: An Example of Seal Monitoring Results

xii

LLIISSTT OOFF FFIIGGUURREESS

Figure 1-1: Impeller-volute terminology. 1-3

Figure 1-2: Cross section of a typical end suction pump. 1-3

Figure 1-3: Pump, system and efficiency curves. 1-4

Figure 1-4: Onset of adverse conditions caused by operating away from BEP[59]. 1-6

Figure 2-1: Types of AE signals [49]. 2-3

Figure 2-2: Traditional features of an AE signal. 2-3

Figure 2-3: SPM Processing. 2-8

Figure 2-4: Damping effect due to sensor location (from [112]). 2-8

Figure 2-5: Process of extracting a traditional envelope spectrum (from [121]). 2-9

Figure 2-6: PeakVue Processing. 2-10

Figure 2-7: Spike Energy Processing. 2-11

Figure 2-8: Typical Spike Energy signals (from [129]). 2-12

Figure 2-9: SWANTech Processing. 2-13

Figure 2-10: SWE in a centrifugal pump (from [16]). 2-13

Figure 2-11: SEE Processing. 2-14

Figure 2-12: Holroyd Instruments Signal Processing. 2-15

Figure 3-1: PCI2 board block diagram (from [6]). 3-2

Figure 3-2: (A) A signal, (B) its envelope calculated using a Hilbert transform, and (C)

the resulting FFT. 3-5

Figure 4-1: Different amplitude-duration characteristics of AE hits. Data has been

processed by a Swansong II filter but probable EMI noise still remains.

(PUMAnalysis software courtesy of Imes Group Ltd.) 4-3

Figure 4-2: Improved trending is obtained by using the telltale-hit criteria to segregate

bursts from residual changes in the continuous emission level. (A) Shows

the normalised AE feature versus flow determined from the complete

xiii

unfiltered dataset, whilst (B) shows the same normalised feature from a

reduced dataset determined by removing telltale hits. (Vertical lines indicate

90% confidence limits, which increase because the number of hits used to

determine the sample mean reduce. Very few hits remain at normalised

flows between 0.9 and 1.1.) 4-5

Figure 4-3: Effect of frequency and wavelet filters on pump AE signal corrupted by a

100Hz frequency sinusoidal signal and high frequency impulsive noise, latter

emanating from a VFD. (a) Original time signal, (b) FFT, (c) 10kHz –

1000kHz bandpass filtered and (d) 10kHz – 200kHz bandpass filtered. 4-7

Figure 4-4: Effect of averaging on identification of harmonic noise elements hidden

amongst a broadband continuous AE pump signal with (A) 8192 samples

and no averaging, (B) 8192 samples and 8 averages, (C) 8192 samples and

64 averages, and (D) 1024 samples and 64 averages. Sampling rate was

5MHz. 4-8

Figure 4-5: Examples of changes in spectral composition that can be observed in a

typical CWT plot (Morlet wavelet). 4-9

Figure 4-6: Effects on Haar wavelet denoising to 12 levels on two different types of

noise: (A-B) EMI/RFI; (C-D) VFD noise. Due to the impulsive nature of

the noise signals, wavelet denoising is not particularly effective in either case. 4-11

Figure 4-7: Effect of wavelet type on denoised signal. (A) Original waveform, (B)

Denoised with a Haar wavelet, (C) Denoised with a Daubechies 6 wavelet

and (D) Denoised with a Biorthogonal 1-3 wavelet. 4-13

Figure 4-8: Effect of threshold estimating algorithm on low signal to noise ratio signal

(VFD noise): (A) original signal; (B) Using Stein’s unbiased risk estimate,

(C) Using √(2logN), (D) using an alternative algorithm from [45]. 4-14

Figure 4-9: Effect of filtering on the final denoised signal. (A) Original signal, (B) After

denoising by Haar wavelet; (C) 100-1000kHz digital filter applied prior to

denoising by Haar wavelet. 4-15

Figure 4-10: (A) Overall signal RMS levels as they change with flow and (B) separated

continuous and discrete RMS parts as they change with flow. 4-16

Figure 5-1: Physical layout of the implemented DBMAS. 5-6

Figure 5-2: Data flow schematic. 5-7

Figure 5-3: Architecture diagram of AEData. 5-8

xiv

Figure 5-4: AEDMAS layout with main tables, relationships, primary keys (PK) and

foreign keys (FK). Primary, secondary and metadata tables are designated

by (P), (S) or (M) respectively after the table name. One-to-many

relationships are designated by the symbols 1 and ∝ respectively. 5-8

Figure 5-5: Session data entry form. 5-9

Figure 5-6: Process Files Form extracts information from DTA files and analyses WFS

files. 5-10

Figure 5-7: Process of extracting and/or processing data. 5-12

Figure 5-8: Relating timestamps to test conditions. 5-13

Figure 5-9: Raw AE data and corresponding secondary data obtained for a pump flow

test. 5-15

Figure 6-1: First two longitudinal modes showing normalised phase and group velocities

(latter are shown as dashed lines) for an 8-mm SS316 infinitely long rod

waveguide. Wave speed is normalised with respect to transverse wave

speed, given in equation (6-1). 6-5

Figure 6-2: First four normalised torsional phase velocities for 8-mm diameter, infinitely

long waveguides of different materials. Wave speed is normalised with

respect to transverse wave speed, given in equation 6-3. 6-5

Figure 6-3: (A) Angled 43mm waveguides for MS, Delrin, SS316 and Alumina (Clockwise

from top left). (B) Pointed 43mm waveguides (NC). 6-6

Figure 6-4: Waveguide holder. 6-8

Figure 6-5: Waveguide holder setup (NC). 6-9

Figure 6-6: Time waveforms for flat waveguide samples. Different y-scales are used to

facilitate better appreciation of waveform profiles. Time scales are identical.

(43x8mm samples) 6-11

Figure 6-7: Amplitude for different flat-faced waveguides. 6-11

Figure 6-8: Amplitude versus face angle for all sets of angled samples. 6-12

Figure 6-9: Effect on amplitude by using a pointed waveguide. 6-12

Figure 6-10: Effect of pointed end on rise-times. 6-13

Figure 6-11: Material effects on duration. 6-14

Figure 7-1: Wavelet transforms at different sampling rates (Face-to-face signals). 7-6

xv

Figure 7-2: Difference in resolution between (A) STFT in dBAE and (B) CWT, log scale

(Delrin 51-mm sample). Computing times were 594ms and 15141ms

respectively. 7-6

Figure 7-3: Comparison of results obtained from the custom written programs in (A)

LabVIEW and the (B) AGU-Vallen wavelet tool. LabVIEW parameters:

600 scales, 100 samples per time increment, ~30s processing time. AGU-

Vallen parameters: frequency resolution of 5kHz, 1000 wavelet samples,

~60s processing time. 7-7

Figure 7-4: Amplitude (in dB) - frequency (kHz) spectra of averaged time signals. (Face-

to-face spectrum shown as a shadow on each graph.) 7-9

Figure 7-5: Increasing rod length decreases frequency of resonant peaks. 1st resonant

peak match values given in Table 7-1 (aluminium samples). 7-9

Figure 7-6: Effects of different types of points on the resulting frequency spectrum (43-

mm long, 8-mm diameter aluminium samples). 7-11

Figure 7-7: Effect of points on wavelet spectrograms. 7-11

Figure 7-8: CWT from 43-mm long, 60˚ waveguides (Data sampled at 10 MHz). 7-12

Figure 7-9: Effect of face angles of FFT spectra (43-mm long SS316 samples). 7-13

Figure 7-10:Changes in wavelet spectrograms for changing face angles (30-mm

aluminium samples). 7-13

Figure 7-11: Changes due to different face angles for 30-mm SS316 waveguides do not

correspond with flexural modes (only one shown for clarity). 7-14

Figure 7-12: Wavelet spectrograms with longitudinal mode frequencies. 43 mm x 8 mm

waveguides (Phase modes are shown as dotted lines, group modes as solid

lines). 7-15

Figure 7-13: Mode separation is clearly evident in Delrin waveguides. 7-16

Figure 7-14: Standard amplitude spectrum showing changes in location of 2nd resonance

peak. 7-16

Figure 8-1: (A) Pump, system and efficiency curves. (B) Onset of adverse conditions

caused by operating away from BEP[59]. 8-2

Figure 8-2: Normalised flow curves for two speeds tested 8-5

Figure 8-3: Changes in dynamic pressure with flow. 8-7

Figure 8-4: Discharge Flange Vibrations. 8-7

xvi

Figure 8-5: Results of axial vibration results showing changes in (A) RMS and (B)

Kurtosis. 8-8

Figure 8-6: Energy of hits (y axis) versus time (x axis) versus number of hits (z axis) as

flow is reduced from full flow to low flow at two different speeds for (A)

suction AE and (B) discharge AE. Total energy is superimposed over each

graph. Flows as a ratio of BEP for each test condition are also shown. 8-9

Figure 8-7: Most mean parameters increase at low flows for all signals, whilst only Pump

and Discharge signals increase noticeably at high flows. (Results from test at

997 rpm shown.) 8-10

Figure 8-8: Effect on hit rate (hits per second) by denoising processes: (A) Swansong

filtering and (B) Wavelets. Missing points signify no hits detected at that

condition. Results have not been normalised. Difference in magnitude is

due to (A) being determined from hit data sets and (B) determined from

waveform streaming files collected periodically. 8-10

Figure 8-9: Effect of Swansong II filter on rise time signals. (Discharge AE only

shown, for test at 997 rpm.) 8-11

Figure 8-10: Changes across frequency octave bands for (A) suction AE and (B)

discharge AE. 8-13

Figure 8-11: Changes across wavelet bands for (A) suction AE and (B) discharge AE. 8-13

Figure 8-12: Selected normalised parameters show little sensitivity to changes in speed:

(A) Energy in 2nd Wavelet Band, Suction AE signal and (B) Background

RMS, Discharge AE signal. 8-13

Figure 9-1: (A) Total head and (B) efficiency curves for Pump 2. 9-5






Figure 9-7: Test 8 at 2236rpm. Green line shows raw voltage from flowmeter whilst

black line shows changes in Absolute Energy for PumpAE. At 0.5-0.9BEP,

flow does not remain constant (the control valve was kept in the same

position). 9-8

xvii

Figure 9-8: Hit intensity plots for (A) Pump 6 at 2263rpm, and (B) Pump 7 at 2263rpm.

These show number of hits (z axis) versus PAC Energy (y axis) versus time

(x axis). The superimposed black lines shows time averaged Absolute

Energy. Red lines shows changes in flow (also labelled as %BEP). To

emphasise low energy changes, small numbers of very high-energy events at

low flow are not shown. 9-10

Figure 9-9: Normalised RMS of PumpAE for (A) Group 1 pumps and (B) Group 2

Pumps. 9-11

Figure 9-10: Normalised RMS of PumpAE after wavelet denoising for (A) Group 1

pumps and (B) Group 2 Pumps. 9-11

Figure 9-11: (A) Normalised Peak and (B) Normalised Energy in 12th Wavelet Band for

Group 2 Pumps. 9-11

xviii

LLIISSTT OOFF TTAABBLLEESS

Table 2-1: Comparison of various Stress-wave technologies. 2-22

Table 3-1: Octave and 1/3-octave band frequencies. 3-10

Table 5-1: Characteristics of conventional and data driven processing 5-3

Table 6-1: Approximate acoustic properties of test materials used in theoretical

calculations. 6-6

Table 6-2: Results of rough amplitude calculations. 6-11

Table 7-1: Approximate natural frequencies of first longitudinal mode in cylindrical

rods with differing lengths. 7-2

Table 9-1: Selected data acquisition settings. 9-3

Table 9-2: Sensor data collected from each test. 9-3

xix

NNOOMMEENNCCLLAATTUURREE

AE Acoustic Emission

AET Acoustic Emission Testing

ANN Artificial Neural Network

ANSI American National Standards Institute (governing standards for

chemical process pumps)

API American Petroleum Institute (governing standards for petroleum

process pumps)

AveDBurstDuration Average duration of all bursts detected after denoising

AveDBurstEnergy Average of all energies from bursts detected after wavelet

denoising

AveDBurstPeak Average all the peaks detected in a datafile after wavelet denoising

BA Vibration signals measured on the bearing housing the axial

direction

BgndRMS RMS of the signal components that have been removed by

denoising

BEP Best efficiency point (a flow)

BR Vibration signals measured on the bearing housing in the radial

direction

BV Vibration signals measured on the bearing housing in the vertical

direction

CBM Condition Based Maintenance

CF Crest factor

CM Condition monitoring

COTS Commercial-Off-The-Shelf

DenRMS RMS of the signal after denoising

DEvents Number of events per second after denoising

xx

DPR2 Background RMS / Max DBurst Peak

E Energy

FaceAE AE signals measured on a sensor attached to a waveguide which in

turn is attached to the back of the stationary face and protrudes

through the gland plate

FFT Fast Fourier Transform

GlandAE AE signals measured on the gland plate.

GlandVib Vibration signals measured on the seal stuffing box

H,h Head of the pump, normally in meters, but sometimes in feet

HI Hydraulic Institute

IFD Incipient failure detection

K Kurtosis

MaxDBurstPeak Highest peak detected in a datafile after denoising

MaxDBurstEnergy Maximum energy of bursts detected after denoising

N, n Number of samples

NPSH Net positive suction head

NPSHA Net positive suction head available

NPSHMR Net positive suction head margin ratio

NPSHR Net positive suction head required

P Power

PeakFreq Amplitude of highest peak in FFT

p∞ Static reference pressure of the fluid, or total inlet pressure

pv Equilibrium vapour pressure of the fluid

PumpAE AE signals measured on the pump casing.

PumpVib Vibration signals measured on pump casing

RMS Root mean square

SNR Signal to noise ratio

SE Suction Energy

SflangeAE AE signals measured on the suction nozzle flange of the pump

xxi

sg, SG Specific gravity

SS Specific Speed

SSS or NSS Suction Specific Speed

SpipeVib Vibration signals measured on suction pipe

SuctionAE AE signals measured on the suction pipe.

STD Standard deviation

U∞ Reference speed (usually inlet relative velocity)

UWA University of Western Australia

VFD Variable frequency drive, synonymous with VSD

VSD Variable speed drive

x[n] Array of length n, usually containing the scaled and digitized AE

time signal

x[i] Individual element of the array x[n], where 0<i<n

ρ Fluid density

μ Sample Mean

σ Sample Standard deviation

σ2 Sample Variance

64kPower Normalised power in 63k third octave band


102kPower Normalised power in 101.6k third octave band




256kPower Normalised power in 256k octave band

322kPower Normalised power in 322.5k octave band




xxii

Progress begins when we stop thinking

a problem is difficult

and start believing it is merely complex.

P.J. Kelly

xxiii

1. IINNTTRROODDUUCCTTIIOONN11

1.1 PRELUDE

This project was initiated by the desire to solve a problem. As a maintenance engineer for an oil

refinery, once a month I found myself hopelessly standing by as a centrifugal pump released hot

bitumen onto the plant floor when its mechanical seal failed on start up. Unfortunately, failure did

not occur every time the batch plant started up, and countless investigations had not uncovered any

conclusive evidence of the cause or the operating conditions that preceded the seal’s demise.

Consequently, this PhD began with the intention of developing a technique for monitoring

mechanical seals in service, and to determine what condition indicators preceded failure. Acoustic

emission (AE) was identified as the most appropriate technology for the task.

As the project continued, it became apparent that the following problems had to be solved before a

mechanical seal monitoring system based on AE, could ever be developed for commercial

application.

(a) Commercially available AE systems have been designed for structural monitoring

applications (eg. pressure vessels, cranes etc). AE signals from pumps are very different and

so traditional AE signal processing techniques may need to be supplemented and/or

modified to ensure that they accurately characterise changes in machinery AE signals.

(b) A variety of noise sources, peculiar to machinery monitoring applications, can appear very

similar to acoustic emission sources, and if not identified and removed will, at best, lead to

1-1

PhD Thesis – Chapter 1, Introduction

confusion when interpreting results or at worst, result in a perfectly healthy seal or pump

being removed from service.

(c) As the time of failure is unpredictable, the amount of data that can be collected is

considerable. If the ultimate aim is to detect incipient failure, all of this data needs to be

analysed and scrutinised. To do this manually would be impractical.

(d) When it is impossible to locate a sensor where required (eg. on a seal face), waveguides can

be used. The effect of waveguide shape and material composition needs to be understood so

that signals can be interpreted correctly.

(e) Acoustic emissions collected from the seal are strongly influenced by the hydraulic

conditions within the pump. Therefore, to be able to differentiate between an incipient seal

failure and an unfavourable hydraulic condition, the latter must be identified. To do this, the

pump’s individual acoustic emission signature needs to be characterised.

(f) Mechanical seal failures need to be initiated in a controlled but realistic environment so that a

statistically significant amount of AE data can be collected, and the precursors to failure

identified.

(g) Electronics for collecting and analysing AE data are currently too expensive and

cumbersome. Physically smaller, more robust and cost effective alternatives are required for

continuous AE monitoring applications.

This thesis then became a process of investigating and overcoming each of these obstacles in turn.

Although sufficient time was not available to develop a seal-monitoring system, as solutions to last

two challenges are still proving elusive, the ultimate goal is now much closer to being realised.

1.2 BACKGROUND

Centrifugal pumps are the most prevalent electrically powered, rotating machine used today. Of

these, most are single-stage with end-suctions as shown in Figure 1-2.

A centrifugal pump moves fluid from one point to another by converting mechanical energy into

hydraulic energy through centrifugal action. The hydraulic energy is referred to as head and is

measured in meters.

1-2


In an end-suction pump, fluid enters the

pump through the suction flange (see

Figure 1-2

) and is directed into the rotating

impeller via the eye (See Figure 1-1). Once

in the impeller, radial vanes help to

accelerate the fluid towards the outer

circumference where the fluid leaves the

impeller with a higher velocity and enters

the volute (part of the pump’s casing). As

the impeller is not located centrally within

the volute, the fluid slows down gradually

and gains pressure (hydraulic energy,

known as head), before exiting the pump

through the discharge.

Figure 1-1: Impeller-volute terminology.

In a double-suction pump the fluid enters the pump through a side mounted suction flange.

Specially designed channels split the flow into two, so that the fluid enters both sides of the

impeller at the same time (hence the name double-suction), theoretically balancing the load on the

impeller. Thereafter the fluid is pressurised in much the same way as in an end-suction pump.

Double-suction pumps tend to be used in high flow, high head applications because of their very

high efficiencies.

Figure 1-2: Cross section of a typical end suction pump.

1-3


Figure 1-3: Pump, system and efficiency curves.

The amount of pressure a particular centrifugal pump-impeller combination can generate depends

on its design and on flow. This head-flow relationship is known as a pump’s characteristic curve, or

more commonly, it’s pump curve. An example is shown in Figure 1-3. Exactly where a pump

operates on its pump curve depends on the amount of system backpressure against which the

pump must deliver the fluid. This also varies with flow, and is called the system curve. The

intersection of both curves is known as the duty point.

Condition based maintenance (CBM) techniques are commonly used to manage pump reliability,

with prediction typically based on results of oil and/or vibration analysis. Monitoring frequency is

generally periodic as continuous monitoring can only be justified on very large or unspared pumps.

These methods are very effective at identifying many pump problems, such as incipient bearing

failure, motor-pump coupling misalignment, impeller unbalance and extreme cavitation.

Consequently, average pump mean times between failures have greatly improved over the past

several decades.

Unfortunately not all pump problems can be detected using traditional condition monitoring

techniques. In particular, mechanical seal failures, which cause fluid to leak from the pump casing,

and hydraulic problems caused by operating away from the pump’s duty point, are still difficult, if

not impossible, to identify or predict. Nowadays, these problems are more often the root cause of

removing pumps from service. Unfortunately, traditional monitoring techniques offer little

capability for feeding back to operators the instantaneous effect of varying a pump’s operating

conditions on its long-term reliability.

1-4


Efficiency of the mechanical-hydraulic energy conversion varies with flow, as shown in Figure 1-3.

Maximum input energy is converted into increased head at the Best Efficiency Point (BEP). In

well-designed systems, the duty point is close to BEP. However, due to system resistance changes,

poor design or control valve throttling, pumps are often forced to operate far away from their BEP.

This causes hydraulic, thermal and mechanical losses to increase, in addition to a number of more

serious adverse effects, as shown in Figure 1-4.

No published data has been obtained regarding the cost of suboptimal pump operation and/or

failure in Australia, however correlations can be made with results of reviews undertaken in other

countries. According to a 2001 report by the AEAT [33], centrifugal pumps in the European

Union use approximately 117TWH of electricity per annum, which is 10% of all electricity

consumed by industry and commerce in the EU, and account for 58MT of CO2 emissions (1998

figures). The report states that the “largest energy savings are to be made through the better design

and control of pump systems” and “efficiency may fall off fast as operation moves from the Best

Efficiency Point”. It also recognizes that “on average, pumps operate at flows below their best-

efficiency flow”. Although the report does not quantify how inefficiently pumps are being

operated, a similar study in Finland claims that of the 1690 centrifugal pumps reviewed in detail,

average efficiency was less than 40%[56]. Typical BEP values range between 55% and 80%; thus it

can be concluded that most pumps studied in the Finnish review were not operating close to their

BEP value.

1-5


Figure 1-4: Onset of adverse conditions caused by operating away from BEP[60].

In practice, low and high flow conditions are only deemed a problem when they result in significant

noise or vibration, or cause the failure of pump components such as bearings and impellers. Yet

much smaller changes in flow away from BEP can result in significant energy losses, as well as

inducing mechanical seal problems in many services. It is surmised that the reason for this lack of

attention to hydraulic instabilities is that they can only be detected indirectly, by the ultimate

degradation of mechanical components (eg. impeller pitting).

In comparison to other rotating and reciprocating machines used by industry, most centrifugal

pumps are relatively small, cheap and reliable (especially single-stage end-suction pumps). It is

their sheer number and infiltration into all areas of industry that justifies ongoing research and

development to improve their operation and maintenance. As a result of their low capital cost,

selected monitoring methods must be inexpensive and preferably non-intrusive. To justify

upgrading existing infrastructure, novel techniques must also offer greater capability for monitoring

operating conditions and incipient failure prediction than existing techniques.

1.3 THESIS STRUCTURE

1.3.1 Overall layout

Each chapter in this thesis addresses one of the problems listed in section 1.1. Several parts

(namely chapters 2 and 5-8) were written as stand alone papers and submitted to (and accepted by)

1-6


referred journals or conferences. The associated publication reference is given at the beginning of

each chapter, followed by its aim.

A literature review covering the body of work relevant to a particular problem statement is

contained within each chapter. As a result, if papers were to be presented herein as they were

published, considerable repetition would ensue. Therefore, for brevity, review sections of papers

containing considerable repetition have been replaced with appropriate linking statements.

Conversely, for the sake of relevance and continuity, some additional content has been added

subsequent to a paper’s publication. This is designated with a ‘NC’ (New Content) after the

relevant section header and predominantly relates to chapter aims and postscripts. Other than

minor linguistic improvements, no content has been removed or altered.

Several papers were co-authored. Consequently, the extent to which these co-authors contributed

to each body of work is detailed below. Furthermore, a set of initials (eg. MH or PK) in a section’s

header designates it as having been written, or the underlying work performed by, the respective

co-author. In all cases, Professors Michael Norton or Jie Pan offered general supervisory guidance

and oversight.

1.3.2 Chapter topics

Most current state of the art systems for detecting incipient machinery faults have been undertaken

by corporately, or privately funded researchers who tend to commercialise, rather than publish their

findings. Consequently, chapter 2 contains a review of systems available on the market (in 2003) by

analysing their underlying patents and research findings. It also suggests why the maintenance

community has not readily accepted AE technology, and suggests how this might be rectified in the

future. Published at a maintenance conference in 2003, this paper helps maintenance professionals

determine how best to interpret and compare results obtained with one system against results

collected with another. It also provided the basis for development of in-house electronics used in

this research project.

Chapter 3 outlines signal processing theory relevant to subsequent chapters. Basic statistical

analysis, wavelet and octave band analysis are discussed. These techniques are not usually applied

to acoustic emission testing, so their appropriate application needs to be carefully considered.

Equations, developed by the author, for computing statistics of long arrays from shorter subsets of

the array are also presented, facilitating faster and more efficient processing of long waveforms.

1-7


As acoustic emission signals are very small, analysis is complicated by the ingress of noise. Chapter

4 describes various techniques that were applied to separate noise from acoustic emission activity of

interest, thus improving the likelihood of detecting trends relating to incipient faults or failure

conditions. These techniques are either not usually applied to the AE signals from machinery or

their application to AE monitoring of machinery has not been detailed before. Examples of how

these techniques were applied to AE data collected by the author from centrifugal pumps are given.

Condition monitoring creates huge volumes of AE data that cannot be processed manually. To

overcome this problem, a relational database was developed with the help of Mr Paul Kelly, an IT

systems architect, for analysing and managing tens of gigabytes of AE data collected during the

course of pump experiments. This system is described in Chapter 5. In the work described

therein, Mr Kelly setup the hardware, provided advice on the overall system architecture and wrote

many of the advanced visual basic and SQL routines used by the database. Conversely, the author

defined the user requirements, designed most of the user interfaces, wrote many of the simpler

code modules and queries to extract data and developed all advanced signal processing routines.

This system allowed the author to process over 250GB of AE data collected during various stages

of this research project.

Waveguides are often necessary to provide simpler transmission paths from otherwise inaccessible

components (eg. mechanical seals), or to separate delicate sensors from hot machinery and/or

strong magnetic fields. However, it is hypothesised that they also complicate analysis significantly.

To clarify these interactions, Chapters 6 and 7 contain two published papers that verify, for the first

time, the temporal and frequency changes to AE simulated signals as they pass through short, solid,

cylindrical rods of varying materials, diameters and face angles. Basic waveguide theory is also

summarised, demonstrating the author’s familiarity with relevant acoustic emission wave theory.

Chapter 8 presents a case study on how AE monitoring, using the signal processing and data

management techniques described in the preceding chapters, can be used to successfully detect

changes in flow within a double-suction pump. It is based on a Comadem conference paper that

was written with Dr Melinda Hodkiewicz, who analysed vibration and dynamic pressure signals

against which AE trends are compared. This is the first time such results have been presented.

The cumulative results of flow tests on a number of small end-suction centrifugal pumps are

presented in Chapter 9. This facilitates analysis of the similarities and differences in AE behaviour

between pumps, which again has not been reported previously. The sensitivity of AE to the testing

process is also discussed and suggestions are offered to improve the accuracy of AE results

obtained from pump experiments.

1-8


Finally, Chapter 10 discusses the overall conclusions from this work and how the preceding

chapters contribute to the overcoming the challenges identified in Section 1-1. Numbered

references from all chapters are listed alphabetically in Chapter 11.

1-9

2. R

2-1

REEVVIIEEWW OOFF AAEE

TTEECCHHNNOOLLOOGGYY 22

Sikorska, J. and Norton, M.P.,

Proceedings of ICOMS 2003,

Perth Australia, Paper 011 pp.1-13

The aims of this chapter are to:

(a) Review signal processing techniques used for the detection of faults in rotating equipment

by analysing published works and relevant patents; and

(b) Identify problems that would need to be overcome for AE monitoring tools to be more

widely accepted by the maintenance community.

Publication Abstract

Acoustic emission (AE) has been used as an NDT technique for incipient fault detection since the

1960’s, but it is not widely applied to machine condition monitoring. Yet research indicates that

AE can be a very powerful tool for detecting impulsive faults, such as wear, lubrication problems

and impacts in bearings, gears and mechanical seals. Various proprietary “black box” instruments,

collectively referred to as stress wave analysis tools, are being offered by vibration equipment

manufacturers. However their operating principles and relationship to acoustic emission is unclear.

This may be a contributing factor to the restricted adoption of AE by the condition monitoring

community. To remove some of these barriers, this paper explains how these instruments measure

and characterise acoustic emissions generated by faults within a machine. It also discusses how

with a better understanding of the underlying principles by users, and a few minor technological

improvements by manufacturers, acoustic emission can become a powerful and affordable incipient

failure detection tool for many rotating equipment components.

PhD Thesis – Chapter 2, Review of AE Systems

2.1 INTRODUCTION

Acoustic emissions (AE) are high frequency elastic waves generated by the release of energy from

micro- and macroscopic defects within a material. They are also known as stress waves and are

generated in discrete packets, the amplitude, direction and polarization of which depend on the

material’s crystalline structure [12]. In general, the amount of elastic energy released is proportional

to the volume of the sample, the speed of deformation and the amount of micro-damage created

[25].

Common sources of stress that can cause AE generation include material fracture, crack nucleation

and growth, phase changes, or variations in external conditions that produce a static imbalance or

phase instability. Consequently, it has been used as a non-destructive monitoring technique for

over thirty years. Condition monitoring of machinery is a more recent application of this

technology, however its extension to the detection of defects in bearings, gearboxes and other

structural components is obvious. After all, most machinery problems result from material

degradation (be it wear, fatigue or overstressing) or extraneous fluid-structure interactions (e.g.

cavitation). These have all been shown to generate acoustic emissions and a brief overview of

relevant research will be given in the following section. Though many of these failure modes can

be detected by vibration monitoring, it has determined that acoustic emissions generated by these

faults precede detectable vibrations [120], particularly in low speed equipment.

A question therefore needs to be asked: “why is AE not being more widely used for detecting

incipient machinery faults?” The answer is that in fact it is, albeit under a variety of different names

and in a rather limited capacity. Unfortunately, not only is it difficult or impossible to compare the

output of one system against another, very few objective comparisons have been undertaken.

Therefore, users have no option but to believe a manufacturer’s sales literature on why their

particular technology variant is superior to any other. Furthermore, many of these systems give

their output in the form of one or more ‘magic numbers’, yet supply little or no information on

how these are derived. It is therefore not surprising that AE is viewed with a degree of scepticism

by the condition monitoring community.

This paper will address some of the confusion surrounding the use of AE for diagnosing incipient

machinery faults. After summarizing initial research in this area, the main commercially available

technologies will be explained. It is important to be aware that these explanations are the authors’

interpretations of available literature and original manufacturers’ patents. However, only the

manufacturers could clarify this any further and then only for their product. Consequently, the

authors believe that by presenting these systems collectively, users can better appreciate the

potential and limitations of available acoustic emission technology.

2-2


2.2 AE RESEARCH

AE signals are described as being continuous,

burst or mixed mode, as depicted in Figure 2-1.

Burst emissions are typically discrete transients

with relatively short decay times and even shorter

rise times. They are often approximated as

exponentially decaying sinusoids. Continuous

emissions are bursts that occur too closely

together to differentiate between individual

events. The highest measurable rate of discrete

bursts is about 50kHz, above which the length of

the bursts usually exceeds the time between

them. Continuous AE generally appears as an

increase in the background noise level, with no

distinguishing features other than its amplitude

and frequency content [12]. Mixed mode AE

contains a number of high individual bursts

above a background level of continuous

emissions. All types of acoustic emission are

generated by discrete processes, however they

differ in the repetition rates of the underlying

faults. Acoustic emission signals generally range from a few kHz up to several MHz. Typical

parameters used for describing AE signals are shown in Figure 2-2

Figure 2-1: Types of AE signals [50].

Figure 2-2: Traditional features of an AE signal.

. AE energy (~RMS2) is

calculated by integrating the signal over time. Generally, amplitude and frequency analysis are more

meaningful parameters for measuring continuous emissions, whilst count rate and energy analysis

are useful for burst emission characterisation [95].

Early work in the use of acoustic emission monitoring for detecting faults in rotating equipment

was undertaken by Exxon in the late 1970s [14]. In a four year study, which measured vibration

and RMS of the AE signal in the 80-120kHz frequency band, a variety of problems were detected

including bearing, gear and coupling defects, cavitation, seal leakage, pump motor misalignment,

extruder screw failures and piping induced misalignment. During a 31-week representative period,

37 actual or potentially serious problems were identified.

Lubrication and wear is a major problem for many machinery components, particularly bearings

and seals. AE has been confirmed as a “relatively easy to implement technique that generates

useful real-time information from the processes occurring at the area of contact between bodies

2-3


subjected to wear” [44]. Relationships have been established between many AE parameters and

sliding contact variables for typical wear tests. Research indicates that the primary wear

mechanism under sliding lubrication conditions can be identified from the time-varying nature of

the AERMS voltage signal: no signal is detectable above background noise when full elasto-

hydrodynamic lubrication occurs, under adhesive wear the RMS changes gradually (either increasing

or decreasing), whereas abrasive wear is characterised by a highly fluctuating signal [57].

Furthermore, by measuring peak amplitude and rise times of demodulated acoustic emission bursts,

it has been concluded that “actual removal of material is characterised by high amplitude, short rise

rate acoustic emission signals, and that running-in wear and plastic flow results in lower amplitude

signals with much longer rise rates” [72].

Others have shown AE to be a very sensitive indicator of lubrication conditions [69] and changes in

lubricant properties [17, 18, 75]. Miettenen used AE monitoring (primarily the number of counts)

to detect contamination concentrations in bearing lubricants as small as 0.02 weight-% and

differentiated between contaminants of different particle sizes and/or hardness [75]. AE levels

decreased when clean lubricant was reintroduced, though not to the level found in a new bearing.

Boness found that when contaminant concentration in the lubricant was changed, the integrated

RMS signal clearly showed the expected change in abrasive wear rates, corroborated by a very

similar change in wear scar volume [17].

Consequently, others have tried to establish if AE can detect bearing failures. A fundamental study

of the causes of ultrasonic bearing vibrations was published in 1977 by Catlin [24]. He recognised

that high frequency signals were produced by a number of different failures mechanisms including:

micro-cracks and cracks from fatigue or local over-stressing; surface roughening from a lack of

lubrication; surface dents from hard contamination; and micro-pitting. Burst emissions were

observed when the bearing passed over sharp discontinuities. These pulses excited many high

frequency resonances of the bearings, local structural components and accelerometers used for

detecting the signals. In initial stages of defect development, the amplitude of excited resonances

was related to defect severity. As these defects grew and their edges were smoothed over, discrete

pulses became obscured by high amplitude, broadband noise (continuous AE). Additionally, high

frequencies were less able to be excited and the overall spectrum shifted to a lower frequency range

as damage progressed. A lack of bearing lubrication was found by Catlin to generate a steadily

increasing, continuous AE signal, due to gradual roughening from continuous metal-to-metal

contact. Unfortunately, this was similar to advanced spalling, which identified the need to detect

bearing problems in their early stages.

An important corollary from Catlin’s work was that excited resonances, from either the structure or

the accelerometer, could be used to improve detectability. This formed the basis for a number of

2-4


patented systems that will be discussed below. Good reviews of bearing vibration monitoring by

such techniques are given in [73].

In the 1980’s and 1990’s a diagnostic algorithm to identify a variety of faults in journal bearings was

developed [102, 103]. They separated the signals into burst and continuous emissions and

characterised frequency information as either wideband or narrowband, and either tuned to the

rotating speed or untuned. Depending on these classifications, the authors were able to correctly

diagnose rubbing, metal wipe, and bearing tilt, as well as rotor cracks, fatigue cracks and radial plane

damage. Although metal wipe could be monitored with conventional AE techniques (clear patterns

could be identified in the time signal and envelope), an envelope spectrum was required to diagnose

running and bearing tilt. A similar model was developed for incipient failures in rotary air

conditioning compressors [104] where the difference between abrasive wear and vane butting were

identified. Turbine rotor rubbing has also been detected [54] at bearing housings up to 2m away

[71].

In addition to frequency analysis of AE signals (or demodulated AE) a variety of other advanced

techniques have been used to better identify bearing defects. These include autoregressive

coefficients [55], Point Process Spectral Averaging [59], adaptive noise techniques [121] and pattern

recognition analysis [65].

Slow speed equipment can be problematic for vibration analysts due to the limitations imposed by

instrumentation (at low frequencies), low energies of fault frequencies and the use of

accelerometers rather than displacement probes for routine measurement. In these applications,

research has shown that AE can detect simulated faults in rolling element bearings rotating between

1 and 100rpm [54, 55, 99], although there does seem to be some disagreement as to the required

size of the fault before detection is guaranteed, particularly for inner race defects where

transmission paths are complex [54, 55, 123]. In Tandon and Nakra’s work, AE counts were also

shown to increase with load [123]. Others have also identified unbalance in low speed equipment

by demodulated resonance analysis [21].

AE has also been shown to be capable of detecting a variety of fluid borne problems in machinery

[53]. In 1977, Derakshan et al showed that “acoustic emissions” are generated during bubble

generation and collapse and determined that cavitation intensity in a hydroturbine was proportional

to the RMS of the acoustic emission signal [29]. Finley also observed that cavitation in machinery

was easily detectable as “high amplitude excursions in the RMS”[34]. More recently, Neill et al

were also able to detect cavitation in centrifugal pumps as well differentiating it from recirculation

[80, 81]. Board has also shown that changes in pump flow conditions (cavitation, recirculation and

best efficiency) can be detected by stress wave energy levels [16] (See Figure 2-10).

2-5


Failure of mechanical seals has also been detected by acoustic emission analysis. A continuous

monitoring system was implemented in a Japanese refinery that measured AERMS and it was found

that one of four patterns was observed in a seal’s AE trends prior to failure [61]. The authors

concluded that instability in the AE signal corresponded to instability in the lubricating film and

that an estimation of wear speed and wear amount is possible by monitoring seal emissions.

Similar results were reported by others who identified leakage, dry running and cavitation in the seal

gap by measuring acoustic emissions (with RMS) [76] and determined that acoustic emission

variations coincided with torque and temperature variations during sliding wear [68].

2.3 EXISTING TECHNOLOGY

2.3.1 Shock pulse – SPM Instruments

Shock pulse was first developed and patented in Sweden in the early 1970’s by SPM Instruments

[112, 114]. It involves measuring and analyzing high frequency shock waves that are generated by a

rotating bearing and detected by a piezoelectric accelerometer resonant at 32kHz. SPM’s initial

research found that shock pulse amplitude is based on rolling velocity, oil film thickness and the

mechanical condition of the mating faces. They also discovered that in an undamaged bearing,

although the maximum level of these stress waves depended on the thickness of the lubricant, the

ratio between peak values and the baseline “noise” level did not change. Surface damage however,

affected this ratio dramatically [78]. Later work determined that amplitude was also determined by

sensor location with respect to the load zone [113].

SPM then undertook a vast amount of testing on control bearings to determine relationships

between:

• Lubricant film parameter, Λ, and shock pulse value in decibels (dBBsv)

• The difference in dBsv under lubricated and dry running conditions

• Lubricant film thickness and dBsv for a fixed sliding velocity

• The baseline dBsv (i.e. in a good bearing) for various bearing types and sliding velocities

Over 9943 sets of real-time data from a variety of bearing types, sizes, manufacturers, lubricants

and loads were collected.

To ascertain the condition of an operating bearing, shock pulse readings can be taken from

operating bearings and the values compared against this known data. Additional processing then

translates this diagnosis into simple parameters that can be trended.

2-6


2.3.1.1 How it works

As already stated, shock pulse measurements are collected with a piezoelectric transducer that has a

natural frequency of 32kHz. The electrical signal is bandpass filtered and separated into its

transient and stationary components, the latter of which is disregarded. Each individual transient

event is then converted to an analogue pulse with an amplitude proportional to the energy in the

transient pulse, measured in decibels. This signal is passed through a series of discriminator and

gate systems, to determine the amplitudes at which four defined count rates occur: 1000 events per

second, 200 events per second, 40 events per second and the largest event in 2 seconds. This is

done, in effect, by changing the value of the discriminator’s gain until the desired occurrence rate is

exceeded.

The effect of rolling velocity is then mitigated by subtracting a known shock value (dBi) for a

bearing of that size and rpm from each of these four count rates to give normalized shock pulse

readings. This requires the user to enter information about the type, size and speed of the bearing.

The resultant four amplitudes gives shock pulse output values referred to as HR (high rate of

occurrence), dBc (dBsv carpet value), LR (low rate of occurrence) and dBm (dBsv maximum)

respectively. In terms of more commonly used AE parameters: HR, dBc and LR are equivalent to

calculating the count rates at three different and increasing thresholds, and dBm is the maximum

peak detected during a 2 second period. dBm can then be used to place the bearing into one of

three empirically derived condition zones (indicated by green-yellow-red colored light) and gives a

simple visual indication of whether the bearing is healthy or at risk of failure.

dBc and the difference between dBm and dBc (ΔdBsv) are used to indicate the existence of a

lubrication problem or bearing fault. This is indicated by a CODE between A and D. If the

system deems that damage has not yet occurred, then it will calculate an oil thickness and return the

value in the form of a LUB No. This is calculated by subtracting ΔdBsv from a known value for an

equivalent lubricated bearing. This is then mathematically converted into a LUB No. One unit

increase of LUB No corresponds to one micro inch increase in oil film thickness, with 0 implying

dry running.

Unfortunately, if a bearing has already suffered some damage, then LUB No cannot be determined.

Instead, the system tries to quantify the damage. This is done by subtracting the normalized values

of HR from LR and converting the result into a COND No (based on more empirical data)

depending on the type of bearing, its speed and diameter. A COND No of around 30 indicates

surface stress and increases with damage severity; a value of more than 45 infers a high risk of

failure. Rapidly increasing COND numbers also imply that the condition is deteriorating quickly

and the bearing is at high risk of imminent failure.

2-7


Figure 2-3: SPM Processing.

New analysers also take into account the sensor’s location with respect to the bearing’s load zone,

to account for any damping that might be occurring because of an indirect transmission path.

Again, these calculations are based on empirical data collected by measuring the effect of position

on expected signal strength (See Figure 2-4).

Some SPM systems now offer additional FFT functionality. A time series is measured and

demodulated (based on the 32kHz being

the AM carrier frequency) before being

digitized and subjected to an FFT, which

has a variety of windowing, averaging and

scaling options. From this enveloped

shock pulse spectrum, the source of high

readings can be verified in much the same

way as traditional demodulated spectra.

However, it is not recommended that this

spectrum be used as a condition indicator,

due to the numerous factors that can

affect the absolute magnitude of SP

spectral lines.

Figure 2-4: Damping effect due to sensor location (from [113]).

2-8


2.3.2 CSI PeakVue

CSI’s PeakVueTM technology uses an ordinary accelerometer to measure the stress waves. Based on

research undertaken by Robinson et al [99] a patent was lodged in 1998 [100]. Subsequently, CSI

(now Emerson Process Management) manufactured this system, which measures peak values of the

accelerometer signal and can determine:

• Amplitude of individual stress wave events

• Approximate duration of individual stress wave events

• Rate of stress wave events

Processing can be undertaken by analog or digital systems to display time and frequency

information. PeakVue is a standard feature of most CSI portable analysers.

PeakVue was developed to overcome the traditional deficiencies of demodulation: namely

difficulties in determining an appropriate bandwidth and reduced amplitude accuracy.


This system is best described by comparing it to a standard enveloping system, commonly used for

identifying bearing faults. Both approaches are based on the premise that impulsive events (eg.

bearing faults, gear tooth interactions) cause regular excitation of natural frequencies in either the

structure or sensor, resulting in a signal where the fault frequency is modulating a higher frequency

carrier. Demodulation (also known as enveloping) involves passing this signal through a series of

filters and rectifying it prior to analysis or digitisation to isolate the fault from its carrier.

2.3.2.2 Envelope Analysis

The process of demodulation is shown in Figure 2-5. Firstly, a high pass (HP) filter removes low

frequency components such as rotating speed and its harmonics from the signal, whilst leaving the

higher forcing frequencies of interest and/or their harmonics. The signal is then rectified before

Figure 2-5: Process of extracting a traditional envelope spectrum (from [122]).

2-9


being passed through a low pass filter to remove the carrier and limit the bandwidth to the

frequency range of interest (containing the fundamental frequencies of the “desired” faults). This

leaves just the envelope of the initial signal.

If the filters have been selected appropriately, it should be possible to determine any fault

frequencies by looking at the time difference between successive peaks. Alternatively, the

demodulated signal can be digitized for storage or processing by an FFT. If so, the final bandwidth

must be less that the Nyquist frequency (sampling frequency/2.56) to avoid aliasing. This

restriction may cause a reduction in the amplitude resolution of the envelope, which can make

analysis difficult; analog filtering effectively averages the signal energy over the time taken to pass

through the filter. The lower the final bandwidth,

the longer the time interval taken for filtering and

hence the lower the final amplitude. Decreasing

the final bandwidth, either because a slow

digitization rate is desired or because the

frequencies of interest are much lower than the

carrier, will cause any peaks to be reduced until

they can no longer be distinguished from the noise

floor.

Theoretically, bearing faults often create a large

number of harmonics, so it should be possible to

find an area of the spectrum that is

uncontaminated by the lower frequencies

associated with rotating speed (that are often much

greater in magnitude). However, for the less

experienced technician, this dependence on correct

filter selection can make envelope analysis quite

difficult or time consuming. Also, reasonable

alarm levels are almost impossible to establish as

the energies associated with specific problems

change in bandwidth as the faults progress.

PeakVue was developed to overcome these

deficiencies. As shown in Figure 2-6, rather than

passing the signal through the final low pass filter

to extract the envelope, PeakVue uses a dual Figure 2-6: PeakVue Processing.

2-10


sample and hold amplifier to determine the signal peak over a particular time increment. The

duration of this increment is equal to the inverse of the sampling frequency (digitization rate),

which is set at 2.56 times the maximum frequency of interest, Fmax. Consequently, no low pass filter

is required to avoid aliasing. Also, the true signal amplitude is extracted rather than an averaged

value, irrespective of Fmax, although frequency resolution is lost at lower bandwidths (illustrated in

[97, 98]). This process continues until the desired number of samples is obtained.

Analysis is undertaken by inspection of the digitized PeakVue time signal, FFT and autocorrelation.

Units are always those of the sensor used to take the measurements. Although averaging can be

undertaken, CSI recommend that it not be done as the unaveraged time signal offers as much

information as the spectrum. Additionally, on variable speed equipment, changes in rotating speed

can reduce peak values as these are unwittingly “averaged out”. To avoid this problem, a signal

can also be collected from a tachometer for synchronous averaging. Finally, it is recommended

that long term trending be undertaken of a number of parameters, for which alarms can then be

established. Guidelines are also given on the anticipated PeakVue time waveforms and spectra for a

variety of faults and problem conditions [97, 98].

2.3.3 Spike Energy – Entek

Spike EnergyTM was developed by IRD International in the late 1970’s and is incorporated into the

Entek dataPAC 1500 analyzer. Like PeakVue, Spike Energy uses a normal accelerometer to

measure the stress wave. Overall signal magnitude is expressed in “gSE” units and instrumentation

provides a Spike Energy spectrum and Spike Energy time waveform for diagnostic analysis.


Illustrated in Figure 2-7, the measured high frequency vibration signal is first filtered by a band pass

filter. The user can select one of six high pass corner frequencies (Fmax) but the low pass corner

frequency is fixed at 65kHz. The signal is then passed through a sample and hold amplifier (similar

Figure 2-7: Spike Energy Processing.

2-11


to PeakVue) which detects the peak

amplitude and applies a decay constant.

The value of this constant is automatically

selected by the instrument and depends on

the Fmax, the higher the frequency, the

shorter the time constant. This results in a

saw-toothed output signal (Figure 2-8) that,

according to the inventors, accurately

represents defect severity. An FFT is then

applied to determine the gSE spectrum.

Spike Energy is sensitive to the transducer’s

resonant response, mounting method and

position. Therefore, different

accelerometers will give different gSE

readings. Additionally, stud mounting is recommended as magnets or hand held probes will cause

too much signal attenuation.

Figure 2-8: Typical Spike Energy signals (from [130]).

Due to the variety of sensors that can be used for these measurements, no severity charts or

guidelines are given. Instead, it is recommended that Spike Energy be collected at the same time as

other vibration and/or process parameters, and correlations between them be derived over time for

specific equipment [130].

2.3.4 Stress Wave Analysis – Swantech Ltd

Stress wave analysis by Swantech is another instrumentation technique for measuring friction,

shock and dynamic load transfer in rotating machinery. It requires the use of a specific Swan

sensor that has a resonant frequency of 38.5kHz, and demodulates the signal prior to digitization

and processing [4]. The three main analysis tools are:

1. Stress wave energy (SWETM) which is a function of the amplitude, shape, duration and rate of

all individual impulsive events that occur during a fixed time period;

2. Stress wave spectrum (SWSTM), which is an FFT of the demodulated signal;

3. Stress wave amplitude histogram (SWAHTM), which is a plot of the number of events versus

amplitude.

2-12


Figure 2-9: SWANTech Processing.

With these techniques, Swantech has been able to detect a variety of problem conditions including

bearing wear and localized surface damage, lubricant degradation, turbine impeller rub, labyrinth

seal wear and foreign object damage [15].


Unlike the systems discussed so far, SWANTech uses a more traditional AE approach, extracting

features from individual impulsive events, as the basis for its analysis. An event is assumed to begin

when a preset threshold is crossed and ends when the enveloped signal drops below that threshold.

Peak amplitude and energy are determined for each event. SWAHTM (number of events versus

amplitude) is then plotted from the calculated amplitudes (in volts). Amplitude is representative of

the intensity of a single friction or shock event, so the histogram is a graphical representation of the

number of events occurring with a certain intensity. Healthy machinery is usually represented by a

narrow bell shaped distribution at the low end of the voltage scale. As machinery deteriorates, the

number of higher intensity events increase and the distribution appears skewed towards higher

voltages. This is particularly prevalent in the case of lubrication problems, particulate

contamination or skidding between rolling element bearings.

SWETM (stress wave energy), which can be considered synonymous with AE energy, is a function

of not only event intensity, but also duration and is therefore useful for giving an overall impression

on the extent and severity of

damage that may be occurring.

It is determined by summing

individual event energies. It is

a simple, trendable parameter

that seems to be sensitive to

lubrication quality, bearing

deterioration and abnormal

preloads, such as those

imposed by misalignment.

Figure 2-10: SWE in a centrifugal pump (from [16]).

2-13


Final failure is normally preceded by a rapid increase in SWE. It can also distinguish between

different types of fluid related problems (Figure 2-10).

2.3.5 SEE – SKF

SKF patented the technology behind their SEE (Spectral Emitted Energy) devices in 1988 [126]1.

Its aim was to overcome the problems incurred when harmonics of fault frequencies overlapped

with the 30-40kHz resonant bandwidth of traditional stress wave devices.


A wideband acoustic emission (piezoelectric) transducer is used to collect the signal that is then

passed through a series of components that give a narrow bandwidth signal around a central

frequency, generally in the 250-350kHz range. Defects will cause an increase in the amplitude of

the filtered signal. It is then passed through four comparators that count how many times the

amplitude exceeds one of four predefined thresholds. The lowest threshold is checked at a rate of

1MHz, the next highest at a rate of 2MHz, the next at 4MHz and the highest at a rate of 8MHz.

These are then totaled together and then divided by the time taken to determine the counts. This

gives an overall weighted count rate, with the higher amplitudes being counted up to eight times

more often than the lowest values. This can be supplied to a digital readout or converted to an

analog voltage for logging.

Figure 2-11: SEE Processing.

1 SKF stopped production of SEE Pens (portable data loggers) in 2003, but the functionality is still included in SKF online analysers. (NC)

2-14


SKF claim that SEE provides useful information on lubrication related problems, undetectable by

vibration analysis, however results can be misleading if collected in isolation. Poor couplant

between the sensor and the machine will result in lower than expected values; also readings can

actually decrease as a defect spreads, similar to the effect on an enveloped signal [2]. Although SKF

does give some general guidelines on how to interpret SEE values [3] they recommend

maintenance decisions are based on higher-than-normal values for a particular operating machine,

rather than these predefined limits.

2.3.6 Acoustic Emission – Holroyd Instruments

Since the mid 1990s, Holroyd Instruments have released a number of products based on their

patented acoustic emission/stress wave sensing technology [51]. Portable instruments include the

MHC-Memo, MHC-Classic and MHC-Solo data loggers. They also have a system for online

analysis. All require the use of special Holroyd sensor and give results in terms of two parameters:

Distress® and dB level.


Holroyd AE transducers have a natural frequency around 100kHz and contain in-built electronics

to amplify and filter the signal into a narrow band around this resonance. A measured signal is then

passed through a series of components, as shown in Figure 2-12 (and described in [52]), which

significantly expand the analyzer’s dynamic range. The subsequent output is an overall voltage level

in decibels that represents the mean RMS of the signal detected by the transducer with minimal

error. Though not clearly stated in the literature, it can be reasonably assumed that the dB level is

some function of this signal over time (e.g. average over ‘x’ seconds).

dB level is an indicator of the general state of the equipment and varies with both operating

conditions and state of the equipment. Therefore, no absolute failure limits can be set. However, it

Figure 2-12: Holroyd Instruments Signal Processing.

2-15


is claimed to be a useful parameter for trending the rate of deterioration of suspect machinery.

Typical data loggers facilitate a dB level range of up to 80-92dB.

The second parameter used by Holroyd equipment is called Distress®. This is determined by taking

the analog dB level signal through a series of analog envelopers that separately extract the means of

the slowly changing (DC), and quickly fluctuating (AC), parts of the signal and convert them into

logarithmic form. Distress® is then determined by subtracting one mean from the other. (As

these are in logarithmic format, this is equivalent to dividing the AC mean by the DC mean.)

The mean AC component of the signal is a function of the number and amplitude of short duration

excursions in the RMS envelope. Therefore, it represents the number and magnitude of impulsive

acoustic emission events occurring in the machinery. Dividing it by the mean DC value acts to

compensate for any changes due to operating conditions or electrical noise. Consequently,

condition levels can be established that are independent of the machine and its operating

environment (see Table 2-1). Distress® is supposedly very sensitive to lubrication problems and

can be used for determining the optimal times to relubricate. Additionally, it can be used to give an

early indication of bearing and gear faults.

2.4 COMPARISONS AND DISCUSSION

The final output given by an analyzer is influenced by a number of factors that relate to the

equipment used to measure and process the acoustic emission signals. These will now be discussed.

2.4.1 Source Characteristics

Although it is assumed that equipment does not actually influence the source it is measuring, the

ability to discriminate desired acoustic emissions (e.g. bearing faults) and undesirable acoustic

emissions (e.g. flow noise) does vary between systems.

To improve detectability, all systems discussed use the transducer’s natural frequency to amplify the

signal, later removing this carrier frequency through some form of demodulation. Systems that use

their own sensors claim that the resonant response has been carefully selected to detect material

faults, yet these resonant frequencies vary from 30 kHz to several hundred kilohertz. Fortunately,

lubrication events tend to cause broadband excitation (due to the varying size of the wear particles)

and so will most likely be detected equally well by any of the systems. Equivalently, bearing and

gear faults generally have sufficient harmonics to excite any of these frequencies.

2-16


Because amplitude attenuates much more at higher frequencies, detection of extraneous, unwanted

signals from other machine components is less likely when using higher frequency sensors,

improving the overall signal to noise ratio. This is particularly important when measuring high-

speed equipment where harmonics of gear mesh and blade pass frequencies could easily move into

the 20-50kHz bandwidth. If these signals are accidentally detected, they may swamp the analyzer’s

dynamic range or make any faults appear larger than they should be. SEE and Holroyd equipment

should be much less prone to these interactions as their resonant frequencies are higher.

Research has indicated that frequency information can be used to discriminate between sources.

Therefore, it would be easier to identify unwanted sources with systems that incorporate some

form of frequency analysis of the time signal (e.g. PeakVue, Spike Energy, SWANTech, SPM with

spectrum options). The user might then be able to collect the data again with a different

bandwidth or sensor location, to remove extraneous signals.

2.4.2 Transducers

As already discussed, signals measured by higher frequency sensors will have a greater SNR.

However, they will most likely be smaller, requiring more sensitive transducers and greater

amplification. Signal conditioning will also have to be less susceptible to instrumentation noise,

but it can be assumed that the manufacturers will have considered these factors. More importantly

from a user’s perspective, at higher frequencies it is crucial to have an adequate couplant (e.g.

silicon grease) between the sensor and work-piece, otherwise most of the signal’s energy will not be

transmitted. Some manufacturers also recommend this when working in the 20-50kHz frequency

range.

Measuring stress waves with an accelerometer seems advantageous: no additional equipment is

required to that used for routine vibration testing. However, there are some problems. As

mentioned earlier, output amplitude is a function of the sensor’s response to an excitation.

Therefore, for the same fault condition, different accelerometers will give different readings. This

prevents any worthwhile comparison being made of components measured by different sensors,

even if all other measuring instrumentation remains unchanged. As even the best sensors can fail

and technology is constantly being updated, there is no guarantee that the same accelerometer type

will always be available. If the relevant sensor was discontinued, years worth of stress wave data

could become worthless and self-defined failure parameters could no longer be used. Although

taking additional measurements with an additional AE (or stress wave) sensor takes more time, it is

likely that the sensor would remain unchanged for the life of the analysis product.

2-17


It is possible to detect stress waves without using the sensor’s resonant response, although signals

are much smaller (10-6V). This approach is used in AE non-destructive testing where tiny signals

are measured by broadband acoustic emission sensors and low noise signal conditioning hardware

Due to the relatively flat response over a larger bandwidth, these sensors are commonly used in

condition monitoring research projects where more features are being extracted from the AE

signal2.

As indicated earlier, sensor location can significantly affect the acoustic emissions measured. If the

transmission path between the source of the acoustic emission and the sensor changes, so will the

amplitude and frequency content of the signal that is detected. Consequently, it is very important

that measurements be taken in the same location every time. If possible, the most direct structural

path between the source and sensor should be used.

2.4.3 Gain and Threshold Level

Most analog systems require the signal to be amplified (or attenuated) so that it is within the

working range of the analog components. This usually requires the user to select appropriate range

or gain settings. If too large a range is selected (compared to the actual range of the signal), then

available resolution will be lost and small signals may be lost in the noise floor. Additionally,

systems that rely on counting the number of bursts above a certain threshold will give erroneously

low values. If too small a range is selected, then the signal will be amplified too much and clipping

of higher amplitude bursts will occur. This will reduce the value of the maximum amplitude

detected to the limit of the range) and effect the frequency content of the signal.

If either the signal amplitude is unknown, or if it varies too much, selection of the correct gain can

be difficult and may result in the loss of signal information. To overcome these problems, Holroyd

instruments have devised a method of accommodating large variations in signal amplitudes without

requiring any gain selection or sacrificing resolution. Remaining technologies may require the user

to try different settings until their signal remains within the limits of the analyzer.

2.4.4 Differences in Functionality

Functionality and the mechanisms by which each system indicate lubrication problems and bearing

faults are summarized in Table 2-1.

2 It is also common to use the AE sensor’s natural frequency to facilitate improved detection of active structural defects. (NC)

2-18


Very few comparisons of available stress wave detection systems have been undertaken. Alfredson

and Mathew evaluated a variety of time domain techniques for monitoring bearings and declared

that although shock pulse was a more reliable indicator of problems than statistical analysis (crest

factor, probability distribution, kurtosis) but required significant operator skill and could not be

used exclusively [9]. Bansal et al compared shock pulse against AE (Peak, RMS and count rate) and

found that the two methods were in good agreement [11]. Both techniques were able to

differentiate between new and reconditioned bearings. Tandon and Nakra concluded that AE was

better than shock pulse and demodulation at measuring very small defects and defects in slow

speed bearings [123]. However, detectability of AE decreased with load, whilst the other

techniques improved. PeakVue has also been reported to provide better detectability that

demodulation [99].

It is well reported that AE levels change with operating conditions, particularly load. Additionally,

as fault conditions deteriorate it is expected that background levels of continuous emission will

increase. In some cases, burst events will remain evident above this background level, whilst at

other times they will disappear into the “noise” floor. For more reliable results, some allowance for

these conditions should be made. SPM and Holroyd equipment do account for these changing

conditions, however in the remaining systems it is up to the user to interpret the data accordingly.

Traditional AE analysis hardware (used by NDT professionals and researchers) has not been

discussed in detail. Some of these alternatives are capable of analyzing AE signals by traditional

vibration tools (e.g. spectra, octave bands etc). Based on DSP hardware, they digitize the analog

signals at 4-20MHz, prior to storage and analysis. This makes them extremely powerful and very

flexible. However, they are not as portable as traditional vibration data loggers, mostly being

incorporated into a desktop PC; the authors are not aware of any currently available variant that can

be integrated into a modern laptop. This should change in the near future, increasing the

availability of AE analysis to a much broader community.

2.5 CONCLUSIONS

Although it is unlikely that acoustic emission will ever replace vibration monitoring, it is a powerful

supplementary tool for detecting incipient faults in machinery. Unfortunately, ensuring reliable,

repeatable and meaningful results can be a complicated process. There are tools on the market that

can simplify these tasks and there is much evidence that these have provided maintenance

professionals with valuable information on many occasions. Black box instruments, like SEE Pen

by SKF and LUB Checker by Holroyd Instruments, are particularly useful for answering simple

questions like "how much lubrication do I need?". The suitability of AE for answering this

particular question has been verified by independent research. Problems begin when these tools

2-19


claim the ability to diagnose more complicated faults with little or no interaction from the condition

monitoring specialist: irrespective of how good the technology, in order to trust the output of these

devices, most technical personnel need to understand what these instruments are doing and how

they derive their results. This can be difficult to find out. The fact that each variant is different but

makes similar claims, adds to confusion and scepticism. Clear explanations of the technology and

comprehensive objective comparisons are needed for these tools to be more widely accepted. This

paper has attempted to initiate this process. Another problem with the present generation of black

box instruments is that they can only identify a limited number of fault types, predominantly

bearing problems. However, future equipment already in development will use more advanced AE

parameters and neural networks to characterise the signals. Automatic diagnoses of machinery

faults should then be significantly more reliable.

The alternative is systems that expose the user to the underlying signals, allowing the analyst to

make a diagnosis. Although this removes the mystery, it does have other drawbacks. Advanced

acoustic emission systems are very expensive and are only produced by a limited number of

manufacturers. They also require the user to become more knowledgeable in acoustic emission

analysis, which has some marked differences from vibration. Issues of where to place sensors, how

to couple them to the equipment and instrumentation sensitivities become even more important

when looking at millivolt or microvolt signals between 10 and 1000kHz. Time waveforms can be

more useful than frequency spectra, so the tools required to analyse them are different.

Furthermore, AE signals are often a function of operating conditions. Finally, to complicate

matters further, there are no standards to guide users in their interpretation of results.

Nevertheless, as the NDT community has shown over the past thirty years, it is possible to

overcome these challenges and the rewards are significant. Advanced AE tools offer the capability

of detecting certain types of faults earlier than currently possible, and can detect problems in

components that presently fail without any warning. This promises safer, cleaner and lower-cost

operating plants. As computers get faster, AE tools will become more numerous and affordable.

When enough collective experience has been gained, standards will be written and standardisation

will follow. Only then, will AE monitoring become as routine as vibration analysis is today.

2.6 POSTSCRIPT (NC)

This review formed the basis of a specification for in-house electronics, DAQ hardware and

software used to collect AE signals from pumps and mechanical seals. Basic requirements

included:

(i) 100-1000kHz bandwidth;

2-20


(ii) 4 channels of simultaneous data digitized at 2-5MHz;

(iii) 20-100dB of gain, some of it programmable;

(iv) The ability to quantify and characterise both discrete and continuous emissions;

(v) The ability to apply and tune the value of floating thresholds;

(vi) Use of generic broadband AE sensors to improve repeatability and independence of results

from sensor response;

(vii) Use of standard or easily understood AE signal features;

(viii) Flexibility to add or change features retrospectively;

(ix) Continuous on-line operation (24 hours a day, 7 days a week) for an indefinite period of

time;

(x) Full exposure to the raw time signal.

The final system, as used to collect some of the data presented in Chapter 9, is summarised in

Appendix B.

2-21

PhD Thesis – Chapter 2, Review of AE Systems PhD Thesis – Chapter 2, Review of AE Systems

2-22 2-22


2-23

3. A

3-1

AEE SSIIGGNNAALL

PPRROOCCEESSSSIINNGG33


(a) Review relevant signal processing theory; and

(b) Introduce the methods used for analysing experimental work presented in later chapters.

Equation Chapter 3 Section 1

3.1 INTRODUCTION

As no two acoustic emission signals will ever be the same, most AE processing is strictly statistical.

In machinery monitoring, this first involves acquiring a large amount of data about an underlying

AE generation mechanism; features are then calculated for this data and plotted; experts peruse the

graphs for changes in feature patterns and make deductions about the underlying conditions (if

known). More recently, neural networks have been used to automate the process of finding

patterns and correlations.

An alternative approach has been taken in this thesis to better relate and quantify AE signals with

test parameters. This involves grouping features into ensembles that correspond to a particular

operating condition or failure mode. Thereafter, statistics describing the feature set are calculated

and trended. Correlations with underlying fault conditions are consequently easy to identify and/or

quantify.

PhD Thesis – Chapter 3, AE Signal Processing

Figure 3-1: PCI2 board block diagram (from [6]).

Types of acoustic emission data can be grouped into: (a) features that describe individual acoustic

emission bursts (or hits) and (b) features extracted about the background continuous signal (where

individual events are inseparable). In modern commercial-off-the-shelf (COTS) AE systems, both

feature groups are extracted by novel digital components, called Field Programmable Gate Arrays

(+s). These are programmed to extract a single feature from a single channel of filtered, digitized

AE data (eg. duration, maximum amplitude); multiple FPGAs are then used to extract a number of

features simultaneously. A block diagram of the COTS AE data acquisition board used in this

thesis (PCI2 by Physical Acoustics Corporation) is shown in Figure 3-1. With the use of FPGAs,

AE systems are theoretically capable of processing up to 20,000 hits per second. Whilst hit features

are extracted from the filtered waveform, continuous signal descriptors are generally extracted from

its envelope.

In addition to extracting signal features, some COTS AE equipment can stream a portion of the

filtered digitized waveform directly to the hard disk. The number of AE channels that can be

simultaneously streamed to disk depends on: (a) the sampling rate (typically 2-5MHz); (b) the

transfer rate of the associated computer bus (which is usually PCI for AE compatible data

acquisition boards); (c) the speed of the receiving hard-disk; (d) the resolution of the digitised signal

(typically 12 to 16 bit) as this defines the number of bytes used to store each numerical value in

memory; and finally (e) the software used to control the transfer.

When generic high-speed data acquisition boards are used (instead of COTS equipment), all

features are usually extracted from digitized waveforms in software. Although this facilitates greater

analysis flexibility, fewer hits can be processed than using state-of-the-art FPGA systems.

3-2


In this research, initial data was collected using generic hardware and custom-written software,

whilst later experiments were conducted with a COTS system. The former option was able to

stream four channels of simultaneous 12/16-bit waveform data at 5MHz, whilst the COTS system

was limited to 2MHz when streaming the same amount of data on similar PC hardware.

3.2 HIT PARAMETERS

Defining a burst as an event that exceeds a certain threshold, traditional features of AE signals used

in this work, as illustrated in Figure 2-2, include:

• Average repetition rate of bursts (interchangeably referred to as event rate, burst rate or hit

rate),

• Amplitude, usually measured in millivolts or dBAE (where 1dBAE=10-6 microvolts)

• Energy (when acquired by the COTS hardware this is called PAC Energy),

• Rise-time (time between first threshold crossing and maximum peak amplitude),

• Counts (number of times an AE burst crosses the threshold),

• Duration.

Values are generally trended to detect abnormalities or visualise rates of change. They may also be

displayed as histograms (number of events versus amplitude, energy, rise time etc), in which

changes in the shape of the curves are analysed; for example, a change from continuous to burst

emission will skew a Gaussian-looking histogram to the right.

Hit data features are a function of the type and value of the selected threshold. Thresholds can be

referred to as fixed, in which case they are set to a specific value for the duration of the test (eg.

40dBAE) or floating, where the threshold level is set as a defined amount (a fixed voltage, or fixed

number of standard deviations) above the background level. Fixed thresholds are typically used

when monitoring static equipment, whilst rotating machinery requires floating thresholds to avoid

swamping acquisition hardware when fluctuations in operating conditions cause the background

signal level to rise.

All experimental work collected for this thesis used floating thresholds. These were set at the

following levels, depending on whether generic or COTS systems were being used:

Generic: 3 standard deviations above the mean

COTS: ASL + 10 dBAE (ASL will be defined shortly)

3-3


3.3 CONTINUOUS SIGNAL PARAMETERS

Traditional AE features for measuring continuous AE signals include RMS, ASL and Absolute

Energy. The manufacturer of the COTS equipment also calculates a feature called Signal Strength.

These terms will be explained in the following sections. All continuous AE features are extracted

by FPGAs from scaled, filtered and enveloped AE signals and require a single input to define the

envelope’s time constant (defined below). Additional statistical and frequency parameters need to

be determined manually from digitized waveforms.

Unless otherwise stated, all features defined in the following sections were calculated for all

experimental data collected.

3.3.1 Traditional Time-dependant Parameters

Traditional analogue demodulation removes high frequency oscillations (the unwanted fast

changing components) from a signal by rectifying it and then applying a time constant that

“smooths out the bumps”. A diagram of this process was given in Figure 2-5. The overall

magnitude and energy of the envelope is related to the magnitude and energy of the raw time signal.

Although significant bursts can be identified, the exact values for arrival times, rise-times and

durations are affected by the time constant; the larger the time constant, the more the signal is

smoothed out and the more envelope values will differ from their true values. Analogue

enveloping can be performed with diodes, peak-hold circuits or operational amplifiers, but the latter

option has a superior dynamic range. In modern COTS systems, FPGAs perform equivalent

processing.

Energy (EN), Power (PN) and RMS (ψN) can be calculated from the digitized signal or envelope (x[N]):

1N2

0

[ ]Ni

E x i=

= ∑−

(3.1)

1

2N

N sE f x i−

= = ⋅0

[ ]Nis

PN t N =⋅ Δ ∑ (3.2)

1 2 1 211 NP −⎡ ⎤ 2

0

[ ]NN

is

x if N

ψ=

⎡ ⎤= = ⋅⎢ ⎥ ⎢ ⎥

⎣ ⎦⎣ ⎦∑ (3.3)

where N is the number of samples acquired from the voltage signal (direct or enveloped) and fs is

the sampling rate for the waveform or digitized envelope.

3-4


Enveloping is particularly

useful if the AE is used as a

carrier for a lower frequency

fault, such as in the case of

bearing or gear defects. By

demodulating the signal, the

higher frequency components

are removed, whilst preserving

the repetition rates of the

underlying faults. In Figure

3-2 demodulation of an AE

signal (shown in A) collected

from a centrifugal pump

highlights a repetitive fault

(envelope is shown in B),

corresponding with the pump’s

speed of rotation. An FFT was

then performed on the

demodulated output to identify

or confirm fault frequencies

(given in C). Many burst features can also be extracted from the envelope, although absolute

magnitudes and rates of occurrence are highly dependant on the time constant (which in analogue

systems depends on the circuitry). As described in Chapter 2, most AE devices for monitoring

machinery use some form of demodulation to extract AE data in real-time.

(A) (B) (C)

Figure 3-2: (A) A signal, (B) its envelope calculated using a Hilbert transform, and (C) the resulting FFT.

Demodulation can also be undertaken by applying a Hilbert transform to the digitised signal. The

discrete Hilbert transform of a discrete signal, x[n], given by [70]:

[ ]{ } [ ],

ˆdi i N

x n x nN iπ =−∞ ≠

= = ⋅1 [ ]x i+∞

−∑H (3.4)

where N is the number of samples and [ ]x i is the scaled voltage at some discrete sampling interval

‘i’ in volts.

Applying this transform causes a 90 degrees phase shift, so in complex terms the original signal

(assuming no negative frequency components) can be written in the form:

3-5


[ ] [ ] [ ] [ ] [ ] (3.5) ˆ j nz n x n jx n E n e ϖ= + =

The envelope of z(n) is then determined by:

[ ] [ ] [ ] [ ]2 ˆ 2E n z n x n x n= = + (3.6)

Squaring the envelope of the signal, E[n], gives its instantaneous power.

The Hilbert transform of a discrete signal is usually calculated by multiplying the signal’s discrete

Fourier transform by sgn(ω) [70]. Therefore, provided that the signal x(n) is band limited3, its

Hilbert transform, , will have the same amplitude spectrum and auto-correlation function as

the original signal. Unlike a Fourier transform however, a Hilbert transform remains time

dependant.

x̂(n)

In this thesis work, demodulation was used in two ways:

(i) COTS hardware used analogue demodulation to determine continuous AE features of an

AE signal. A time constant of 200ms was typically set.

(ii) Envelopes were used in custom LabVIEW routines that automatically located and

quantified bursts in AE waveform files.

3.3.1.1 PAC Terminology

COTS equipment used (from Physical Acoustic Corporation) refers to continuous AE descriptors

as time dependant features (because they rely on a time constant value from the user). These include

RMS (as defined previously), Absolute Energy, PAC Energy, Signal Strength and Absolute Signal

Level (ASL).

Signal strength is defined as the absolute value of a detected AE signal. It has the same mathematical

description as energy, given in Equation 3.1, but is calculated from the enveloped signal.

PAC Energy also has the same mathematical definition as signal strength and generic energy, but

with reduced resolution. It is a 2-byte feature derived from the “integral of the rectified voltage

3 Given most discrete signals are passed through an anti-aliasing filter prior to digitisation, this assumption

is generally valid.

3-6


signal (i.e. the envelope) over the duration of the AE hit” [5] (rather than 8-bit floating point

numbers used for other energy features).

Absolute Energy is not determined from the envelope, but rather is a measure of the true energy in

the voltage signal, and thus is equivalent to Equation 3.1.

Absolute signal level (ASL) is defined as the time averaged amplitude of the envelope4. In

mathematical terms:

1N

0

[ ]i

E iASL

N==∑

−

(3.7)

where E[i] is the signal envelope in dBAE. The COTS hardware manufacturers calculate floating

thresholds by adding a user-defined fixed amount (typically set to 10dBAE in this work) to the ASL.

Unfortunately, not all features can be accurately determined from a signal’s envelope. Variance and

kurtosis, as well as all frequency descriptors, must be calculated from the filtered, digitized

waveform.

3.3.2 Calculating Statistical Parameters from Raw Waveforms

A discrete sample record is defined as a set of (usually equally spaced) samples collected over a

finite time interval. The duration of the sample record is therefore equal to the time between

successive samples (which equals the inverse of the sampling frequency, fs) multiplied by the

number of samples, N.

Mean, μN, of a signal is a measure of its DC offset and is calculated by:

[ ]11 N

μ−

= ⋅0

Ni

x iN =

∑ (3.8)

where N is the number of samples and [ ]x i is the scaled voltage at some discrete sampling interval

‘i’ in volts.

4 This is deduced from the manufacturer’s literature as no clear definition is given.

3-7


Variance, σ2N (=Std deviation2, σN) is the second statistical moment and indicates the deviation of the

signal around the mean. It is therefore indicative of the spread of burst amplitudes that are smaller

and larger than the baseline level. As the signal mean approaches zero, the standard deviation

approaches the RMS.

[ ]{ }1 22 1 N

Nσ μ−

= ⋅ −0

Ni

x iN =

∑ (3.9)

Kurtosis, KN, (fourth statistical moment) provides an indication of the proportion of time that is

spent at higher amplitudes and is described in its non-dimensional form.

[ ]{ }1 41 N

Nμ−

= ⋅ −40

NiN

K x iNσ =

∑ (3.10)

Kurtosis is commonly calculated by vibration analysts when highly impulsive signals with very small

amplitudes are expected; these types of signals are an early indicator of bearing faults. In AE work,

Kurtosis could potentially describe the type of AE signal being generated; the more impulsive a

signal, the less time is spent at higher amplitudes and the higher the Kurtosis value. On the other

hand, a Gaussian distribution, by which many continuous AE signals can be described, has a

Kurtosis value of three (3). A significant increase in Kurtosis (1-2 orders of magnitude) could

indicate a change from continuous to burst emissions.

Crest Factor, CFN, is the ratio of signal peak to RMS and is also used in vibration analysis for the

detection of highly impulsive bearing faults.

( [ ])max x n

NN

CFψ

= (3.11)

The crest factor of a sine wave is 1.42, whilst a Gaussian distribution is tends to have a value of 1.5.

3.3.2.1 Dealing with very large sample sizes

The equations listed in the previous section require that the entire signal array, x[n], is processed in

one pass. As signals get longer (much greater than 10,000 samples) these equations become more

highly memory intensive, causing computation times to increase substantially. Therefore, as part of

this thesis, the author derived algorithms to calculate the statistics for long signals from smaller

subsets of the signal. The final equations are listed in this section, with full derivations given in

Appendix A.

3-8


If a signal is comprised of N samples, these can be broken up into M sub-sections with S samples

in each. Values for mean, variance, kurtosis and crest factor can be calculated from the following

equations:

1M

μ−

0

[ ]Sk

N

k

Mμ ==

∑ (3.12)

where

1S −

0

[ ][ ] i

S

x ik

Sμ ==

∑ (3.13)

11 1 2M S− −

2

0 0

1 [ ]k i

N

x iS

Mψ = =

⎡ ⎤⎧ ⎫⋅⎨ ⎬⎢ ⎥

⎩ ⎭⎢ ⎥=⎢ ⎥⎢ ⎥⎣ ⎦

∑ ∑ (3.14)

2 2 2 Nσ N Nψ μ= − (3.15)

1 1 1 14 3

3

M S M S

N N N2 2 40 0 0 0

4

[ , ] [ , ]1 4 6k i k i

N NN

x i k x i kK

N Nμ μ ψ μ

σ= = = =

⎪ ⎪⎪ ⎪= ⋅ − ⋅ + −⎨ ⎬⎪ ⎪⎪ ⎪⎩ ⎭

∑∑ ∑∑− − − −⎧ ⎫

(3.16)

where N is the total number of points in the sample record, S is the total number of points in a

section of the sample record, M is the number of sections (N=M × S), i is the position in the

sample within the sub-section (between 0 and M) and k is the number of the sub-section (between

0 and S).

Optimal block size for best performance was determined (by trial and error) to range between 1024

and 4096 points (in powers of 2).

3.3.3 Frequency Parameters

Traditionally, frequency spectra have not been used in acoustic emission analysis because AE is

generally considered to be broadband (energy is distributed across a wide range of frequencies) or

unduly influenced by the natural frequency of the acoustic emission sensor. However, when

looking for rotating equipment faults, demodulated frequency spectra can be useful, particularly if

signals have been collected with wideband sensors (that have no strong resonant frequency).

3-9


Fc Fmin-Fmax

1/3 Octave bands Octave bands

64kHz: 57.0-71.8kHz 45.3-90.5kHz

81kHz: 71.8-90.5kHz

101.6kHz: 90.5-114.0kHz

128kHz: 114.0-143kHz 90.5-181.0kHz

161.3kHz: 143.7-181.0kHz

203.2kHz 181.0-228.1kHz

256kHz: 228.1-287.4kHz 181.0-362.0kHz

322.5kHz: 287.4-362.0kHz

406.4kHz: 362.0-456.1kHz

512kHz: 456.1-574.7kHz 362.0-724.1kHz

645.1kHz: 574.7-724.1kHz

812.8kHz: 724.1-912.3kHz

1024kHz: 912.3-1149kHz 724.1-1448kHz

Table 3-1: Octave and 1/3-octave band frequencies.

Additionally, some researchers have observed that certain types of faults change the relative

distribution of energies between different frequency bands [79].

After Fourier transformation, the most widely used methods of analysing energy in different

frequency bands are octave band analysis and 1/3 octave band analysis, as described in ANSI

S1.11-1986. These involve filtering the signal into defined frequency ranges and calculating total

power (energy per unit time) for each band. Octave band energies can also be determined by

summing the associated third-octave-band energies. Relevant centre frequencies (Fc) and their

associated frequency limits (Fmin - Fmax) for acoustic emission analysis are listed in Table 3-1.

In this work, octave band energies were determined for the first set of experimental data (collected

using generic hardware), whilst third-octave band energies were calculated for later experimental

data sets (using COTS equipment). Initial data could have been reprocessed if necessary, but this

was not required.

3-10


3.3.4 Joint-time-frequency analysis

Fast Fourier Transformation (FFT) destroys all time information: a FFT calculation returns all

frequencies that are present in the signal irrespective of whether they exist for its entire duration or

are only transitory. Therefore, in general, FFT analysis should only be performed on stationary and

ergodic signals.

Burst AE signals are neither ergodic, nor stationary. In many cases, nor are continuous emissions.

Therefore over the last decade many researchers have attempted to characterise AE signals by more

advanced techniques such as wavelets [43, 62, 87, 119, 125, 127], Wigner-Ville functions [32], short-

time Fourier Transforms (STFT) [131] and Kalman filtering [88], of which wavelets seem to be the

most popular. Known collectively as “Joint Time Frequency Analysis” (JTFA), these algorithms

analyse how frequencies change with time and are more computationally intensive than standard

frequency analysis.

For all JTFA techniques, the amount of time-frequency information being analysed (known as the

time-bandwidth product) is limited by the uncertainty principle [13]:

1

4t f

πΔ ⋅ Δ ≥ (3.17)

This relationship necessitates that for a given sample set, any improvement in time resolution can

only occur as a reduction in frequency resolution and vice versa.

STFTs can be thought of as simply an extension of traditional FFTs; the time-frequency space is

effectively divided into a number of equal blocks and then a windowed Fourier transform is

performed on each block. Blocks are overlapped slightly to overcome errors caused by applying

the window function. Results are typically displayed on a contour plot, called a spectrogram, or a 3-

D x-y-z (time-frequency-amplitude) plot. Unfortunately, ultimate resolution is limited by equation

3.17, so the only way to increase temporal resolution is to decrease frequency resolution and vice

versa. STFTs are useful because they are easy to compute and interpret.

Continuous-time wavelet transforms (CWT) on the other hand, are special mathematical functions

that split a signal into non-uniform sections (windows), localised in time and scale [94]. Scale is

related to, but not synonymous with, frequency. To determine values for each window, wavelet

transformation dilates and translates a basis function (called a mother wavelet). The CWT is simply

the cross correlation of the time signal, x[n], with this dilated and translated wavelet. The results of

this process are then displayed as a spectrogram or xyz plot.

3-11


A wavelet transform has the (discrete) form:

11

2N

x i bψ−

−,

0

[ ]a bi

W a x ia=

−⎛ ⎞= ⋅ ⎟⎜⎝ ⎠

∑ (3.18)

where ( )tψ is the basis function, a is the scale and b is the translation in time.

The centre frequency of the mother wavelet can be determined from its frequency response, but

can be roughly approximated as the Nyquist frequency (half of the sampling rate in Hertz).

Thereafter the centre frequency decreases (is translated) by a factor of two.

Energy in a particular wavelet band, or band energy (Ea), can be calculated by:

2x

bE W= ∑ ,a ab

(3.19)

As wavelet transformation preserves the energy of the original signal, total energy, E, can be

determined by integrating (or in discrete terms, summing) the individual bands.

2xWE ,

2a baE

a a= =∑ ∑∑ (3.20)

CWT generally has superior resolution over STFT because of its non-uniform division of time and

scale: at higher scales, signals are divided into very short blocks of time, allowing identification of

very quick, impulsive events, whilst at lower scales they are highly localised in scale, facilitating

identification of specific frequencies. This makes them incredibly useful for identifying sharp (i.e.

high frequency) bursts and discontinuities that may be masked by larger stationary signals or noise

(low frequency). Unfortunately, a CWT generally takes more time to calculate than an STFT.

A computationally lighter alternative to CWT is the discrete wavelet transformation (DWT), called

the Mallat algorithm, which uses special orthogonal filter banks to divide the signal into scale bands.

Many wavelets also facilitate perfect reconstruction of the original signal from its component parts

which means that any signal can be split up into sections, unwanted or unused parts of the signal

then removed, and the signal rebuilt without affecting the important or desired information. This is

not possible with traditional band-pass filters (i.e. octave band analysis, STFT). Discrete wavelet

analysis and its application to signal denoising is discussed further in Chapter 4.

In this thesis, discrete wavelets were used to separate AE bursts from continuous emissions. A

flowchart of the process used to extract denoised AE bursts is given in Appendix C.1.

3-12


National Instrument’s Signal Processing Toolbox for LabVIEW includes various algorithms for

denoising and detrending signals depending on the type of background noise that is

“contaminating” the signal [94]. Unfortunately, there are no clear guidelines as to which wavelet

should be applied. Selection is data dependant and therefore the most appropriate wavelet function

can only be ascertained through trial and error. Most prior AE work used various Daubechies [64,

77, 87, 127] or Haar wavelet functions for discrete wavelet processing [92], and Gabor (otherwise

known as Morlet) or Meyer continuous wavelet functions [43, 117]. Some guidelines developed

for this research are presented in Chapter 4.

Unless otherwise stated, in this thesis waveform signals were processed by applying either 0.1-

1MHz or 0.1-0.4MHz bandpass digital filters and then denoised using a Haar wavelet function.

Denoised signals were subsequently enveloped by the Hilbert function to simplify automated burst

identification. However, burst parameters (peak, RMS etc.) were extracted from the denoised

signals and not their envelopes.

The Haar mother wavelet has the form:

1 0 0.5t+ ≤ ≤⎧( ) 1 0.5 1

0w t t

otherwise

⎪= − < ≤⎨⎪⎩

(3.21)

( )For daughter wavelets: , ( ) 2 2aa bw t w t b for all a,b= ⋅ − (3.22)

3.4 ENSEMBLE STATISTICS

An ensemble is a collection of sample records. In signal processing of time based data, statistics

can either be determined on the sample records directly, in which case the statistics describe

features of the underlying signal at a temporal location, or on an ensemble, which then describes

how the signal varies across the sample records.

In this thesis, ensemble averages are calculated for all AE features. A bar over a parameter denotes

an ensemble, rather than sample statistic (i.e. mμ is the sample record mean based calculated on n

samples, whilst mμ denotes the ensemble mean calculated on m ensembles).

3.4.1 Confidence limits

A signal is described as being stationary if its sample record statistics are temporally static. In

mathematical terms:

3-13


( ) [ ] ME x n = μ (3.23)

( ) 2 2ME σ σ= (3.24)

where M is the number of sample records in the ensemble.

A signal is termed ergodic if any single ensemble statistic can be used to describe the signal. Acoustic

emission signals are rarely ergodic and often non-stationary, especially if signal blocks are short

(<100,000 samples). Therefore, for ensemble averaging to be useful, confidence limits are required

to quantify the spread of values that make up the sample records in the ensemble.

All statistical equations quoted so far assume that values being analysed (the underlying data in the

case of sample record statistics, or the statistics themselves in the case of ensemble averages) vary

according to a normal distribution. Based on the same assumption, confidence limits for the

respective ensemble means and ensemble variances can also be established.

Two sided 100%(1-α) confidence limits on the mean of a normal distribution, with unknown

population variance can be determined from the t distribution:

/2, 1 /2, 1M M M M MM

t tM M

α Mαμ σ μμ− σ−− +

≤ ≤ (3.25)

where Mμ is the mean of the ensemble statistics, S is the standard deviation of the ensemble

statistics, M is the number of records in the dataset and tα/2,M-1 is the t distribution with M-1 degrees

of freedom for the desired confidence interval.

The less variable the statistics between each sample set, the tighter the confidence limits.

Conversely, large confidence limits indicate significant variability in the statistics between measured

features. This is either attributable to an insufficient number of features calculated, or due to an

increased variability in the underlying data as may be caused by very impulsive, highly non-

stationary acoustic emission signals.

In this thesis, unless otherwise stated, line graphs show the ensemble mean and 90% confidence

limits of a particular AE feature. As baseline levels vary significantly between different pumps

and/or speeds, ensemble averages are normalised by the mean ensemble value that corresponds to

particular condition (eg. BEP or max-NPSHA).

3-14

4. DDEENNOOIISSIINNGG

AAEE DDAATTAA 44 The aim of this chapter is to present methods of denoising various forms of AE data so that

information pertinent to unfavourable operating conditions or incipient mechanical failure can be

detected and accurately quantified.

Abstract

Acoustic emission (AE) data collected from operational equipment comprises of genuine

information, pertaining to AE sources of interest, and noise. A variety of signal processing

techniques have been developed over the past 20 years to denoise condition-monitoring data, with

varying degrees of success when applied to AE analysis. This is, in part, because what constitutes

noise and what is regarded as genuine data changes from one application to the next; thus it is

impossible to develop a single set of processing techniques that will be applicable in all situations.

Consequently, publications relating to noise management tend to be application specific, and most

relate to traditional AE applications such as crack monitoring. To simplify the process of acquiring

useful AE information from rotating machines, this paper presents examples of how hit filtering,

averaging and wavelet denoising were used by the author to improve the signal to noise ratio of AE

data collected from centrifugal pumps.

4.1 INTRODUCTION

Post processing techniques are typically applied to extract either the source signal of interest, or

some characteristic of this signal, from waveform data or a feature-set. Assuming that noise is

4-1

PhD Thesis – Chapter 4, Denoising AE Data

defined as any unwanted part of the data set, in centrifugal pump monitoring (and presumably other

rotating machinery as well), it has been identified by the author as originating from:

(a) Residual, steady background noise from electronic components;

(b) Aperiodic transient events, such as valve activity or EMI/ RFI;

(c) Quasi-periodic transients events that occur approximately the same interval apart, such as

rubbing of seals or bearing faults;

(d) Low frequency modulation of the background continuous AE level, also common in

reciprocating machinery signals or due to misalignment;

(e) Genuine AE from mechanisms not of interest to the analysis being performed.

In many cases, noise signals appear very similar to genuine acoustic emissions. Therefore it is not

unusual to spend considerable time and effort analysing data that may in fact have no bearing to the

conditions being analysed. Although good acquisition practices can reduce the risk of noise

infiltration, there is no way to eliminate it entirely. Therefore methods to identify noise during the

testing process, and to reduce its consequences to subsequent analysis, are required. To this end, a

variety of techniques were trialled as part of the author’s PhD research on centrifugal pump

monitoring and the results of this process are presented in the following sections.

Unless otherwise explicitly stated, the author collected all AE data, used to illustrate denoising

techniques, from centrifugal pumps; no simulated sources or artificially generated signals are used.

4.2 HIT FILTERING

As the name suggests, hit filtering involves removing data points (normally associated with features

of discrete AE events) that have certain characteristics from the analysis set; it is used extensively in

structural AE testing for separating different generation mechanisms (eg. fibre breakages and

delamination in composite pressure vessel monitoring). Hit filtering can be applied to the feature

set at any time, including during pre-processing.

This type of filter has obvious denoising applications when features of the noise bursts differ from

genuine AE bursts. For example, it is often possible to remove EMI/RFI by filtering on total

duration, rise-time, amplitude and frequency concurrently because EMI/RFI tends to have very

high amplitudes, very short durations and broad frequency spectra.

Similarly, frictional sources and fluid noise are separated from propagating cracks by amplitude-

duration characteristics [10]. Based on this, a “Swansong II” filter was presented in references [35,

4-2


36] to extract crack-generated AE from fluid noise when verifying tank and pressure vessel

integrity; this filter removes all data within 0.5 seconds of a “telltale” hit, defined by the following

criteria:

Amplitude (in dBAE) Duration

< 5dB above threshold AND >2000 μs

OR <10dB above threshold AND >3500 μs


Figure 4-1 shows how hit filtering can be used in a traditional AE application (crack propagation in

a pressure vessel). Red and green points signify AE hits measured during a pressure test of a steel

cylinder. Different types of noise were removed by identifying hits with certain combined

amplitude, duration or energy characteristics. These are shown as red points. (The coloured lines

indicate probable varieties of noise.) Remaining hits are shown in green; these were subsequently

analysed to ascertain the nature, severity and location of the potential crack. In this example,

probable EMI has not been removed by hit filtering and thus needs to be removed by alternative

processing.

Figure 4-1: Different amplitude-duration characteristics of AE hits. Data has been processed by a Swansong II filter but probable EMI noise still remains. (PUMAnalysis software courtesy

of Imes Group Ltd.)

SATURATION: removed by filtering hits

with Duration ≥106

Probable EMI noise

FLUID NOISE: removed by

Swansong filter

FRICTION NOISE (removed by

Swansong filter)

Either side of a tell-tale hit

SMALL EVENTS: removed by filtering hits with Amplitude <46dB

4-3


Hit filtering is also useful in denoising AE data acquired from rotating machinery and has been

used successfully by the author in two ways. In machinery applications, the background AE level

can fluctuate during a single shaft rotation or due to changes in operating conditions. Although a

floating threshold is used to avoid collecting hits when this background AE level increases, it may

not be possible to adjust this threshold sufficiently quickly, resulting in a large number of false hits.

Using the aforementioned Swansong criteria, noise events are removed very effectively.

Additionally, hydrodynamic machines have a number of AE generation mechanisms. Certain

mechanisms are typified by defined, short duration, high amplitude AE bursts (such as incipient

cavitation bubbles), whilst others have protracted, quasi-continuous profiles (very-low flow

pressure surging). The Swansong criteria can therefore be used to segregate both types of events,

thus allowing the severity of each underlying generation mechanism to be quantified separately.

This is exemplified in Figure 4-2. Results from a pump test showed no apparent relationship

between the AE feature and normalised flow. However, by applying the telltale-hit criterion, quasi-

continuous events (which constituted most of the data points) were removed and the trend became

evident.

Minor improvements were observed by extending the filter to segregate additional higher amplitude

bursts as per the following (complete) criteria:

Amplitude (in dBAE) Duration

< 5dB above threshold AND >2000 μs




OR < 25dB above threshold AND >6500 μs




Unlike the original Swansong II filter, when applied by the author to AE signals collected from

centrifugal pumps, only hits meeting these criteria were removed and not those 0.5 seconds either

side of a telltale event.

4-4


Figure 4-2: Improved trending is obtained by using the telltale-hit criteria to segregate bursts from residual changes in the continuous emission level. (A) Shows the normalised AE feature versus flow determined from the complete unfiltered dataset, whilst (B) shows the same normalised feature from a reduced dataset determined by removing telltale hits. (Vertical lines indicate 90% confidence limits, which increase because the number of hits used to determine the sample mean reduce. Very few hits remain at normalised flows between 0.9 and 1.1.)

4.3 FREQUENCY FILTERS

If noise and AE data signals differ substantially in their frequency content, and these bandwidths do

not overlap, it is possible to remove noise by applying frequency filters; these may be digital or

analogue. Frequency filters are particularly effective at removing low frequency noise emanating

from mechanical sources (eg. minor misalignment or unbalance) that may be superimposed on

higher frequency AE, as shown in Figure 4-3.

Unfortunately, many types of noise overlap the typical AE bandwidth and therefore cannot be

removed by frequency filters. An example of a signal corrupted by noise emanating from a VFD is

given in Figure 4-3. Although the low frequency sinusoidal noise can be removed, the higher

frequency noise associated with the impulsive bursts ranges from 300kHz to above 1.2MHz. Even

limiting the bandpass range to 10-300kHz does not eliminate the noise bursts, indicating that it is

broadband (see Figure 4-3d) across the AE monitoring bandwidth.

As frequency filters merely attenuate the out-of-band components rather than extract them from

the signal, particularly large signals will possibly not be removed entirely. EMI and RFI bursts are

typical noise events that often remain after frequency filtering.

4-5


4.4 AVERAGING

Averaging is commonly employed in vibration monitoring to reduce the effects of random noise

and accentuate stationary components, but rarely if ever applied to AE signal analysis. In AE

testing of rotating machinery, stationary components may be due to incipient mechanical faults, but

are more commonly caused by noise. Therefore averaging can be used to detect certain noise

artefacts in AE signals, such as those associated with exceeding the frequency-amplitude

characteristics of on-board amplifiers. These harmonics cannot be removed by post-processing, so

the root cause (generally too much gain) must be rectified immediately, otherwise data will be

completely unusable.

The relationship between random error in the power spectrum (i.e. non-stationary signal

components), εr, and the number of averages, n, is given by . Therefore, doubling the

number of averages effectively quadruples the detectability of stationary spectral peaks.

0.5r nε −=

Figure 4-4 shows the effect of increasing the number of averages on the detectability of noise

harmonics infiltrating the acquisition system. Without averaging (Figure 4-4A), it is very difficult to

identify whether any stationary components are present in the signal. (The 100kHz local maximum

is a feature of the sensor’s response.) By increasing the number of averages, harmonics become

obvious; these are first noticeable in the region between the Nyquist frequency (half the sampling

frequency) and filter cut-off (Figure 4-4B), but with more averages become evident across the 0-

1MHz monitoring bandwidth (Figure 4-4C).

A large number of samples are also required (per average) to ensure that frequency resolution is

adequate to isolate individual frequencies: sf f NΔ = where N is the total number of samples per

average, Δf is the frequency resolution and fs is the sampling frequency. In Figure 4-4 C, 8192

samples were used in each average and clear spectral peaks are seen in the final graph. Figure 4-4

D represents the same data but this time only 1024 samples were used in each of the 64 averages;

this time individual harmonics are much harder to identify at all but the very highest frequencies.

4-6


(A)

VFD Noise bursts

Frequencies relating to VFD Noise bursts

(B)

(C)

VFD Noise bursts still remain

(D)

Figure 4-3: Effect of frequency and wavelet filters on pump AE signal corrupted by a 100Hz frequency sinusoidal signal and high frequency impulsive noise, latter emanating from a VFD. (a) Original time signal, (b) FFT, (c) 10kHz –1000kHz bandpass filtered and (d) 10kHz – 200kHz bandpass filtered.

4-7


Harmonics difficult to spot amidst the noise

Noise decreases and harmonics become apparent

Extent of harmonic infiltration is now

clearly visible

Fewer samples mean harmonics are harder

to identify

Anti-aliasing filter cutoff

(A)

(B)

(C)

(D)

Figure 4-4: Effect of averaging on identification of harmonic noise elements hidden amongst a broadband continuous AE pump signal with (A) 8192 samples and no averaging, (B) 8192 samples and 8 averages, (C) 8192 samples and 64 averages, and (D) 1024 samples and 64 averages. Sampling rate was 5MHz.

4-8


4.5 WAVELET DENOISING

4.5.1 Continuous Wavelets

Wavelets are special mathematical functions that identify frequencies present in a signal and when

they occur (see Figure 4-5). Using basis functions called mother wavelets, continuous wavelet

transformation (CWT) works by a process called shifting: the amplitude and frequency of this basis

function is scaled (resulting in daughter, or baby, wavelets) before being applied to incremental time

blocks. Signals are divided into smaller frequency blocks at low frequencies, and shorter time

blocks at high frequencies. Consequently, wavelet transformation can separate sharp (i.e. high

frequency) bursts and discontinuities from larger, lower frequency stationary signals and/or

background noise. More details have been presented in Chapters 3 and 5.

Continuous wavelets, as shown in Figure 4-5, are used to display frequency-time changes

simultaneously and can help identify noise events of very short durations with unusual frequency

contents. Unfortunately, they are computationally intensive, present results as a visual 3-D graph

(which may or may not have adequate resolution to see the intermittent information) and are

therefore only applicable for manual interpretation.

In this thesis, continuous wavelets were used for visual identification of noise sources and for

general qualitative evaluation of signal frequency contents.

Figure 4-5: Examples of changes in spectral composition that can be observed in a typical CWT plot (Morlet wavelet).

4-9


4.5.2 Discrete Wavelets

A computationally lighter and more useful alternative to CWT is discrete wavelet transformation

(DWT), which uses orthogonal digital filter banks to divide the signal into scale bands. These can

facilitate perfect reconstruction of the original signal from its component parts, which means that

any signal can be split up into sections, unwanted or unused parts of the signal removed, and the

signal then rebuilt without affecting the important or desired information. Perfect reconstruction is

not possible with traditional band-pass filters (i.e. octave band analysis, STFT).

Wavelet denoising works by equating or scaling all coefficients in each band, which are less than

some threshold value, to zero. Thresholds can be set manually, or by sophisticated algorithms that

self-calculate an appropriate value to separate individual impulsive events from noise. Various

algorithms are available for determining which components of the signal should be retained or

rejected, and their selection depends on the type of noise present in the signal.

For a given noise distribution, wavelet-denoising performance depends on:

(a) the type of wavelet,

(b) the algorithm used to calculate thresholds,

(c) signal bandwidth,

(d) the total number of samples being denoised at one time,

(e) the number of bands, and

(f) the number of samples in each band.

Unfortunately, deciding on the most appropriate wavelet for the given noise environment is not

straightforward. This may explain the inconsistent results reported from this process [62, 80, 81,

87, 105, 119, 125, 127].

Fortunately, trialling different wavelets and denoising parameters has been simplified by the release

of advanced toolsets for common engineering and DAQ programs such as LabVIEW, Matlab and

Mathematica. Yet even with these tools, the process of determining the best approach through trial

and error is a time consuming task. As part of this PhD research, a number of guidelines were

developed for using wavelets to denoising AE signals collected from centrifugal pumps and these

are presented in the following sections.

4-10


(A)

(B)

Figure 4-6: Effects on Haar wavelet denoising to 12 levels on two different types of noise: (A-B) EMI/RFI; (C-D) VFD noise. Due to the impulsive nature of the noise signals, wavelet denoising is not particularly effective in either case.

(C)

(D)

4-11


4.5.2.1 Type of Noise

Discrete wavelets are not very effective at denoising signals when the noise is highly impulsive, such

as EMI and interference generated by variable frequency drives (see Figure 4-6). In Figure 4-6B an

EMI spike is perfectly retained by the denoising process, as are the repetitive spikes indicative of

VFD switching interference, shown in Figure 4-6D.

Many thresholding rules assume that the noise has a Gaussian distribution, with a standard

deviation equal to 1. When this is not the case, the threshold must be multiplied by an appropriate

factor. If the noise is Gaussian, but has a standard deviation that is not equal to one, the scaling

factor for all levels can be determined from the first level. For non-Gaussian noise (eg. background

AE from typical hydraulic sources in a pump) scaling for each level must be determined separately.

There also appears to be no disadvantage to this approach when denoising Gaussian infiltrations.

4.5.2.2 Type of wavelet selected

When applying wavelets to denoising AE from centrifugal pumps, Haar wavelets were most

effective at retaining AE bursts, whilst removing virtually all intervening background noise, due

their more impulsive profile (see Figure 4-7). A review of the literature also indicates that lower

order Daubechies (2-3) and Haar wavelets are more commonly applied to denoising AE signals [38,

40, 67, 106, 124] than other wavelet functions, although higher-level Daubechies wavelets [45, 64,

93] have also been used by some researchers. In the author’s research, Haar wavelets were chosen

because they were most effective at retaining individual bursts and removing all noise between

them, which simplified the subsequent process of automated burst identification.

4.5.2.3 Threshold algorithms

Best results for denoising pump AE signals were obtained by using LabVIEW’s heursure option [94].

This uses a Stein's Unbiased Risk Estimate [30] for waveforms with high signal-to-noise-ratios,

whilst at low signal-to-noise-ratios, when the first option becomes very noisy, thresholds are

calculated from √(2logN) (where N is the number of samples). Also tested, and found to be equally

successful was an algorithm presented in [46] for denoising digestive sounds. The effect of

different thresholding algorithms on a typical pump AE signal is illustrated in Figure 4-8.

4.5.2.4 Effect of bandwidth

Filtering the signal prior to denoising can radically alter the denoised results because the wavelet

filter has a smaller number of data-rich bands from which to identify discrete signal changes. In

Figure 4-9 a time signal (shown in A) is denoised by a Haar wavelet. In (B) no digital filter is

applied to the time signal, whilst in (C), a 100-1000kHz bandpass filter is applied prior to denoising.

4-12


(A)

(B)

(C)

(D)

Figure 4-7: Effect of wavelet type on denoised signal. (A) Original waveform, (B) Denoised with a Haar wavelet, (C) Denoised with a Daubechies 6 wavelet and (D) Denoised with a Biorthogonal 1-3 wavelet.

4-13


(A)

(B)

Figure 4-8: Effect of threshold estimating algorithm on low signal to noise ratio signal (VFD noise): (A) original signal; (B) Using Stein’s unbiased risk estimate, (C) Using √(2logN), (D) using an alternative algorithm from [46].

(C)

(D)

4-14


(A)

(B)

(C)

Figure 4-9: Effect of filtering on the final denoised signal. (A) Original signal, (B) After denoising by Haar wavelet; (C) 100-1000kHz digital filter applied prior to denoising by Haar wavelet.

In the latter bandwidth VFD noise dominates and is retained by the wavelet algorithm, whilst (by

comparison) less impulsive, genuine AE data is removed.

4.5.2.5 Signal length

Signals must be sufficiently long so that statistical properties of the background noise can be

properly estimated. Windows need to be much longer than individual discrete bursts, and therefore

in acoustic emission work, the author has found that waveforms must have at least 16384 samples

(fs= 2-5MHz) for wavelet denoising to have a reasonable likelihood of success.

4-15


(A) (B)

Figure 4-10: (A) Overall signal RMS levels as they change with flow and (B) separated continuous and discrete RMS parts as they change with flow.

4.5.2.6 Denoised features

In addition to quantifying the energy in each wavelet level, the total energy in the post-denoised

signal is a measure of the severity of discrete events. By subtracting this value from the energy in

the original signal, the level of background noise can also be quantified. An example of this

approach is given in Figure 4-10 for data obtained during pump testing. In Figure 4-10B wavelet

denoising has separated the energies associated with the discrete and continuous signal parts,

enabling both to be trended separately; in this case, the more significant trend associated with burst

activity was not detectable in the original signal due to its much smaller total energy contribution.

4.6 CONCLUSIONS

This chapter was by no means a comprehensive discussion on the AE signal denoising. Not only

have commonly applied noise reduction methods such as adaptive filtering and neural networks

been omitted, so too have less well known processing algorithms that may be incredibly useful in

select applications. Instead, only those techniques found useful to the problem of noise reduction

in AE signals collected from centrifugal pumps, have been examined. However, as was discussed

herein, there are many types of noise for which these techniques are not suitable. The most

important to the author’s application is eliminating noise from high frequency, high power,

oscillating electro-magnetic fields, such as those created by the coupling of variable frequency

drives with electric motors, for which no technique trialled to date has been successful. Work in

this area is ongoing. Nevertheless, many other types of noise observed when undertaking pump

monitoring are removed by wavelet denoising in the case of waveform data, and advanced hit

filtering in the case of feature-sets. Coupled with better onsite noise identification and

management, sufficiently noise-free data can be extracted to make meaningful diagnoses of

4-16


conditions within centrifugal pumps. This will be the subject of further discussions in Chapters 8

and 9.

4-17

5. DDAATTAA

MMAANNAAGGEEMMEENNTT

Sikorska, J., Kelly, P. and Pan J.,

Mechanical Systems and Signal Processing,

(2005) In press.

55The aim of this chapter is to present a system that can administer and process, using techniques

discussed in previous chapters, an unlimited amount of AE data efficiently, effectively and with

rigorous quality assurance.


The use of acoustic emission testing in continuous monitoring applications presents the analyst

with the problem of managing substantial volumes of generated data. To better administer this

information, a system has been developed to catalogue, store and process AE data. Based on a

relational database, it is independent of the software and hardware used to acquire the data. This

paper describes the system and how it is used. Guidance is also given on how to optimise the

database for performance, robustness and ease of data extraction. Finally, examples of its data

mining capabilities are also presented.

5.1 INTRODUCTION

A data driven approach to signal processing involves cataloguing primary data according to a set of

descriptive parameters (known as meta- or tertiary data) and using these parameters to drive

automated, or semi-automated, signal processing and reporting. Results of processing generate

additional, derivative datasets (known as secondary data), which may also be stored and catalogued.

In the application presented here, acoustic emission (AE) hit information, waveform arrays and

time-logged parameters form the primary data, whilst statistics describing this primary data, such as

RMS, kurtosis and wavelet level energies, comprise the secondary data. Tertiary (meta-) data then

5-1

PhD Thesis, Chapter 5 – Data management

consists of information such as equipment TAG numbers, plant, channel information, dates and

operating conditions (flow, NPSH etc). Traditional file-based and data-driven management systems

are compared in Table 5-1.

The primary reason for the authors adopting a data driven methodology was the volume of

information generated by the application of AE to condition monitoring of rotating equipment; the

aim of which is to detect incipient, random failure of internal machinery components or the

damaging conditions that initiate failure. Whereas an AE structural test, conducted over a finite

time period, may result in a few thousand bursts and possibly 10’s of megabytes of waveform data,

a typical day’s continuous monitoring can produce between one hundred thousand and three

million AE events and around 2 gigabytes of waveform data. A single condition monitoring

session can last for days, weeks or months.

Dealing with data on this scale presents a number of challenges:

(1) Filing and retrieving data: Whilst it is feasible to have a folder on a computer file-system

containing several thousand files, few computers can present their users with a list of millions

of files in a timely and useful manner. Amongst other things, a file-system can only assign one,

meaningful attribute (the file name) to a piece of data. This is inadequate in a large-scale

system. Storing metadata in a searchable database provides a rich mechanism for identifying

and retrieving primary data, irrespective of whether the source file is stored on disk or archived

offline.

(2) Dealing consistently with data: In an application where collected data is to be post processed by an

evolving population of techniques, keeping track of what algorithms have been applied to what

datasets requires significant time and effort. By updating the metadata automatically as each

piece of primary data is analysed, an audit trail is created and the integrity of the analysis

assured.

(3) Modularity: Any attempt to analyse large amounts of data will, invariably, involve a mosaic of

techniques, often drawn from disparate applications or libraries. A data-driven approach, such

as the one presented here, provides a modular infrastructure into which post- processing

modules can be easily integrated and applied to sets of primary data. This facilitates direct

comparisons of techniques across statistically significant numbers of samples.

5-2


Conventional file-based approach

Data driven approach

Volume of primary data sets <10,000 >100,000

Primary data storage Files/folders with meaningful names

Numbered files either online, distributed, or archived

Post-processing time <8 hrs for all data >>8 hrs: selective post processing required

Processing management Manual auditing of processing (versions of algorithms etc.)

Automatic auditing

Volume of secondary data points <1,000 >10,000

Secondary data management Collation of secondary data by hand or simple scripts.

Automated collation of secondary data. Large number of secondary data iterations.

Graphing and reporting One off graphs and reports Automated graph and report generation with multiple iterations

Level of training required to establish system

Very low Above average

Level of training required to manage and use system

Minimal Average

Level of IT support required to manage system per data point

High Low

Table 5-1: Characteristics of conventional and data driven processing.

Additionally, in AE testing, a file-based approach has further limitations:

(4) Limited post-processing and graphing capabilities: When data has been collected using commercially

available hardware/software it may be stored in a proprietary format limiting the signal

processing and graphing options to those provided by the vendor. These are not well suited to

all applications and/or may restrict research options.

(5) Comparison of results: Proprietary file structures differ not only between manufacturers, but also

between hardware products by the same manufacturer. The most commonly used AE vendor

software packages have limited capability to analyse multiple tests concurrently, making it

difficult to compare results from different tests, or data collected across differing hardware

/software platforms. In the latter case, this also limits the user’s ability to validate objectively

the performance of a particular piece of hardware/software.

(6) Missing data: Information required for the interpretation of data may not always be contained

within the vendor’s datafile; using a database facilitates easy and accessible storage of any

metadata deemed necessary by the designers and operators of the system.

5-3


(7) Restrictions to accessing data: If data is collected using a commercially available software package,

analysing it on other machines may not be possible without purchasing additional licenses of

the software. This restricts data sharing and can represent a significant cost to the user.

An acoustic emission data management and analysis system (DBMAS), named ‘AEData’, was

developed to overcome these restrictions. Based on a Relational Database Management System

(RDBMS) implemented in Microsoft SQL Server, AEData catalogues all data files created by any

data acquisition system, extracts the data contained therein, co-ordinates batch post-processing

routines and collates the results of this post-processing, before arranging the results into a format

suitable for graphing.

Although some literature on the use of relational databases for managing condition monitoring data

has been published, the proposed models are primarily, either for storing and mining data [22, 23,

115], or as configuration management tools [129]. No reports have been found where these

systems are used to drive the analysis of data in a structured way. Commercial software packages

for administrating condition-monitoring data (eg. EXAKT, Ivara EXP, Avantis.CM) also use

relational databases, but these programs have their own in-built processing capabilities and tend not

to facilitate the addition of custom processing modules. Therefore, they do not offer the flexibility

required for research or development applications. Furthermore, commercial database

management systems are often prohibitively expensive.

5.2 DATABASE THEORY (PK)

A database is a file or a set of files in which information is stored. Within the file, data is grouped

in tables; each table holds a set rows; each row is a set of fields that collectively represent one

instance of that datum.

Rows in one table are usually related to rows in other tables either directly or transitively. These

relationships are implemented as part of the structure of the database and it is from them that the

term “Relational Database Management System” or RDBMS is derived. To appreciate the

difference between these types of relationships, consider the following scenario: waveform data

collected from a particular pump is stored in one table, details about the test in another table and

information about the pump (including its manufacturer) is contained in a third table.

Linguistically, the relationships between the data can be described as:

“These waveforms are from test A which used pump B from manufacturer C.”

5-4


The relationship between waveforms and pump A is direct, whilst the relationship between

waveforms and manufacturer C is transitive.

The importance of good table design should not be underestimated. By breaking data into subsets

and storing these in separate tables with relationships properly defined (a process referred to as

“normalizing”), data duplication is reduced or eliminated. Data also becomes less susceptible to

corruption caused by failure to ensure that all copies of redundant information are

updated/inserted.

As a database grows large, good table design becomes increasingly important because it becomes

impossible to visualize the dataset in its entirety; mistakes are thereby more likely, particularly

during development. Often these are difficult, if not impossible, to detect. Good design and a

structured implementation will reduce the frequency and decrease the consequences of these

mistakes. Although, to the novice database developer, this upfront design effort may seem of little

value, normalising data is a design task that will, ultimately, pay large dividends. Guidelines on how

best to normalise data are beyond the scope of this paper, but are covered well in various database

texts [83, 90, 118].

When extracting data from an RDBMS, queries are used to join related tables and present

information in a variety of orders and groupings. Queries may perform mathematical operations,

update metadata, create tables, and export data to other applications. They are the primary interface

to data stored in an RDBMS.

5.3 HARDWARE SETUP USED

Figure 5-1 shows the overall hardware layout used for collecting and processing AE data.

Although it would be possible to do all collection, processing and analysis on one computer,

multiple PCs have been used for convenience and to enhance capability.

Acquisition, batch processing and data analysis PCs are all standard clones with more RAM (up to

512MB) and larger IDE hard drives. For acquiring up to 4 channels of AE data simultaneously, the

data acquisition PC is fitted with either one National Instruments NI-6110 DAQ PCI card, in

which case data acquisition is controlled by custom LabVIEW routines, or two Physical Acoustics’

PCI2 DAQ cards, controlled by AEWin. As testing generally occurs offsite, this computer is

transported as required.

5-5


Figure 5-1: Physical layout of the implemented DBMAS.

The database is hosted on a 1.2GHz DELL Optiplex PC running Microsoft Windows 2000 Server

and SQL Server 2000. It is managed remotely with the use of TightVNC freeware. SQL server is

configured to use 128Mb of RAM at start up.

The typical flow schema from data collection to analysis output is given in Figure 5-3. Its

relationship to the DBMAS and its various stages will be discussed through the remainder of this

paper.

5.4 IMPLEMENTATION DETAILS

For reasons discussed earlier, AEData was developed to:

a) Catalogue and file AE Data collected from multiple acquisition systems so that it could be

readily mined,

b) Batch process this data using custom LabVIEW routines,

c) Assimilate analysis results into a volume that could be more easily interrogated for new trends

and patterns, and

d) Facilitate appropriate graphing of processed results;

User interface elements (forms) were only implemented when required to simplify data entry,

ensure input data integrity or as a mechanism for executing batch code. In AEData, this code was

written in LabVIEW, C++ or VB, depending on accessibility and the task being performed.

5-6


Figure 5-2: Data flow schematic.

5-7


For example, advanced signal processing

routines for wavelet and octave band analysis

were written in LabVIEW using its Advanced

Signal Processing Toolset, whilst low-level file

parsing code was implemented in C++ for

improved performance. Visual Basic was

generally used to “bolt it all together”. This

modularity is shown in Figure 5-3.

Following the philosophy discussed in part 5-

2, information in the database is separated

into one or more: (a) setup tables, which store

information about the data acquisition

session/test (meta-data), (b) primary data tables containing data extracted from acquisition software

files (i.e. Hit information from vendor software) and (c) secondary data tables, containing the

results of all custom analysis and post-processing. Descriptions of various primary and secondary

Figure 5-3: Architecture diagram of AEData.

Figure 5-4: AEDMAS layout with main tables, relationships, primary keys (PK) and foreign keys (FK). Primary, secondary and metadata tables are designated by (P), (S) or (M) respectively after the table name. One-to-many relationships are designated by the symbols 1 and ∝ respectively.

5-8


tables are given in Appendix D. Figure 5-4 shows how data in these tables relate to one another.

5.4.1 Metadata Input

Details regarding a particular data acquisition session (that are constant for the entire session) are

entered into the database via a form. Wherever possible, list-boxes help the user to select a value

from a set of common options. This helps ensure data entry consistency and efficiency, which is

particularly important when entering text information that may be used for querying the data. It

also reduces the time to enter data and thus increases the likelihood that it is actually recorded.

Data that is collected for each test (and usually entered via the form shown in Figure 5-5) is stored

as a single row in AE_Session table, including the session start date, plant and equipment number of

the test subject, and the name of the file directory into which data will be (or has been) saved.

Entering session information before logging files to the database simplifies subsequent cataloguing

and processing significantly.

Metadata for a particular session may vary in type or quantity from one session/location/pump to

the next. To enable the recording of disparate collections of session data, this variable information

Figure 5-5: Session data entry form.

5-9


is saved to a separate table (AE_Session_Parameters). Here the value of each piece of data

(ParameterValue) is stored along with its associated parameter name (ParameterName) as separate

rows. There is no limit to the number of parameters that can be recorded about a particular

session. Channel information (name, number, gain, type, sensor serial number etc) about either

AE or Parametric (non-AE) channels is stored in a third table, again as separate rows, facilitating

cataloguing of details from any number of channels (AE or parametric) per session. If the

acquisition computer is not connected to the database, session data (and associated channels and

parameters) are stored locally and synchronised later.

Once a session has ended and the primary data collected, another form is used to upload and

catalogue all files in the database. This form initiates routines that copy all filenames for one or

more acquisition sessions, along with their type and creation date to the Files table. Data can also

be transferred from one hard-drive to another, which is useful when separate PCs collect and

process the data. With custom written data acquisition software, logging file information can be

performed at the time of data capture (although data transfer is still initiated later). Using

commercial software this is rarely possible and filenames must be copied separately. The

implementation of this process ensures that all files collected are catalogued and accounted for. In

an environment where a session may produce more than 10,000 files it can be difficult to obtain

this level of rigorous quality assurance manually.

Figure 5-6: Process Files Form extracts information from DTA files and analyses WFS files.

5-10


5.4.2 Primary data input

Once AE data files have been catalogued and filed by the system, any data contained therein can be

extracted or processed. Again, a form (Figure 5-6) is used to drive this process. Visual Basic code

associated with the ‘GO’ button calls LabVIEW routines that have been compiled into a dynamic

link library, to perform file parsing and data extraction. Separate routines are implemented for

each file type. For example, a processing library was created in LabVIEW to read DTA files

created by AEWin for Physical Acoustics’ PCI2 data acquisition hardware. This code extracts: (a)

information about the acquisition setup (HDT, HLT, whether it is a new file or a continuation of a

previous file etc), (b) hit features, (c) time-dependant AE and parametric data, and (d) timestamps,

as well as any metadata not recorded elsewhere. Each data group is saved in a separate table for

reasons discussed previously.

The location of any burst waveforms within the DTA file and information required to interpret

them is determined and stored. This includes the waveform channel, number of samples, sampling

rate and total gain, which facilitates very quick data mining for waveforms thereafter (eg. when

averaging waveforms and/or determining continuous wavelet transforms). It should be noted that

Visual Basic, Matlab or a C variant could have be used as an alternative to LabVIEW and may be

judged by some to be more appropriate. Irrespective of the language used for file parsing, this

task’s level of complexity depends on the accuracy of documentation provided by the supplier of

the acquisition code. Unfortunately, this may not always be complete or accurate. If the originator

is unwilling or unable to provide this information, reverse engineering of file formats may be

required using low-level tools (eg. a Hex Editor) and a good deal of patience.

5.4.3 Secondary data input

One of the requirements for the database was to enable waveform-streaming files (typically

containing unscaled waveforms between 1 and 2 million samples long) to be subjected to advanced

and automated signal analysis. Accordingly, LabVIEW/C++ routines were developed to read,

scale and process the files using wavelet and spectral functions supplied with the LabVIEW Signal

Processing Toolbox (version 2). These programs were then compiled into libraries that could be

invoked by the batch processing routines within AEData. The results of each processing method

are stored in separate tables, with links to the originating file record. (In the context of this

DBMAS, the raw waveforms represent primary data, albeit they are not stored in the database,

whilst the analysis results are the secondary data.)

Figure 5-6 shows the batch processing control screen. A user selects a session in the list box, along

with the types of analysis/post-processing required. The database identifies the files from the tests

5-11


that make up the session, evaluates which of these files have already been processed and then

invokes the appropriate library to process remaining files, storing the resultant data in tables. If a

new analysis routine is developed, files are processed by the supplementary method and results

inserted into a fresh table. Existing tables are thus unaffected, ensuring that an auditable trail is

available for the comparison of results.

The ability to batch process one or more sessions, which may involve tens of thousands of files and

result in days of continual processing, is one of the key benefits of this DBMAS. This volume of

advanced analysis could never have been coordinated manually. Due to its significance, a schematic

of this process is provided in Figure 5-7. Examples of processing that have been performed with

AEData in this fashion include determining third-octave-band and discrete wavelet-level energies

(by various types of wavelets), denoising and calculating a variety of statistics that historically have

Figure 5-7: Process of extracting and/or processing data.

5-12


not been applied to AE signals due to the traditional processing overhead.

5.4.4 Defining session subsets

Sometimes it is necessary to extract a subset of data points from a particular AE_Session, for

further analysis or averaging. To do this, AEData defines a Test, which is merely an increment of

time (up to the length of an entire session) for which certain conditions are kept constant. Tests

and their conditions are stored in much the same way as AE_Sessions and AE_Session_Parameters.

Consequently, any Test can have an infinite number of associated conditions. Each Test belongs

(is linked) to a single AE_Session, has an ID, start and end timestamps and one or more test

conditions. Timestamps can either be created at the time of data acquisition (in which case they are

extracted from the acquisition files) or added later. Again, a form (shown in Figure 5-8) is used to

set the Test boundaries and instigate table population.

5.4.5 Further analysis and graphing

Once data and analysis results have been stored in the database, further processing and analysis is

performed using queries. In AEData, these are written with the help of Microsoft Access Query

Analyser or directly in Transact-SQL as stored procedures, which execute on the SQL Server.

Figure 5-8: Relating timestamps to test conditions.

5-13


(Transact-SQL is an SQL dialect, the language by which all systems communicate with a RDBMS.)

A common requirement of researchers and analysts is to determine ensemble statistics (mean,

variance etc) of one or more datasets. With a RDBMS, this becomes trivial. Routine aggregate

functions are incorporated into SQL and custom routines can be written for additional

functionality. One such function determines the mean confidence limits, using a student-t test, for

a generic selection of records (eg. RMS per channel for a certain Test).

Although MSAccess does have some limited graphing capability, the authors have found this

inadequate for presentation of AE results. Advanced graphing packages such as Origin or

SigmaPlot are substantially more useful, and latest versions can interrogate databases directly.

Scripts can also be written from within these graphing packages to automate database querying and

display results. However, if readily available graphing software does not have this advanced

functionality, data can always be exported from the database into text files. These are then

imported into the graphing space. Alternatively, advanced graphing objects can be incorporated

into the DBMAS, a solution commonly implemented when building application-grade software.

Results extracted from the database were presented using Origin Versions 7.0 and 7.5.

5.5 EXAMPLES OF RESULTS EXTRACTED FROM AEDATA

Figure 5-9 gives an example of the types of results that can be extracted when using a data driven

approach to managing AE information. The purpose of the monitoring session from which these

results were obtained was to measure changes in AE signals as the flow through a double suction

pump was altered (by throttling a discharge control valve). Two tests were conducted at different

motor speeds. More details on this test is presented in [107] and Chapter 7.

Figure 5-9 (a) shows a time extract of some raw data collected from this pump, as flow was

decreased in intervals from full flow to low flow. (Colours/intensity indicate the number of points

in a particular time-parameter interval.) Another parameter is superimposed. As this is a different

type of data the vendor’s software was unable to present it on the same graph.

Periods of constant flow were marked with Timestamps created during the monitoring session, as

discussed earlier. Flow values corresponding to these various blocks of time were entered into the

database once acquisition was complete. Using this information (i.e. the metadata), queries were

executed to: (a) extract data corresponding to all steady-state conditions, (b) summarise the

parameters accordingly and (c) normalise the summarised statistics based on a set of predefined

rules. Figure 5-9 (b) presents the resulting output for one such parameter, at both motor speeds.

5-14


(A) (B)

Figure 5-9: Raw AE data and corresponding secondary data obtained for a pump flow test.

5.6 OPTIMISING PERFORMANCE (PK)

5.6.1 Hardware

The speed of an RDBMS is only a second order function of CPU speed, so upgrading the CPU will

only provide a marginal improvement in database performance. Other techniques are therefore

required to promote efficient interaction when databases increase in size or complexity.

As tables become large, queries that involve joins will begin to slow. Initially, and intuitively, this

will correlate with an increase in CPU utilization on the database server. There will come a point

however, where a step increase in processing time is observed, coinciding with a step decrease in

CPU utilization. This happens because the storage space required for the temporary table created

by a join is larger than the amount of RAM available to the RDBMS. Consequently, the RDBMS

builds the temporary table on disk rather than in memory. Accessing information on a computer

hard-drive is two orders of magnitude slower than accessing information stored in RAM; hence the

server reports 100% disk utilization and a low CPU load. This phenomenon is referred to as

bottlenecking. The system, in the described state, would be “disk bound” and this is why

production, multi-user database servers have several gigabytes of memory installed.

If a dedicated computer is to be used for the RDBMS then it is recommended that consideration be

given to providing it with as much memory and as fast hard disks as budget permits. Adding an

additional 1-2GB of memory will deliver a profound increase in performance. Furthermore, if the

database is to be hosted on a centrally managed server, data and processing requirements should be

5-15


reviewed periodically with the database administrator so that the system can be tuned and

appropriate times chosen to run batch-processing jobs.

The RDBMS component of AEData was originally installed on a server with 256MB of memory

and a standard hard drive. Performance was adequate until one or more tables started exceeding a

couple of million records. At this point it became necessary to upgrade the amount of RAM and

install a RAID disk controller.

5.6.2 Indexes

The most significant improvement in data processing speed comes from the creation and

management of indexes. Indexes increase performance because the RDBMS will endeavour to

store the associated data in RAM rather than on disk. A table may have several indexes defined.

To create an index, a subset of fields within a table (that can be used to select and sort records from

that table) is nominated. Some database tools will analyse a query and make recommendations

about indexing which can be useful for novice database developers.

In AEData indexes generally included table ID fields (SessionID, TestID, Fileref etc) and

Channel_number. Additionally, fields commonly used for generating queries (parameter_name,

test_description) were also defined as Indexes.

5.6.3 Stored Procedures

Stored procedures are scripts that run on the database server. They are usually written in a version

of SQL. Where bulk updates of data are needed, these scripts generally run about twice as fast as

the equivalent code implemented in a client based application such as MSAccess. This occurs

because there is no overhead of client/server communication. It also reduces load on the network

significantly.

5.7 CONCLUSIONS

Although the system described was developed for use on AE data collected from rotating

machinery, it is equally applicable to any other type of AE testing. In fact, the general approach is

relevant to any signal source that results in large volumes of data requiring storage, processing

and/or data mining. From a design perspective, a data-driven management system such as this has

no inherent functional limitations on number of channels, tests or post-processing routines. The

number of concurrent users depends primarily on the database management software. If budget

5-16


constraints dictate, it can be (and has been) implemented on relatively inexpensive desktop

hardware, albeit configured appropriately.

The biggest advantage of this approach however, is that millions of data points can be analysed as

desired, with auditable rigour, and minimal manual intervention. Also crucially important, accessing

and mining the data, is simplified greatly because it is now independent of both the hardware and

software that was used for its creation/acquisition. This frees the analyst or researcher to process

and graph their data as required, instead of being hampered by the limitations of vendor software.

Once primary data has been extracted from data acquisition files, or analysis routines performed on

raw waveform data, files can be archived, reducing storage overheads associated with massive

amounts of raw data. Collating results in one place also simplifies the process of backing up and

archiving. Research effort and time can then be spent developing and assessing new signal

processing techniques on statistically significant sample sizes, rather than simply administrating the

information.

5.8 POSTSCRIPT

The database management system described in this chapter was fundamental to all work

subsequently described in this thesis. Processing modules were written (primarily by the author) in

Visual Basic, TransactSQL or LabVIEW for all techniques described in chapters 3 and 4. Queries

were developed (by the author or Mr Kelly) to extract subsets of these processed results for

ensemble averaging and/or plotting in Origin. Even graphs of individual waveform files used the

database management system to locate desired files. Although resulting images were not stored or

catalogued in the database, this would be an obvious extension to the current system and may be

incorporated into future versions.

5-17

6. EEFFFFEECCTT OOFF WWAAVVEEGGUUIIDDEESS

-- PPAARRTT OONNEE

Sikorska, J. and Pan, J.,

Journal of Acoustic Emission,

2004, Vol 22, pp. 264-273.

66The aims of this chapter are to:

(a) Identify how waveguides affect the transmission of simulated acoustic emission signals to

the sensor;

(b) Illustrate some of the issues relating to AE signal transmission that can complicate AE data

analysis.


This paper presents the effects of varying waveguide material and/or shape on traditional acoustic

emission characteristics of pulsed events. Forty-eight solid cylindrical waveguides were fabricated

from either alumina ceramic, mild steel, stainless steel (SS316), 2024-T3 aluminium or extruded

Delrin. Three different lengths, two diameters and four sensor face angles were tested. The effects

of a point contact at the source end of the waveguide were also verified. Results show that

although ceramic guides attenuate the signals more than commonly used grades of steel and

aluminium, waveform fidelity is less affected. Point contacts at the source end also had a negative

effect on all AE features and waveform profiles.


6-1

PhD Thesis – Chapter 6, Effect of Waveguides: Part One

6.1 RELEVANT ULTRASONIC THEORY

The solutions to the general wave equation for acoustic waves travelling through infinitely long

isotropic cylinders are detailed in references [39, 74, 96]. However, there are several key points

applicable to this research work that will now be summarised.

Longitudinal (axially symmetric) wave modes in an infinitely long cylinder vary with radial and

longitudinal distance along the cylinder but are symmetrical around the circumference. The

relationship between longitudinal phase velocity and frequency is known as the Pochammer-Chree

equation and resembles the Lamb dispersion relationship in plates. Its solution is well documented

and can be found in the aforementioned references.

Group velocities also vary with frequency in a similar fashion. Figure 6-1 depicts the first two

longitudinal phase and group velocities for one of the materials tested in this study. It can be seen

that the fundamental mode approaches a constant speed as frequency approaches zero. This

limiting speed corresponds to the longitudinal rod speed,

Lc E ρ= (6.1)

where E is the Young’s modulus for the material and ρ is the density. As frequency increases, the

first longitudinal modes asymptote to the Rayleigh velocity. Rayleigh waves are surface acoustic

waves in which longitudinal and shear modes couple together and travel at a common speed; it is

always less than the transverse speed by 5 to 15% [4]. They also decay rapidly with depth and

therefore displacements are confined to within one order of Rayleigh wavelengths from the surface.

Higher longitudinal modes have cut-off frequencies below which they will not propagate, and

above which they are dispersive. Torsional waves (through an infinite cylinder) propagate with one

constant phase velocity equal to the shear (or transverse) speed:

1E2(1 )Tc ρ ν

= ⋅+

(6.2)

where ν is the Poisson’s ratio. Again, higher order phase modes are dispersive and each has a

unique cut-off frequency below which these modes will not propagate, as evident in Figure 6-2.

Cut-off frequencies are a function of the rod diameter (due to the changing velocity), but

independent of length. All torsional group velocities are dispersive. Similar, but mathematically

more complex relationships govern the dispersive characteristics of flexural (anti-symmetric) waves

in rods; these are described in reference [86].

6-2


At very narrow diameters, cut-off effects act like a high pass filter, leaving only lower order modes

as the means by which low frequency longitudinal, torsional and flexural waves can be transmitted.

Conversely, above a certain frequency, which is a function of rod diameter and material properties

(namely E, ρ, and ν) speeds approach their limiting values and dispersion will no longer occur. In

this interim region however, where wavelength is in the same order as the rod diameter, significant

dispersive effects can be expected.

In addition to dispersion, AE bursts are attenuated as they travel through the material. Some of

this energy loss is caused by friction and thermal conductivity and occurs in all materials,

independent of frequency. However in compounds with larger grain sizes, such as plastics or

composites, frequency-dependant attenuation can increase significantly (by up to 3 orders of

magnitude) when compared to metals. Attenuation is also generally greater for transverse waves

than for longitudinal modes. Finally, it is important to recognise that AE sensors are not equally

sensitive to all wave modes, causing further apparent signal reduction of various modes.

One side effect of dispersion is the tendency of AE bursts, which are made up of numerous waves

of different polarizations and consequently different speeds, to lengthen in time as they move

further from the source. Called mode separation, continuous transformation between modes

exacerbates this effect [26].

In a finite rod, wave motion becomes even more complicated due to reflections from, and

resonance at, both ends [85]. Acoustic waves are partially reflected and transmitted at every

interface, the relative amounts of which depend on the ratio of the relative characteristic

impedances of each pair of boundary materials. This impedance is a function of material density

and the speed of the wave hitting the interface. If material attenuation is low, these reflections will

combine with late arriving components of the original signal, thus complicating feature

identification.

For non-zero angles of incidence, the character of the wavefront will also change as it hits the

interface. Called mode conversion, the change in angle of the wavefront will convert part of the

incoming wave from longitudinal to shear and vice versa. Flexural waves that may not have been

present in the original waveform can also be excited.

Finally, cylinders have natural frequencies that are a function of their length and wave velocity. As

an AE burst is made up of numerous waves of different frequencies, resonance can be expected.

Another interesting effect of dispersion is that higher order resonant modes are no longer integer

multiples of the fundamental frequency. This can complicate identification of these artefacts.

6-3


In traditional AE processing, various AE features are used for describing AE signals, as described

in previous chapters. Amplitude and frequency analysis are more meaningful features for measuring

continuous emissions, whilst count rate and energy analysis are useful for burst emission

characterisation [95]. Defining a burst as an event that exceeds a certain threshold, the following

features will be extracted and analysed in Part 1 of this study:

• Peak amplitude

• Burst energy

• Duration and/or total counts

• Rise time and/or counts to peak

• Signal strength.

6.2 EXPERIMENTAL METHOD

A variety of short waveguides were tested to illustrate the effects of: (a) material properties; (b)

length; (c) diameter; (d) face angle and (e) source point on the transmission and reception of pulsed

AE bursts.

Forty-eight (48) waveguides were tested in total, fabricated from five materials: Delrin acetal resin

(white extruded bar, exact grade unknown), common mild steel (exact grade unknown), SS316,

2024-T3 aluminium and sintered 99.9% alumina ceramic (α-Al2O3). Approximate material

properties are summarised in Table 6-1. Sensor face angles were 0˚, 30˚, 45˚ or 60˚; lengths were

30, 43 or 51 mm; and diameters were 5 or 8 mm. Unless otherwise stated, the rod face attached to

the pulser was flat.

Test specimens were grouped as follows:

(a) 5 sets of 8-mm diameter waveguides with face angles of 0˚, 30˚, 45˚ and 60˚ from each

of the four materials (20 waveguides in total). Lengths were set at 43 mm between

centres (at which pulser and receiver were located). (Four sets shown Figure 6-3A.)

(b) 3 sets of 8-mm diameter, shorter (30 mm) waveguides with face angles of 0˚, 30˚, 45˚

and 60˚; from Delrin, 2024-T3 aluminium and SS316.

(c) 2 sets of narrower (5 mm) mild steel rods, 30mm and 43mm in length, with face angles

of 0˚, 30˚, 45˚ and 60˚.

6-4


Figure 6-1: First two longitudinal modes showing normalised phase and group velocities (latter are shown as dashed lines) for an 8-mm SS316 infinitely long rod waveguide. Wave speed is normalised with respect to transverse wave speed, given in equation (6-1).

Figure 6-2: First four normalised torsional phase velocities for 8-mm diameter, infinitely long waveguides of different materials. Wave speed is normalised with respect to transverse wave speed, given in equation 6-2.

6-5


(A) (B)

Figure 6-3: (A) Angled 43mm waveguides for MS, Delrin, SS316 and Alumina (Clockwise from top left). (B) Pointed 43mm waveguides (NC).

(d) 3 longer (51 mm) rods from Delrin, 2024-T3 aluminium and SS316, 0˚ sensor face angle.

(e) 5 8-mm diameter, 43-mm long rods from mild steel (1 of), SS316 (1 of) or AL2024-T3 (3

of), each with one end tapered (45˚) to a point (pulser end). The sensor end was always

flat. More specifically the details of each (shown in Figure 6-3B) were:

• 1 x MS waveguide with dull point

• 1 x SS316 waveguide with sharp point

• 1 x AL2024-T3 waveguide with a flat point (1.6 mm in diameter)

• 1 x AL2024-T3 waveguide with dull point

• 1 x AL2024-T3 waveguide with sharp point

The following designation is used to describe samples:

Material Abbreviation Density (g/cm3)

cL (mm/μs)

cT (mm/μs) ZL

DR 1.42 1.8 0.9 Delrin * 2.6

Mild Steel MS 8.00 5.6 3.1 45

SS316 SS 7.89 5.7 3.0 45

AL2024-T3 AL 2.77 6.4 3.1 17.7

ALM 3.9 10.6 3.5 40.6 α-Al2O3

Table 6-1: Approximate acoustic properties of test materials used in theoretical calculations.

* Properties estimated for Delrin, as exact grade was unknown [31].

6-6


Material x OD in mm x Length in mm “PT” (optional).

6-7


To quantify the effect of each waveguide, data

was also collected from the sensor mounted

directly on the pulser face. The resulting

values were used to normalise all data.

Longitudinal pulses, representative of AE

bursts, were generated by a DECI Model

SE25-P sensor driven by a DECI Model 600

pulser, set at 150V5. A waveguide holder

(Figure 6-4) was fabricated to ensure that the

relative positions of the sensor, waveguide and

pulser were kept constant (Figure 6-5). The

pulser was attached to the underside of the

holder with a Delrin collar containing 8 rare

earth magnets and two locating pins to ensure

repeatable orientation. Incorporated into the

sensor mount was a screw and spring that held

the sensor firmly against the top waveguide

face. This could be oriented to any angle.

O-rings prevented the waveguides from contacting the sides of the sample holder. Silicone grease

was used as a couplant between the pulser-waveguide and waveguide-sensor faces.

Figure 6-4: Waveguide holder.

Bursts were detected using a B1080 Digital Wave Corp. single-ended wideband AE sensor

connected to a PAC 2/4/6 preamplifier supplying a gain of 20 dB and fitted with 100-1000 kHz

bandpass filters. Data was recorded by a PAC PCI-2 card, housed in a 2.8-GHz Pentium-IV

desktop computer. Prior to digitisation, signals were bandpass filtered onboard at 1-3000 kHz. A

fixed trigger threshold was set at 50 dBAE. For each waveguide, between 128 and 300 triggered

waveforms, each consisting of 8192 samples with 256 pre-trigger samples, were collected and

digitised at 10 MHz.

Various burst features were collected using AEWin software for each triggered event, including

peak amplitude (in dBAE), AE energy, signal strength, duration, rise time and AE counts. These

5 Locating the pulser at the geometric centre of the waveguide avoided the excitation of torsional or

flexural waves. (NC)

6-8


Figure 6-5: Waveguide holder setup (NC).

were then averaged and graphed using the system described in [108] and Chapter 6. All values used

in any particular average varied by less than 2% indicating high pulser repeatability. Individual

waveforms were also captured and averaged to remove the effects of random instrumentation

noise.

6.3 RESULTS AND OBSERVATIONS

6.3.1 Burst Profiles

As can be seen clearly in Figure 6-6, waveform profiles being received by the sensor vary

significantly between materials. Alumina ceramic appears by first inspection most like the face-to-

face signal but greatly reduced in amplitude. On the other hand, waveforms associated with both

steels are significantly protracted and full of broadband artefacts. Aluminium wavefront also

appears filled with similar broadband artefacts. This loss of waveform fidelity was also observed by

Ono, in his tests on longer samples of the same material using lasers as a source of AE [84]. Mode

separation is most obvious in the Delrin sample, which consequently lost the pulse’s sharp

wavefront.

6.3.2 Amplitude and Energy

The effects on amplitude were very similar for most waveguides as can be seen from Figure 6-7 to

Figure 6-9. The trend for energy was similar. These show that both features decreased with length,

increased diameter and increased face angle. As expected, attenuation was higher for α-Al2O3 and

Delrin. For the three materials from which different length waveguides were tested, approximate

6-9


values for attenuation were calculated (see Table 6-2). Although these numbers are very

approximate, they do indicate that Delrin’s broadband attenuation is highly non-linear and in this

application (across the frequency range studied) significantly less than predicted by the literature.

6-10


Figure 6-6: Time waveforms for flat waveguide samples. Different y-scales are used to facilitate better appreciation of waveform profiles. Time scales are identical. (43x8mm samples)

Figure 6-7: Amplitude for different flat-faced waveguides.

6-11


Figure 6-8: Amplitude versus face angle for all sets of angled samples.

Figure 6-9: Effect on amplitude by using a pointed waveguide.

6-12


6.3.3 Rise time, Duration and Counts

Rise-times (and counts to peak) were greatest for Delrin, due to its much lower speed, and obvious

mode separation. Otherwise material effects on rise-time were minimal. There were also no clear

relationships between length, face angle and rise-time. The same cannot be said for pointed

waveguides, as shown in Figure 6-10, which caused rise-times to increase significantly in all three

materials tested.

Durations (and AE counts) were affected by material properties, as shown in Figure 6-11. Results

support initial observations from time waveforms, with steel samples showing the highest

durations, which increased with both length and diameter.

Figure 6-10: Effect of pointed end on rise-times.

Average peak amplitudes of samples in dBAE Attenuation

30mm 51mm Δ dB/m

SS316 107.7 101.0 6.7 515

AL2024-T3 105.6 100.2 5.4 415

Delrin 95.3 90.5 4.8 369

Table 6-2: Results of rough amplitude calculations.

6-13


Figure 6-11: Material effects on duration.

Results for other materials were less consistent. Although alumina appeared to have a very similar

waveform to the originating pulser burst, its much smaller duration value suggests that its

calculation is as much a feature of the fixed threshold value as it is the waveform profile: a small

signal takes less time to drop below the fixed 50 dBAE trigger level. (A smaller threshold, closer to

the true noise floor, may have lessened this affect.) Neither face angles, nor end points had any

appreciable effect on durations.

6.4 DISCUSSION

6.4.1 Effect of Material

Little difference in the time domain can be seen between mild steel and stainless steel waveguide

responses, especially when compared to other material samples. AE features analysed were also

consistent between the two steel grades, indicating that neither material has any obvious acoustic

superiority over the other. Durations and counts were substantially higher in both grades of steel

and both features increased with rod length. The reason for this is not immediately obvious from

the tests in this paper but is the subject of investigations presented in Part 2.

Of the metallic specimens, 2024-T3 aluminium samples transmitted pulses with the least signal

distortion. However waveforms still appeared protracted and filled with broadband artefacts not

present in the original signal, although not as much as in either steel. Alumina ceramic was the least

6-14


distorted of all materials, with signals appearing similar to the originating bursts. Delrin, due to its

much slower wave speed, showed the clearest mode separation of all materials.

As predicted, overall attenuation of both energy and peak amplitude was greatest in non-metallic

waveguides, however rudimentary calculations based on drop in peak amplitude for different

lengths indicated that this was not as severe as was expected. Furthermore, any attenuation present

in Delrin seemed to be non-linear and therefore longer waveguides caused proportionally less

attenuation than shorter guides across the traditional AE bandwidth.

6.4.2 Effect of Length, Diameter and Face angle

Attenuation also increased with length and diameter (where measured). Reasons for the latter are

not clear, and were only tested for one material (mild steel) but may be due to the increased number

of modes present in wider rods (below cut-off). This will be discussed further in Part 2. In all

materials, attenuation increased significantly with face angle, particularly above 30 degrees. This is

probably because mode conversion is causing redirection of energy into modes that the transducer

is less able to detect.

6.4.3 Effect of Pointed Source End

Pointed 43-mm waveguides attenuated the signals by over 12 dB when compared to similar flat

waveguides with adequate couplant. This was greater than the attenuation quoted by other

researchers [84, 128]. The type of point also made a significant difference on amplitude and energy

attenuation with finer (sharper) points causing greater signal reduction than rounder or flat points.

This may explain the discrepancy between these values and those published in other work.

6.5 CONCLUSIONS

Although signals passing through metallic waveguides were not attenuated significantly, signals

were protracted and appeared as if contaminated by reflections and resonance artefacts. Conditions

deteriorated with length and diameter. It is possible that much longer waveguides (>100mm) might

avoid this problem due to the additional attenuation that would be incurred; however this requires

further investigation. On the other hand, α-Al2O3 did not suffer from this problem, with primary

bursts being easily separated from subsequent reflections. Although alumina had a higher overall

attenuation, measured bursts showed significantly less distortion and (except for amplitude and

energy due to its increased attenuation) best-matched AE features of face-to-face signals.

6-15


Unfortunately, use of waveguides in AE testing cannot be avoided, however their implementation

must be properly considered. Whenever feasible, specific designs should be tested prior to

installation to determine actual transmission characteristics.

Further effects on transmission of frequency and modal information were investigated in reference

[110], which has been included as the next chapter of this thesis.

6-16

7. EEFFFFEECCTT OOFF WWAAVVEEGGUUIIDDEESS

-- PPAARRTT TTWWOO

Sikorska, J. and Pan, J.,

Journal of Acoustic Emission, 2004,

Vol 22, pp. 274-287.

77 The aims of this chapter are to:

(a) Identify how waveguides affect the transmission of frequency information corresponding

to simulated acoustic emission signals to the sensor;

(b) Illustrate some of the issues relating to AE signal transmission that can complicate AE data

analysis.



This paper presents the effects of varying waveguide material and/or shape on the frequency and

joint-time-frequency characteristics of pulsed acoustic emission (AE) events. Forty-eight solid

cylindrical waveguides were fabricated from either alumina ceramic, mild steel, stainless steel

(SS316), 2024-T3 aluminium or extruded Delrin. Three different lengths, two diameters and four

sensor face angles were tested. The effects of a point at the source end of the waveguide were also

verified. Analysis methods included continuous wavelets, short time Fourier transforms and

standard Fourier transforms. Results show that detected signals were dominated by the resonant

harmonics of the waveguides, particularly with commonly used waveguide materials such as

aluminium and steel.

7-1

PhD Thesis – Chapter 7, Effect of Waveguides: Part Two

7.1 RELEVANT THEORY

The solutions to the general wave equation for acoustic waves travelling through infinitely long

isotropic cylinders are detailed in references [26, 39, 74, 96]. Additionally, key points applicable to

this research work were summarised in Chapter 6 [109].

7.1.1 Effect of Finite Ends

Finite cylinders have natural frequencies for axially symmetrical longitudinal waves with clamped-

clamped or free-free boundary conditions, according to the following:

2nncfL

= (7.1)

where c is the wave speed, n is an integer and L is the length of the waveguide. For materials

analysed in this study, the corresponding values based on the longitudinal wave speed are given in

Table 7-1.

7.1.2 Frequency Signal Processing

Traditionally, frequency analysis has not been used to describe AE signals because material induced

AE is generally broadband (energy is distributed across a wide range of frequencies) or unduly

influenced by the natural frequency of the sensor. However, in certain cases amplitude spectra can

be useful to quickly ascertain the presence of specific frequencies. Additionally, some researchers

have observed that certain types of faults change the relative distribution of energies between

different frequency bands [79].

fn in kHz cL Material

for length in m (mm/μs) (Abbreviation)

0.03 0.043 0.051

Delrin (DR) 1.8 30 21 18

Mild Steel (MS) 5.6 93 65 55

SS316 (SS) 5.7 95 66 56

AL2024-T3 (AL) 6.4 107 74 63

10.6 177 123 104 α-Al2O3 (ALM)

Table 7-1: Approximate natural frequencies of first longitudinal mode in cylindrical rods with differing lengths.

7-2


7.1.3 Advanced Joint-Time-Frequency Techniques

Two frequently applied joint-time frequency techniques are Short Time Fourier Transforms (STFT)

and wavelets. These were introduced in Section 3.2.6.

The most commonly used wavelet in AE analysis is the Gabor (otherwise known as Morlet)

function [43], also used by the AGU-Vallen freeware. Its function is defined by the following:

(7.2) 2( ) exp( /2)cos(5 )w t t t= −

The relationship between scale, a, and centre frequency, fc, for the Morlet wavelet is given by the

following [91]:

52

s

c s

f where af f a>0

aπ

⎧⎪= ⎨⎪⎩

=0 (7.3)

where fs is the sampling rate in Hertz. The relationship between centre frequency and bandwidth

therefore becomes:

5 /2 5 0.81/ 2s sf a fcentre frequency

bandwidth aπ

π= = ≈ (7.4)

Similarly, the translated position on the x-axis, b, is determined by:

xb

s

b Nx where b=0,1,2...,N/Nf x

×= (7.5)

where Nx is the number of time samples per time division and N is the total number of samples.

For efficient computation, these should both be integral powers of 2 (eg. 64, 1024, 8192 etc).

7.1.4 Modal AE

It is assumed that specific faults propagate by a defined set of one or more dispersive modes, the

characteristics of which are a function of material properties and dimensions [109]. Modal AE

involves acquiring and digitizing AE time signals, from which individual propagation modes are

then identified and extracted. Information about particular modes can then be quantified in an

attempt to identify and/or source the underlying faults.

7-3


7.2 EXPERIMENTAL METHOD

A variety of short waveguides were tested to illustrate the effects of material properties, length,

diameter, face angle and source point on the transmission and reception of pulsed AE bursts.

Details of samples tested and experimental method used were described in Part 1.

7.2.1 Additional Post-Processing

After capturing burst data using AEWin (Physical Acoustics software for PCI-2 board), 32

sequential bursts of 8192 samples each (digitised at 10 or 40 MHz) were synchronously time-

averaged and post-processed using programs written in National Instrument’s graphical

programming language, LabVIEWTM (Version 6.1) and the data management system described in

[108] and Chapter 6. Due to the highly transient nature of the AE data, no window or overlapping

was applied prior to performing an FFT; as bursts were always captured completely within each

dataset, spectral leakage was considered highly unlikely.

To exclude the effect of sensor-pulser response, cross-spectra were performed against the face-to-

face averaged signal and results determined in terms of the magnitude and phase output.

Coherence was checked, but always found to be equal to one. Paradoxically, the effect of the

sensor-pulser was easier to appreciate by superimposing its FFT over waveguide amplitude spectra,

so this is the selected presentation method implemented here.

STFT and CWT analysis was performed using programs written in LabVIEW, which utilised

standard modules from the LabVIEW Signal Processing Toolset (Version 7)[6]. Wavelet results

were also compared with the output from AGU-Vallen freeware and found to be comparable (see

Fig. 3). Improved resolution obtained by increasing sampling rate from 10 MHz to 40 MHz can be

seen in Figure 7-1. Unless otherwise indicated, wavelet spectrograms in this paper were digitized at

40 MHz.

Wavelet processing used a Morlet wavelet as the mother wavelet and for this work computation

was based on 600 scales and 100 samples per time window. STFTs were undertaken using 1024

frequency increments and a Hanning window. As the simultaneous time-frequency resolution of

the STFT has functional limitations, two sets of graphs with time increments of 128 or 512 samples

per time window, to extract improve temporal and spatial resolution respectively, were created for

each sample. As expected, CWT provided better simultaneous time-frequency resolution (see

Figure 7-2) than any one STFT but took significantly longer to compute; the respective STFT and

7-4


CWT functions behind the spectrograms shown in Figure 7-2 took 594 and 15141 ms to calculate

respectively (i.e. wavelets took almost 30 times longer). Nevertheless, wavelets offer a visually more

7-5


(A) 10 MHz (B) 40 MHz

Figure 7-1: Wavelet transforms at different sampling rates (Face-to-face signals).

(A) STFT (B) CWT

Figure 7-2: Difference in resolution between (A) STFT in dBAE and (B) CWT, log scale (Delrin 51-mm sample). Computing times were 594ms and 15141ms respectively.

7-6


(A)

(B)

Figure 7-3: Comparison of results obtained from the custom written programs in (A) LabVIEW and the (B) AGU-Vallen wavelet tool.

LabVIEW parameters: 600 scales, 100 samples per time increment, ~30s processing time. AGU-Vallen parameters: frequency resolution of 5kHz, 1000 wavelet samples, ~60s processing time.

7-7


economic display of time-frequency information and therefore are predominantly shown herein. It

should be noted however, that computing multiple STFTs, with different time and frequency

intervals provides the same information in a much shorter time (albeit in two graphs), and therefore

would be more appropriate for real-time feature extraction.6

7.2.2 Modal Analysis

A commercially available software package called DISPERSE (Version 2.0.16c)7 was used to

generate individual modes for the materials being analysed. These were then superimposed over

the signals using a standard drawing package; frequency and time scales were matched

appropriately.

Unfortunately, exact material properties for the various materials (particularly Delrin and mild steel)

were not known and due to limited access to DISPERSE, examining slightly different values to

obtain better modal matching was not possible. Therefore some discrepancies can be expected,

particularly for Delrin and mild steel.

7.3 FREQUENCY RESULTS

7.3.1 Flat Waveguides

Resulting averaged amplitude spectra for all flat 43-mm waveguides are given in Figure 7-4. The

face-to-face spectrum is also plotted to show the effect of its non-uniform frequency response on

other results. (Wavelet plots are given in Figure 7-12.) Spectra of metallic waveguides are

dominated by harmonics of longitudinal rod frequencies. The first of these peaks (~60 kHz)

matches the results listed in Table 7-2. This is confirmed by observing the changing location of

these peaks for different rod lengths (see Figure 7-5).

Alumina is also affected by rod resonance, but due to its much higher wave speed, there are fewer

harmonics permeating the bandwidth of interest. Additionally, the material’s frequency-dependant

damping seems to reduce the effect of rod resonance above 600 kHz. Similarly, Delrin’s high

damping [109] is probably the reason no resonant effects can be seen in its amplitude spectra.

6 This may also be a function of the algorithms used by National Instruments, Signal Processing Toolset for

LabVIEW. (NC)

7 Propagation modes using DISPERSE were determined by Dr. Mike Lowe from the NDT Lab at Imperial College in

London . (NC)

7-8


Figure 7-4: Amplitude (in dB) - frequency (kHz) spectra of averaged time signals. (Face-to-face spectrum shown as a shadow on each graph.)

Figure 7-5: Increasing rod length decreases frequency of resonant peaks. 1st resonant peak match values given in Table 7-1 (aluminium samples).

7-9


7.3.2 Angled Waveguides

Changes in spectral content resulting from different sensor face angles can be seen in Figure 7-9

and Figure 7-10. (Results from 43 mm x 8 mm SS316 samples are shown, but similar findings were

found for other materials). As angle increases, overall amplitudes decrease and resonant peaks

become more pronounced. For metallic guides new spectral components become noticeable,

particularly above 45˚. These were not obvious in narrow (5-mm diameter) mild steel rods, alumina

or Delrin guides. In the latter case, virtually no changes to spectral content were observed as angles

increased. Narrow MS rods in fact showed patterns very similar to alumina waveguides, implying

that these additional spectral artefacts are indeed due to higher order modes below their cut-off

frequencies.

7.3.3 Pointed Waveguides

The effects of points at the source of the waveguide are depicted in Figure 7-6 and Figure 7-7. This

shows that all points reduced low and high frequency content and amplified rod resonances,

contrary to other reported results [84]. The magnitude of these effects depended on the type of

point, with smaller, sharper points being most detrimental.

7.3.4 Modal data

Modal information is presented in Figure 7-11 and Figure 7-12. It was difficult to relate mild steel

and Delrin theoretical modes with experimental results, implying that the actual material properties

differed to those given (Table 6-2). (Mild steel’s spectrogram is not shown, but the analysis of the

43-mm samples showed it was almost identical to that of SS316.) Although no correlation with

theoretical results was obtained for Delrin, it is deduced that its CWT spectral peaks were

nevertheless most likely from transmission of the originating burst via two different dispersive

modes, because samples clearly demonstrated mode separation (see Figure 7-13) and no discernible

change in frequency.

Early arrival of high frequency information (~700 kHz), which cannot be matched to any modes, is

most obvious in the SS316 spectrogram, but can also be seen in plots for other materials. This is

also present in the face-to-face spectrogram (Figure 7-1), indicating that these are probably

frequencies from the pulser or resonance in the sensor (neither being truly broadband). Similar

peaks can sometimes be seen at approximately 110 kHz, 350 kHz and 580 kHz, depending on the

amplitudes of neighbouring spectral highlights.

7-10


Figure 7-6: Effects of different types of points on the resulting frequency spectrum (43-mm long, 8-mm diameter aluminium samples).

Figure 7-7: Effect of points on wavelet spectrograms.

The actual frequency of the second longitudinal rod resonance peak also reduced slightly as angle

increased (Figure 7-14), probably because the absolute waveguide length increased (centre-centre

distances are kept constant).

As already discussed, face angles caused significant mode conversion and energy redistribution

from the non-zero face angles, contrary to the authors’ expectations. Instead most energy was

7-11


Figure 7-8: CWT from 43-mm long, 60˚ waveguides (Data sampled at 10 MHz).

contained in the first longitudinal group mode. Interestingly, no appreciable energy was observed

in flexural modes at first few rod resonances, whilst in the case of metallic waveguides, new spectral

artefacts were also visible (Figure 7-11). These were much more prevalent in wider guides, at face

angles of 45˚ and 60˚ and may be due to the excitation of radial waveguide resonances. For 8-mm

diameter SS316 rods, theory predicts this to be around 400 kHz, which corresponds with

frequencies seen in Figure 7-11d).

7.4 DISCUSSION

7.4.1 Effect of material

In all metallic and alumina waveguides, the first longitudinal group mode was the mechanism by

which bursts were transmitted through the waveguides. Modal identification in Delrin samples was

inconclusive.

Of all materials tested, alumina ceramic best retained the shape and frequency characteristics of the

originating (face-to-face) waveform, albeit with significantly reduced amplitude (discussed in Part

1). Although some reflections could be seen in its CWT spectrogram, these were easily separated

from the original pulse. Some excitation of the waveguide’s first and second resonant frequencies

was also observed. Narrow (5mm) steel waveguides were almost as effective in retaining original

waveform shape, without the attenuation.

7-12


Figure 7-9: Effect of face angles of FFT spectra (43-mm long SS316 samples).

Figure 7-10:Changes in wavelet spectrograms for changing face angles

(30-mm aluminium samples).

7-13


Figure 7-11: Changes due to different face angles for 30-mm SS316 waveguides do not correspond with flexural modes (only one shown for clarity).

Of all materials tested, alumina ceramic best retained the shape and frequency characteristics of the

originating (face-to-face) waveform, albeit with significantly reduced amplitude (discussed in Part

1). Although some reflections could be seen in its CWT spectrogram, these were easily separated

from the original pulse. Some excitation of the waveguide’s first and second resonant frequencies

was also observed. Narrow (5mm) steel waveguides were almost as effective in retaining original

waveform shape, without the attenuation.

7-14


Figure 7-12: Wavelet spectrograms with longitudinal mode frequencies. 43 mm x 8 mm waveguides (Phase modes are shown as dotted lines, group modes as solid lines).

Lower damping and slower wave speeds caused larger diameter metallic waveguides to lose all

waveform fidelity; late arriving components of the first longitudinal group mode merged with

subsequent reflections and numerous harmonics of the waveguide’s natural frequency. The result

was a protracted and irregular waveform, with a highly resonant spectrum.

Delrin waveguides showed little or no resonant effects due to the material’s high damping, although

waveforms had little resemblance to the originating pulses in either temporal appearance or

frequency content.

7-15


Figure 7-13: Mode separation is clearly evident in Delrin waveguides.

Figure 7-14: Standard amplitude spectrum showing changes in location of 2nd resonance peak.

7-16


7.4.2 Effect of length and diameter

Longer rods accentuated mode separation and decreased the frequency of resonant modes.

Unfortunately, harmonics of fundamental resonant frequencies permeated the entire bandwidth,

affected only by the material’s inherent damping, implying that lengthening rods to reduce these

frequencies is unlikely to be successful. Unfortunately, shortening rods so that fundamental

frequencies are above the frequency range of interest is impractical in all but the fastest acoustic

materials (even a 10-mm alumina waveguide would still have a fundamental frequency at around

530 kHz). Modal content was otherwise unaffected by length.

Reducing diameter had a significant effect of the frequency and modal content of resulting

waveforms by limiting the number of modes by which signals could be transmitted. Although

individual modes (above the first longitudinal group mode) could not be identified in CWT

spectrograms, narrower waveguides had significantly fewer spectral artefacts than their wider

counterparts.

7.4.3 Effect of face angle

Face angles of 45˚ or 60˚ caused energy to be transferred from the first longitudinal velocity to

longitudinal and/or radial resonant modes. Only waveforms being transmitted through Delrin

waveguides remained unaffected by face angle, although amplitudes did decrease as shown in

Chapter 6.

7.4.4 Effect of a pointed source end

Contrary to results published by others, pointed waveguides affected waveform profiles and

frequency content significantly. Not only were amplitudes and energies attenuated, resonant effects

were amplified and certain high and low frequency regions were filtered. Steels and aluminium

were affected similarly. The severity of changes depended on the shape of the waveform tip, with

sharper points being most detrimental.

7.5 CONCLUSIONS

This work indicates that waveform integrity is affected by the number of modes available for signal

propagation. This, in turn, is directly related to acoustic wave speed and inversely proportional to

waveguide diameter. Therefore, narrow waveguides fabricated from materials with very high

acoustic velocities (ceramics) should best retain waveform properties.

7-17


In commonly used materials such as aluminium, mild steel and stainless steel, longitudinal

resonance artefacts dominate frequency responses in all but very narrow guides. These effects are

magnified by application of a point on the end of a waveguide, or by angling the sensor face, which

also excites radial resonances. Consequently, both should be avoided if trying to maintain

waveform integrity.

Although use of waveguides in AE testing cannot be avoided, their implementation must be

properly considered; dispersion, attenuation, mode conversion and/or waveguide reflections

occurring within the waveguide will affect the signals being detected by the sensor. If possible,

ends should be square and parallel and materials should be selected depending on the transmission

characteristics required for a particular monitoring task. If possible, specific designs should be

tested prior to installation so that transmission characteristics can be verified.

7-18

8. FFLLOOWW MMOONNIITTOORRIINNGG OOFF

AA DDOOUUBBLLEE--SSUUCCTTIIOONN PPUUMMPP

Sikorska, J. and Hodkiewicz, M.,

Proceedings of the Comadem 2005,

pp. 192-202.

88


(a) Verify whether changes in hydraulic conditions in a double suction pump can be

monitored using acoustic emission; and

(b) Determine if acoustic emission can detect these conditions more accurately that dynamic

pressure or vibration monitoring.


Measurements from acoustic emission, vibration and dynamic pressure transducers are presented

for a range of flows and two pump speeds on a double-suction centrifugal pump. These show the

AE signals are more sensitive to change in flow condition away from the pump’s best-efficiency

point than either vibration or dynamic pressure measurement. Results indicate that AE monitoring

may therefore be appropriate for determining the location of best efficiency and for detecting the

onset of undesirable hydraulic conditions such as recirculation and cavitation.

8.1 INTRODUCTION

A centrifugal pump moves fluid from one point to another by converting mechanical energy into

hydraulic energy, which is referred to as head and is measured in meters. The efficiency of the

mechanical-hydraulic energy conversion varies with flow (Figure 8-1a). Input energy is most

efficiently converted into increased head at a single flowrate called the Best Efficiency Point (BEP).

8-1

PhD Thesis – Chapter 8, Flow Monitoring of a Double Suction Pump

In well-designed systems, the operating (or duty) point is close to BEP. However, due to system

changes, poor design or process control necessities (eg. control valve throttling), pumps are often

forced to operate away from their BEP. This loss of efficiency can as high as 30%, resulting in

significant energy losses when considering the total power consumption of centrifugal pumps.

As flow deviates from BEP, hydraulic, thermal and mechanical losses increase, and a number of

more serious adverse effects, as shown in Figure 8-1b, may develop. The flows at which these

occur vary from pump to pump and can rarely be determined analytically. Therefore experimental

means of detection are required. The most well known, and potentially troublesome, hydraulic

problems are cavitation at low suction pressures and recirculation at low flows. Mechanical

component failures often result from operating under these conditions.

Cavitation is the formation (and subsequent collapse) of bubbles caused by fluid pressure dropping

below the vapour pressure of the liquid. It reduces equipment’s hydraulic performance and can

(A)

(B)

Figure 8-1: (A) Pump, system and efficiency curves. (B) Onset of adverse conditions caused by operating away from BEP[60].

8-2


result in significant noise, vibration, pressure pulsations and erosion to metal surfaces. To avoid

cavitation damage, manufacturers recommend that a pump operates with suction pressure that

exceeds its minimum net positive suction head required (NPSHR), which is an experimentally

determined value derived from the pressure at which total discharge head drops by 3% (flow kept

constant). From the perspective of the research work described herein, it is important to recognise

that cavitation starts occurring (called incipient cavitation) at suction pressures significantly above

NPSH; most pumps routinely operate satisfactorily with some level of incipient cavitation.

Cavitation effects increase with flow due primarily to additional friction losses incurred in suction

piping reducing the available suction pressure at the pump’s inlet.

At low flows, fluid reverses back from the impeller into the inlet pipe, travelling along its outside

diameter creating a rotating field. When this reversal occurs at the inlet eye it is know as suction

recirculation, whilst if it occurs at the discharge tips of the impeller vanes it is called discharge

recirculation. The amount of flow reversal increases as flow is reduced. In most pumps, the

capacities at which suction and discharge recirculation commence differ, although generally (but not

always) suction recirculation will commence at a higher flow than discharge recirculation [37].

Damage associated with either form of recirculation is more likely in high energy pumps [20].

8.2 DETECTION OF ADVERSE HYDRAULIC CONDITIONS

Determining where a pump is operating on its curve at any instant is typically done by measuring

flow or discharge pressure and then comparing readings to the pump’s head-flow curve. A typical

curve is shown in Figure 8-1. In practice there are a number of difficulties with this approach: (i)

the pump curve is only valid for a new pump, (ii) the shape and location of the curve may change

due to wear, and (iii) only a small percentage of pumps have associated flow measuring devices.

A number of methods have therefore been trialled over the past fifty years to detect damaging

hydraulic conditions in centrifugal machines. These include visual observation, dynamic pressure,

vibration, sound (both internally and externally) and more recently acoustic emission. Only the

latter three methods do not require special pumps or intrusive instrumentation. The key findings of

this body of previous work are now summarised.

8.2.1 Stable Cavitation (low NPSHA)

Stable cavitation is generally observed as high frequency pulses in vibration, acoustic and pressure

signals, of random duration and amplitude, superimposed over a level of broadband noise from the

pump [42, 66]. As the degree of cavitation increases, the amplitude of casing vibrations, noise and

dynamic pressure fluctuations rise significantly. The magnitude of pressure pulsations caused by

8-3


stable cavitation are dependant on the pump-impeller geometry [19, 27]. Furthermore, overall

noise and vibration levels are a function of operating conditions (NPSHA, flow and speed) and

pump design [63, 101, 111]. Under stable cavitation, broadband vibrations can also modulate

specific frequencies (running speed, vane pass frequency, 0.5 x vane pass frequency). Increases at

these specific frequencies have been measured with vibration [7, 58] and acoustic sensors [28, 111].

The generation of acoustic emissions during bubble formation and collapse was first identified by

Derakshan in 1977, who also determined that cavitation intensity in a hydro-turbine was

proportional to the RMS (root mean square) of the acoustic emission signal [29]. Since then others

have reported that cavitation in machinery can be detected by “high amplitude excursions in the

RMS” [34]. Crest factors (ratio of the signal peak to its RMS) have been suggested as more reliable

indicators of incipient cavitation than RMS because during cavitation, acoustic emission bursts

typically have very high peak values with respect to the background noise level [66, 80, 81], but no

work in this has been published.

Frequency analysis of AE spectra has also been used to identify cavitation. In work by Neill et al,

the AE spectrum was very similar in shape to that under rated duty, but up to 40dBAE higher in

amplitude (depending on the degree of cavitation and sensor location). A sudden decrease in the

high frequency energy (0.5-1MHz) was also observed as cavitation developed [80, 81].

8.2.2 Recirculation

Collective research suggests that discharge recirculation can appear very similar to cavitation:

broadband pressure pulsations increase, as do levels at specific frequencies such as the blade-pass

frequency [41]. Suction recirculation on the other hand tends to increase mainly the low frequency

pressure pulsations (<125Hz) and dampens high frequencies [41]. The amplitude of pressure

pulsations increases with the intensity of recirculation. Pressure pulsations at 10-75Hz were also

observed by others under low flow conditions, who were unable to determine if these were due to

suction or discharge recirculation [47]. Unfortunately, dynamic pressure measurements are very

sensitive to sensor location, reducing with distance from the pump.

Under low flow conditions, low frequency axial vibrations have also been measured on the non-

drive end bearing housing of single stage, double suction horizontal pumps [49, 82].

Limited research on the use of AE for detecting low flow conditions has been reported. These

indicate that raw AE signals during recirculation tend to be mixed mode (bursts superimposed over

a background level of broadband AE), with little change in the background levels from non-

recirculating conditions, but with much larger discrete events [80, 81]. As for cavitation, Neill

8-4


reported that the recirculation AE spectrum appeared very similar in shape to the normal spectrum,

but this time only 10dBAE (rather than 40dBAE) greater in amplitude. Similar results were

published that showed RMS levels increased at low (and high) flows [8]. In this work, AE

measurements were also affected by sensor placement, however similar trends were seen on both

suction flange and pump casing AE signals.

8.2.3 Identifying BEP

Collective work indicates that absolute noise and vibration levels depend on the size, speed and

load of the pump. Nevertheless for a given pump design, operating with sufficient NPSH margin,

most research indicates that overall acoustic noise, vibration, stress wave (low frequency AE) and

acoustic emission levels are lowest when the pump operates at, or near BEP [8, 63, 101, 111].

8.3 EXPERIMENTAL SETUP

8.3.1 Pump Rig

Data was collected from a

Worthington 12 LN/H21-B axially

split, double suction, single stage

centrifugal pump. This model was

old and had been modified and

rebuilt many times by the asset

owner. Exact impeller design and

internal dimensions could not be

measured, so theoretical points of

suction and discharge recirculation

were not determined. Efficiency curves were however calculated for each motor speed and

normalised with respect to best efficiency; these are presented in Figure 8-2

Figure 8-2: Normalised flow curves for two speeds tested

.

The facility into which the pump was installed for data gathering is located at Blakers Pump

Engineers in Perth, Australia and was designed to meet Hydraulic Institute, ISO and Australian

standards for pump testing; it consists of a main storage tank, pressurised suction vessel and

associated inlet and outlet pipes, valves and fittings.

Vibration, dynamic pressure, AE and performance data was collected on the pump, operating at

two motor speeds (887rpm and 997rpm) and at discrete operating points over a range of flows

0.40-1.20 Q/QBEP. Speed was controlled by a variable frequency drive whilst flows were controlled

8-5


by manually adjusting the position of a valve downstream of the pump. The results of these tests

are reported in the following section. Additional tests involving reducing the suction pressure to

the pump were also conducted and associated results will be reported in a separate publication.

8.3.2 Vibration and Pressure (MH)

At each operating point 150 seconds of vibration and dynamic pressure sensor data was digitally

recorded using a Sony Tape recorder from B&K 4396 Deltatron accelerometers and PCB 112A22

dynamic pressure transducers. Performance data providing flow, pressure, motor power and pump

efficiency was also recorded. One accelerometer was placed on the non-drive end pump bearing

housing in line with the pump shaft axis and another was mounted on the discharge flange of the

pump. Dynamic pressure transducers were placed in ports located directly on the suction and

discharge nozzles of the pump.

Signals from the digital tape recorder were subsequently down-sampled at 2048 Hz using a National

Instrument PCI-4552 board providing four channels of simultaneous 16-bit data. Signal processing

thereafter used purpose-built LabVIEW and Matlab programs as described in [48, 49] to determine

statistical parameters of the signals and perform joint time frequency analysis.

8.3.3 AE Measurement

AE signals were measured using three Physical Acoustics wideband sensors (WD/WDI/WDIS),

each with 46dB total analog gain. These were placed on the suction and discharge flanges and

pump casing, held in place by magnetic mounts. Silicon grease was used as the couplant.

Signals were collected using two Physical Acoustics PCI2 boards installed into a 2.8GHz clone PC,

running AEWin for PCI2. In addition to all hit information, time dependant AE parameters were

collected every 10 seconds and waveform files (2048576 samples) every 30 seconds. A floating

trigger was set at 20dB (above the ASL). All channels were passed through analog bandpass filters,

set at 100-1000kHz, and signals were then digitised at 2MHz. Data was extracted and analysed

using a SQL server based data management and analysis system as detailed in [108].

Swansong II [35] and inverse Swansong II filters were applied to hit data sets to separate impulsive

burst data from continuous AE hits. Unfiltered hit data was also analysed to determine the

effectiveness of these filters.

AE waveform files were post-processed to determine: (a) traditional statistics (eg. RMS, Peak,

Kurtosis, Crest Factor), (b) 1/3rd octave band powers and (c) discrete wavelet level energies. In the

8-6


latter case, the Haar wavelet was used and signals decomposed to 12 levels. Waveform files were

also denoised, again using the Haar wavelet, to extract burst events from amongst the background

noise; standard AE hit features (eg. peak, energy, duration, rise time, counts etc) were then

determined for the extracted hits. Unless otherwise indicated, all parameters were averaged at each

constant flowrate (data collected as valves were being adjusted have been removed) and normalised

by the mean of the parameter at BEP. Graphs also show 90% confidence limits for these point

estimates, determined using a standard student-t test.

All analysis algorithms were implemented using code written in LabVIEW Version 6.1, and

incorporated standard VIs (virtual instruments, i.e. code modules) from the Signal Processing

Toolset for LabVIEW.

8.4 RESULTS AND DISCUSSION

8.4.1 Dynamic Pressure and Vibration Results (MH)

Contrary to the general findings discussed in the review section of this paper, dynamic pressure (0-

2kHz) measurements from both the suction and discharge flanges increased marginally with flow

(Figure 7-3). This is also contrary to data collected from other double-suction pumps by the

authors [47] and may reflect some of the modifications made to the internal configuration of this

particular unit. Vibrations measured on the discharge flange of the pump showed an opposite

trend (Figure 8-4) to dynamic pressure (Figure 8-3). Axial vibration showed the clearest changes at

low flows, increasing in magnitude and impulsiveness, indicated by a rise in the RMS and Kurtosis

values respectively (Figure 8-5).

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.60.50

0.75

1.00

1.25

1.50

1.75

2.00

2.25

2.50

2.75

3.00 889 rpm 997 rpm

Dis

char

ge F

lang

e Vi

b RM

S, m

m/s

2

Q/QBEP

Figure 8-4: Discharge Flange Vibrations.

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.62.0

2.5

3.0

3.5

4.0

4.5

5.0

5.5

6.0

6.5

7.0

Dyn

amic

Pre

ssur

e, k

Pa

Q/QBEP

Suction pressure, 889 rpm Suction pressure, 997 rpm Discharge pressure, 889 rpm Discharge pressure, 997 rpm

Figure 8-3: Changes in dynamic pressure with flow.

8-7


8.4.2 AE Results

Energy (unnormalised) as a function of time, flow and number of bursts are shown in Figure 8-6

(A) – (B) for suction flange AE and discharge flange AE respectively. The first observation is that

individual burst energy distributions differ markedly between sensors; both the number of hits and

the total energy is significantly greater on the discharge flange than for either of the other AE

sensors. As expected, these unnormalised parameters increase with pump speed. Pump casing and

discharge flange AE signals show some increase in total energy at high flow rates (>1.2BEP), but

this is not as severe as for partial flow.

Figure 8-6 also shows that flow at, or around, BEP is characterised by a small number of medium

energy bursts, whilst low flow conditions are accompanied by a much larger number of lower

energy bursts, that increase in number and energy as flow is reduced further, and have a much

greater spread of energies. This change is most pronounced in the suction AE signal, however

similar changes can also be detected in the other two signals (pump casing results not shown).

Further results indicate that most AE parameters analysed, on all channels, changed significantly at

low flows (see Figure 8-7). Only crest factor, kurtosis and hit counts showed no consistent trends.

Although third-octave energies showed similar trends to other parameters, confidence limits were

significantly wider, implying more variability in the underlying signals. Of the waveform features,

wavelet parameters (at all levels) showed the smoothest trends with tightest limits.

Energy related features of AE signals collected at the pump suction flange generally increased quite

significantly at low flows with only minor increases at high flows (Figure 8-12); the same features

from signals collected at the pump outlet increased at both low and high flows. It is hypothesised

(A) (B)

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.60.5

1.0

1.5

2.0

2.5

3.0

Brg

Axia

l Vi

b RM

S in

mm

/s2

Q/QBEP

889 rpm 997 rpm

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.62.0

2.5

3.0

3.5

4.0

4.5

5.0

5.5

6.0

6.5

7.0 889 rpm 997 rpm

Brg

Axia

l Vib

, Ku

rtos

is

Q/QBEP

Figure 8-5: Results of axial vibration results showing changes in (A) RMS and (B) Kurtosis.

8-8


that low flow AE activity is due to stress waves initiated by recirculation, whilst high flow AE

activity is due to incipient cavitation caused by reduction in NPSH margin. Signals mounted on the

pump casing indicate that its features were a combination of those observed at the suction and

discharge of the pump. For most parameters, in this particular pump, the local

minimum/maximum occurred at a flow that slightly exceeded BEP.

Figure 8-6: Energy of hits (y axis) versus time (x axis) versus number of hits (z axis) as flow is reduced from full flow to low flow at two different speeds for (A) suction AE and (B) discharge AE. Total energy is superimposed over each graph. Flows as a ratio of BEP for each test condition are also shown.

(A)

(B)

8-9


As flow was reduced from BEP, bursts also increased in amplitude and number but reduced in rise

time (Figure 8-9) and duration (almost identical to rise time trend so therefore not shown),

indicating a more impulsive profile. Denoising by Wavelets or Swansong II filtering removed most

events from around BEP, whilst leaving hits at low flows (Figure 8-8).

On signals collected on the suction and pump casing, individual mean hit energies decreased

slightly, before increasing as flow was reduced further (Figure 8-7A). This “dip” is presumed to be

(A) (B)

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6

1

10

100

Q/QBEP

Suction Flange Discharge Flange Pump Casing

Mea

n Ab

solu

te E

nerg

y of

Hit

s

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6

1

10

100

Q/QBEP

Suction Flange Discharge Flange Pump Casing

Mea

n W

avel

et E

nerg

y Le

vel 3

Figure 8-7: Most mean parameters increase at low flows for all signals, whilst only Pump and Discharge signals increase noticeably at high flows. (Results from test at 997 rpm shown.)

(A) (B)

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6

0.01

0.1

1

10

100

1000

Un-

norm

alis

ed H

it R

ate

Afte

r Sw

anso

ng II

Filt

erin

g

Q/QBEP

Suction flange, 889rpm Discharge flange, 889rpm Pump casing, 889rpm Suction flange, 997rpm Discharge flange, 997rpm Pump casing, 997rpm

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.61E-3

0.01

0.1

1

10

Un-

norm

alis

ed H

it R

ate

Afte

r W

avel

et D

enoi

sing

Q/QBEP

Suction flange, 889rpm Discharge flange, 889rpm Pump casing, 889rpm Suction flange, 997rpm Discharge flange, 997rpm Pump casing, 997rpm

Figure 8-8: Effect on hit rate (hits per second) by denoising processes: (A) Swansong filtering and (B) Wavelets. Missing points signify no hits detected at that condition. Results have not been normalised. Difference in magnitude is due to (A) being determined from hit data sets and (B) determined from waveform streaming files collected periodically.

8-10


the start of suction recirculation. A similar turning point is not seen on the discharge AE sensor:

mean hit energies increased with virtually any decrease in flow from BEP. It is therefore

hypothesised that discharge recirculation was probably occurring to some extent at all flows below

BEP. Although not calculated, this is plausible as this pump has an unusually large discharge.

As flow was reduced further (<0.5BEP) to very low flows (where thermodynamically induced surge

could start occurring), amplitude and energy of bursts in the pump suction decreased slightly.

However, of all waveform parameters, only several high frequency third-octave band parameters

were sensitive to this change (see Figure 8-11), indicating the decrease in overall energy and

amplitude is probably due to a loss of high frequency signal components. This hypothesis is

supported by the collective knowledge gained from measuring vibration and dynamic pressure

(albeit at lower frequencies) that generally associates low flow recirculation with low frequency, high

amplitude pressure pulsations. The underlying vortices would be expected to attenuate high

frequencies. Interestingly, this decrease was not observed in wavelet energies associated with

similar frequency bands (see Figure 8-11). This phenomenon was also not seen in the pump

discharge signal parameters, which continued to increase as flow reduced.

Discharge AE signals also became more intense as flow was increased above BEP, but not to the

same extent as seen at low flows. The rate of increase however, was substantially more severe than

due to flow reduction. Although basic energy parameters show measurable changes as flow is

deviated from BEP, filtering of these signals by wavelet denoising or Swansong II filtering to split

up signals into continuous and impulsive components generally increased the accuracy of trends

and decreased sensitivity to changes in pump speed. Wavelet parameters were also insensitive to

changes in pump speed (Figure 8-12).

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.60.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6 No Filter Post Swansong II filter Post Inverse Swansong II filter

Mea

n H

it R

ise

Tim

e

Q/QBEP

Figure 8-9: Effect of Swansong II filter on rise time signals.

(Discharge AE only shown, for test at 997 rpm.)

8-11


8.5 CONCLUSIONS AND RECOMMENDATIONS

This work indicates that acoustic emission can be a powerful and sensitive tool for detecting off

duty conditions in double suction pumps, presumably because the acoustic emission events relate

directly to pressure pulsations occurring within the fluid. Expected changes associated with off-

duty conditions can be clearly observed in hit-time intensity plots, which as shown by previous

research, are affected significantly by sensor position. Although basic energy parameters show

measurable changes as flow is deviated from BEP, further post processing of these signals

improves the accuracy of trends and decreases sensitivity to changes in pump speed. Results

indicate that on this particular double suction pump, collecting acoustic emission signals from

suction and discharge flanges can clearly identify the changes occurring within the pump that may

be associated with the onset of suction recirculation, discharge recirculation and incipient

cavitation. This research also shows that acoustic emission can be a more useful tool for analysing

hydraulic activity than traditional techniques such as dynamic pressure and/or vibration

monitoring. However, further testing is required on a variety of centrifugal pumps to determine if

they all show similar trends.

8-12


(A) (B)

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6

1

10

100

Mea

n En

ergy

in W

avel

et B

ands

Q/QBEP

d1 d2 d3 d4

Decreasing frequency

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6

1

10

100

Ener

gy in

Wav

elet

Ban

ds

Q/QBEP

d1 d2 d3 d4

Decreasing frequency

Figure 8-10: Changes across frequency octave bands for (A) suction AE and (B) discharge AE.

(A) (B)

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6

1

10

100

Mea

n En

ergy

in T

hird

Oct

ave

Band

s

Q/QBEP

128k Third-octave band 256k Third-octave band 512k Third-octave band 645k Third-octave band

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6

1

10

100M

ean

Ener

gy in

Thi

rd O

ctav

e Ba

nds

Q/QBEP

128k Third-octave band 256k Third-octave band 512k Third-octave band 645k Third-octave band

Figure 8-11: Changes across wavelet bands for (A) suction AE and (B) discharge AE.

(A) (B)

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6

1

10

889rpm 997rpm

Q/QBEP

Ener

gy in

2nd

Wav

elet

Ban

d

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6

1

10 889 rpm 997 rpm

Q/QBEP

Back

grou

nd R

MS

(pos

t W

avel

et d

enoi

sing

)

Figure 8-12: Selected normalised parameters show little sensitivity to changes in speed: (A) Energy in 2nd Wavelet Band, Suction AE signal and (B) Background RMS, Discharge AE signal.

8-13

9. FFLLOOWW MMOONNIITTOORRIINNGG OOFF

EENNDD--SSUUCCTTIIOONN PPUUMMPPSS


(a) Present the results of flow tests conducted on a number of single stage centrifugal pumps;

(b) Discuss the sensitivity of AE results to the testing process.

Abstract

99

A review of literature discussing the suitability of acoustic emission monitoring for detecting and

identifying off-duty and potentially damaging hydraulic conditions in centrifugal pumps was

presented in [107] and Chapter 8. The experimental and data processing methods were also

outlined therein. This chapter presents the results of flow tests conducted on a variety of smaller,

new, end suction pumps so that potential similarities and differences can be identified. All data was

collected on pumps being tested at Blakers Pumps’ test facility and processed using the techniques

described previously in this thesis.

9.1 INTRODUCTION

Part of the pre-commissioning process for many ANSI and API pumps is a pre-acceptance test to

verify the head-flow and efficiency-flow characteristics of the specific unit. These performance

tests are conducted in accordance with relevant pump standards [1, 116] on specially designed test

rigs at the supplier’s or manufacturer’s premises and may be witnessed by the customer. Testing

involves installing a new pump on the test rig and running it at various flows. Pressures, flows,

motor current, supply voltage, motor speed, VFD frequency and fluid temperatures are measured at

9-1

PhD Thesis – Chapter 9, Flow Monitoring of End-suction Pumps

each flow and the resulting head-flow and efficiency-flow curves are calculated. These are then

checked against allowable acceptance criteria. This is the easiest and most convenient time to

collect condition-monitoring data for the purposes of benchmarking new pumps, so it was

considered useful to determine if AE monitoring during this process was practical.

As will be discussed in this chapter, it was very difficult to obtain good quality AE data during

acceptance testing for a number of reasons. Typically such results would be removed from the data

set and not presented in published results. However these problems affect the applicability of AE

as a practical monitoring technique and suggest a further challenge, in addition to those listed in the

introduction to this thesis that must be overcome before valid, repeatable and reproducible results

are obtained from every pump test.

9.2 EXPERIMENTAL SETUP

AE data was collected from six new Gould’s single-stage end-suction centrifugal ANSI 3196 or API

670 process pumps at Blakers’ pump test facility. These are designated as Pumps 2, 4, 5, 6, 7 and 8.

Details of each pump are given in Appendix E.

Tests were conducted in two groups: Group 1 consisted of Pumps 2, 4 and 5 whilst Group 2

contained Pumps 6 to 8. Group 1 tests were performed at one motor speed, whilst Group 2 tests

were undertaken at two different motor speeds.

The author preferred not to acquire data during customer witnessed acceptance testing as this was

generally conducted very quickly and very little AE data could be collected from each flow

condition. If time permitted, AE data was collected immediately proceeding, or after the official

tests whilst the pump was installed on the test rig. Unfortunately, it was not uncommon for Blakers

to need to test multiple pumps in one day, so this additional time was often unavailable, thereby

necessitating the collection of data during witness-tests.

9.2.1 DAQ Hardware and Software

Group 1 data was collected using purpose- built AE electronics, a generic data acquisition board by

National Instruments (PCI-6110) and data acquisition software developed in LabVIEW (described

in Appendix B), whilst Group 2 data was collected using commercial off the shelf (COTS)

hardware and software by Physical Acoustics (as described in [107] and Chapter 8). The primary

differences between the acquisition platforms were (a) noise immunity, (b) maximum hit data

capture rates and (c) pre-triggering capability; in all cases the COTS system was superior.

9-2


Experience gained during Group 1 testing affected settings used for Group 2 testing, and therefore

results were expected to be more consistent8. Selected data acquisition settings that could

potentially affect results are given in Table 9-1.

Group 1 Group 2

Waveform sample rate (Maximum rate possible for 4 channels)

5MHz 2MHz

Number of points in waveform 1048576 ~2197152

Bandwidth 100-900kHz 100-1000kHz

Threshold type Floating Floating

Threshold calculation (hardware dependant)

ASL+10 RMS+3σ

Table 9-1: Selected data acquisition settings.

9.2.2 Sensors

Unfortunately, physical limitations of the (very small) pumps prevented certain sensors from being

positioned where desired; therefore sensor types and placements differed between tests. As the

pumps were to be delivered to Blakers’ customers immediately following testing, it was not possible

to modify the pumps so waveguides could be installed.

Table 9-2 summarises the locations of AE sensors mounted on each pump. Only signals collected

from the pump casing are considered in this chapter.

Pump casing

Gland Plate

Suction Pipe

DischargePipe

Suction Flange

Discharge Flange

PumpAE GlandAE SuctionAE DPipeAE SFlangeAE DFlangeAE

Test 2

Test 4

Test 5

Test 6

Test 7

Test 8

Table 9-2: Sensor data collected from each test.

8 Both sets of pumps were tested prior to work on the double-suction pump described in the previous

chapter.

9-3


9.3 OBSERVATIONS

9.3.1 Test 2

Pump 2 could only be tested during witnessed testing and with the motor controlled by a large

VFD (variable frequency drive). Although some noise contamination was suspected at the time of

data collection, its severity only became apparent during post-processing using techniques described

in Chapter 4.

Flow was increased gradually and incrementally from zero flow to maximum flow. Its head-flow

and efficiency-flow curves are given in Figure 9-1. As can be seen from Figure 9-1B, only a few

points were acquired and the curve at high flow rates appears quite flat. This complicated the

identification of best efficiency point. Fortunately, selecting the wrong point against which the

remaining data would be normalised would only change the relative position of the feature trends

with respect to other pumps; it would not affect the shape of feature trends.

9.3.2 Test 4

Pump 4 was tested on an air driven motor (due to the customer’s requirements); therefore results

are relatively free of VFD interference. Unfortunately some minor contamination remains because

power for the instrumentation was unwittingly taken from the VFD supply line.

Steady flow conditions were interspersed with reduced suction pressure conditions (as data from

NPSH tests was also being collected) but a few minutes of data could be collected at each flow.

Unfortunately, due to the limitations of the motor, high flow data could not be collected and the

maximum flow achieved was only 95% of BEP.

Head-flow and efficiency-flow curves for Pump 4 are given in Figure 9-2.

9.3.3 Test 5

AE data collection was not undertaken during witnessed testing. Flow could therefore be increased

gradually from zero flow to the pump’s maximum flow of approximately 126% BEP. Resulting

head-flow and efficiency-flow curves are given in Figure 9-3. Again, BEP was difficult to select

given the flat nature of the efficiency-flow curve at high flows Figure 9-3(B).

9-4


(A) (B)

0.00 0.25 0.50 0.75 1.00 1.250

60

65

70

75

80

85

90

95

100

105

Pump 2, 1202rpm

Perc

ent

of M

axim

um H

ead

Normalised Flow 0.00 0.25 0.50 0.75 1.00 1.25 1.50

0.0

0.2

0.4

0.6

0.8

1.0

1.2 Pump 2, 1202rpm

Effi

cien

cy

Normalised Flow

Figure 9-1: (A) Total head and (B) efficiency curves for Pump 2.

(A) (B)

0.00 0.25 0.50 0.75 1.00 1.250

60

65

70

75

80

85

90

95

100

105

Pump 4, 2693rpm

Perc

ent

of M

axim

um H

ead

Normalised Flow

0.00 0.25 0.50 0.75 1.00 1.25 1.50

0

5

10

15

20

25

30

35

40

45

50

55 Pump 4, 2693rpm

Effi

cien

cy

Normalised Flow


(A) (B)

0.00 0.25 0.50 0.75 1.00 1.250

60

65

70

75

80

85

90

95

100

105

Pump 5,2940rpm

Perc

ent

of M

axim

um H

ead

Normalised Flow

0.00 0.25 0.50 0.75 1.00 1.25 1.50

5

10

15

20

25

Pump 5,2940rpm

Effi

cien

cy

Normalised Flow


9-5


Unfortunately, the VFD was required for speed control, resulting in significant noise contamination

that predominantly affected low amplitude signals.

9.3.4 Test 6

Flows at two pump speeds were tested: 2236rpm and 2980rpm. The VFD was again used for

motor speed control, but new COTS electronics appeared less susceptible to VFD noise ingress

and features were relatively unaffected.

At 2236rpm, testing commenced at 57%BEP and increased to 130%BEP. It was then dropped to

BEP, before being reduced in steps to 27% BEP.

Data at 2980rpm was collected whilst Blakers were conducting pre-acceptance performance testing.

Consequently, flows were not changed gradually and incrementally. Instead, various conditions

were selected until enough data was obtained to verify the pump curve. As the pump was designed

to operate on light hydrocarbon rather than water, current limitations of the motor restricted the

maximum flowrate to slightly less than BEP (assumed to be BEP so that data could be normalised).

Only one to two minutes of data could be collected at each flow. Cavitation could be heard at

maximum flow; reducing flow very slightly stopped this noise.

Resulting head-flow and efficiency-flow curves for both motor speeds are given in Figure 9-4.

Efficiency-flow curves had obvious maxima, so best efficiency points were easy to estimate.

9.3.5 Test 7

Again, flows at two pump speeds were tested: 2263 rpm and 2673 rpm. Both tests were conducted

after acceptance testing, so two to five minutes of data was obtained at each flow, in decreasing

flow steps.

At the higher speed and maximum flow, cavitation could be heard. No other unusual events were

noted during either test.

Head-flow and efficiency-flow curves for Pump 7 are given in Figure 9-5. Again, these had

relatively defined maxima and therefore BEP was easy to estimate.

9.3.6 Test 8

Flow tests were conducted at pump speeds of 2236 rpm and 2427 rpm.

9-6


(A) (B)

0.00 0.25 0.50 0.75 1.00 1.250

60

65

70

75

80

85

90

95

100

105

Pump 6, 2236rpm Pump 6, 2980rpm

Perc

ent

of M

axim

um H

ead

Normalised Flow

0.00 0.25 0.50 0.75 1.00 1.25 1.5035

40

45

50

55

60

65

70

75 Pump 6, 2236rpm Pump 6, 2980rpm

Effi

cien

cy

Normalised Flow


(A) (B)

0.00 0.25 0.50 0.75 1.00 1.250

20

40

60

80

100


Perc

ent

of M

axim

um H

ead

Normalised Flow

0.00 0.25 0.50 0.75 1.00 1.25 1.50

10

15

20

25

30

35

40 Pump 7, 2263rpm Pump 7, 2673rpm

Effi

cien

cy

Normalised Flow


(A) (B)

0.00 0.25 0.50 0.75 1.00 1.250

20

40

60

80

100


Perc

ent

of M

axim

um H

ead

Normalised Flow

0.00 0.25 0.50 0.75 1.00 1.25 1.50

10

15

20

25

30

35

40 Pump 8, 2236 rpm Pump 8, 2427 rpm

Effi

cien

cy

Normalised Flow


9-7


Pump 8 was suffering an acoustic resonance of the pump’s drain line, which resulted in incredibly

high emission levels that may have overshadowed any changes in flow dynamics. A surprise

consequence of this was that it was very difficult to actually obtain steady flow conditions between

50 and 110% of BEP, and therefore the actual flows for each data set may not be correct.

Unstable pump operation can best be seen in the raw voltage signal coming from the flowmeter

(green line in Figure 9-7). Although the control valve position remained constant for the periods

designated by the horizontal red bars, at values corresponding to 50-90%BEP flow conditions do

not remain static. (Once the acoustic resonance had been rectified, this behaviour no longer

occurred.)

Interestingly, at first glance, resulting head and efficiency curves, given in Figure 9-6, do not seem

to illustrate this erratic behaviour. However, when they are compared to curves from an identical

pump (eg. Pump 7 shown in Figure 9-5) differences in curve shape at low flow can be seen: the

total head in Pump 8 actually decreases slightly at low flows resulting in a local maximum at 40-

50%BEP, rather than increasing to a maximum at zero flow. It also continues to drop between

50% and 90% of BEP, which is relatively flat in Pump 7. This type of curve is highly undesirable

for the reasons described (it is very difficult to maintain a constant flowrate). Unfortunately, due to

time constraints on the test rig, no AE data was obtained after the resonance problem was rectified.

Figure 9-7: Test 8 at 2236rpm. Green line shows raw voltage from flowmeter whilst black line shows changes in Absolute Energy for PumpAE. At 0.5-0.9BEP, flow does not remain constant (the control

valve was kept in the same position).

9-8


At high speed (2427rpm) and very low flows (<50% BEP) the discharge hose was seen to “jump

around” and pulsations could be heard emanating from the test rig.

9.4 AE RESULTS

As can be seen in Figure 9-9 and Figure 9-11, PumpAE trends varied widely between pumps.

For data collected from the pump casing, the following features showed the most consistent trends

and tightest confidence limits:

• All RMS and energy features

• Peak amplitudes

• Octave and third-octave band powers, in particular those having centre frequencies

between 256kHz and 512kHz inclusive.

• All wavelet energies

Kurtosis, crest factor, and most hit features (including rise time) displayed no recognisable trends or

had confidence limits that exceeded the trends they encompassed. For brevity, these results are not

included. Swansong filtering or wavelet denoising did not alter results significantly.

Unless otherwise indicated, feature graphs shown can be considered representative of other trends.

9.4.1 Changes at High flow

In all but one test, features increased at flows above BEP (see Figure 9-9 to Figure 9-11),

presumably due to the onset of incipient cavitation. Changes at high flow were evident in virtually

all feature trends and hit contour plots (see Figure 9-8) appeared similar to those presented in

Chapter 7.

In many tests, high flow was also characterised by an increase in the variability of AE Energy

signals, as illustrated in Figure 9-8A: Absolute Energy (the black line) is relatively smooth at low

flows, but becomes less consistent as flow increases above BEP. This would be expected after

commencement of cavitation due to the formation and collapse of vapour bubbles. The changes

are less obvious in Figure 9-8B because Absolute Energy is displayed using a log scale.

9-9


(A)

(B)

Figure 9-8: Hit intensity plots for (A) Pump 6 at 2263rpm, and (B) Pump 7 at 2263rpm. These show number of hits (z axis) versus PAC Energy (y axis) versus time (x axis). The superimposed black lines shows time averaged Absolute Energy. Red lines shows changes in flow (also labelled as %BEP). To emphasise low energy changes, small numbers of very high-energy events at low flow are not shown.

9-10


(A) (B)

0.00 0.25 0.50 0.75 1.00 1.25 1.500.10

1.00

10.00

Test2, 1202rpm Test4, 2693rpm Test5, 2940rpm

Nor

mal

ised

RM

S

Normalised Flow 0 0.25 0.5 0.75 1 1.25 1.5

0.10

1.00

10.00

Test6, 2236rpm Test6, 2980rpm Test7, 2263rpm Test7, 2673rpm Test8, 2236rpm Test8, 2427rpm

Nor

mal

ised

RM

S

Normalised Flow Figure 9-9: Normalised RMS of PumpAE for (A) Group 1 pumps and (B) Group 2 Pumps.

(A) (B)

0.00 0.25 0.50 0.75 1.00 1.25 1.500.10

1.00

10.00

100.00

Test2, 1202rpm Test4, 2693rpm Test5, 2940rpm

Normalised Flow

Nor

mal

ised

Den

oise

d RM

S

0 0.25 0.5 0.75 1 1.25 1.5

0.10

1.00

10.00

100.00


Normalised Flow

Nor

mal

ised

Den

oise

d RM

S

Figure 9-10: Normalised RMS of PumpAE after wavelet denoising for (A) Group 1 pumps and (B) Group 2 Pumps.

(A) (B)

0.00 0.25 0.50 0.75 1.00 1.25 1.500.10

1.00

10.00


Normalised Flow

Nor

mal

ised

Pea

k

0.00 0.25 0.50 0.75 1.00 1.25 1.50

0.01

0.1

1

10

100

1000

10000 Test6, 2236rpm Test6, 2980rpm Test7, 2263rpm Test7, 2673rpm Test8, 2236rpm Test8, 2427rpm

Normalised Flow

Nor

mal

ised

Wav

elet

Ene

rgy

- Le

vel

12

Figure 9-11: (A) Normalised Peak and (B) Normalised Energy in 12th Wavelet Band for Group 2 Pumps.

9-11


Hit intensity plots (Figure 9-8) also show a sudden drop in the number of very low energy (VLE)

hits as flow rises above BEP, whilst the number and amplitude of large energy hits increase.

9.4.2 Changes at Low Flow

All pumps were designated by HI standards as having ‘Low’ suction energies (listed in Appendix

E). Damaging recirculation and increasing AE levels at low flows were therefore not expected.

Significant variability was observed in all features at flows below BEP, indicating that changes in the

hydrodynamic behaviour at low flow were not consistent. This was supported by qualitative

observations reported at the time of testing, summarised in the section 9.3. As could be expected,

the smoothest trends were obtained (a) for pumps with flat head-flow curves and (b) when a

reasonable amount of data could be acquired at each condition (i.e. Test 7, designated by the blue

lines in Figure 9-9 to Figure 9-11).

Only Pump 8 (suffering from an acoustic resonance problem) showed any increase in parameters at

low flow. This is the most pronounced in the Denoised RMS signal (Figure 9-10B) where results

appear similar to those presented in Chapter 8, increasing at low and high flows. At the lowest

flows AE features dropped and the discharge hose was clearly seen to be pulsating. Thus it is

concluded that this decrease in AE features is probably due to significant low flow recirculation

attenuating acoustic emissions.

Pump 5 had another unexpected trend at low flow. Of all end-suction pumps tested, this had the

highest suction-specific-speed, suction energy (albeit still considered ‘Low’) and suction

recirculation factor, indicating that of any of these six pumps, it would be the most likely to suffer

from recirculation. However, it was also the smallest unit tested so signals might be hard to detect

above the noise floor at low flow. It is also possible, but less likely depending on the volume of

bubbles and exact sensor placement with respect to the location of any recirculation, that

vaporisation could be attenuating signals further.

Conversely, the trend could be genuine. A recent industry publication reported two small pumps

(1x1.5-7 and 3x4-15) for which normalised overall vibration results also had a positive or flat slope

to the left of BEP [89]. The author concluded this was due to the pump’s hydraulic design, but no

further details on the pump were given. He also suggested that the “bathtub” vibration curve

suggested by API 610 [1] (that was seen in AE results presented in Chapter 8), in which overall

levels increase at low and high flows (presumably caused by hydraulic excitations), is not

representative of all pumps. Test 5 results could be viewed as supporting these conclusions.

Further testing is required to identify what types of pumps fall into this category.

9-12


One final observation of AE data at low flows can be derived from the hit-intensity plots given in

Figure 9-8. BEP (Region 3) was characterised by small number of very low energy (VLE) hits

which increased in number but decreased in severity as flow reduced. At very low flows (<50%

BEP) they started to decrease in number and increase in amplitude. Unfortunately, because of their

very small energy, they contributed little to time-averaged AE features, which showed little change

at intermediate flows (eg. Absolute Energy, shown as black lines on Figure 9-8). Averaged trends

were also unaffected. These VLE patterns are much clearer in Pump 6 (Figure 9-8A) than in Pump

7 (Figure 9-8B), which is the larger of the two pumps, and are similar to the results presented in

Chapter 8.

9.4.3 Results at BEP

As already discussed, hit intensity plots (particularly in larger pumps) show a distinct change in the

energy and number of individual AE events at, or around, BEP (Region 3 on Figure 9-8).

Contrary to the reports of others reviewed in the previous chapter, in most pumps tested as part of

this thesis the local minimum (or maximum) in feature curves occurred below BEP. Three reasons

are suggested for these results. Firstly, the location of BEP in many pumps was very difficult to

verify. Secondly, some of these pumps (particularly 2 and 5) were specifically designed for low flow

operation and therefore the point at which the fluid enters the pump with the least shock would be

expected to occur at flows below BEP. Finally, AE features do not discriminate between emission

sources, but rather are an accumulation of all hydrodynamic conditions within the pump.

Therefore, the process of averaging favours sources with larger amplitude emissions, or those

which occur in much greater numbers, than smaller, less numerous events. In small pumps, many

of the relative changes that occur with varying flow rates are subtle, so may be hidden by the

averaging process.

9.4.4 Other observations

It was not uncommon for AE levels to take some time to stabilise, particularly at low or

intermediate flows. This is illustrated in Figure 9-8: at flows between 80 and 96% of BEP, AE

levels (black line) kept increasing for some time after a new flow condition (red line) was

established. It is not as noticeable at flows above BEP, or at very low flows. This tendency was

also much less likely in pumps that have flat head-flow curves, as illustrated in Figure 9-8; when

testing Pump 6, which had a highly sloping head-flow characteristic curve, AE values took much

longer to stabilise (Figure 9-8 A) than when testing Pump 7 (Figure 9-8 B), which had a much

flatter curve.

9-13


9.5 DISCUSSION & CONCLUSIONS

9.5.1 Detection of changes in hydraulic conditions

Some preliminary conclusions about hydraulic conditions in pumps can be derived from the data

presented. High flow conditions above BEP result in an increase in AE energy and amplitude

levels. Signal variability also increases. This supports the work of others, reported in Chapter 8,

which verified that AE levels increase substantially with the onset of incipient cavitation. At high

flows, this condition can be expected and thus it is concluded that the rise in AE features above

BEP is due to this phenomenon.

Although it unlikely that any of the healthy pumps (i.e. except for Pump 8) were suffering from

damaging recirculation due to their low suction energies, some changes in the acoustic emissions

were observed. Hit intensity plots show that low flow conditions, not associated with severe

pulsations, are characterised by a very large number of very small energy events. These decrease in

number and increase in amplitude as flow increases towards BEP. This supports the findings of

the previous chapter. At very low flows, AE hits also increase in energy and decrease in number

but much more rapidly than at flows approaching BEP. High flows, in addition to the reasonable

number of high energy events expected with incipient cavitation, are characterised by very few

small-energy discrete events. Therefore, it is likely that the relative size and energy of an individual

acoustic emission event is related to the size and energy of the originating vapour bubble.

These last observations have not been reported in previous work on AE monitoring of centrifugal

pumps. They also suggest that the best type of feature to identify off-duty conditions may be based

on extracting and quantifying these low energy events. Unfortunately, the extent of these changes

varies substantially with pump size and design and in some cases changes are barely noticeable.

Therefore further development and processing is required to identify how best to quantify and

normalise these low energy changes.

9.5.2 Pump Testing Process

Ordinarily, most of this data would have been omitted from a publication or thesis, justified by

claiming that the experimental procedure was flawed in some way. Instead, “good” results would

be used to illustrate the hypothesis that AE is a useful measure of flow conditions in a pump. As

Chapter 8 showed, there is some evidence to support this argument. However, “good” results are

only part of the story. If, instead, the hypothesis was to determine whether AE could be used by

industry to reliably identify hydraulic conditions in a pump, then it is important to appreciate when

and how such results can, and cannot, be obtained.

9-14


This chapter is therefore included to illustrate two points; firstly, that AE measurements are highly

sensitive to the testing environment and secondly, to propose that actually setting up a pump test to

acquire good AE data is not a simple task. The author suggests that smaller pumps are particularly

sensitive to hydraulic test conditions because acoustic emission energy variations with differing flow

conditions are relatively small. In larger pumps, such as the double suction pump discussed in the

previous chapter, changes in hydraulic conditions are sufficiently large as to (a) have a dominant

hydrodynamic mechanism and (b) swamp the inherent variability of AE signals.

This variability does not affect verification of basic head-flow and efficiency-flow characteristics

used to measure static conditions, such as pressure, flow and motor power because static

transducers are relatively immune to small-scale hydraulic disturbances. Acoustic emission sensors

on the other hand are very sensitive to this variability and therefore to acquire accurate flow-AE

feature curves, it is extremely important to obtain steady-state conditions at each flow tested. If an

extreme condition has been measured (eg. high flow cavitation or low flow recirculation) sufficient

time must be allowed for conditions to stabilise and any vapour bubbles to condense back into

liquid. Pumps must also be free of acoustic or hydraulic problems that could overshadow changes

in flow.

In essence, AE testing must be much slower and more deliberate than during traditional pump

acceptance tests. Only then can the AE analysts and pump owners be comfortable they have

obtained a true indication of the pump’s hydraulic behaviour.

9-15

10. CCOONNCCLLUUSSIIOONNSS &&

RREECCOOMMMMEENNDDAATTIIOONNSS 1100

10.1 DISCUSSION

The biggest challenge of this project was its multi-disciplinary nature. To develop a feasible seal or

pump monitoring solution based on AE requires knowledge not only of acoustic emission, pumps,

seals and their hydro-mechanical failure mechanisms, but also of computing, electronics design,

advanced signal processing, fluid dynamics and structural mechanics.

Although any final system developed for monitoring mechanical seals and/or hydraulic conditions

in industrial centrifugal pumps needs to be small, mobile and cost effective, the requirements of

such a system must first be established. This necessitated collecting data with more flexible (and

less mobile) equipment that could monitor a large number of parameters; the most suitable features

for early warning of incipient failures could then be determined.

As AE hardware was not initially available, a review of alternative systems then on the market for

monitoring acoustic emissions from machine components was conducted, based on their

underlying research and patents. The aim of this work was to: (a) determine whether purchasing

existing equipment for monitoring failures in seals and pumps was appropriate, (b) identify the

similarities and/or differences in these technologies so that a base set of requirements for in-house

electronics could be developed and (c) establish whether results of these different systems were

directly comparable. This work was presented at ICOMS in 2003 and is included as Chapter 2 of

this thesis.

From the perspective of this research project the results of the review were disheartening. Firstly, it

determined that although each manufacturer effectively measured some form of acoustic emission

10-1

PhD Thesis – Chapter 10, Conclusions & Recommendations

signal, each coined their own name for the stress wave or resulting output. Secondly, an analysis of

the signal processing techniques underpinning these systems determined that enough disparity

between variants, in sensor type, signal condition performed and output algorithms, precluded their

direct comparison. Furthermore, no one system was obviously superior to any other. Thirdly,

only one of the products considered was flexible enough to monitor different types of failure from

those for which it was originally designed (and in 2002 that option was still unavailable in Australia);

the remainder were specifically geared to detect one particular fault type (usually rolling element

bearing problems), and although the marketing literature stated they could were more generic, there

was little evidence to support these claims. Most systems were also not suitable for continuous on-

line monitoring.

For these various reasons, the author decided that the most appropriate option for this research

project was to develop a system in-house. This would allow generic signal processing techniques to

be used, with methods and results being expressed in well-understood terminology.

The final in-house system used commercially available AE sensors and preamplifiers, generic data

acquisition hardware, custom built filters and amplifiers, and software programs written by the

author in LabVIEW. The greatest limitation of this system was insufficient immunity against noise

emanating from variable frequency drives, inadequate hit capturing capability and the inability to

collect information prior to event triggering. Consequently, when commercial-grade equipment

was made available, this in-house system was forsaken9. Interestingly, although the new COTS

hardware did provide improved noise immunity, VFD interference still remains a problem at low

signal levels.

A pump-rig was also designed and constructed at UWA in order to collect experimental data from

mechanical seal failures under controlled conditions. Space and budget restrictions limited the size

of the facility and pump. This ultimately ensured that modern, well-designed mechanical seals

installed therein survived all punishment the author could inflict, including: heating the seal

chamber jacket to 130 degrees Celsius, cavitating the pump by throttling the suction valve, running

the seal completely dry for several days and injecting ceramic oxide particles into the seal chamber

via the flush. With the help of the seal manufacturer, it was determined that the pump was simply

too small and slow to generate conditions that would exceed the actual (rather than published)

design limits of typical mechanical seals. Some preliminary findings were published in [106], an

extract from which is included in Appendix F.

9 Work has continued by the designer, Mr Paul Kelly, with the aim of ultimately providing a cost-effective

electronics platform for any seal and pump AE monitoring algorithms developed by the author.

10-2


Despite the inability to initiate seal failures, the process was useful for identifying a number of other

issues that needed to be resolved before seal failure signals could be correctly interpreted and

actioned. These were listed in the introduction and how this thesis addressed these problems are

hereby summarised.

(a) Commercially available AE systems have been designed for structural monitoring

applications (eg. pressure vessels, cranes etc). AE signals from pumps are very

different and so traditional AE signal-processing techniques may need to be

supplemented and/or modified to ensure that they accurately characterise changes in

machinery AE signals.

AE signals in pumps are comprised of impulsive and continuous emissions. Depending on the

nature of the problem, traditional AE applications generally focus on one or other signal form.

Crack monitoring for example, is concerned with discrete events, whilst leak detection seeks

continuous emission signals. Rotating machines, such as centrifugal pumps, have a variety of

failure mechanisms or failure inducing hydrodynamic conditions, each comprising of differing

amounts of burst and continuous emissions.

COTS AE systems are primarily designed for structural monitoring and therefore have significant

capacity to acquire, characterise and store information pertaining to discrete AE events.

Unfortunately, as already mentioned, their capability to monitor continuous AE signals over

extended periods of time is limited. Therefore, to supplement the features available, time

waveforms were collected and subjected to further analysis. Additional signal descriptors, as

outlined in Chapter 3, were selected for their accepted suitability to vibration analysis of rotating

machinery signals. These included statistical variables, frequency octave-band energies and discrete

wavelet energies. As longer waveforms were required to overcome the inherent non-stationarity of

AE signals, processing algorithms to calculate these additional features from shorter subsets of the

signal were also devised by the author. In total, for all data collected from centrifugal pumps, 42

features were calculated (or extracted from COTS system data files) and logged.

Structural AE work, and the COTS systems developed for them, generally display test results

graphically. This is acceptable when analysing data from a time-limited test and when acceptance

criteria are known in advance. However, when new applications for extended monitoring are being

developed, some way of assimilating the data is required so that the acceptance criteria can be

established. To do this, ensemble averages with 90% confidence limits were calculated for all 42

features. As shown in chapters 8 and 9, this can be a very effective mechanism for detecting and

comparing trends but is not widely used by the AE community. This is probably because digesting

10-3


data in this way is not possible when relying solely on functionality built into vendor supplied

software packages.

(b) A variety of noise sources, peculiar to machinery monitoring applications, can appear

very similar to acoustic emission sources, and if not identified and removed will, at

best, lead to confusion when interpreting results or at worst, result in a perfectly healthy

seal or pump being removed from service.

Acoustic emission signals are very small (in the order of a few microvolts to a few millivolts).

Although amplifiers are used to increase the size of the signals throughout the measurement chain,

signals are still very sensitive to noise contamination. As frequencies get higher, amplitudes tend to

decrease and vulnerability to noise contamination increases. Furthermore, certain types of noise

only become problematic at higher frequencies. AE signals from rotating machines are even more

at risk because they are often collected using broadband non-resonant sensors measuring signals up

to 1-2MHz. Consequently, noise identification, management and mitigation needs to be very

carefully considered.

Problems are compounded when machinery is driven by variable frequency drives. These control

the speed of a motor by changing the voltage frequency. Unfortunately, they emit pulses of

broadband, high frequency electro-magnetic radiation. This couples to the equipment or AE

instrumentation resulting in very large, broadband, impulsive noise bursts superimposed onto AE

signals at a switching frequency of 10-50kHz. Noise bursts appear very similar to genuine acoustic

emissions, particularly when viewed in isolation (i.e. hit data) and are captured by AE

instrumentation as if they were genuine AE hits.

Furthermore, hydro-mechanical machines such as centrifugal pumps have a number of genuine

mechanisms that generate acoustic emissions. Some of these ultimately lead to failure, whilst others

are of little consequence and can probably be ignored. Therefore an important part of any fault

identification process is separating wanted from unwanted signals.

Chapter 4 reviewed a number of denoising techniques for identifying noise in AE signals at the

time of data collection, or removing the effects thereafter. Examples based on actual AE signals

collected from centrifugal pumps by the author, were given to demonstrate how each could be used

to improve legitimacy of analysed data. Although none of these techniques were new, in the case

of Swansong filtering and averaging, their application to noise management of machinery AE had

not been reported previously. Discrete wavelet analysis is a widely applied denoising technique, but

recommendations on how best to apply it to AE machinery monitoring could not be found. As a

result, the author spent considerable time and effort evaluating different processing options before

10-4


determining the best set of parameters. To simplify this process for others, the author’s

experiences are collated into a set of guidelines and recommendations included in this chapter. In

essence, Haar wavelets were found to be the most effective at removing background noise whilst

retaining individual events. Improved performance was obtained when threshold values were

determined individually for every level, and when digital filters were not applied. Energy in the

background and retained signals was quantified separately and treated as additional AE features that

related to the continuous and discrete parts of the signal respectively.

All techniques included in Chapter 4 were used to successfully identify or denoise signals collected

in later chapters. As shown in Chapter 8, Swansong filtering and wavelet denoising can improve

the extraction of trends that are otherwise swamped by noise. Unfortunately, no method tested

was found to reduce the effects of VFD contamination. Work in reducing the susceptibility of AE

instrumentation to VFD noise ingress is therefore required.

(c) As the time of failure is unpredictable, the amount of data that can be collected is

considerable. Furthermore, given that the ultimate aim is to detect incipient failure, all

of this data needs to be analysed and scrutinised. To do this manually would be

impractical.

Continuous pump monitoring resulted in 2-3GB of waveform and feature data every day. This

volume of information could not be analysed manually. Very quickly, even accompanying process

data and experimental conditions became difficult to track. Further complicating analysis, COTS

equipment stored its data in proprietary files that could not be read on a second computer without

purchasing additional software licences.

Although databases are a ubiquitous part of commercial life, for some reason they are not often

used to manage engineering research data. With the help of a systems architect, a database

management system was designed and implemented to manage metadata about AE tests, extract

features from proprietary COTS systems, analyse waveform data as per the algorithms described in

previous chapters and store results in a minable format. Originally devised as an Access database, it

very quickly exceeded that software’s capacity (about 2 million rows in any one table) and was

upgraded to a SQL Server system, with a Microsoft Access user interface. A paper was published

in the Journal of Mechanical Systems and Signal Processing regarding the development of this

database management system and is included as Chapter 5.

The biggest impact of the database management system on this research project was the ease with

which new processing algorithms could be trialled on any data collected and results incorporated

into the database. As soon as a new technique was developed (in any software language) and

10-5


compiled into an appropriate format, data was selected and automated processing commenced.

Results were automatically logged into the database, eliminating errors associated with manually

entering data into spreadsheets or tables. Depending on the nature of the analysis algorithms, up to

20GB of data could be processed in one day, limited primarily by the low cost processing and SQL

server hardware being used.

Another advantage of this system was that it offered huge flexibility in the selection of methods for

analysing, processing and presenting data. Graphs could be displayed using scientific graphing

packages such as Origin or SigmaPlot and signal processing could be performed using the

mathematical platform of choice (which in the author’s case was generally LabVIEW). When speed

or access to system function calls necessitated using lower level languages such as C++, these too

could be seamlessly incorporated into the system. Tools were used for the purpose they were

designed, improving functionality, increasing reliability, decreasing frustration and increasing the

proportion of time that was spent developing new analysis techniques and looking for trends and

patterns in the data.

Managing and maintaining the system, which by the end of this research project comprised of

tables with over 50 million rows, was no longer an onerous task. Archiving and backing-up became

straightforward. If no new analysis techniques were required, analysed raw data could be

permanently archived to optical disk; all interaction with the data now occurred via the database.

With this system, between 250 and 300 Gigabytes of raw AE data was extracted, analysed and

logged during the course of this research project (most of it relating to continuous mechanical seal

operation). The database itself is now over 11GB in size. Yet extracting pertinent results, even for

tests that were processed years ago, is a trivial process.

Whether data is collected by a single researcher in a university using a COTS AE system, or from a

collection of black boxes remotely transmitting extracted features from the field, this type of

database management system will ensure maximum information is extracted from any data

collected. Collectively, this data then becomes the basis for the next generation of expert-based

diagnosis systems.

(d) When it is not possible to locate a sensor where desired, waveguides can be used. The

effect of waveguide shape and material composition needs to be understood so that

signals can be interpreted correctly.

Mechanical seals are contained within a housing (generally referred to as a stuffing box) that is

incorporated into the pump. Acoustic emission generated by the seals must therefore travel via a

10-6


convoluted path to reach a sensor mounted on the stuffing box or pump casing. Alternatively, a

waveguide can be mounted through the stuffing box to the seal, giving signals a more direct

transmission path to the sensor. Waveguides are also useful when pumps contain very hot fluid

(e.g. bitumen pumps) as direct contact with a hot casing could damage AE sensors.

In Chapters 6 and 7, the effects of different waveguide dimensions, material compositions and face

angles were analysed on the temporal and frequency characteristics of simulated AE signals. Prior

research in this area is limited and relates solely to flat-faced, much longer, metallic guides.

Techniques used for analysis included basic feature comparisons, frequency averaging, short-time-

frequency analysis and continuous wavelet transformation. The latter two techniques also

facilitated basic modal analysis.

The work described in chapters 6 and 7 was published as two papers in the Journal of Acoustic

Emission and determined that in the case of short rods, with dimensions of the same order as the

AE wavelengths, signals received by the sensor were affected significantly by the natural

longitudinal frequency of the rod. This was particularly problematic in metallic guides (Mild steel,

SS316 and Aluminium) where strong resonant frequencies dominated responses; effects increased

with length and diameter. Whilst Alumina Oxide ceramic and Delrin rods attenuated signals

significantly, resonant effects were also minimised.

In addition to strongly attenuating signals, pointed metal guides (as opposed to flat faces with

adequate couplant) acted as a bandpass filter, whilst retaining or amplifying the undesirable

harmonic artefacts in the pass band. The amount of attenuation increased with the sharpness of

the point, explaining why previous reports disagreed on the amount of attenuation (if any) imposed

by using a pointed guide. Angled waveguides increased the spectral and modal complexity of

responses, probably due to the excitation of bending modes.

These findings indicate that short, cylindrical, solid waveguides should be avoided if at all possible.

Where absolutely necessary, rods should be as narrow and short as possible, made from very high

acoustic speed materials with inherent damping (i.e. ceramics) and have flat ends attached with a

good couplant.

(e) Acoustic emissions collected from the seal are strongly influenced by the hydraulic

conditions within the pump, and therefore to be able to differentiate between incipient

seal failure and an unfavourable hydraulic condition, the latter must be identified. To

do this, a pump’s individual acoustic emission signature needs to be characterised.

10-7


Although seal failures could not be invoked, observations during extended pump and seal

monitoring indicated that seal AE signals were strongly affected by the hydraulic conditions within

the pump. This, in turn, was affected by how far the pump operated away from its best efficiency

point. It is also well known that suboptimal pump operation affects mechanical seal longevity.

Consequently, it is crucial for any automated seal monitoring device to know where a pump is

operating on its curve so that it can: (a) identify conditions that lead to poor seal reliability and (b)

differentiate between AE signals caused by incipient failure and signals caused by inappropriate

operation.

To determine how AE is affected by moving away from best efficiency point, AE data was

collected from a large double suction pump and a number of new small single suction pumps at a

pump supplier’s test facility. Data was analysed using the processing techniques described in

Chapters 3 and 4 and managed by the system described in Chapter 5.

This work determined that energy features showed the most consistent trends, with tightest

confidence limits. In a large double-suction pump, denoising hit data with a Swansong filter and

waveform data with discrete wavelets improved trends significantly. These techniques were less

successful in smaller pumps for reasons that are described below. Octave-band and wavelet-band

energies showed similar trends to overall energies. Differences between bands were very subtle, if

any.

From the data it can be concluded that average features increase (normalised with respect to the

mean level at BEP) in magnitude above BEP due to the onset of incipient cavitation. Once severe

cavitation is established, AE levels drop. This supports the findings of previous work by others.

Average feature trends below BEP differ substantially between pumps. In the small end-suction

pumps, levels at low flows increased only slightly, from a minima between 0.6 and 0.9BEP, whilst

in one case levels actually decreased. In the much larger, and higher energy double-suction pump,

with a known tendency towards recirculation, features increased substantially at low flows. This

suggests that recirculation can increase AE levels, but probably only in high-energy pumps. More

testing on larger pumps is required to confirm this conclusion.

In addition to averaged feature trends, hit intensity plots were also analysed. These provided

additional information not obvious from in any of the 42 averaged features trended. Flows above

BEP are characterised by a large spread in the energies of individual hits. Flows below BEP on the

other hand, are typified by a very large number of very low energy (VLE) events. These increase in

number, but decrease in amplitude as flow reduces, until at very low flows (<50%BEP) the reverse

trend starts to occur (VLE hits decrease in number and increase in magnitude). This observation

has not been reported elsewhere and may be related to the formation of small recirculation vortices.

10-8


It is hypothesized that at very low flows, these consolidate, grow and collapse due to the varying

pressure distributions within the pump, releasing shock waves that emanate as much larger acoustic

emission events. Preliminary results indicate that in small, low energy pumps, these pulsations are

too small to affect overall signal features, whilst in larger pumps features increase substantially.

More testing is required to determine how pumps can be grouped by their AE-flow characteristics;

this would facilitate more definite conclusions about the causal mechanisms for AE within

centrifugal pumps, particularly at low flows.

Although denoising techniques, trialled as part of this thesis, seem to be very effective at separating

out discrete events in large pumps, they appear ineffective for small pumps. This is probably

because both techniques use fixed criteria for accepting or rejecting hits. As pumps reduce in size,

the amplitude and energy of these events decrease. Based on these observations, a variable filter

will be developed in follow-up work for separating out these events. It is hypothesized that this

filter would provide better identification of low flow in pumps than the AE features trialled to date.

Another conclusion derived from testing is that AE results are much more sensitive to test

condition stability than traditional condition monitoring parameters. Although accurate head-curve

and efficiency-curve data can be obtained from a relatively speedy testing process, AE results

require significantly more care and thus, more time. Flows need to be increased or decreased in

steps, from one extreme to the other so that conditions within the pump change gradually. This

reduces the time for AE levels to stabilise. Larger pumps with steep head-flow curves also take

longer to stabilise because small changes in flow result in larger deviations in the hydrodynamic

pressure field. In retrospect, these observations seem obvious given the known sensitivity of AE

to bubble dynamics, however its effect on the quality of pump test results has not been reported

previously.

Significantly more test work is required to be confident in the conclusions pertaining to

relationships between acoustic emission features and hydraulic conditions in centrifugal pumps.

However, with the systems established as part of this project and issues already overcome, acquiring

and processing this data is limited only by ready access to large numbers of variable size pumps and

the time to collect accurate AE data.

Finally, brevity dictates that some research has been omitted from this final thesis. Work

undertaken in areas of noise identification and mitigation, analysis of cavitation data and

relationships between acoustic emission features and mechanical seal operation is incomplete and

therefore has been omitted. This research is continuing and will be the subject of future

publications.

10-9


10.2 SUMMARY OF FINDINGS

Although the conclusions of this research were detailed in the previous section, the major findings

can be summarised by the following key points:

(1) Vibration monitoring techniques are well suited to analysing acoustic emission waveforms.

(2) Current AE hardware does not provide sufficient noise immunity against interference from

variable frequency drives. This can not be removed during post-processing as signals have

very similar characteristics to genuine AE events.

(3) Denoising techniques, such as wavelets (with Haar wavelets) for waveform data and

Swansong filtering for hit features, are very effective at dividing mixed-mode acoustic

emission signals into discrete and continuous parts.

(4) Database systems simplify the process of trialling new analysis techniques and collating or

interrogating results independently of software and hardware used to collect the data.

When applied to AE monitoring, database servers should be used as the amount of

information will very quickly exceed the limits of desktop database software.

(5) Cylindrical waveguides with dimensions in the same order as the acoustic wavelengths

should be avoided for acoustic emission work, particular when waveform fidelity is

required. If absolutely necessary, rods should be as narrow and short as possible, made

from ceramic material, with flat ends and attached with a good couplant.

(6) Acoustic emission analysis appears very effective at identifying changes in hydraulic

conditions within large centrifugal pumps. Energy features show the most consistent

trends. Overall energy features are much less consistent in smaller pumps, particularly at

low flows.

(7) By measuring AE signals from both the suction and discharge flanges of a pump

simultaneously, automating the process of flow identification in large pumps should be

possible using energy features from both sensors.

(8) Hit energy is a better indicator of hydraulic conditions in small pumps; distinct changes

occur in the distributions of very low energy events with changing flow. Similar patterns

are also observed in large pumps.

10-10


(9) Although this work was applied to a particular type of rotating machine, and in fact

originally a particular component, the approaches described herein can be applied to

continuous monitoring of any rotating or reciprocating machine.

10.3 RECOMMENDATIONS FOR FUTURE WORK

As has already been mentioned, this work is by no means over. The following issues still need to

be resolved before a system for detecting incipient failure of mechanical seals in centrifugal pumps

can be developed:

(1) Improve instrumentation immunity against VFD noise, or develop denoising techniques

that can subsequently remove data coinciding with the drive’s switching frequency from

the analysis dataset;

(2) Design a filter to quantify the magnitude and severity of VSE events typical of low flow

conditions;

(3) Benchmark a large number of new pumps to determine if these can be categorised by the

acoustic emission signature;

(4) Determine how acoustic emission signatures change as a pump deteriorates with age;

(5) Develop a system for failing seals in realistic and controlled conditions so that acoustic

emission data can be collected prior to failure. (A reasonably large pumping system,

running at high speed, should be used to increase the likelihood of success. Working with

mechanical seal manufacturers at their facilities may also increase the chance of obtaining

useful and interpretable incipient failure data.)

(6) Develop small, portable and cheap electronics design that can be permanently installed on

pumps to monitor changes in flow and acoustic emission during life. Current COTS AE

systems are far too large, not particularly portable and prohibitively expensive.

10.4 FINAL WORDS

Ironically, the ultimate goal for maintenance professionals is to make themselves redundant, by

improving equipment reliability and eliminating the need for maintenance all together. Realistically

(and fortunately for those of us at the start of our maintenance careers), that is highly unlikely to be

realised in the foreseeable future. The next best thing is to know exactly how and when equipment

10-11


will fail so that its operational and economic impact can be minimised, or even mitigated

completely.

I started this PhD with the aim of developing such a system for mechanical seals. Despite not

achieving this objective, I am satisfied with progress that has been made. With future input from

pump and seal manufacturers, it is only a matter of time before such a system can be and will be

developed. With the added ability to assess how a pump is performing, my maintenance colleagues

and I may also be able to improve the efficiency of our processing plants and reduce our electricity

demands; this not only returns an economic benefit to our employers but also reduces the damage

these plants inflict on our environment. We can then spend time pondering other obstacles to our

early retirement.

10-12

11. RREEFFEERREENNCCEESS

1111 1. (1995) "API 610 - 1995: Centrifugal Pumps for Petroleum, Heavy Duty Chemical and Gas Industry

Services", 8th ed: American Petroleum Institute.

2. (1995) "Application Note CM3014: What are Enveloping and SEETM?" SKF Condition

Monitoring, pp.1.

3. (1996) "Application Note CM3020: Taking and Interpreting SEETM Pen Readings", SKF

Condition Monitoring, pp.1.

4. (2002) "SWANTM Technology Overview", SWANTECH LLC, pp.1-5.

5. (2003) "PCI-2 Based AE System User's Manual, Part # 6301-1000", Princeton Junction NJ:

Physical Acoustics Corporation.

6. (2003) "PCI-2 Product Bulletin 102803", Physical Acoustics Corporation: Princeton Junction

NJ, pp.1-4.

7. Abbot, P.A. (1989) "Cavitation detection measurements on Francis and Kaplan hydroturbines" in

Proceedings of the ASME Proceedings of International Symposium on Cavitation and

Noise Erosion in Fluids Systems: ASME-Fluids Engineering Division, pp.55-61.

8. Alfayez, L., Mba, D. and Dyson, G. (2005) "The application of Acoustic Emission for detecting

incipient cavitation and the best efficiency point of a 60KW centrifugal pump; case study", NDT&E. In

press.

11-1

PhD Thesis – Chapter 11, References

9. Alfredson, R.J. and Mathew, J. (1985) "Time domain methods for monitoring the condition of rolling

element bearings", Mechanical Engineering Transactions, IEAust: pp.102-107.

10. Amaya, M. (1994) "Recent applications of acoustic emission testing for plant equipments" in

Proceedings of the Progress in Acoustic Emission, VII: Proceedings of the 12th

International Acoustic Emission Symposium, Sapporo, Japan, pp.305-312.

11. Bansal, V., Gupta, B.C., Prakash, A. and Eshwar, V.A. (1990) "Quality inspection of rolling

element bearings using acoustic emission technique", Journal of Acoustic Emission, 9(2): pp.142-

146.

12. Beattie, A.G. (1983) "Acoustic Emission, principles and instrumentation", Journal of Acoustic

Emission, 2(1): pp.95-127.

13. Bendat, J.S. and Piersol, A.G. (2000) "Random Data - Analysis and Measurement Procedures",

3rd. Edition ed, New York: John Wiley & Sons.

14. Bloch, H.P. (1978) "Acoustic incipient failure detection systems are successful at Exxon", The Oil

and Gas Journal, Feb. 6: pp.66-72.

15. Board, D.B. (2000) "Stress Wave Analysis of Turbine Engine Faults" in Proceedings of the

IEEE Aerospace Conference, Big Sky, MT, pp.79-94.

16. Board, D.B. (2002) "Stress Wave Analysis: Presentation to the Vibration Institute" [online], Ed.

Available from: [Accessed: 20/02/2003].

17. Boness, R.J. and McBride, S.L. (1991) "Adhesive and abrasive wear studies using acoustic emission

techniques", Wear, 149: pp.41-53.

18. Boness, R.J., McBride, S.L. and Sobczyk, M. (1990) "Wear studies using acoustic emission

techniques", Tribology International, 23(5): pp.291-295.

19. Brownell, R.B., Flack, R.D. and Kostrewsky, G.J. (1985) "Flow visualisation in the tongue region

of a centrifugal pump", The Journal of Thermal Engineering, 4(2): pp.35-45.

20. Budris, A.R. (1993) "The shortcomings of using pump suction specific speed alone to avoid suction

recirculation problems" in Proceedings of the 10th International Pump Users Symposium,

Texas A&M University.

21. Buzzacchi, G., Cartoceti, M., De Mechelis, C. and Sala, C. (1983) "In-field experience in

condition monitoring of rotating machinery by demodulated resonance analysis", Journal of Acoustic

Emission, 2(1): pp.11-18.

11-2


22. Carletona, C.J., Dahlgrena, R.A. and Tate, K.W. (2004) "A relational database for the monitoring

and analysis of watershed hydrologic functions: I. Database design and pertinent queries", Computers &

Geosciences. In press.

23. Carletona, C.J., Dahlgrena, R.A. and Tate, K.W. (2005) "A relational database for the monitoring

and analysis of watershed hydrologic functions: II. Data manipulation and retrieval programs",

Computers & Geosciences. In press.

24. Catlin, J.B.J. (1983) "The use of ultrasonic diagnostic techniques to detect rolling element bearing defects"

in Proceedings of the Machinery Vibration Monitoring and Analysis Meeting, pp.123-130.

25. Cempel, C. (1991) "Vibroacoustic condition monitoring", 1st ed, England: Ellis Horwood Ltd.

26. Cheeke, J.D.N. (2002) "Fundamentals and Applications of Ultrasonic Waves", Boca Raton,

Florida: CRC Press.

27. Chu, S., Dong, R. and Katz, J. (1995) "Relationship between unsteady flow, pressure fluctuations and

noise in a centrifugal pump - Part B: Effects of blade-tongue interactions", ASME Journal of Fluids

Engineering, 117: pp.30-35.

28. Cudina, M. (2003) "Detection of cavitation phenomenon in a centrifugal pump using audible sound",

Mechanical Systems and Signal Processing, 17(6): pp.1335-1347.

29. Derakhshan, O., Houghton, J.R., Jones, R.K. and March, P.A. (1991) "Cavitation monitoring

of hydroturbines with RMS acoustic emission measurement", ASTM STP 1077: pp.305-315.

30. Donoho, D.L. and Johnstone, I. (1995) "Adapting to unknown smoothness via wavelet shrinkage",

Journal of the American Statistical Association, 90: pp.1200-1224.

31. Dupont "Delrin Design Guide - Module III" [online], Ed. Available from:

http://plastics.dupont.com/plastics/pdflit/americas/delrin/230323c.pdf [Accessed:

15.03.2004].

32. Dvoracek, J., Petras, J. and Pazdera, L. (2000) "The phase of contact damage and its description by

help of acoustic emission", Journal of Acoustic Emission, 18: pp.81-86.

33. ETSU and AEAT PLC (2001) "Study on improving the energy efficiency of pumps", European

Commission, pp.70.

34. Finley, R.W. (1980) "Computerized preventative maintenance systems using modified acoustic emission

techniques", Materials Evaluation, 38: pp.15-20.

11-3

http://plastics.dupont.com/plastics/pdflit/americas/delrin/230323c.pdf


35. Fowler, T.J. (1993) "Structural integrity testing of process piping with acoustic emission" in

Proceedings of the 1st International Symposium of Process Industry Piping, Orlando,

Florida: National Association of Corrosion Engineers.

36. Fowler, T.J., Blessing, J.A., Conslisk, P.J. and Swanson, T.L. (1989) "The MONPAC system",

Journal of Acoustic Emission, 8(3).

37. Fraser, W.H. (1981) "Flow recirculation in centrifugal pumps" in Proceedings of the 10th Annual

Turbomachinery Symposium, Texas, USA: Turbomachinery Laboratory, Texas A&M

University, pp.95-100.

38. Ganesan, R., Das, T.K., Sikder, A.K. and Kumar, A. (2003) "Wavelet based identification of

delamination defect in CMP (Cu-Low k) Using Nonstationary acoustic emission signal", IEEE

Transactions on Semiconductor manufacturing, 16(4): pp.677-685.

39. Graff, K.F. (1975) "Wave motion in elastic solids", London: Oxford University Press.

40. Grosse, C., Finck, F., Kurz, J. and Reinhardt, H.W. (2004) "Improvement of AE technique using

wavelet algorithms, coherence functions and automatic data analysis", Construction and Building

Materials, 18: pp.203-213.

41. Guelich, J.F. and Bolleter, U. (1992) "Pressure Pulsations in Centrifugal Pumps", ASME Journal

of Vibration and Acoustics, 114: pp.272-279.

42. Guelich, J.F., Schwarz, D. and Carney, B. (1990) "Pump vibrations caused by cavitation" in

Proceedings of the Institution of Mechanical Engineers, Fourth European Congress, Fluid

Machinery for the Oil, Petrochemical and related industries, pp.55-60.

43. Hamstad, M.A., Downs, K.S. and O’Gallagher, A. (2003) "Practical Aspects of Acoustic

Emission Source Location by a Wavelet Transform", Journal of Acoustic Emission. "Practical

Aspects of Acoustic Emission Source Location by a Wavelet Transform," M. A. Hamstad, K. S. Downs

and A. O’Gallagher, submitted to J. of Acoustic Emission, 2003.

44. Hanchi, J. and Klamecki, B.E. (1991) "Acoustic emission monitoring of the wear process", Wear,

145: pp.1-27.

45. Hang, W., Kexiong, T. and Deheng, Z. (1997) "Extraction of partial discharge signals using

wavelet transform" in Proceedings of the 5th International conference on properties and

applications of dielectric materials, Seoul: IEEE, pp.322-325.

46. Hibbard, M. (2003) "Automated Detection of Aerophagia", Honours Thesis, UWA & Sir

Charles Gardiner Hospital.

11-4


47. Hodkiewicz, M. (2005) "The effect of partial-flow operation on the axial vibration of double-suction

pumps", PhD Thesis, University of Western Australia.

48. Hodkiewicz, M. and Pan, J. (2004) "Identification of transient axial vibration on double-suction

pumps during partial flow operation", Acoustics Australia, 32(1).

49. Hodkiewicz, M.R. and Norton, M.P. (2002) "The effect of change in flow rate on the vibration of

double suction centrifugal pumps", Proceedings of the Institute of Mechanical Engineers Part E:

Process Mechanical Engineering, 216: pp.47-58.

50. Holroyd, T. (2000) "Acoustic Emission & Ultrasonics", Oxford, UK: Coxmoor Publishing

Company.

51. Stresswave Technology Ltd (1991) "Method and apparatus for detecting variations in a process by

processing emitted acoustic signals", USA Patent No. 5005415, Final.

52. Holroyd Instruments Ltd. (1995) "Enhanced means of processing signals used to interpret the

condition of machinery", USA Patent No. 5473315, Final.

53. Holroyd, T.J., Tracey, T.E., Randall, N. and King, S.D. (1991) "Stress wave sensing - affordable

AE for industry", Acoustic Emission, Current Practice and Future Directions, ASTM STP

1077: pp.25-34.

54. Jamaludin, N. and Mba, D. (2002) "Monitoring extremely slow rolling element bearings: Parts I and

II", NDT&E, 35: pp.349-666.

55. Jamaludin, N., Mba, D. and Bannister, R.H. (2001) "Condition monitoring of slow-speed rolling

element bearings using stress waves", Proceedings of the Institution of Mechanical Engineers.

Part E. Journal of Process Mechanical Engineering, 215: pp.245-271.

56. Jantunen, E., Miettinen, J. and Ollola, A. (1993) "Maintenance and downtime costs of centrifugal

pumps in Finnish Industry" in Proceedings of the The 13th International Pump Technical

Conference, Pumps for a Safer Future, London, UK, pp.9-20.

57. Jiaa, C.L. and Dornfeld, D.A. (1990) "Experimental studies of sliding friction and wear via acoustic

emission signal analysis", Wear, 139: pp.403-424.

58. Jong, J. (1996) "Coherent phase wide band demodulation technique for turbomachinery cavitation

detection and monitoring" in Proceedings of the 50th Meeting of the Manufacturing Failure

Prevention Group, pp.621-632.

11-5


59. Jong, J.-Y., Dorland, W.D., Smith, R.A. and Zoladz, T. (2000) "Advanced bearing condition

diagnosis using acoustic emission techniques" in Proceedings of the 54th Meeting of the

Manufacturing Failure Prevention Group.

60. Karassik, I.J., Krutzsch, W.C., Fraser, W.H. and Messina, J.P. (1986) "Pump Handbook", 2nd

ed: McGraw Hill.

61. Kataoka, T., Yamashina, C. and Komatsu, M. (1987) "Development of an incipient failure

detection technique for mechanical seals" in Proceedings of the Fourth international pump

symposium, pp.121-129.

62. Khalfallah, R. and Simard, P. (1996) "An algorithm for the automatic segmentation of acoustic

emission signals" in Proceedings of the 14th World Conference on Non Destructive Testing

(14th WCNDT), New Delhi, pp.2555-2558.

63. Konno, D. and Yamada, Y. (1984) "Does impeller affect NPSHR?" in Proceedings of the 1st

International Pump Users Symposium, Texas A&M University, pp.29-35.

64. Kwon, O. and Joo, Y. (1997) "Source location in highly dispersive media by wavelet transform of AE

signals" in Proceedings of the Fourth Far East Conference on NDT (FENDT '97), Korea.

65. Li, C.J. and Li, S.Y. (1995) "Acoustic emission analysis for bearing condition monitoring", Wear,

185(1-2): pp.67-74.

66. Li, S.C. ed. (2000) "Cavitation of Hydraulic Machinery". ed. Li, S.C. Vol. 1. 2000, Imperial

College Press: London.

67. Li, X., Dong, S. and Yuan, Z. (1999) "Discrete wavelet transform for tool breakage monitoring",

International Journal of Machine Tools & Manufacture, 39: pp.1935-1944.

68. Lim, S., Lee, W. and Choi, M. (1994) "The condition monitoring of mechanical seals: Relation

between AE and torque" in Proceedings of the 12th Intl AE Symposium: Progress in AE VII,

Japan, pp.589-593.

69. Lingard, S., Yu, C.W. and Yau, C.F. (1993) "Sliding wear studies using acoustic emission", Wear,

162-164: pp.597-604.

70. Long, D.G. (2004) "Comments on Hilbert Transform Based Signal Analysis" [online], Ed. Provo

Utah:Brigham Young University, Dept of Electrical and Computer Engineering, Available

from: http://www.mers.byu.edu/docs/reports/MERS0401.pdf [Accessed: 30/06/2005].

11-6

http://www.mers.byu.edu/docs/reports/MERS0401.pdf


71. Mba, D. and Hall, L.D. (2002) "The transmission of acoustic emission across large-scale turbine

rotors", NDT&E, 35: pp.529-539.

72. McBride, S.L., Boness, R.J., Sobczyk, M. and Viner, M.R. (1989) "Acoustic emission from

lubricated and unlubricated rubbing surfaces", Journal of Acoustic Emission, 8(1-2): pp.S192-

S196.

73. McFadden, P.D. and Smith, J.D. (1984) "Vibration monitoring of rolling element bearings by the

high frequency resonance technique - a review", Tribology International, 17(1): pp.3-10.

74. Meeker, T.R. and Meitzler, A.H. (1964) "Guided wave propagation in elongated cylinders and

plates", Physical Acoustics, 1(A): pp.112-169.

75. Miettinen, J. and Andersson, P. (2000) "Acoustic emission of rolling bearings lubricated with

contaminated grease", Tribology International, 33: pp.777-787.

76. Miettinen, J. and Siekkinen, V. (1995) "Case study : Acoustic emission in monitoring sliding contact

behaviour", Wear, 181-183: pp.897-900.

77. Moholkar, V.S., Huitema, M., Rekveld, S. and Warmoeskerken, M.M.C.G. (2002)

"Characterization of an ultrasonic system using wavelet transforms", Chemical Engineering Science,

57: pp.617-629.

78. Morando, L. (1996) "Technology overview: Shock Pulse Method" in Proceedings of the 50th

Meeting of the Manufacturing Failure Prevention Group, pp.811-820.

79. Neill, G. (1998) "PC Based diagnostic system for the condition monitoring of rotating machines", PhD

Thesis, Heriot-Watt University.

80. Neill, G.D., Brown, E.R., Reuben, R.L., Sandford, P.M., et al. (1998) "Detection of recirculation

in pumps using acoustic emission" in Proceedings of the COMADEM, Melbourne, Australia,

pp.651-660.

81. Neill, G.D., Reuben, R.L., Sandford, P.M., Brown, E.R., et al. (1997) "Detection of incipient

cavitation in pumps using acoustic emission" in Proceedings of the IMechE, Part E: Journal of

Process Mechanical Engineering, pp.267-277.

82. Nelson, W.E. (1987) "Pump vibration analysis for the amateur" in Proceedings of the 4th

International Pump Users Symposium, Texas: Turbomachinery Laboratory, Texas A&M

University, pp.109-119.

11-7


83. Norton, P. and Andersen, V. (2000) "Guide to Microsoft Access 2000 Programming",

Indianapolis USA: SAMS.

84. Ono, K. and Cho, H.-J. (2004) "Rods and Tubes as AE Waveguides" in Proceedings of the

26th European Conference on Acoustic Emission Testing, Berlin: DGZfP, pp.593-603.

85. Onoe, M., McNiven, H.D. and Mindlin, R.D. (1962) "Dispersion of Axially Symmetric Waves in

Elastic Rods", Journal of Applied Mechanics - Transactions of the ASME, 29(12): pp.729-

734.

86. Pao, Y.-H. and Mindlin, R.D. (1960) "Dispersion of Flexural Waves in an Elastic, Circular

Cylinder", Journal of Applied Mechanics - Transactions of the ASME, 27(9): pp.513-520.

87. Paul, D. (1994) "Detection of change in processes using wavelets" in Proceedings of the IEEE-SP

International Symposium on Time-Frequency and Time-Scale Analysis, pp.174-179.

88. Perez, L.V., D'Attellis, C.E. and Ruzzante, J.E. (1997) "A model for acoustic emission signals and

burst occurrence estimation", Insight, 39(2): pp.83-87.

89. Perez, R.X. (2005) "Operating Centrifugal Pumps Off-design", Pumps and Systems Magazine,

April: pp.20-25.

90. Peterson, J.V. (2002) "Absolute Beginner's Guide to Databases", Indianapolis USA: Que.

91. Phillips, W.J. (2003) "Wavelets and filter banks course notes" [online], Ed. 1, Available from:

http://www.engmath.dal.ca/courses/engm6610/notes/notes.html [Accessed: 15-07-2004].

92. Pineyro, J., Klempnow, A. and Lescano, V. (2000) "Effectiveness of new spectral tools in the

anomaly detection of rolling element bearings", Journal of Alloys and Compounds, 310: pp.276-

279.

93. Qi, G. (2000) "Wavelet-based AE characterization of composite materials", NDT&E, 33: pp.133-

144.

94. Qian, S. (2002) "Introduction to time-frequency and wavelet transforms", Upper Saddle River, NJ:

Prentice Hall PTR.

95. Raj, B. and Jha, B.B. (1994) "Fundamentals of acoustic emission", British Journal of NDT, 36(1):

pp.16-23.

96. Redwood, M. (1960) "Mechanical Waveguides", London: Pergamon Press.

11-8

http://www.engmath.dal.ca/courses/engm6610/notes/notes.html


97. Robinson, J.C. (2001) "Detection and severity assessment of faults in gearboxes from stress wave capture

and analysis" in Proceedings of the Society for Machinery Failure Prevention Technology

(MFPT) 2001 Annual Conference.

98. Robinson, J.C. and Berry, J.E. (2001) "Description of PeakVue and illustration of its wide array of

applications in fault detection and problem severity assessment" in Proceedings of the Emerson

Process Management Reliability Conference 2001, pp.94.

99. Robinson, J.C., Canada, R.G. and Piety, R.G. (1996) "Vibration monitoring on slow speed

machinery: new methodologies covering machinery from 0.5 to 600rpm" in Proceedings of the 5th

International Conference on Profitable Condition Monitoring - Fluids and Machinery

Performance Monitoring: BHR Group Ltd, pp.169-182.

100. CSI Technology, Inc. (1999) "Machine fault detection using vibration signal peak detection", USA

Patent No. 5895857, Final.

101. Rzentkowski, G. (1996) "Generation and Control of Pressure Pulsations emitted from centrifugal

pumps: A review", ASME Pressure Vessel and Piping Division publication on Flow induced

vibration, 328: pp.439-454.

102. Sato, I., Yoneyam, T., Sato, K., Tanka, T., et al. (1991) "Applications of acoustic emission

techniques for diganosis of large rotating machinery and mass production products", Acoustic Emission:

Current Practice and Future Directions, ASTM STP 1077: pp.287-304.

103. Sato, I., Yoneyama, T., Sasaki, Y. and Sazuki, T. (1983) "Rotating machinery diagnosis with

acoustic emission techniques", Journal of Acoustic Emission, 2(1-2): pp.1-10.

104. Sato, I., Yoneyama, T., Sato, K., Tanaka, T., et al. (1988) "Diagnosis of rotating slides in rotary

compressors using acoustic emission technique", Journal of Acoustic Emission, 7(4): pp.173-178.

105. Serrano, E.P. and Fabio, M.A. (1996) "Application of the wavelet transform to acoustic emission

signals processing", IEEE Transactions on Signal Processing, 44(5): pp.1270-1275.

106. Sikorska, J.Z. (2004) "Monitoring the efficiency of centrifugal pumps with acoustic emission: Grant

B252M Final Report", Office, S.E.D., Editor, University of Western Australia: Perth WA.

107. Sikorska, J.Z. and Hodkiewicz, M. (2005) "Comparison of acoustic emission, vibration and dynamic

pressure measurements for detecting change in flow conditions on a centrifugal pump" in Proceedings of

the COMADEM, Cranfield University: Comadem.

108. Sikorska, J.Z., Kelly, P.J. and Pan, J. (2005) "Development of an AE data management and

analysis system", Mechanical Systems and Signal Processing. In press.

11-9


109. Sikorska, J.Z. and Pan, J. (2004) "The effect of waveguide material and shape on AE transmission

characteristics, Part 1: Traditional features", Journal of Acoustic Emission, 22: pp.264-273.

110. Sikorska, J.Z. and Pan, J. (2004) "The effect of waveguide material and shape on AE transmission

characteristics, Part 2: Frequency and Joint-Time-Frequency Characteristics", Journal of Acoustic

Emission, 22: pp.274-287.

111. Simpson, H.C., Macaskill, R. and Clark, T.A. (1966) "Generation of hydraulic noise in centrifugal

pumps" in Proceedings of the Conference on Vibrations in hydraulic pumps and turbines,

England: IMechE, pp.84-108.

112. SPM Instruments Inc. (1971) "Method and Arrangement for Determining the Mechanical State of

Machines", USA Patent No. 3554012, Final.

113. Sohoel, E.O. (1984) "Shock pulses as a measure of the lubricant film thickness in rolling element

bearings" in Proceedings of the Condition Monitoring '84, pp.148-161.

114. SPM Instruments Inc. (1985) "Method and Instrument for Determining the Condition of an

Operating Bearing", USA Patent No. 4528852, Final.

115. Staeb , J.A., Epema , O.J., van Duijn , P., Steevens , J., et al. (2002) "Automated storage of gas

chromatography–mass spectrometry data in a relational database to facilitate compound screening and

identification", Journal of Chromatography A, 974: pp.223-230.

116. Standard (2000) "Standard for Centrifugal Tests", ANSI/ Hydraulic Institute.

117. Suzuki, H., Kinjo, T., Hayashi, Y., Takemoto, M., et al. (1996) "Wavelet transform of acoustic

emission signals", Journal of Acoustic Emission, 14(2): pp.69-84.

118. Syverson, B. (2002) "Murach's SQL for SQL Server", Fresno, California: Mike Murach &

Associates Inc.

119. Takemoto, M., Nishino, H. and Ono, K. (2000) "Wavelet Transform - Applications to AE

Signal Analysis", Acoustic Emission - Beyond the Millennium: pp.35-56.

120. Tan, C.C. (1990) "Application of acoustic emission to the detection of bearing defects" in Proceedings

of the Institution of Engineers Australia Tribology Conference, Brisbane: IEAust, pp.110-

114.

121. Tan, C.C. (1991) "Adaptive noise cancellation of acoustic noise in ball bearings" in Proceedings of

Asia-Pacific Vibration Conference on Machine Condition Monitoring, pp.28-35.

11-10


122. Tandon, N. and Choudhury, A. (1999) "A review of vibration and acoustic measurement methods for

the detection of defects in rolling element bearings", Tribology International, 32: pp.469-480.

123. Tandon, N. and Nakra, B.C. (1992) "Comparison of vibration and acoustic measurement techniques

for the condition monitoring of rolling element bearings", Tribology International, 25(3): pp.205-212.

124. Tian, Y., Lewin, P.L., Davies, A.E., Swingler, S.G., et al. (2002) "Comparison of on-line partial

discharge detection methods for HV cable joints", IEEE Transactions on dielectrics and electrical

insulation, 9(4): pp.604-615.

125. Tognola, G., Ravazzani, P., Ruohonen, J. and Grandori, F. (1995) "Time-frequency analysis of

acoustic emissions: a wavelet approach" in Proceedings of the IEEE 17th Annual Conference of

the Engineering in Medicine and Biology Society, pp.1067-1068.

126. SKF Engineering and Research Centre (1986) "Method and means for detecting faults in defects in

moving machine parts", USA Patent No. 4768380, Final.

127. Wang, Q. and Chu, F. (2001) "Experimental determination of the rubbing location by means of

acoustic emission and wavelet transform", Journal of Sound and Vibration, 248(1): pp.91-103.

128. Wood, B.R.A., Flynn, T.C., Harris, R.W. and Noyes, L.M. (1991) "The use of waveguides in

acoustic emission monitoring projects", Acoustics Australia, 19(3): pp.87-89.

129. Wua, Y., Grahama, G., Lub, X., Afaqa, A., et al. (2004) "Configuration monitoring tool for large-

scale distributed computing", Nuclear Instruments and Methods in Physics Research A, 534:

pp.66-69.

130. Xu, M. (1995) "Spike Energy and its Applications", Shock and Vibration Digest, 27(3): pp.11-

17.

131. Ziola, S. and Searle, I. (1997) "Automated source identification using modal acoustic emission",

Review of progress in quantitative non-destructive evaluation, 16: pp.413-419.

11-11

PhD Thesis – Appendix B J.Z. Sikorska

B-1

PhD Thesis – Appendices

AAPPPPEENNDDIIXX AA DDEERRIIVVAATTIIOONN OOFF SSTTAATTIISSTTIICCSS

A.1 MEAN

Let x0,..,x N-1 be a sample group where N is the number of samples in the group.

1

0

[ ]N

iN

x i

Nμ

−

==∑

(A1.1)

Divide N into M subsets, each having S=N/M samples.

If we define ‘l’ as the position in the current array, and ‘k’ as the number of the subset (both

starting from zero) then:

(A1.2) 1 1

0 0

[ ] [ , ]M S

k l

x i x k l− −

= =

= ∑∑

So:

1 1

0 0

1 1

0 0

1

0

[ , ]

[ , ]

M S

k lN

M S

k l

M

Sk

x k l

N

x k l

M S

M

μ

μ

− −

= =

− −

= =

−

=

=

=×

=

∑∑

∑∑

∑

(A1.3)

where

1

0

[ ]S

lS

x l

Sμ

−

==∑

(A1.4)

A-1


A.2 VARIANCE

212

01

2 2

0

1 1 12 2

0 0 0

1 12 2

0 0

1 12

20 0

0

( [ ] )1

1 ( [ ] 2 [ ] )

1 [ ] 2 [ ]

1 [ ] 2 [ ]

[ ] [ ]2

[ ]

NN

iN

N Ni

N N N

N Ni i i

N N

N Ni i

N N

i iN N

SQk

x iN

x i x iN

x i x iN

x i x i NN

x i x i

N N

X k

μσ

μ μ

μ μ

μ μ

μ μ

−

=

−

=

− − −

= = =

− −

= =

− −

= =

=

−=

−

= ⋅ − +

⎧ ⎫= ⋅ − +⎨ ⎬⎩ ⎭⎧ ⎫= ⋅ − +⎨ ⎬⎩ ⎭

⎛ ⎞⎜ ⎟⎜ ⎟= − ⋅ +⎜ ⎟⎜ ⎟⎝ ⎠

=

∑

∑

∑ ∑ ∑

∑ ∑

∑ ∑

1 1

20

[ ]2

M M

Sk

N N

X k

M Mμ μ

− −

=

⎛ ⎞⎜ ⎟⎜ ⎟− ⋅ +⎜ ⎟⎜ ⎟⎝ ⎠

∑ ∑

(A1.5)

where

12

0

[ ][ ]

S

lSQ

x lX k

S

−

==∑

(A1.6)

and

1

1

[ ][ ]

S

lS

x lX k

S

−

==∑

(A1.7)

A-2


A.3 SKEWNESS

31

30

12 2

30

1 1 1 13 2 2

30 0 0 0

1 1 13 2 2 3

30 0 0

3

( [ ] )1

1 ( [ ] 2 [ ] ) ( [ ] )

1 [ ] 3 [ ] 3 [ ]

1 [ ] 3 [ ] 3 [ ]

1

NN

iN

N N Ni

N N N N

N Ni i i i

N N N

N Ni i i

x iSkN

x i x i x iN

x i x i x iN

x i x i x i NN

X

μσ

μ μ μσ

μ μσ

μ μ μσ

σ

−

=

−

=

− − − −

= = = =

− − −

= = =

−= ⋅

= ⋅ − + ⋅ −

⎧ ⎫= ⋅ − + −⎨ ⎬

⎩ ⎭⎧ ⎫

= ⋅ − + −⎨ ⎬⎩ ⎭

=

∑

∑

∑ ∑ ∑ ∑

∑ ∑ ∑

3N

N

μ

1

2 30

[ ]3 2

M

CUk

N N N

k

Mμ ψ μ

−

=

⎧ ⎫⎪ ⎪⎪ ⎪− ⋅ +⎨ ⎬⎪ ⎪⎪ ⎪⎩ ⎭

∑

(A1.8)

where

13

0

[ ][ ]

S

lCU

x lX k

S

−

==∑

(A1.9)

12

0

[ ][ ]

S

lSQ

x lX k

S

−

==∑

(A1.10)

1

0

[ ][ ]

i

S

lS

x lX k

S

−

==∑

(A1.11)

and ψN is the RMS of the full sample set.

A-3


A.4 KURTOSIS

41

40

1 1 1 1 14 3 2 2 3 4

40 0 0 0 0

1 1 1 14 2 2 3 4

40 0 0 0

1

04

( [ ] )1

1 [ ] 4 [ ] 6 [ ] 4 [ ]

1 [ ] 4 [ ] 6 [ ] 4 [ ]

[ ]1

NN

i

N N N N N

N N N Ni i i i i

N N N N

N N Ni i i i

M

QUk

x iKN

x i x i x i x iN

x i X x i x i x i NN

X k

μσ

μ μ μ μσ

μ μ μσ

σ

−

=

− − − − −

= = = = =

− − − −

= = = =

−

=

−= ⋅

⎧ ⎫= ⋅ − + − +⎨ ⎬

⎩ ⎭⎧ ⎫

= ⋅ − + − +⎨ ⎬⎩ ⎭

=

∑

∑ ∑ ∑ ∑ ∑

∑ ∑ ∑ ∑1 1 1

2 30 0 0

1 1

2 2 40 04

[ ] [ ] [ ]4 6 4

[ ] [ ]1 4 6 3

M M M

CU SQ Sk k k

N N N

M M

QU CUk k

N N N N

X k X k X k

M M M M

X k X k

M M

μ μ μ

μ μ ψ μσ

− − −

= = =

− −

= =

⎧ ⎫⎪ ⎪⎪ ⎪− ⋅ + − ⋅ +⎨ ⎬⎪ ⎪⎪ ⎪⎩ ⎭⎧ ⎫⎪ ⎪⎪ ⎪= − ⋅ + −⎨ ⎬⎪ ⎪⎪ ⎪⎩ ⎭

∑ ∑ ∑ ∑

∑ ∑

4Nμ

(A1.12)

where

14

0

[ ][ ]

S

lQU

x lX k

S

−

==∑

(A1.13)

and remaining variables were defined previously.

A-4


AAPPPPEENNDDIIXX BB

IINN--HHOOUUSSEE DDAAQQ SSYYSSTTEEMM

B.1 HARDWARE

B.1.1 Sensors

Commercially available AE sensors were used for all test work.

B.1.2 Signal conditioning equipment

As DC power supply to the preamplifiers was coupled to the signal line, an AC-decoupling network

was incorporated into specially designed signal conditioning electronics; these also encompassed

additional 20-40dB amplification and 8th order 1MHz Butterworth low pass anti-aliasing filters.

The total analog gain could be selected in 20dB increments ranging from 20 to 100dB for the

B1080 sensor, between 40 and 60dB (originally 60-80dB) for the WD sensor channel and between

40 and 80dB for the WDI channel. Differences arose because of the limitations associated with

each sensor’s preamplifier.

B.1.3 Computing & DAQ

Data was acquired on a self-assembled clone PC, fitted with a Genuine Intel PC single processor

(Pentium IV, 833MHz) motherboard that also incorporated network and sound hardware, 480MB

RAM and 40GB IDE hard drive. An additional case fan was fitted to extract heat generated by the

DAQ hardware (detailed below). The operating system was Windows 98 SE and further software

installation was minimized to prevent sources of corruption. Remote monitoring was managed by

TightVNC freeware.

The PC was fitted with two National Instruments PCI data acquisition cards: PCI-6110, capable of

simultaneously sampling 4 analog input channels at up to 5MHz, which was used for AE signal

acquisition; and PCI-6024, which was used for collecting multiplexed data from 8 other analog

inputs at combined rates of up to 200 kHz. A National Instruments BNC-2110 break-out box

B-1


facilitated the connection of conditioned pressure and temperature signals to this board. A

specially made cable was fabricated to connect the AE signal-conditioning unit to the PCI-6110

directly.

Both the PCI-6110 and PCI-6024 include a number of on-board analog amplifiers for additional

gain. These can be set from the National Instruments software program Measurements and

Automation Explorer (designed for configuration of NI hardware and generally referred to as ‘MAX’)

or programmatically from LabVIEW. MAX can also setup relationships between the incoming

analog signals and actual physical parameters for automatic scaling. This functionality was used for

all non-AE data channels, as it was not envisaged that gain would need to be altered during the

course of an experiment. The dynamic range of AE signals however, would most likely exceed the

optimum span of any one amplifier and consequently on-board gain selection was incorporated

into the main acquisition program. Additionally, pre-board gain was also included as a user control,

ensuring that all data was saved in terms of sensor units and therefore independent of the actual

amplifiers selected.

Once the system was configured and operating satisfactorily, an image was taken. This was then

reloaded prior to every new experiment to ensure stable and repeatable operation.

B.2 AE SOFTWARE

B.2.1 LabVIEW programs summary

Data acquisition and analysis programs were written predominantly in LabVIEW 6.02 (National

Instruments programming software) with a few key routines written (by others) in C++ to optimize

performance.

Subroutines for high-speed data-capture were based on the NIHSDL (high speed data logging)

API, which performed data streaming of the digitized data directly to disk, where it was stored in

compressed (16-12bit) little endian binary form. Functionality to facilitate software triggering and a

timeout was also included. A text header was written to the top of every file so that the sampling

and testing parameters could be viewed through a standard text editor.

B-2


B.2.2 Acquisition functionality

During pump operation, data was continuously acquired from three AE sensors and various

pressure and temperature transducers. Controlled by the LabVIEW program ‘RecordData3.vi’,

several processes occurred in parallel. These included:

1) Measuring RMS (and standard deviation) of the AE signal from one channel (typically the

sensor at the seal face and measured every 30 seconds), which is then logged to a text file;

2) Periodic recording of the time signals from all channels in long or medium length files (typically

1048576 and 131072 scans, at 60-minute and 30-second intervals respectively);

3) Recording a short time file (typically 4096/8192 scans) each time a threshold was exceeded on

the same channel on which RMS is being logged. The threshold could either be set to a fixed

voltage or as a certain number of standard deviations above RMS. Theoretically 10 up to

triggered acquisitions could be recorded per second.

Note: Unfortunately, the PCI-6110 does not support pre-triggering, so no information can be

obtained about a particular event prior to the threshold being exceeded. Therefore it is likely

that rise times and durations will be less accurate that using traditional AE hardware, which

does support hardware pre-triggering.

4) Measuring and logging the averaged values indicated by a variety of other transducers. This

occurred immediately preceding every AE acquisition and at periodic intervals in between.

These were then averaged and included in the AE file header, as well as being logged to a

separate text file.

5) Displaying the most recent AE file and its FFT on screen (updated periodically);

6) Displaying a log of current instrumentation on screen;

7) Issuing a message based on the current error status to an external computer, via TCP/IP, when

requested by this external computer. This formed the basis of remote alarming so that any

problems with the acquisition PC could be immediately rectified, thus minimizing the risk of

missing failure events due to hardware or software failures.

B-3


8) Logging acquisition parameters and filenames in the database management system described in

Chapter 5.

Sample rates, file intervals and file lengths (number of samples) could be changed by the user

without stopping the program. Similarly, gain for AE channels could also be controlled through

the RecordData3 interface during program operation. This allowed the user to change gain settings

when the signal amplitude increased or decreased, without risking loss of acquisition time.

B.2.3 Basic Processing Functionality

Data processing was managed by a database management system described in Chapter 5.

An envelope routine was written to extract one or more individual events from the background

noise. After determining AE burst features of each event in a particular file, the Event Statistics were

calculated and recorded. Burst parameters were also recorded after pre-processing with a Haar

wavelet-denoising algorithm. These are collectively referred to as Denoised Statistics. Additionally, a

number of other statistical and frequency parameters, as defined in part 2, were calculated on the

longer files (as a large number of points are required for these to be statistically valid) and are

referred to as General Statistics and Frequency Statistics respectively. The actual parameters in each

group include the following:

Event Statistics

- Number of events (usually only 1 for triggered data files)

- Event rate

- Average Rise time

- Average Event Duration

- Maximum peak

- Average event peak

- Event energy

Denoised Statistics

- DEvents = Number of events per second after denoising

- AveDBurstPeak = Average all the peaks detected in a datafile after denoising

- MaxDBurstPeak = Highest peak detected in a datafile after denoising

- AveDBurstEnergy = Average of all energies from bursts detected after denoising

B-4


- MaxDBurstEnergy = Maximum energy of bursts detected after denoising

- AveDBurstDuration = Average duration of all bursts detected after denoising

- DBurstRMS = RMS of the signal after denoising

- BackgroundRMS = RMS of the signal that has been removed by denoising

- DPR2 = Background RMS / Max DBurst Peak

General Statistics

- Mean

- RMS

- Variance

- Skewness

- Kurtosis

- Crest Factor

Frequency Statistics

- PeakFreq = Amplitude of highest peak in FFT

- 63kPower = Normalised power in 63k octave band




- 1MPower = Normalised power in 1024k octave band

The latter 7 frequency parameters were based on an RMS averaged amplitude spectrum.

All parameters were imported, with respective confidence limits, variances and max and mean

values were imported into Origin (Version 7) for graphing.

B-5


C-1

AAPPPPEENNDDIIXX CC

SSIIGGNNAALL PPRROOCCEESSSSIINNGG AALLGGOORRIITTHHMMSS


C.1 CALCULATE WAVELET DENOISED STATISTICS

C-1


C-2


D-1

AAPPPPEENNDDIIXX DD

AAEEDDAATTAA TTAABBLLEESS


D.1 DETAILS OF PRIMARY AND METADATA TABLES

Table Name Primary key(s) Additional indexes

Foreign key(s) Description Populated by:

AE_Session SessionID Info-table that lists parameters describing a data collection session and are entered for every session (either prior to, or subsequently).

Includes plant/location, TAG, data directory, start date, general comments, archive DVD number, DAQ hardware used and acquisition software version.

Manual entry into ‘Session_Info’ Form

AE_Session_Parameters SessionID, ParameterName

AETPID SessionID Info-table that lists parameters that are dynamic between sessions, but are constant for the duration of one test (eg. Flow, pressure, number of channels).


Channel_Info SessionID ChanNum AE/Parametric

SessionID Data setup for each AE or parametric channel (eg. Name/location, sensor serial number, gain)


Tests NewTestID SessionID Data-table containing the description, start and end times/time-marks of each Test. ‘Manage Timestamps’ Form +VB/SQL code

Test_parameters NewTestID ParamName ParamValue

PARID NewTestID Details of parameter names and their values that are kept constant for the Test. (eg. Flow, pressure, waveguide material, length)

‘Manage Timestamps’ Form +VB/SQL code

Files Filename Channel Filetype

Fileref SessionID File information (eg. name, filetypes, file start time, PAC test start time, numbers of hits/waves etc in the file, processed flags, error flags)

Enter_Files Form + VB code

DTA_HitData Fileref Channel LVRelTime Filepos

HitID Fileref Hit data that has been extracted from the PAC ‘DTA’ files. VB/Labview code behind “Process Files” form. (Process DTA option)

DTA_Waves Fileref Filepos Channel LVRelTime

WaveID Fileref Stores location of each wave within the ‘DTA’ file and the data to interpret it (eg. Gain, sampling rate, samples, channel.)

VB/Labview code behind “Process Files” form. (Process DTA option)

DTA_Timestamps Fileref LVRelTime

TSID Fileref Time and value of each PAC timestamp extracted from the ‘DTA’ file. Additional timestamps can be entered manually.


DTA_TDData Fileref Channel TDDTime

TDDID Fileref Time, value and channel of each PAC time dependant data parameter extracted from the ‘DTA’ file.


PAC_WF_Info Fileref Channel

Fileref (1:1) Information required to interpret and compare waveform data results and/or analysis. (eg. sampling rate, number of samples.)

VB/Labview code behind “Process Files” form. (Process WFS option)

D-1


D-2

D.2 DETAILS OF SELECTED SECONDARY DATA TABLES

Table Name Primary key(s) Additional indexes

Foreign key(s) Description Populated by:

GenStats Fileref Channel

Fileref Results of traditional statistical analysis performed on a particular channel of a NI-6110 binary (‘bin’) file.

VB code & Labview dll behind “Process Files” form. (Process GenStats option)

PAC_WF_GenStats Fileref Channel

Fileref Results of traditional statistical analysis performed on a particular channel in a ‘WFS’ file. VB code & Labview dll behind “Process Files” form. (Process WFS option)

PAC_WF_FreqStats Fileref Channel

Fileref Results of 1/3 octave band analysis performed on a particular channel in a ‘WFS’ file. VB code & Labview dll behind “Process Files” form. (Process WFS option)

PAC_WF_FreqPeaks Fileref Channel

Fileref Frequencies and amplitudes of the highest frequency peaks (up to 10 per FFT) in the averaged frequency spectrum performed on a particular channel in a ‘WFS’ file.

VB code & Labview dll behind “Process Files” form. (Process WFS option)

PAC_WF_DenStats Fileref Channel dBThreshold

Fileref Results of burst feature extraction post-wavelet-denoising performed on a particular channel in a ‘WFS’ file.


PAC_WF_WaveletStats Fileref Channel

WaveletRefNum Fileref Results of wavelet decomposition energy analysis performed on a particular channel in a ‘WFS’ file.


PAC_WF_WaveletEnergies Fileref Channel Dlevel

EnergyNum Fileref Energies for each wavelet detail level determined during wavelet decomposition energy analysis performed on a particular channel in a ‘WFS’ file.



AAPPPPEENNDDIIXX EE

PPUUMMPP DDEETTAAIILLSS

E.1 GROUP 1 PUMPS

Test Name Test2 Test4 Test5

Type Description End Suction End Suction End Suction

Pump Model No. 4x6-16N 1.5x3-13 1.5x3-11N

Standard class 3700 3196 3700

Curve No. 2750-3 4995-2 2854-2

Motor (kW) 15.0 37.0 15.0

BEP flow (m3/hr) 122 92 34

BEP head (m) 29 102 77

Rated Duty Flow (m3/hr) 68 60 29

Rated Duty NPSHR (m) 0.70 1.70 1.60

Test rpm 1202 2963 2950

Impeller diameter 400 302 248

VFD Frequency 40 50 NA

Impeller eye area (cm2) 120 31.6 60

Impeller eye radius (mm) 61.8 31.7 33.3

Suction Specific Speed (imperial units) 11,699 8,503 13,460

Suction Energy, x 106 69 63 137

HI SE Classification Low Low Low

Suction Recirculation Factor, x 106 11.6 15.7 20.3

E-1


E.2 GROUP 2 PUMPS

Test Name Test6 Test7 Test8

Type Description End Suction End Suction End Suction

Pump Model No. 3x6-13 1x3-13B 1x3-13B

Standard class 3700 3700 3700

Curve No. 3893-2 2727-2 2727-2

Motor (kW) 55 30 30

BEP flow (m3/hr) 115/150 16.2/17.8 18/21

BEP head (m) 66 82/124 67/116

Rated Duty Flow (m3/hr) 121 20 20

Rated Duty NPSHR (m) 1.70 2** 2**

Test rpm 2236/2980 2263/2673 2267/2427

Impeller diameter 316 316 316

VFD Frequency 40/50 38/45 38/45

Impeller eye area (cm2) 126.5 45.2 45.2

Impeller eye radius (mm) 63.5 37.9 37.9

Suction Specific Speed (imperial units) 9,700 10,859 10,859

Suction Energy, x 106 108/145 73/87 73/79

HI SE Classification Low Low Low

Suction Recirculation Factor, x 106 14.1/18.8 16.0/18.9 16.0/17.2

E-2


AAPPPPEENNDDIIXX FF

SSEEAALL MMOONNIITTOORRIINNGG RREESSUULLTTSS

It should be remembered when viewing these un-normalised results from different sensors that

signal magnitude is largely affected by sensor type and therefore the relative severity of the signals

can only be compared to results from that sensor.

FaceAE = sensor mounted on a waveguide through the gland plate to the back of the (fixed)

stationary face

GlandAE = sensor mounted on the gland plate

PumpAE = sensor mounted on the pump casing

F.1 SEAL FAILURE RESULTS

No catastrophic failures occurred during the testing period, although minor damage was incurred

on the carbon face of three separate seals. Extensive discussion with the manufacturer’s design

engineers in the UK finally concluded that due to the size and speed of the pump (and hence seal

face), it was not realistic to create conditions at the interface that would exceed the actual limits of

the face materials (rather than the conservative limits documented in the literature for which the

pump rig was designed).

F.2 SEAL CONDITION

Acoustic emissions generated at the seal face repeatedly matched local fluctuations in stuffing box

and seal face temperatures (see Figure F-1). These seem to be internally generated, because

increases in pump rig main line and jacket temperatures lagged behind corresponding rises at the

seal face/stuffing box. Furthermore, AE signals did not correlate with externally driven

temperature changes, such as fluctuations in circulating fluid temperature or from heat added by the

jacket circulation system (see light blue line in Figure F-1). This supports the conclusion that both

temperature changes and AE events are most likely caused by changing friction conditions

generated between the mating faces, presumably from changing seal face wear.

F-1


Seal face temperature fluctuations are most closely matched by FaceAE (pump rig seal only),

although very similar signals are also detected by GlandAE. Under most conditions, GlandAE

signals also bear some resemblance to the corresponding PumpAE signal, implying that acoustic

emissions being detected at the gland plate come from sources within the pump casing as well as

from the seal. This does not occur however, under low NPSH conditions caused by throttling the

suction valve (see Figure F-2), probably because the seal chamber pressure falls to below

atmospheric pressure and air is ingested into the seal, thereby eliminating the fluid coupling

between the seal and the pump. As air is well known to be a very strong attenuator of AE signals,

this suggests that the transmission of AE bursts from the pump chamber to the gland plate occurs

via the fluid, rather than through the pump casing or structure.

Figure F-2 also shows that seal AE signals are highly dependant on the operating history of the

seal. Early running at BEP (and also low flow conditions) did not cause significant distress to the

seal as indicated by relatively smooth FaceAE and GlandAE trends. However, after the pump had

been subjected to low NPSH conditions and dry running, BEP signals were substantially more

erratic. High flow (127% BEP) also caused some large peaks in all trends, but in-between the

signals were relatively stable.

23/04 12:00 24/04 12:00 25/04 12:00 26/04 12:00 27/04 12:0025

30

35

40

45

50

55

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

FaceTemp

degr

ees C

RMS

RMS

, Sea

l Fac

e

Main line temperature

Figure F-1: Seal AE RMS and temperature fluctuations

F-2


Only AE bursts accompanying extreme seal face temperature fluctuations are observed on the

PumpAE signal. On the other hand, this signal does seem to identify low NPSH conditions

(caused by valve throttling) very effectively; in Figure F-2 the PumpAE signal increases by almost

500% when the valve is throttled and remains at that level until BEP flow is resumed (see Figure

F-3).

Figure F-2: RMS values from FaceAE (blue), GlandAE (red) and PumpAE (green) for various flow conditions. Face temperature fluctuations are shown in grey. Flow conditions are indicated in the top graph.

F-3


Throttled suction valve

Dry running

Figure F-3: Zoomed view of previous test showing high PumpAE levels during Low NPSH conditions caused by valve throttling.

F-4


F-5

Documents

the application of acoustic emission monitoring - the UWA