Upload
others
View
5
Download
0
Embed Size (px)
Citation preview
An Adaptive Estimation Algorithmfor Aircraft Engine Performance Monitoring
Olivier Leonard, Sebastien Borguet and Pierre Dewallef
University of Liege - Turbomachinery Group
Chemin des chevreuils 1 - 4000 Liege - Belgium
Abstract
In the frame of turbine engine performance monitoring, system identification procedures areoften used to adapt a simulation model of the engine to some observed data through a set of so-called health parameters. Doing so, the values of these health parameters are intended to representa faithful image of the actual health condition of the engine. The Kalman filter has been widelyused to achieve the identification procedure in real time, on-board, applications. However, in orderto achieve a proper filtering of the measurement noise, the health parameters must be supposedto vary in time relatively slowly preventing any abrupt accidental events to be tracked effectively.This contribution presents a procedure called adaptive filtering and based on a covariance match-ing method, intended to automatically release the health parameters once an accidental event isdetected. This enables the Kalman filter to deal with both continuous and abrupt fault conditions.
INTRODUCTION
In the last years, improving the availability of aircraft turbine engines as well as mini-
mizing their maintenance costs turned out to be key issues for airline companies. In this
context, engine manufacturers are more and more involved in research programs intended
to dynamically adapt the maintenance planning based on the actual engine health condition
rather than using fixed schedule based on the number of running hours.
Basically, the monitoring of the engine health condition can be carried out by observing
the drifts from some reference values of a set of measurements performed on-board, while
the engine is in service. However, as a modification in the engine health condition generally
involves a drift on nearly all of the measurements at the same time, the identification of
the underlying component fault turns out to be uneasy. Moreover, the measurement drifts
are corrupted by random errors whose amplitude is comparable to the measurement drifts
induced by the component faults of interest. To improve the significance of the measurement
drifts, it is appropriate to translate them into a set of numerical parameters intended to
1
represent the health condition of the main engine components.
The link between the measurement drifts and the health parameter deviations is done by
an aero-thermodynamical simulation model. Such models are developed for design, perfor-
mance analysis or control purposes. They are based on the resolution of a set of equations
expressing mass, momentum and energy balances inside the engine. Besides their capability
to predict most of the measurements taken along the gas path of the engine in its whole
operating range, these models may also be parameterized. The nature of these parameters
may vary from one manufacturer to another but they usually consist in efficiencies and flow
capacities related to the main engine components. The simulation model is therefore able to
predict the value of the measurements corresponding to specific ambient conditions (e.g. the
ambient temperature and pressure, the humidity, . . . ), control conditions (e.g. the power
setting) and health conditions through a given set of health parameters. Doing so, the
performance monitoring task is achieved by the resolution of an inverse problem where the
health parameter deviations are determined based on the observation of the measurement
drifts.
Within this framework, system identification techniques bring an answer to the rather
poor physical meaning of the raw measurement drifts. Indeed, these techniques are intended
to adapt a simulation model of the engine to the observed data through the set of health
parameters. Accordingly, the set of adapted health parameters represents a faithful image of
the engine health condition. Moreover, system identification techniques provide the user with
a large set of methodologies intended to filter the random measurement noise. In the turbine
engine literature, the use of system identification techniques together with measurements
performed along the engine gas path is usually referred to as Gas Path Analysis (GPA). The
interested reader is merely referred to [1] for a simple but yet complete introduction of the
GPA approach to turbine engine health condition monitoring.
Among the large amount of available system identification methods, the Kalman filter is
considered herein as an appropriate approach. The Kalman filter provides an optimal update
of the health parameters (in the least-squares sense) each time new data are available, at
the cost of a very low computational load. However, the correct filtering of the random
measurement errors requires proper constraints to be set on the variation rate of the health
parameters. In the frame of engine health monitoring, the transition model of the health
parameters is based on the assumption of a relatively slow variation. Consequently, the
2
Kalman filter tracks engine wear with a good accuracy. Engine wear is the primary cause of
deterioration ; it affects all components at the same time and evolves smoothly with time.
On the other hand abrupt faults, due for instance to foreign object damage, may happen
too. These events generally impact one (at most two) component at a time. The Kalman
filter is less capable of following short-time-scale variations and consequently the estimated
fault is spread on several components. This behavior is referred to as a “smearing effect” in
the literature [3].
Different strategies have already been considered in the turbine engine literature to over-
come this issue. One of the most advanced consists in determining whether or not the mea-
surement drifts are related to an abrupt component fault before using them in the Kalman
filter. When an abrupt component fault is detected, a bank of Kalman filters is used, each
of which considering a specific accidental event. Provided that at least one of these specific
filters correctly fits the situation at hand, the abrupt fault can be isolated and accomodated
by the simulation model and the generic Kalman filter can be used again. Even if succesful
applications of this procedure are found in the literature (see for example [1]), the setup of
the fault partition and of the corresponding bank of specific Kalman filters, as well as the
selection of the most relevant one, turn out to require a significant experience from the user.
The present contribution investigates a different approach in which the constraints on
the health parameter variation rate are adapted on-line such that they best fit the actual
situation. Therefore, both abrupt and continuous faults can be processed by a single Kalman
filter.
THE SIMULATION MODEL
In the framework of diagnosis, the purpose of the engine simulation model is to reproduce
the value of some variables measured along the engine gas path. Among the measurable
quantities, some are independent variables, such as the uncontrollable ambient conditions
(temperature and pressure, humidity, . . . ) or the command variables (e.g. the power set-
ting). In the present document, no distinction is made between the command variables and
the uncontrollable inputs. All operating conditions are aggregated into a column vector
denoted uk where k is a discrete time index. The other measurable quantities are dependent
variables and are stored in a m-dimensional vector yk simply called the measurements.
3
Besides their dependency upon the operating conditions uk, the measurement yk are also
a function of the actual engine health condition1.
To take advantage of this dependency, the simulation model is parameterized by a set of
so-called health parameters denoted by a p-dimensional vector wk. Doing so, the simulation
model is a vector function noted G(uk,wk). In this contribution, the simulation model is
restricted to the steady-state prediction of gas path measurements (temperatures, pressures,
thrust, air mass flow rates, . . . ) and rotational spool speeds. Due to the nature of the model,
the health parameters are restricted to the efficiency, the flow capacity or the effective cross-
section of the main components of the engine.
As a deterministic model is a simplification of the real-world process, it is unlikely to
perfectly match the random errors corrupting the observed measurements. To establish the
equality, a random additive error term εk is defined according to:
yk = G(uk,wk) + εk (1)
As the value of εk cannot be known beforehand, only its probability to fall in certain
ranges can be described analytically. A common assumption consists in modeling the be-
havior of εk by a zero-mean (E(εk) = 0), white and Gaussian random variable with covariance
matrix defined by:
E(εkεTl ) =
Rr if k = l
0 if k 6= l(2)
With some abuse of notations, the operator E(·) represents the mathematical expectation
of a random variable2 and the superscript T is the vector-matrix transpose. The specification
of the equation (1) together with the covariance matrix Rr is called the statistical system
model because it combines the deterministic system model G(·) to the stochastic description
of the random measurement errors.
In the frame of turbine engine diagnosis the quantity of interest is the difference between
the actual engine health condition and a reference one. In this contribution, this reference
1 Strictly speaking, the measurement yk are also a function of the transient effects taking place inside theengine (i.e. mechanical inertia, heat transfers and gas dynamic) but, as only steady-state behaviors areconsidered herein, these effects are not taken into account in the present document.
2 See [4] for a more proper statement of the concept of mathematical expectation of a random variable.
4
value is represented by a so-called prior value which designates a value of the health pa-
rameters w−k which is available to the user prior to the knowledge of the measurements yk.
To take this consideration into account, a residual is defined as the difference between the
measurements yk and those predicted by the model using the prior value of the parameters
w−k :
rk , yk − G(uk, w−k ) (3)
In order to reduce the computational burden a first order linearization of the simulation
model around the prior health parameter values is used (see [7] for a discussion about the
use of linear models in turbine engine diagnosis):
Gk ,∂G(uk,wk)
∂wk
∣∣∣∣wk=bw−k (4)
The multiplying factor Gk is a m× p influence matrix which indicates how the measure-
ments are drifting from the assumed values when each health parameter moves away from
its prior value. According to (3) and (4) the model equation (1) becomes:
εk = rk − Gk(wk − w−k ) (5)
As the influence matrix Gk depends upon the operating conditions, its estimation requires
p additional runs of the simulation model at each time step if it is calculated by forward or
backward differences (2p for central differences). From a computational point of view, this
situation is not favorable since the processors currently available onboard of aircraft engines
are not yet able to achieve such tasks in real-time. However, for the application detailed
further in the document, the matrix Gk does not change significantly for the considered range
of operating conditions and the influence matrix is computed beforehand for nominal health
parameters and cruise flight conditions. Yet, when large ranges of operating conditions are
considered, it may be advantageous to exploit the nonlinearities of the simulation model
with respect to the operating conditions by processing important ranges of flight conditions,
as outlined in [8]. The influence matrix can be interpolated between a bank of matrices
computed beforehand for a set of control conditions (usually the engine pressure ratio or
the non-dimensional low pressure spool speed) and corrected for the actual inlet conditions
(typically the total pressure and temperature) which provides a high accuracy together with
5
a low computational load.
THE KALMAN FILTER
The most objective way to solve the estimation problem would consist in a maximum
likelihood approach in which the health parameters are considered as unknown quantities
that can take any possible value. Within the present Gaussian framework, this comes down
to determine the value of wk minimizing the inner-product εTk R−1
r εk. However, due to the
fact that the number of health parameters to be determined is higher than the number of
available measurements yk for each data sample, the maximum likelihood approach leads to
a non-unique solution.
To overcome this issue while keeping a sample by sample data processing, the health pa-
rameter deviations are considered, symmetrically to the random errors, as Gaussian random
variables with mean and covariance matrix respectively defined by w−k = E(wk|yk−1) and
P−w,k = E[(wk − w−k )(wk − w−k )T |yk−1
]. Again with some abuse of notations, the operator
E(·|yk−1) designates the conditional mathematical expectation of a random variable before
the data sample at time step k is used to update the health parameters [4].
Once the current data sample yk is observed, the application of the Bayes’ rule allows to
build the posterior probability density function describing the probability of occurrence of
the health parameters at time step k. The appealing property of using Gaussian random
variables together with the first order Taylor expansion of equation (4) is that the posterior
probability distribution is still Gaussian, which enables the estimation problem to be solved
analytically.
Mathematically, the estimation problem comes down to the determination of the posterior
mean and covariance matrix of the health parameters respectively defined by wk = E(wk|yk)
and Pw,k = E[(wk − wk)(wk − wk)
T |yk
]. The use of the posterior conditional mathematical
expectation E(·|yk) underlines the fact that the latter quantities are obtained once the current
measurement sample yk is used.
As for Gaussian random variables the mathematical expectation is equivalent to the
maximum of the posterior probability density function (MAP estimator), the posterior mean
6
wk is computed by the resolution of a maximization problem3. After some algebraic steps,
the maximization problem yields the following quadratic minimization problem (see [6] for
more details):
J (wk) = (wk − w−k )T (P−w,k)−1(wk − w−k ) + εT
k R−1r εk (6)
The first term on the right hand side in the previous expression simply expresses that
the values of wk must remain in a neighborhood of the prior values w−k (the shape of the
neighborhood being specified by the covariance matrix P−w,k).
The resolution of the quadratic minimization problem only involves basic linear algebra
operations. Indeed, it comes down to find the value of wk which cancels the derivatives of
the quadratic objective function with respect to wk which yields:
wk = w−k + Kkrk with Kk = P−w,kGTk
(GkP
−w,kG
Tk + Rr
)−1(7)
From the definition of the prior and posterior covariance matrices and exploiting the
previous update rule (7), it yields:
Pw,k = (I−KkGk)P−w,k−1 (8)
The use of the previous relations poses no serious problems provided that proper values for
w−k and P−w,k are available. A very interesting feature of the Kalman filter is that it is able to
exploit a so-called transition model for the health parameters. This enables the computation
of proper values for w−k and P−w,k from the results obtained at the previous time step and
enables the update rules (7) and (8) to be used recursively.
As very few information is usually known beforehand about the way the health parameters
will vary in time, the health parameter transition equation is often restricted to a quasi-
stationary process described by:
wk = wk−1 + ωk (9)
where ωk is called the process noise and is modeled by a zero-mean (E(ωk) = 0), white and
3 It has to be noted that the present approach is not the most generic methodology used to deduce theKalman filter. Yet, the interested reader is referred to [2,5] for a more generic approach.
7
FIG. 1: Performance monitoring tool based on an Kalman filter for the case of a turbofan engine
Gaussian random variable with covariance matrix defined by:
E(ωkωTl ) =
Rw,k if k = l
0 if k 6= l(10)
From the definition of w−k and ωk and noting that E(wk−1|yk−1) = wk−1, it is relatively
straightforward to deduce that w−k = wk−1 and that P−w,k = Pw,k−1 + Rw,k.
In order to provide the reader with an overall picture of the whole procedure, the update
mechanism of the health parameters is shortly depicted in FIG. 1 and more precisely in
algorithm 1 where it is referred to as the Kalman filter. At the current time step k, the
previously estimated health parameters are used by the simulation model to generate a
prediction of the measured values. The comparison between the model outputs and the
observed measurements gives a residual rk which is an image of the distance between the
health parameters and the actual engine health condition. The Kalman filter determines the
new estimate wk through relation (7) and updates the health parameter covariance matrix
(not represented in the figure).
8
Algorithm 1 The Kalman filter
Require: w0, Pw,0, Rr, Rw,k
1: for k > 0 do2: w−k = wk−1 and P−w,k = Pw,k−1 + Rw,k
3: acquire uk and yk
4: rk = yk − G(uk, wbl)− Gk(w−k − wbl)
5: compute the influence matrix through (4)
6: K = P−w,kGTk
(GkP
−w,kG
Tk + Rr
)−1
7: wk = w−k + Krk and Pw,k = (I−KGk)8: end for
PERFORMANCE MONITORING
In the present application it is assumed that the health parameters vary independently
such that the covariance matrix Rw,k is strictly diagonal. Therefore, the covariance matrix is
completely determined by the p diagonal terms called herein the variance of the process noise.
Even if the resulting transition model (9) appears very simple, the covariance matrix Rw,k
enables to control the stochastic character of the time series formed by the health parameters
wk: low variances of the process noise mean slow variations while high variances suppose
fast variations. In normal operation, the health parameters are varying in a relatively
predictable manner and the process noise variance can be set according to the maximum
expectable degradation rate of the health parameters. Of course, the more important the
data acquisition rate is, the lower the process noise variance can be which allows the random
measurement noise to be appropriately filtered without penalizing the faithfulness of the
performance tracking.
However, optimizing the performance tracking for gradual degradations also strongly
deteriorates its behavior when accidental events are encountered (e.g. a foreign object
ingestion). Indeed, the latter events involve large deviations on a small subset of the health
parameters in a very short period of time (typically less than a second). As the health
parameter variations necessary to adapt the simulation model to the actual engine health
condition exceed the range specified by the health parameter variance, the Kalman filter
spreads the fault on all of the health parameters. This behavior, providing the user with a
poor diagnosis, is sometimes referred to as a smearing effect in the turbine engine literature
[3]. Practically, a value for the health parameter variance giving a good balance between
fast tracking abilities and efficient random noise filtering is very difficult to find and the
9
resulting algorithm is more likely to behave poorly in both situations.
ADAPTIVE ESTIMATION
To improve the tracking abilities of the filter without sacrificing the reliability of the
estimation, the present contribution investigates a so-called adaptive estimation procedure.
By adaptive estimation, it is meant that the estimation algorithm is able to automatically
adapt the health parameter variance to both abrupt and continuous engine degradations.
The methodology used in this contribution is strongly inspired from the one developed by
Jazwinski in [9] and is intended to ensure consistency between the predicted residuals rk
and their statistics. The present methodology is developed for the specific kind of transition
model specified by relation (9). The interested reader will find the developments for a more
general family of transition models in [9].
To do so, the update of the health parameters is delayed of N time steps. Therefore,
a buffer of N + 1 residuals, denoted {rk−N , . . . , rk}, is formed by comparing the effective
measurements yk to those obtained through the simulation model when the prior value
w−k−N is used. This gives the following definition for the residuals:
rk−l , yk−l − G(uk, w−k−N) = Gk−l(wk−N−1 − wk−N−1) + Gk−l
N∑i=l
ωk−i + εk−l (11)
Since, by definition, E(wk−N−1|yk−N−1) = wk−N−1, E(εk) = 0 ∀ k and E(ωk) = 0 ∀ k,
it yields that E(rk−l|yk−N−1) = 0. This conditional mathematical expectation indicates
that the residuals are computed using the most recent health parameter estimate (wk−N−1).
Similarly, it can be proved, after some algebraic steps, that:
E(rk−lrTk−m|yk−N−1) = Gk−lPw,k−N−1G
Tk−m+Rrδm,l+(N+1−max (m, l))Gk−lRw,kG
Tk−m (12)
where max (m, l) is equal to m if m ≥ l and l otherwise and δm,l is equal to 1 if m = l and 0
otherwise. Practically, the off-diagonal terms of the matrix rk−lrTk−m are extremely sensitive
to the measurement noise. To make sure that the data are statistically representative,
it is chosen to restrict the present method to the diagonal terms of the previous matrix
equality. Therefore, ensuring the consistency of the residuals with their statistics is done by
10
determining Rw,k such that:
diag(rk−lr
Tk−m
)= diag
(E(rk+lr
Tk+m|yk−N−1)
)(13)
where the operator diag (·) designates the vector made of the diagonal values of a square
matrix. In a previous publication by the authors [6], the delay N is set to zero in order to
avoid a delayed health parameter estimation. In this way, the matrix Rw,k is determined
such as the squared residuals r2k match the diagonal terms of the right hand side of relation
(13) when N = m = n = 0. However, such an approach neglects the fact that, in tur-
bine engine performance monitoring, the level of measurement noise is relatively high and
Rr & GkRw,kGTk . As a consequence, the estimation of Rw,k is mainly driven by the random
measurement noise and the resulting adaptive procedure turns out to be too sensitive to
the random errors. In [6], this effect is overcome with the help of a second Kalman filter
estimating the diagonal terms of Rw,k. Even though it effectively stabilizes the estimation
procedure, the setup of the complete algorithm involves many parameters to be tuned which,
in the end, turns out to be uneasy and does not fulfill the simplicity requirement stated in
the introduction section.
In our search to simplify the complete estimation procedure, and according to the ap-
proach devised in [9], it appeared that the best approach consists in averaging the residuals
rk−l over the N + 1 samples of the buffer according to:
rk =1
N + 1
N∑l=0
rk−l (14)
Doing so, the health parameter estimation is delayed of N time steps but the mean rk is
also expected to be less sensitive to the measurement noise. Indeed, it can be shown that
the mean rk is a white and Gaussian random variable with zero-mean (E(rk|yk−N−1) = 0)
and covariance matrix given by (after some algebraic steps using equations (12) and (14)):
E(rkrTk |yk−N−1) = Gk,NPw,k−NG
T
k,N + Rr +N∑
l=0
Gk,lRw,kGT
k,l (15)
11
where the m× p matrix Gk,l and the m×m matrix Rr are defined as:
Gk,l =1
N + 1
l∑i=0
Gk−i and Rr =1
N + 1Rr (16)
Applying equation (15) to the covariance matching (13) leads to verify the following
equality:
dk , r2k − diag
(Gk,NPw,k−NG
T
k,N + Rr
)= Bkfk (17)
where, for sake of simplicity, the matrix Rw,k is reduced to a vector containing the diagonal
values according to:
fk = diag (Rw,k) and Bk =N∑
l=0
G2
k,l (18)
Since in turbine engine diagnosis there are often more health parameters than available
measurements, the matrix Bk is rectangular (m < p) and the previous equation has no
unique solution. Consequently, the resolution of the previous equation is best done by a
maximum a posteriori approach where the prior value of fk is simply set to fmin = diag (Rminw )
with a covariance matrix Pf . Therefore, an adaptive estimation procedure is obtained by
setting the diagonal terms of the covariance matrix Rw,k to the vector fk given by:
fk = fmin + max[0 ,
(P−1
f + BTk Bk
)−1BT
k dk
](19)
The covariance matrix Pf may be easily built by setting its diagonal terms as fmax/3
where fmax are the maximum allowed values for fk. Loosely speaking, fmax reflects the
squared maximum amplitude of the expected abrupt events. Additionally to the diagonal
terms, it is often found very helpful to set some correlations into the covariance matrix Pf
in order to decrease the sensitivity of the estimation to the measurement noise. Indeed, it
may happen that the health parameter variations generated by abrupt component faults are
not totally independent. For example, the flow capacity and the efficiency of the fan usually
vary together when a fault occurs. Such correlations are set into the covariance matrix by
following the method detailed in [1]. It consists in setting the off-diagonal terms of the
matrix Pf according to:
Pf (i, j) = Pf (j, i) = ρi↔j
√Pf (i, i) · Pf (j, j) (20)
12
FIG. 2: Structure of the adaptive estimation algorithm
where 0 ≤ ρi↔j ≤ 1 is the correlation factor. Theoretically, nothing prevents from specifying
negative values for the correlation factor, but this makes no sense in this application since the
values of fk always rise together whatever the direction of the health parameter deviations.
There is no rule of thumb for setting the correlation factors and they strongly depend upon
the application at hand. However, a value of 0.5 is often found to be a good compromise if
the only available information is that some health parameters “may usually vary together”.
The estimation algorithm described in FIG. 1 with the Kalman filter of algorithm 1 can
be now updated to achieve adaptive estimation. This newly developed procedure is sketched
in FIG. 2. Basically, the residuals are stored in a buffer of size N + 1 whose purpose is to
provide the residual rk−N and to compute the mean residual rk. The latter is used in turn by
the adaptive estimation, namely the relation (19), to determine a proper covariance matrix
Rw,k. Finally, the updated covariance matrix Rw,k and the residual rk−N are used by the
Kalman filter of algorithm 1 to update the estimated health parameters.
Even if the procedure is relatively straightforward, the data buffering, represented by
a simple box in FIG. 2, needs to be commented. Indeed, as the estimation is delayed
of N samples, the residuals rk−l given by the buffer would be related to the prior health
parameters w−k−l−N for l = 0, . . . , N , and not to the most recent prior estimate (namely
w−k−N) as required by both the adaptive estimation and the Kalman filter. To correct this
13
effect, without re-estimating the whole buffer at each time step, the residuals are updated
by the following rule:
rk−l = rk−l − Gk−l(wk−N − w−k−N) ∀ l = 0, ..., N (21)
STABILIZATION OF THE ALGORITHM
Since the square of a Gaussian random variable is no longer Gaussian, the residuals dk
defined in (17) are not Gaussian. As a consequence, the use of the estimation formula (19)
may result in very noisy values for fk with small values of N (i.e. N ' 10). This generates
spurious covariance rise which decreases the stability of the estimation with respect to the
measurement noise. Of course, the estimated values fk can be filtered on L time steps, as
advised in [9], but the corresponding average does not completely correct the problem.
A better approach consists in testing whether or not it is purposeful to apply the value
of fk. If not, a covariance matrix is used which corresponds to normal degradations. This
matrix is denoted Rminw . The test used herein is the well-known χ2 (read chi-squared)
test (see [4]). It is based on the assumption according which, if no abrupt fault occurs,
rk must be an m-dimensional Gaussian random variable with zero mean and covariance
matrix E(rkrTk |yk−N−1) given by relation (13) where Rw,k = Rmin
w . If the latter covariance
matrix is denoted Pr,k, the Mahalanobis distance defined by qk = rTk (Pr,k)
−1 rk follows a
χ2m probability density function with m degrees of freedom. The test compares the scalar
qk with the integral Xα of the χ2 distribution; if qk is equal to the quantile Xα of the χ2m
distribution defined by α =∫ Xα
0χ2
m (x)dx then the probability of obtaining qk or a larger
value in the “null hypothesis” (i.e. no abrupt component fault is present) is given by 1−α.
In the present application, Xα is computed beforehand by setting 1− α to a low value (e.g.
10−6). The use of the statistical test comes down to set the diagonal terms of the covariance
matrix Rw,k according to relation (19) if qk ≥ Xα and to Rminw otherwise.
APPLICATION ON A TURBOFAN ENGINE
The aforementioned methodology is applied to a turbofan layout representative of cur-
rently available commercial aircraft propulsors. It is a two-spool, mixed flow turbofan whose
14
fan
lpc hpc combustorhpt lpt
inlet
SW12RSE12
SW2RSE2
SW26RSE26
SW41RSE41
SW49RSE49 A8IMP
nozzle
2 26 3
13
41 49 6 8
12
1
Wfuel
FIG. 3: The turbofan layout used as an application test case. The symbols hpc and lpc referrespectively to the high and low pressure compressors while hpt and lpt refer to the high and lowpressure turbines.
general structure is sketched in FIG. 3. This turbine engine is a virtual, but still realistic,
engine used in the frame of the OBIDICOTE project4. The simulation model used further
in the application is basically an aero-thermodynamical model parameterized by a set of 11
health parameters represented in the boxes in FIG. 3. These health parameters are the flow
coefficients (SWiR) and the efficiency drops (SEi) of the main components of the engine,
namely the fan, the high and low pressure compressors (hpc and lpc) and the high and low
pressure turbines (hpt and lpt), together with the nozzle area (A8IMP). The simulation
model is described in [10] to which the interested reader is referred for more details.
Additionally to the simulation model, a set of 14 typical accidental component faults is
also supplied. The fault cases, extracted from [11] and detailed in TABLE I, are intended
to be representative of some of the most likely component faults that can be expected on a
current turbofan engine. Since they have been used by many authors in the turbine engine
literature, these test cases can be considered as benchmarks.
To identify the component faults, a set of 7 measurements is used. The positions and
the accuracies of these measurements are also extracted from [11] and are listed in TABLE
II. These measurements are intended to represent those available on-board of a commercial
4 A Brite/Euram project concerned with on-board identification and control of turbofan engines.
15
a -0.7% on SW2R -0.4% on SE2 fan, lpc-1% on SW12R -0.5% on SE12
b -1% on SE12c -1% on SW26R -0.7% on SE26 hpcd -1% on SE26e -1% on SW26Rf +1% on SW41R hptg -1% on SW41R -1% on SE41h -1% on SE41i -1% on SE49 lptj -1% on SW49R -0.4% on SE49k -1% on SW49Rl +1% on SW49R -0.6% on SE49m +1% on A8IMP Nozzlen -1% on A8IMP
TABLE I: Accidental fault cases of a turbofan engine.
index name Description Accuracyu(1) T1 Inlet total temperature 2Ku(2) P1 Inlet total pressure 100Pau(3) Pamb ambient pressure 100Pau(4) WFE Fuel mass flow rate 2g/sy(1) T13 Afterfan total temperature 2Ky(2) P13 Afterfan total pressure 100Pay(3) T3 HPC outlet total temperature 2Ky(4) P3 HPC outlet total pressure 5000Pay(5) Nlp Low pressure spool speed 4rpmy(6) Nhp High pressure spool speed 12rpmy(7) T6 LPT outlet total pressure 2K
TABLE II: Operating conditions uk and measurements yk available on-board of the turbofan layoutconsidered in the present application. The accuracies are 3 times the standard deviation.
turbofan.
The measurements used further in the applications are fictive measurements generated
by the simulation model for constant cruise flight conditions (altitude=10800m, Mach=0.8,
∆Tiso=0K). Random measurement errors are added to all of the simulated measurements
and to all of the operating conditions by using a zero-mean Gaussian random variable with
standard deviations set to the third of the accuracies mentioned in TABLE II. The sequence
16
is 5000 seconds long with a data acquisition rate of 2 samples per second.
In order to represent a slow degradation due to normal operation, a reference fault case is
considered where nearly all of the health parameters are varying simultaneously. This fault
case, referred to as the 9p case further in the document, has already been investigated in [12]
and consists of the following degradations occurring progressively in 5000 seconds: -1.5%
on SW12R, -1.2% on SE12, -1% on SW2R, -1% on SE2, -2.3% on SW26R, -1.4% on SE26,
+0.88% on SW41R, -1.6% on SE41, -1.3% on SE49. To underline the benefits of the newly
developed method, the accidental fault cases listed in TABLE I are successively added, one
at a time, to the reference slow degradation mentioned above. Each abrupt component fault
is added to the slowly drifting fault at time t=2500 s. Doing so, the behavior of the diagnosis
tool is illustrated for both slow and fast drifting faults.
APPLICATION RESULTS
A good indicator of the quality of the performance tracking is the maximum RMS (root
mean square) estimation error (the average is done on the whole flight sequence) expressed in
percent of deviation from the nominal values. If only slowly drifting faults have to be tracked,
the algorithm 1 can be used effectively with a constant covariance matrix Rw,k = Rminw . Such
a situation is represented in FIG. 4. The actual values of the health parameters are drawn in
plain lines with the corresponding health parameter name on the right axes. FIG. 4 shows
clearly that identified values, plotted with the symbols, agree with their actual values. The
maximum RMS error erms remains below 0,15% which hints at the good tracking ability of
the Kalman filter for normal degradations.
However, if an abrupt component fault occurs, such a procedure fails to track and isolate
the fault. This is well illustrated by adding the fault case ‘a’ from TABLE I to the previous
slowly drifting faults. The results given by the previous diagnosis tool are summarized in
FIG. 5. The plain lines representing the actual values of the health parameters exhibit a
brutal drift of 4 health parameters, namely the efficiency and the flow capacity of both
the fan and the low pressure compressor at t=2500 seconds (i.e. SW12R, SE12, SW2R,
SE2). In this case, the Kalman filter is unable to follow the abrupt deviation of the 4 health
parameters. As a result, the maximum RMS estimation error rises above 0.4% which is
visualized in FIG. 5 by the spreading of the fault on several other components such as the
17
FIG. 4: Identification results obtained on a slowly drifting fault on nearly all of the health param-eters and without adaptive filtering
high pressure compressor (SW26R) and the nozzle (A8IMP). The present test case is a good
illustration of the Kalman filter divergence in the presence of a rapid measurement drift.
Even at the end of the sequence, the fault is still not isolated which provides the user with
a wrong diagnosis.
To illustrate the advantage of using an adaptive estimation algorithm, the same fault case
is used but now enabling the update of the process noise covariance matrix Rw,k depicted in
FIG. 2 with a delay of 50 samples (i.e. N = 50). To improve the results, some correlation
factors are set in the covariance matrix Pf according to (20). A correlation of 0.7 is set
between SW12R, SE12, SW2R and SE2, an other of 0.7 between SW26R and SE26, one
of 0.5 between SW41R and SE41 and one of 0.1 between SW49R and SE49. The results
are presented in FIG. 6. The first thing to notice is that the use of an adaptive estimation
algorithm does not deteriorates the accuracy when only slowly drifting faults are present.
18
FIG. 5: Identification results given without adaptive filtering when both slowly drifting faults andan abrupt component fault (fault case ’a’) are present
However, when an abrupt fault occurs, the diagonal terms of the covariance matrix Rw,k
related to the fan and the low pressure compressor are increased. Consequently, the Kalman
filter puts more emphasis on more recent data and the health parameters can be quickly
adapted. As a result, the identified values remain close to their actual ones and no smearing
effect is observed. Accordingly, the maximum RMS error computed for the present test case
remains below 0.15% for the whole sequence.
To have a more general idea of the efficiency of the newly developed diagnosis tool, the
set of 15 test cases is run with and without adaptive filtering. These results are summarized
in TABLE III in terms of maximum RMS error. The values mentioned in the table are
averaged over 10 independent runs in order to ensure their reliability. Since the standard
deviations of the estimated health parameters (i.e. the square root of the diagonal terms
of the covariance matrix Pw,k) are approximately 0.1%, a test case whose maximum RMS
19
FIG. 6: Identification results given when both slowly drifting faults and an abrupt component fault(fault case ’a’) are present and using adaptive filtering
error remains below 0.25% is considered as successful. For sake of simplicity, this is denoted
by a checkmark in TABLE III. The first line in TABLE III refers to the situation where
only the continuous fault is present. As it can be expected, both algorithms give the same
results. However, when an abrupt component fault is introduced at t=2500 seconds, the
results given by the adaptive filtering are far better than those obtained without adaptive
filtering. Indeed, without adaptive filtering only 4 test cases (namely cases ’d’, ’f’, ’i’ and
’m’) are solved satisfactorily while with adaptive estimation, all of the test cases (but one)
are successful.
CONCLUSIONS
In the present contribution, an adaptive estimation algorithm has been developed to
achieve the performance monitoring of aircraft turbine engines for both normal degradations
20
Withoutadaptivefiltering
With adaptivefiltering
9p 0.17% X 0,18% X9p + a 0.41% - 0.18% X9p + b 0.27% - 0,17% X9p + c 0.39% - 0,24% X9p + d 0,16% X 0.17% X9p + e 0,35% - 0.18% X9p + f 0,21% X 0.16% X9p + g 0,31% - 0.17% X9p + h 0,25% X 0.16% X9p + i 0,23% X 0.17% X9p + j 0,52% - 0.57% -9p + k 0,46% - 0.23% X9p + l 0,44% - 0.23% X9p + m 0,16% X 0.18% X9p + n 0.34% - 0.18% X
TABLE III: Comparison of the results obtained with and without adaptive filtering. The resultsare the maximum RMS errors averaged on 10 successive runs.
and abrupt accidental events. The improvements of the new methodology with respect to
a more generic tool, based on a Kalman filter, have been illustrated on a typical turbofan
application. The accurate performance estimation achieved by the adaptive estimation al-
gorithm allows an efficient performance monitoring and a better component fault isolation.
Moreover, the implementation of the estimation algorithm is relatively straightforward and
requires only basic matrix operations. The computational burden is also very limited which
is compatible with the requirements of an on-board application.
The only drawback of the method is that the estimation is delayed of N time steps. High
values for N gives a more important delay but a more accurate abrupt fault detection while
lower delays give faster but less accurate results. In practice, values of N from 10 to 100
are usually sufficient to achieve a very accurate performance tracking. As a consequence,
the improvement in terms of convergence speed of the performance tracking is usually more
important than the time delay induced by the data buffering.
21
ACKNOWLEDGEMENTS
This work is funded by the Walloon Region of Belgium in the frame of its Program FIRST
and is supported by Techspace Aero, Safran Group.
BIBLIOGRAPHY
[1] A. Volponi. Foundation of gas path analysis (part i and ii). In Von Karman Institute LectureSeries, number 01 in Gas Turbine Condition Monitoring and Fault Diagnosis, 2003.
[2] R.E. Kalman. A new approach to linear filtering and prediction problems. In Transactions ofthe ASME, series D, Journal of Basic Engineering, 82, 1960.
[3] M.J. Provost. Kalman filtering applied to gas turbine analysis. In Von Karman InstituteLecture Series, number 01 in Gas Turbine Condition Monitoring and Fault Diagnosis, 2003.
[4] A. Papoulis. Probability, Random Variables, and Stochastic Processes. Electrical & ElectronicEngineering Series. McGraw-Hill International editions, third edition, 1998.
[5] S. Haykin. Kalman filters. In Kalman filtering and neural networks. Wiley series on adaptiveand learning systems for signal processing, communications and control, 2001.
[6] P. Dewallef. Application of the Kalman Filter to Health Monitoring of Gas Turbine Engines:A Sequential Approach to Robust Diagnosis. PhD thesis, University of Liege, 2005.
[7] Ph. Kamboukos and K. Mathioudakis. Comparison of linear and non-linear gas turbine per-formance diagnostics. ASME paper number GT2003-38518, 2003.
[8] Ph. Kamboukos and K. Mathioudakis. Assessment of the effectiveness of gas path diagnosticschemes. ASME Journal of Engineering for Gas Turbines and Power, volume 128, January 2006.
[9] A.H. Jazwinski. Adaptive filtering. Automatica, volume 5:475-485, 1969.
[10] A. Stamatis, K. Mathioudakis, J. Ruiz, and B. Curnock. Real-time engine model implemen-tation for adaptive control and performance monitoring of large civil turbofans. ASME papernumber GT2001-362, 2001.
[11] B. Curnock. Obidicote project - word package 4: Steady state test cases. Technical ReportDNS62433, Rolls-Royce, 2000.
[12] Ph. Kamboukos, K. Mathioudakis, and A. Stamatis. Turbofan performance deteriorationtracking using nonlinear models and optimization techniques. In Transactions of the ASME, volume124 of Journal of Turbomachinery, pages 580-587. October 2002.
22