AAArrrtttiii fffiiiccciiiaaalll NNNooossseee TTTeeeccchhhnnnooolllooogggyyy::: TTThhheee WWWiii –––NNNooossseee AAA PPPrrrooofff iii tttaaabbbiii lll iii tttyyy aaannnddd MMMaaarrrkkkeeettt AAAnnnaaalllyyysssiiisss fffooorrr ttthhheee DDDeeevvveeelllooopppmmmeeennnttt ooofff AAArrrtttiii fff iiiccciiiaaalll NNNooossseee TTTeeeccchhhnnnooolllooogggyyy tttooo MMMooonnniii tttooorrr ttthhheee FFFeeerrrmmmeeennntttaaatttiiiooonnn PPPrrroooccceeessssss iiinnn WWWiiinnneee SSShhhaaawwwnnnaaa MMM... LLL iiinnneeehhhaaannn SSSaaarrrooossshhh NNN... NNNiii zzzaaammmiii

TABLE OF CONTENTS

ABSTRACT…………………………………………………………………………………………1

BACKGROUND

Introduction………………………………………………………………………….……2

What is an E-Nose and How does it work? ……………………...………………………2

CURRENT ARTIFICIAL NOSE APPLICATIONS …………………………………………..…………5

FERMENTATION ……………………………………………………………………………..…….7

IMPORTANCE OF MONITORING FERMENTATION

Malolactic Fermentation Control………………………………………………………..8

Champagne and Sparkling Wine Production……………………………………………9

Champagne and Sparkling Wine Bottling………………………………………….……9

Additives to Effect Wine Characteristics……………………………………………..…10

THE WI – NOSE DESIGN

General Design………………………………………………………………….………10

Sensor Choice……………………………………………………………………..…….13

SENSOR DATA…………………………………………………………………………………….16

NEURAL NETWORK MODELING ……………………………………………………………….….17

MARKET ANALYSIS ………………………………………………………………………………29

CUSTOMER SATISFACTION MODEL

Model Development……………………………………………………………………..29

Accuracy…………………………………………………………………………………31

Device Weight and Size…………………………………………………………………34

Superiority Function Generation………………………………………………………34

INTEGRATION OF CUSTOMER SATISFACTION MODEL…………………………….……………35

RISK ANALYSIS…………………………………………………………………..………………38

FUTURE CONSIDERATIONS .....................................................................................................…...39

CONCLUSIONS…………………………………………………………………………………....40

REFERENCES…………………………………………………………………………………..…42

APPENDIX……………………………………………………………………………………..….44

ABSTRACT

This paper is concerned with the development of an artificial nose to monitor the fermentation process in wine. The design objective of the Wi – Nose was to provide the wine industry with a light weight, compact, and accurate sensing device. Other design factors that were considered included easy of installation, level of maintenance, and lifetime of the device. Data was generated from sensor output versus concentration plots and this data was then classified with the use of a neural network model, specifically NeuroSolutions 5, the most powerful and easy to use neural network simulation environment on the market today. Data was classified into three outputs, stage 1, stage 2, and stage 3 of fermentation. The data was trained, cross validated, and ultimately tested. The optimum percentage for these parameters were determined to be 80% training, of which 10% was cross validation, and 20 % testing. This model was used classify the data, giving accuracy results of 100% for all three fermentation stages. A customer satisfaction model was developed by varying design characteristics. This model ultimately resulted in superiority functions that were used to calculate product demands for varying product prices. These demands were then used to develop plots of net present worth’s as functions of product price to determine the optimal design in terms of consumer satisfaction and profitability. The optimal design was determined to be a device with a 100% correct classification rate, with optimum dimensions of 36 cc and a weight of 1 lb. This design resulted in a TCI of approximately $6.5 million, a ROI of approximately 49%, and a NPW or $11.2 million. A risk analysis was performed varying the cost of raw materials, and it was determined that there is a 90% probability that the Wi – Nose will have an ROI between 40.6% and 57.9%. Thus, production and marketing of the Wi – Nose will be a profitable venture.

BACKGROUND

Introduction

Researchers and manufacturers alike have long envisioned creating devices that can “smell”

odors in many different applications. Thanks to recent advances in organic chemistry, sensor

technology, electronics, and artificial intelligence, the measurement and characterization of

aromas by electrical noses (or e-noses) has become a commercial reality. “While electronic noses

were initially developed as laboratory instruments, science is now moving the technology out of

the laboratory and into the workplace, enabling measurements based on smell to be made at the

source,” says Steven Sunshine, president and CEO of CyranoSciences (Pasadena, CA), one of

several companies that have commercialized the technology.

What Is An E-Nose and How Does it Work?

Electronic noses are artificial smelling devices that identify the specific components of an odor

and analyze its chemical makeup to identify it. They can learn to recognize almost any

compound or combination of compounds. They can even be trained to distinguish between Pepsi

and Coke. Like a human nose, the electronic nose is amazingly versatile, yet it's much more

sensitive. "E-Nose can detect an electronic change of 1 part per million," says Dr. Amy Ryan

who heads the E-Nose project at JPL (Jet Propulsion Laboratory).

The e-nose looks nothing like a human olfactory systems but works quite similar to one. The two

main components of an electronic nose are the sensing system, similar to receptors in our nasal

passages, and the automated pattern recognition system, mimicking neurons and our brain. An

odor is composed of molecules, each of which has a specific size and shape. Each of these

molecules has a correspondingly sized and shaped receptor in the human nose. When a specific

receptor receives a molecule, it sends a signal to the brain and the brain identifies the smell

associated with that particular molecule. Electronic noses based on the biological model work in

a similar manner, albeit substituting sensors for the receptors, and transmitting the signal to a

program for processing, rather than to the brain. Electronic noses are one example of a growing

research area called biometrics, or biomimicry, which involves human-made applications

patterned on natural phenomena.

The sensing system can be an array of several different sensing elements (e.g., chemical sensors),

where each element measures a different property of the sensed chemical, or it can be a single

sensing device (e.g., spectrometer) that produces an array of measurements for each chemical, or

it can be a combination. Generally, electrical noses use a collection of different sensing elements.

These sensors are specially designed to conduct electricity. When a substance -- such as the stray

molecules from a glass of soda -- is absorbed into these sensors, the sensors expand slightly, and

that changes how much electricity they conduct. Because each sensor is made of a different

substance, each one reacts to each substance, or analyte, in a slightly different way. And, while

the changes in conductivity in a single sensor wouldn't be enough to identify an analyte, the

varied changes in various sensors produce a distinctive, identifiable pattern.

Each chemical vapor presented to the sensor array produces a signature or pattern characteristic

of the vapor, a digital “fingerprint” of the specific odor. By presenting many different chemicals

to the sensor array, a database of “fingerprints”, or more appropriately “smell-prints”, is built up.

This database of labeled signatures is used to train the pattern recognition system. The goal of

this training process is to configure the recognition system to produce unique classifications of

each chemical so that an automated identification can be implemented. Artificial neural networks

(ANNs), which have been used to analyze complex data and to recognize patterns, are showing

promising results in chemical vapor recognition.

An ANN is an information processing system inspired by the way the mammalian brain

processes information. ANNs are collections of mathematical models that attempt to emulate

some of the observed properties of biological nervous systems. An ANN consists of a large

number of highly interconnected processing elements – essentially equations known as “transfer

functions” – that are analogous to neurons and tie together with weighted connections that are

analogous to synapses. A processing unit takes weighted signals from other units, possibly

combines them, and gives a numeric result. The behavior of neural networks – how they map

input data to output data – is influenced primarily by the transfer functions of the processing

elements, how the transfer functions are interconnected, and the weights of those

interconnections. Learning typically occurs by example – through exposure to a set of input-

output data, where the training algorithm adjusts the connection weights (synapses). These

connection weights store the knowledge necessary to solve specific problems.

In general, ANNs are well suited to problems that people are good at solving but computers are

not, including pattern recognition and forecasting. ANNs, like people, learn by example.

However, unlike the human capability in pattern recognition, the ANN’s capability is not

affected by subjective factors such as working conditions and emotional state. Most commercial

electronic noses entering the market today employ some sort of ANN for pattern recognition.

This is because ANNs employ a large number of interconnected processing elements working in

unison to solve specific problems, much like biological nervous systems. ANNs are very general

pattern-recognition systems that one can configure, through a learning process, for specific

applications, such as identifying a chemical vapor. When an ANN is combined with a sensor

array, the number of detectable chemicals is generally greater than the number of sensors. Also,

less selective sensors which are generally less expensive can be used with this approach. Once

the ANN is trained for chemical vapor recognition, operation consists of propagating the sensor

data through the network.

This combination of a sensing system with a pattern-recognition system enables e-noses to

process new odors based on patterns of aromas created by earlier experiences, which is much the

same way the human olfactory system works. According to Sunshine, the human nose uses a

complex system of interconnected receptors and neurons, which conduct signals directly to the

brain’s limbic system. “When an aroma is sensed, the molecules from the vapor interact with

numerous receptors, causing them to send a different signal to the brain. The pattern of signals is

then recognized and interpreted by the brain based on prior training.” The human brain is an

incredibly impressive information processor, even though it "works" quite a bit slower than an

ordinary computer. Think of a sort of "analogy" between the complex webs of interconnected

neurons in a brain and the densely interconnected units making up an artificial neural network

(ANN), where each unit--just like a biological neuron--is capable of taking in a number of inputs

and producing an output.

Figure 1 - Neural Network Model

CURRENT ARTIFICIAL NOSE APPLICATIONS

Research in artificial nose technology began as a NASA projected aimed at detecting ammonia

leaks onboard the space station. Because the human olfactory system is not capable of detecting

numerous toxic chemicals until their concentrations are dangerously high, an alternative

detection method was needed. Although initially began as a detection system for toxic chemical

leaks, artificial nose technology has spread to many other industries. Current artificial nose

applications include those found in environmental monitoring, explosives detection, medical

diagnostics, and the food industry.

Environmental monitoring has exploded to the forefront of political, social, and economic

concerns. Artificial nose technology has been used in various environmental monitoring

applications such as air quality monitoring, identification of oil leaks, analysis of fuel mixtures,

identification of household odors, ground water analysis, as well as monitoring of factory

emissions. The electronic nose, or E-nose, provides a safe, economical way to monitor

environmental changes and prevent air, water, and soil contamination, as well as identify

hazardous situations before they become harmful. E-noses have even been implemented in the

oil and gas industry for the detection of harmful gas buildups in onshore and offshore rigs.

Artificial nose technology has proved exceedingly useful in the detection of explosive materials.

E-noses have been able to correctly identify bombs, landmines, TNT, and other explosive

devices, and when used in combination with bomb sniffing dogs, the E-nose becomes extremely

effective. Many governmental divisions, such as Homeland Security, as well as the military, are

interested in artificial nose technology. It is predicted that E-noses will become a staple in

airport security around the world in the near future.

Artificial nose technology is also making huge strides in the medical diagnostics industry.

Although it is not yet an FDA approved method of disease diagnosis, researches have proven its

effectiveness in various applications such as the detection of bacterial infections as well as the

diagnosis of gastrointestinal disorders, diabetes, liver problems, as well as tuberculosis.

Recently, a group of researchers from the Cleveland Clinic Lerner Research Institute have

demonstrated its capability in detecting lung cancer. Researchers used an E-nose to screen

cancer patients as well as monitor the effectiveness of their cancer treatments. The E-nose

provides physicians and clinics with an easy, economical, non-invasive medical diagnostic tool

to screen incoming patients by smell. It has been predicted that E-noses will play an even great

role in disease diagnostics in the foreseeable future as more and more relevant applications are

discovered.

Perhaps the largest and most diverse business sector to be touched by artificial nose technology

is that of the food industry. E-noses have been implemented in various food processing

applications, including assessment of food production, quality control, as well as control of

cooking processes. Specifically, E-nose technology can be found in the inspection of seafood

products, the grading of whiskey, inspection of cheese composition and flavor, and the

monitoring of such processes as fermentation and distillation. E-noses provide the unique

capability of being able to test a product before distribution, an important characteristic in an

industry notorious for status crippling product recalls resulting in huge capital losses as a result

of a few poorly processed or mishandled batches. Although E-nose technology is a relatively

new field, it has already made huge impact on many products consumers use every day. From

detecting hazardous gas buildups on off shore rigs to monitoring cheese composition, E-noses

technology is becoming more and more integrated into industries whose products shape our

everyday lives.

FERMENTATION

Fermentation in wine is the process whereby yeast converts sugar into Carbon Dioxide and Ethyl

Alcohol (Ethanol).

C6H12O6--->2CO2 + 2C2H5OH

An interesting fact is that the atomic weights of the two products are almost the same, so as you

see the carbon dioxide bubbling off you can have the satisfaction of knowing that the same mass

of alcohol is being produced in the wine. This also gives a chemical means of calculating the

alcohol content of the wine.

There are 3 Stages of fermentation:

• Primary or Aerobic (with air) Fermentation

• Secondary or Anaerobic (without air) Fermentation

• Malolactic Fermentation (possible third stage)

During the primary fermentation of wine, the two grape sugars, glucose and fructose are

converted to alcohol (ethanol) by the action yeast. Carbon dioxide is also produced, and leaves

the solution in the gaseous form, while the alcohol is retained in mix. The by-products of primary

fermentation are aromas, flavors, and heat. This stage generally lasts for about a week and is a

critical stage for yeast reproduction. On average, 70% of fermentation activity will occur during

these first few days. There is vigorous action as a result of the rapid fermentation that often

results in considerable foaming on the surface. Primary fermentation is usually conducted in an

open container, hence the aerobic term, covered with a clean tea towel known as the 'primary

fermentation vessel'.

Secondary fermentation is where the remaining 30% of fermentation activity will occur. It is a

much more gentle process that usually lasts anywhere from 2-3 weeks to a number of months,

depending on the amount of nutrients and sugars still available. Secondary fermentation takes

place in a fermentation jar fitted with an airlock, hence the term anaerobic. The occasional

bubbling of the airlock is often all to show that things are still happening.

Malolactic fermentation is a possible 3rd stage that can occur some time after the original

fermentation process has ended, even a year after bottling. A continuation of the fermentation in

the bottle is to be avoided as it can result in a buildup of carbon dioxide which can cause bottles

to burst. Furthermore, it often results in a semi-carbonated wine that does not taste good.

Therefore, this stage is often induced after secondary fermentation but before bottling, by

inoculating the wine with bacteria. The bacteria, lactobacilli, will convert malic acid into lactic

acid. . A small amount of carbon dioxide will be released which in extreme circumstances can

result in a semi-sparkling wine, but the main result lies in the fact that lactic acid has about half

the acidity of malic acid. This results in a somewhat less acidic wine with a much cleaner,

fresher flavor.

IMPORTANCE OF MONITORING FERMENTATION

Malolactic Fermentation Control

There are various situations where monitoring the fermentation process of many products can be

extremely valuable. For example, malolactic fermentation is not a desirable process in bottle

because it often produces a semi carbonated wine of poor quality and taste; however, the process

is often initiated pre-bottling to convert lactic acid to malic acid, to produce a much softer,

fresher tasting product. It is imperative for wineries who employ this flavor controlling process

to be able to initiate malolactic fermentation early enough to avoid its continuation after bottling,

and consequently avoid the production an inferior product that would ultimately result in a loss

of profit.

Champagne and Sparkling Wine Production

Champagnes and sparkling wines are produced by adding rock sugar and supplementary yeast to

partially fermented wine after the primary stage of fermentation has occurred, but prior to the

onset of the secondary stage. This combination of ingredients results in the characteristic

carbonation found in champagnes and sparkling wines. In order to minimize the production time

and thus maximize profit, it is imperative that champagne and sparkling wine producers know

precisely when their batches have completed primary fermentation so that they can add the

appropriate ingredients to achieve their final product sooner and then proceed with the

production of a new batch with as little down time as possible.

Champagne and Sparkling Wine Bottling

In order for champagnes and sparkling wines to be bottled, the fermentation process must be

stopped or severely reduced to avoid carbonation buildup in-bottle that often results in an unsafe

product. Bottles have been known to explode due to carbonation buildups and many times a

chain reaction will be initiated where one bottle after another explodes causing the entire stock to

be destroyed. Because the wine and sparkling wine industries are centered around a single

year’s production, from the yearly grape harvest to the final product that must mature for a

certain amount of time, often years, before being ready for distribution, a loss like this can be

economically devastating. If a year’s stock is ruined, the winery must wait a full year before

beginning the next production process, resulting in an entire year’s loss of revenue in the future.

Knowing precisely when the fermentation process has ceased would be extremely beneficial to

wine and sparkling wine producers. This knowledge would insure that unsafe levels of

carbonation did not build up in-bottle and would ultimately give more economic security to a

relatively risky industry.

Additives to Effect Wine Characteristics

Two major characteristics of wine are sweetness and alcohol content. Control of these

characteristics can be achieved by adding additional nutrients and/or adding yeast retarding

products to the mix, respectively. Sweetness can be increased by adding additional sugars to the

partially fermented stock, where as adding yeast retarding agents impairs and ultimately destroys

yeast cells causing fermentation to cease, resulting in a product with a specific alcohol content.

In order to achieve a specific, desired product, it is crucial to be able to use these additives at the

appropriate time. Waiting too long or adding them too soon will result in either a bitter product

with too high of an alcohol content or an overly sweet product with a low alcohol content,

neither of which are desirable by the majority of wine consumers. Being able to precisely know

when to use additives would be beneficial because it would allow for the production of a

specifically defined product, one that was designed based on key characteristics desired by the

majority of the market. By appealing to a larger consumer base, product sales would begin to

increase and profit would be maximized.

THE WI – NOSE DESIGN

General Design

The design objective of the Wi – Nose was to provide the wine industry with a light weight,

compact, and accurate sensing device. Other design factors that were considered included easy

of installation, level of maintenance, and lifetime of the device. The Wi-Nose consists of a rigid,

plastic hemisphere that contains the workings of the artificial nose. The main components of the

Wi-Nose include a sensor array, a microprocessor with RAM, a wireless transmitter, and a

pneumatic pump. The sample of the test gas enters the device through a small intake. The

pneumatic pump provides the pressure differential necessary for the sample to enter the device.

The sample is then directed towards the sensor array, where it accumulates in the device

headspace. Once in the headspace, the sample gas interacts with the sensor array and is then

expelled through the sample exhaust. The sensor array contains three sensors, two metal oxide

sensors for the detection of ethanol and one electrolyte sensor for the detection of carbon

dioxide.

Figure 2 - Cross sectional view of the Wi-Nose

Upon interaction with the sample gas, the sensors send their respective output signals to the

microprocessor where the information stored in the device’s RAM and then transformed and sent

to the wireless transmitter. The wireless transmitter sends the data to the hub computer where

the information is classified with the aid of an artificial neural network. The proposed system

would allow for the addition of multiple devices for the monitoring of multiple products. By

having all devices transmit their signals to the same computer, the consumer will be only be

responsible for purchasing a single artificial neural network, a program that accounts for a large

percent of the device’s cost.

Figure 3 - Top View of Wi – Nose Device

The Wi - Nose is designed to be easy to install. Four screws and hex nuts are included with the

device, as well as multiple rubber O-rings to prevent moisture from contacting the metal

components. Because most of these units will be installed in metal fermentation vats or tanks, it

is important to reduce the risk of rusting. Moisture control issues are bound to be encountered,

and the Wi – Nose is designed specifically with these potential problems in mind. Moisture

buildup is directed away from the metal interface and sensitive device components due to the

apparatus’s unique shape. The hemisphere design allows for condensation to run down outside

of the plastic cover and drip off the device without damaging the water sensitive components

housed inside.

This design also promotes device longevity. The Wi – Nose has a minimum lifetime of 5 years.

The device should be tested on a semi-regular basis throughout the first five years to guarantee

that it is providing reliable results, however the sensors should be replaced and the device should

be serviced every six to seven years. Usually, the only parts that should have to be replaced at

this time are the sensors; however, they are of relatively low cost and replacing them on this

timetable should not amount to much cost to the consumer.

Tin Oxide Sensor

Sensor Array Board Microprocessor/RAM

Installation Screw

Sensor Choice

Three different sensors are used in the Wi – Nose design. The first is Figaro’s TGS 822 for the

detection of ethanol. The TGS 822 is a tin oxide sensor that features the following

characteristics:

� High sensitivity to organic solvent vapors such as ethanol

� Unresponsive to carbon dioxide

� High stability and reliability over a long period (lifetime ≥ 5 years, up to 200 ºC)

� Long life and low cost

It uses a simple electrical current to produce a resistance output in response to a detectable gas’s

concentration (ppm). Because the TGS 822 is unresponsive to carbon dioxide, this allowed for

its flawless integration into the Wi – Nose. This characteristic eliminates carbon dioxide’s effect

as an interference gas, and ultimately removes carbon dioxide hindrance from the design

parameters.

Figure 4 - a) Figaro’s TGS 822 b) respective resistance versus concentration plot

c) sensor structure

The Wi – Nose also features another metal oxide sensor for the detection of ethanol, the Figaro

TGS 2060. This sensor is based on an alumina substrate and features the following

characteristics:

� Low power consumption

� High sensitivity to alcohol and organic solvent vapors

� Unresponsive to carbon dioxide

� Long life and low cost

The TGS 2060 also utilizes a simple electrical circuit; however, this sensor produces a

conductivity output signal in response to a detectable gas’s concentration. Again, this sensor was

chosen because of its insensitivity to carbon dioxide, thus allowing for the output signal to be

purely based on the concentration of ethanol and not on the concentrations of both ethanol and

carbon dioxide.

Figure 5 - a) TGS 2060 b) Respective output signal versus concentration

The final sensor used in the Wi – Nose is Figaro’s TGS 4160. Figure 6 -TGS 2060

This sensor is utilized in the Wi – Nose for the detection of Structure

carbon dioxide. The TGS 4160 has the following features:

� High selectivity for carbon dioxide

� Unresponsive to ethanol

� Compact size

� Long life

The TGS 4160 differs from both the TGS 822 and the TGS

2060 because it is an electrolyte type sensor. It utilizes

electromotive force to create a signal output that corresponds to

a detectible gas’s concentration. Unlike the previous two

sensors, the TGS 4160 is unresponsive to ethanol, making it an

excellent choice for the Wi – Nose design for the same reason

that the TGS 822 and the TGS 2060 were good choices based

on their insensitivity to carbon dioxide.

Figure 7 - a) TGS 4160 b) Respective output signal versus concentration

SENSOR DATA

Each sensor’s output versus concentration plot was reproduced in Microsoft Excel by fitting

sample data points with the following model:

( ) 1−= nionConcentratmOutput Equation 1

where m and n were parameters that were allowed to vary while the sum of the square of the

difference of output and calculated output was minimized in the Excel Solver add in.

TGS 822 Conc. (ppm) Rs/Ro Rs/Ro calc Squ.Diffe. Sum of sq.

50 2.8 2.8 0.000650 0.0036 100 1.9 1.9 0.000379 200 1.3 1.3 0.002341 300 1 1.0 0.000186 1000 0.48 0.5 0.000040 5000 0.19 0.2 0.000001

m 28.11885857 Rs/Ro=m*(Concentration)n-1 n 0.412643009

Figure 8 -Example of methodology used to reproduce

Sensor TGS 822’s output versus concentration plot

Figure _ represents a typical reproduced plot for the development of the Microsoft Excel

spreadsheets used with the ANN. These plots were used to generate output data for thousands of

ethanol and carbon dioxide concentrations. All reproduced output signal versus concentration

plots and their respective m and n values can be found in the Appendix.

TGS 822 Sensor

y = 28.119x-0.5874

R2 = 1

0.01

0.10

1.00

10.00

10 100 1000 10000 100000

Concentration of Ethanol (ppm)

Rs/

Ro

Figure 9 - Output Signal versus Concentration of Ethanol (ppm) for TGS 822

NEURAL NETWORK MODELING Figure 10 -TGS 4160 Structure

The program used to develop the neural network model

was NeuroSolutions 5, the most powerful and easy to

use neural network simulation environment on the

market today. Initially an evaluation version of an add-

in to the program, NeuroSolutions for Excel 5, was

used. This allowed access to NeuroSolutions within a

familiar spreadsheet environment. The goal was to train

a neural network to classify the stages of fermentation

as 1st, 2nd, or 3rd. Data was collected from 2458

samples: 1741-1st stage, 692-2nd stage, and 25-3rd stage.

The samples were randomized using a built-in

preprocessing data function. Three columns were

defined as input variables: TGS 822 Rs/Ro, TGS 2620

Rs/Ro, and TGS 4160 EMF. Three columns were defined as the desired output: Stage 1, Stage 2,

and Stage 3. A certain percentage of the rows were tagged as training samples. A certain

percentage of the rows were tagged as cross-validation. Cross-validation is a very useful tool to

prevent what is known as over-training. Finally, a certain percentage of the rows were tagged as

testing to define samples to be used for testing the trained network.

Figure 11 - Neural Network Model for Wine Fermentation Stages Determination

Five different runs were completed, each varying the number of training, cross-validation, and

testing percentages.

Run # Training Cross-Validation Testing

1 0.6 0.15 0.25 2 0.7 0.1 0.2 3 0.5 0.1 0.4 4 0.6 0.1 0.3 5 0.8 0.1 0.1

Figure 12 - Run Parameters of Neural Network Modeling

The following is the procedure followed with each of these runs, giving the sample results for

Run #1. The same method was repeated for the other 4 runs.

First, a network was trained using 1000 epochs, which are steps in the training process of

an artificial neural network. A report summarizing the training results was generated.

The first plot shows the learning curve of training and cross-validation data. Next is a table with

minimum and final mean-squared errors (MSE).

Figure 13 - Training Results

The learning curves were then examined to see if the trained neural network did a good job of

learning the data. This was demonstrated if the MSE value approached 0. However, to verify this

conclusion, it was required to run a testing set through the trained neural network model.

First, the classification performance of the “Training” data set was determined. Then, the

classification performance of data that network had never seen was tested. This revealed whether

the neural network simply memorized the training data or truly learned the underlying

relationship. A classification report was then generated that included a confusion matrix,

summarizing classification results in an easy to interpret format, and a table listing various

performance measures.

Figure 14 -Training Data Set Testing Results

Stage 1 and Stage 2 had high accuracy for the training data set, at 100% and 97.8%, respectively.

However, the one Stage 3 sample was incorrectly classified as Stage 2, leading to 0% accuracy

for this stage.

The true test of a network, however, is how well it can classify samples that it has not seen

before. Thus another classification report was generated with confusion matrix and table, using

the testing data set. This was to see if a good model for the data had been

developed.

Figure 15 - Testing Data Set Testing Results

Stage 1 accuracy remained at 100%. Stage 2, though lowered to 92%, is still considered

accurate.For the testing data set, there were no stage 3 samples and thus, accuracy cannot be

determined for this stage.

Unlike a linear system, a neural network is not guaranteed to find the global minimum.

A neural network can actually arrive at different solutions for the same data given different

values of the initial network weights. Thus, in order to develop a statistically sound neural

network model, the network must be trained multiple times. Networks were trained 3, 4, and 5

times, with 1000 epochs for each training run. The first graph gives average of multiple training

runs along with the standard deviation boundaries. Two tables were also generated, with the

average of Minimum MSEs & Average of Final MSEs, and information about best over all

network. The second graph is a plot of learning curves for each of the runs.

Figure 16 - Average MSE with Standard Deviation Boundaries for 5 Runs

Figure #.# Training Multiple Times Results 1

Figure 17 - Training Multiple Times Results 2

The goal was to try and find a neural network model for which multiple trainings approached the

same final MSE. For this particular instance it can be seen that with 5 runs the MSE converged

to approximately 0, and thus the number of optimal runs is 5.

To develop an overall optimized neural network solution, it was required to train the process

multiple times while varying any one of the network parameters to see which gave the best

results. These parameters included hidden layer processing elements, step size, and momentum

rate. The “Vary A Parameter” training process was used to determine the optimum number of

hidden processing elements for learning the sensor data. The number of hidden processing

elements varied from 1 to 4. Each run was for 1000 epochs while the network was run ‘n’ times

for each parameter value. ‘n’ was the optimal training runs number previously found.

Networks do not generally fully learn the problem with only 1 processing element in the hidden

layer. Increasing the number of hidden processing units to 2 results in a significant improvement

in the minimum MSE. Further increasing the number of processing elements eventually results in

the final MSE converging to same value. In general, the network with more processing elements

tends to learn faster. A report summarizing the results of varying number of hidden processing

elements was generated, similar to the training multiple times report.

Fig. #.# Varying Hidden Unit Result 1

Figure 18 - Average of Minimum MSEs with Standard Deviation Boundaries

Figure 19 - Varying Hidden Units Results 2

Now that an optimal neural network model had been found, the data set tagged as “Testing” was

used to test performance of this best network. The testing report and

confusion matrix should have improved results in learning to classify fermentation

stages.

Figure 20 - Optimal Network Test Results

In this particular instance, the percentage classification wasn’t improved, and actually remained

the same. This may be due to the small number of samples actually used, which is a consequence

of using the evaluation version of the Excel add-in. The limitations of this version and its affects

on the results will be discussed in a subsequent section.

After completing all 5 runs, the following were the results:

Figure 21 - Final Results Summary

Based on the high stage accuracies for stage 1 and 2, numerous processing elements, and low

optimal runs required, Run #2 was determined to be the best. Training was set at 80%, of which

10% was cross validation, and testing was set at 20%. Literature confirms that these parameters

are commonly used as the optimal setting for a neural network model.

There are some limitations with the model found, however. For instance, since the evaluation

version of NeuroSolutions for Excel was used, the maximum number of exemplars was 300.

Thus, only 12% of all the data could be used. High accuracies were still obtained for stage 1 and

stage 2, but stage 3 did not have the same results. In fact, with the randomized samples used for

runs 1, 2, and 5, there were not any stage 3 data in the test set, thus a value could not even be

obtained. Even for runs 3 and 4, only one or two stage 3 data were present, which were

incorrectly classified as stage 2.

To fix this, another run was added. This time, instead of randomizing all the rows and then using

that data, the data was selectively chosen such that it contained all stage 3 data, data slightly

before and after the stage 1 – stage 2 break, and data slightly before and after the stage 2 – stage

3 break. This new selectively chosen optimal data of 300 samples was then randomized, trained,

and tested. The training, cross-validation, and testing percentages were the same as our previous

optimal run #2, 70%, 10%, and 20%, respectively. The optimal runs and processing elements

were 3 and 4 respectively. The following is the classification matrix obtained:

Output / Desired Stage 1 Stage 2 Stage 3

Stage 1 27 2 0

Stage 2 0 26 0

Stage 3 0 0 5

Performance Stage 1 Stage 2 Stage 3 MSE 0.025508606 0.029689604 0.004642599 NMSE 0.103065073 0.119288588 0.060775836 MAE 0.090097282 0.09653214 0.041523939 Min Abs Error 0.006176751 0.000895366 0.00119283 Max Abs Error 0.793634176 0.814344117 0.388967564 r 0.947323103 0.938487737 0.970966187 Percent Correct 100 92.85714286 100

Figure 22 - Classification Matrix for Optimal Run

Stage 1 accuracy remained at 100%. Stage 2 decreased to 93% but can still be considered

accurate. Stage 3 now has great results, with the neural network model correctly classifying all

stage 3 data.

Towards the end of the project, the full version of NeuroSolutions, albeit not the Excel-add in,

was able to be accessed. Utilizing this and NeuroBuilder, another add-in, all the data samples, if

chosen, could be used. Among the differences between NeuroBuilder and the Excel add-in is that

the file must be text delimited ASCII rather than a spreadsheet. The inputs and desired variables

were the same. Using 80% training, of which 10% was cross-validation, and 20% testing, the

entire data set was trained and tested. As expected, the results for stage 1 and 2 were quite

accurate, with 100% classification for stage 1 and 99% classification for stage 2. However, the

original problem still remained – all stage 3 data was classified as stage 2. Thus the initial

assumption that this setback resulted from not using all the data was incorrect, since the full

version of the program allowed use of all data. As a final try, the “optimized” data set was used,

which included all of the stage 3 data and the portions of stage 1 and 2 that were at the stage

boundaries, to create another neural network model, using the same previously defined

parameters, with NeuroSolutions-NeuroBuilder. This ended up giving the best results of all, with

a 100% classification rate for all 3 stages. The optimal neural network model had been found.

MARKET ANALYSIS

The number of wineries in the U.S. is approximately 4,740. The proposed market is California,

because it accounts for 90% of American wine production. The relatively small number of

wineries in California implies that information about the Wi – Nose can and will be spread

quickly. This will allows an alpha value of 1 to be reached within the first year. This can be

accomplished through various advertising techniques. The first will be placing ads on the

website, www.WineBusiness.com. This page was chosen because it is the most highly trafficked

website for the wine industry. The magazine which maintains the site, WineBusiness Monthly

(WBM), will also be a venue through which advertising will take place. WBM was specifically

chosen because in contrast to the majority of wine magazines available, which are made for wine

enthusiasts, this one is directed expressly towards the wine industry. It is the industry’s leading

publication for wineries and growers and enables advertisers to reach the entire market in the

most cost-effective way. WBM keeps readers up-to-date on the latest developments and trends in

the global business of making wine, with an emphasis on best practices and new products. As a

final means of advertising, the Wi – Nose will be presented at the Unified Wine and Grape

Symposium (UWGS). This is the largest wine and grape show in the nation, with an established

reputation for providing outstanding, current and breaking news, as well as information and

research. The show represents the collective experience, knowledge and background of the entire

industry and thus is an ideal place to market the Wi – Nose.

CUSTOMER SATISFACTION MODEL Model Development Consumer satisfaction is based not only on demand but on the quality of the product. Consumer

satisfaction can be represented as follows:

βα21 ddS += Equation 2

where d1 is the demand for the Wi – Nose and d2 is the demand for the competitor’s product.

The parameters α and β represent the inferiority function and the superiority function,

respectively. The inferiority function is related to consumer’s knowledge for the product of

interest. The inferiority function is a function of time and can vary with differing amounts of

advertisement costs. This parameter can vary from zero to one.

The superiority function represents the consumer’s preference for the product of interest in

comparison to the competitor’s product.

The parameter Y represents the consumer’s budget and can be represented as follows:

2211 dpdpY +≤ Equation 3

where p1 is the price of the product of interest and p2 is the competitor’s price. Consumer

satisfaction should be maximized while still satisfying the consumer’s budget, resulting in the

following relationship:

= β

α

αβ2

12211 d

dpdpd Equation 4

In the case of the Wi – Nose, the inferiority function α is assumed to be 1 due to the relatively

small market of U.S. wineries which implies that information about the Wi – Nose can and will

be spread quickly and will saturate within the first year. The superiority function β is defined as

the ration of the consumer preference functions for the competitor’s product and the product of

interest. This relationship is illustrated as follows:

=

1

2

H

Hβ Equation 5

The consumer preference functions are defined as the summation of the product’s design

characteristics scores and the weights each of these characteristics carries with the consumer.

This relationship is represented as follows:

∑= iii ywH Equation 6

Where wi is the product design characteristic weight and yi is the product design characteristic

score. The design characteristic weights and scores are often generated by informal surveys in

which the consumer ranks the design characteristics on the order of importance and provide

preference information that relates customer satisfaction to product attributes. This illustrates

that customer satisfaction is a function of product design.

Design characteristics that were altered in the development of the Wi – Nose were accuracy,

device size, and device weight. The weights are calculated using a point system in which votes

for most important design characteristic received the most points and votes for the least

important device characteristic received the least points, in the case of the Wi – Nose, the most

important attribute was determined to be accuracy because it had the most votes. There for the

number of votes for accuracy was multiplied by three. The second most important design

attribute was determined to be device weight because it had the second highest number of votes.

These votes were awarded 2 points each. Finally, the least important design attribute was

determined to be device size, which votes received only one point. The total number of points

was calculated and each attribute’s weight was represented as a percent of the total. The

informal customer preference survey resulted in the following design characteristic weights:

Design Characteristic Weight

Accuracy 0.43 Size (cc) 0.23 Weight (pounds) 0.34

Figure 23 - Design characteristics’ weights

Customer preference functions for each device characteristic were also generated from the

customer preference survey.

Accuracy

Figure 24 represents the customer’s satisfaction based on the accuracy of the device.

This accuracy information is then related to actual device percent correct classification in Figure

25. Finally, customer satisfaction is related to actual device percent correct classification in

Figure 26.

% Happiness vs Accuracy of Device% Happiness vs Accuracy of Device% Happiness vs Accuracy of Device% Happiness vs Accuracy of Device

Always AccurateMostly AccurateSemi AccurateNot Accurate

0

0.2

0.4

0.6

0.8

1

AccuracyAccuracyAccuracyAccuracy

% H

appiness

% H

appiness

% H

appiness

% H

appiness

Figure 24 - Percent Happiness versus Accuracy of Device

Percent correct classification rate was used to evaluate accuracy. It was determined that any

percent correct classification rate below 60 % would define not accurate, where as any percent

correct classification rate in the range of 60-75% would define semi accurate, in the range of 76-

90% would define mostly accurate, and in the range of 91-100% would define always accurate.

Accuracy of Device vs Accuracy of Device vs Accuracy of Device vs Accuracy of Device vs

% Correct Classification Rate% Correct Classification Rate% Correct Classification Rate% Correct Classification Rate

Not Accurate

Semi Accurate

Mostly Accurate

Always Accurate

0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1

% Correct Classification Rate% Correct Classification Rate% Correct Classification Rate% Correct Classification Rate

Accura

cy

Accura

cy

Accura

cy

Accura

cy

Figure 25 - Accuracy of Device versus % Correct Classification Rate

A device with a higher percent classification rate, and thus a greater accuracy, will generate a

higher customer satisfaction than a device with a lower percent classification rate, a relationship

illustrated in Figure 26. A percent classification rate of 60 % is assigned a value of zero

customer satisfaction where as a percent classification of 100% is assigned a customer

satisfaction value of 1.

% Happiness vs % Happiness vs % Happiness vs % Happiness vs

% Correct Classification Rate% Correct Classification Rate% Correct Classification Rate% Correct Classification Rate

y = 2.5x - 1.5

0

0.2

0.4

0.6

0.8

1

0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1

% Correct Classification Rate% Correct Classification Rate% Correct Classification Rate% Correct Classification Rate

% Happiness

% Happiness

% Happiness

% Happiness

Figure 26 - Percent Happiness versus Percent Correct Classification

Device Weight and Size

The above methodology was followed for both the device weight and device size.

Customer satisfaction was related to device weight descriptions, and those same device weight

descriptions were then related to actual device weight measurements (lbs). These actual device

weight measurements were then correlated with customer satisfaction. Similarly, customer

satisfaction was related to device size descriptions, and those descriptions were then related to

actual device size measurements (cc). In turn, these actual device size measurements were

correlated to customer satisfaction. The customer satisfaction plots for both of these device

characteristic can be found in Appendix I.

Superiority Function Generation

The superiority function β was generated by first calculating H1, as illustrated in Equation

5. H2 is the competitor’s consumer preference function and was assumed to be .6 based on

information provided for the Cyranonose 320, and E-nose produced by CyranoSciences. The

Cyranonose 320 has a device weight of 5 lbs compared to the Wi – Nose device weight of 1 lb.

The Cyranonose 320 is approximately the size of a walkie-talkie, where as the Wi – Nose is

much smaller at only 36 cc.

Device Characteristic

Our Device

yi Our Device Weights H1 H2 Beta

Accuracy 1 1 0.43 0.77 0.6 0.779221 Size (cc) 28 0 0.23 Weight (lbs) 1 1 0.34

Figure 27 - Table used to determine customer preference functions

The resulting β value is shown in Figure 27. Multiple β values were generated in this manner by

varying the design characteristics. These β values were then used to determine the demand for

the product of interest at different prices, p1. These demands were then used to determine the net

present worth’s associated with each product design. The net present worth’s were used to

determine the most profitable product design. This design was used as the Wi – Nose final

device design.

INTEGRATION OF THE CUSTOMER SATISFACTION MODEL

Once all the demands for each price within each β had been determined, it was necessary to

determine the net present worth for each of these β’s. This was accomplished by utilizing the

following equation:

Total Capital Investment (TCI) = Fixed Capital Investment (FCI) + Working Capital (WC)

Equation 7

Both FCI and WC are a function of the purchased equipment. Purchased equipment included a

part feeder, automated cleaning and mounting machine, brazing and soldering equipment,

automatic insertion machine, marking equipment, device testing equipment, packaging

equipment, and finally clean room applications. When totaled, the purchased equipment cost

was approximately $1.3 million and this resulted in an FCI of $5.5 million and a WC of $0.97

million. The TCI was calculated to be approximately $6.5 million.

The total annual value of products was calculated using the following equation:

Total Annual Value of Products = Price/unit * Amount of Units Equation 8

where amount is the demand determined for that particular price.

The Total Annual Cost of Raw materials was the summation of the price for each material times

the amount for that material. These amounts depended on the beta value, since size and weight

affected the amount of raw materials needed. Raw materials included the sensors, soldering

material, cover, board, wiring, and software. Depreciation was calculated using a default 5 year

MACRS method. In this analysis, depreciation is a fraction of FCI. All the costs were summed to

give a Total Product Cost of approximately. Finally, an assessment of profitability could be

made.

Return On Investment (ROI) = (Annual Net Profit) / (TCI) Equation 9

Payback Period = (FCI) / (Annual Operating Cash Flow) Equation 10

Net Present Worth (NPW) = Present Worth of All Cash Flows – Present Worth of All Capital Investments

Equation 11

Two NPWs were calculated. The first used annual end of year cash flows and discounting, while

the second used continuous cash flows and discounting. Ultimately, NPW was graphed versus

product price for each of the betas to determine the optimum product price, demand, and

qualities.

-15

-10

-5

0

5

10

15

6500 7000 7500 8000 8500 9000 9500

Price ($/unit)

NP

W 1

(10

^6 $

)

Beta 1

Beta 2

Beta 3

Beta 4

Beta 5

Beta 6

Beta 7

Beta 8

Beta 9

Beta 10

Beta 11

Beta 12

Beta 13

Beta 14

Beta 15

Figure 28 -NPW 1 vs. Product Price

-15

-10

-5

0

5

10

15

6500 7000 7500 8000 8500 9000 9500

Price ($/unit)

NP

W 2

(10

^6 $

)

Beta 1

Beta 2

Beta 3

Beta 4

Beta 5

Beta 6

Beta 7

Beta 8

Beta 9

Beta 10

Beta 11

Beta 12

Beta 13

Beta 14

Beta 15

Figure 29 - NPW 2 vs. Product Price

Beta 3 proved to be the most profitable product. The characteristics of this Beta were 100%

accuracy, a size of 36 cc, and a weight of 1 pound. This led to a Total Annual Cost of Raw

Materials of $2.07 million, with a Total Capital Investment (TCI) of $6.514 million. The price of

the product was $8,000 with a demand of 1651 units. This led to a Total Annual Value of

Products of $13.21 million. The Return on Investment (ROI) was 49.2% with a Payback Period

of 1.5 years. NPW 1 was $11.15 million, while NPW 2 was $11.97 million.

Generally, the optimal happiness product is not the most profitable because of the costs

associated in giving it these best characteristics. With the Wi – Nose, however, the optimal

happiness product was also the most profitable. This is because the only characteristics that were

varied (size & weight) had very little costs associated with them (cover-$2/unit, board-

$1.50/unit, wiring- $2/unit). This is unlike other cases in which the product has numerous

properties that are manipulated, and the costs associated with these variables are much more

significant.

RISK ANALYSIS

To evaluate the risks associated with the Wi – Nose, a risk analysis was performed using

Palisade Decision Tools @Risk 4.5 for Excel. Since the raw materials were what differed with

each product design, a 20% variability in raw material costs, using a normal distribution, was

used. The desired output for the risk analysis was ROI. Using MonteCarlo Sampling type and

10,000 iterations, a histogram and risk curve were produced.

Distribution for Return on investment, ave.%/y/D30

0.000

0.010

0.020

0.030

0.040

0.050

0.060

0.070

0.080

Mean=49.20834

25 35 45 55 65 7525 35 45 55 65 75

5% 90% 5% 40.5921 57.8622

Mean=49.20834

Distribution for Return on investment, ave.%/y/D30

0.000

0.200

0.400

0.600

0.800

1.000

Mean=49.20834

25 35 45 55 65 7525 35 45 55 65 75

5% 90% 5% 40.5921 57.8622

Mean=49.20834

Figure 30 - Risk Analysis Graphs

There is a 90% probability that the Wi – Nose will have an ROI between 40.6% and 57.9%. Thus

under the given setting and conditions, this is a profitable venture.

FUTURE CONSIDERATIONS

An option for future endeavors is to, in addition to manipulating the size and weight parameters,

to vary the number of sensors as well. This would give different values of accuracy. However, it

may then be possible to design a device that will give a higher NPW but is not the “perfect”

product in terms of happiness. This is a possibility because sensor and software costs (sensors-

$250/unit, software-$500/unit) are much more significant than the size and weight costs.

In addition, in order to be more accurate in terms of the financial breakdown, an H2 should be

determined, rather than assumed. At the time of the project, the competitor used was

CyranoSciences. Research into the company revealed the main website to be down and no

further information could be found. However, using an old business profile and making some

phone calls, it was determined that new company, Smiths Detection, has bought out

CyranoSciences. Smiths Detection now oversees all of CyranoSciences’ previous products and

endeavors. Thus, in order to determine what the competition’s current accuracy, weight, and size

are, it would be useful to contact the new parent company.

CONCLUSIONS

• An artificial sensing device, the Wi – Nose, can be successfully designed to monitor the

fermentation stages in wine.

• The sensors will consist of Figaro products:

o TGS 822 – to detect ethanol

o TGS 2620 – to detect ethanol

o TGS 4160 – to detect carbon dioxide

• The pattern recognition system will be an artificial neural network

• The most profitable product is also the “best” product in terms of happiness

• The characteristic and financial breakdown of this product is:

o Accuracy = 100%

o Size = 36 cc

o Weight = 1 pound

o Price = $8,000

o Demand = 1651 units

o Total Capital Investment (TCI) = $6.514 million

o Total Annual Value of Products = $13.21 million

o Total Annual Cost of Raw Materials = $2.07 million

o Return on Investment (ROI) = 49.2%

o Payback Period = 1.5 years

o Net Return = $2.22 million

o NPW 1 = $11.15 million

o NPW 2 = $11.97 million

• Risk Analysis with Monte Carlo Sampling confirms the product is profitable

o 90% probability that the Wi – Nose will have an ROI between 40.6% and 57.9%.

• Future work still to be done however, in order to have more complete analysis

REFERENCES

“Electronic Nose.” <http://science.nasa.gov/headlines/y2004/06oct_enose.htm>

“Electronic Nose Overview.” <http://www.cfi.lu.lv/ionics/CFI_lapai_eng/e-nose.htm>

“Fermentation 101.” <http://www.eckraus.com/wine-making-101.html>

“Figaro USA, Inc. Product Information: TGS 822 – for the detection of Organic Solvent

Vapors.” <http://www.figarosensor.com/products/822pdf.pdf>

“Figaro USA, Inc. Product Information: TGS 2620 – for the detection of Solvent Vapors.”

<http://www.figarosensor.com/products/2620pdf.pdf >

Frauenfelder Mark. “Company Profile: Cyrano Sciences, Inc. – Electronic Nose Sniffs Out New Markets for California Firm.” <http://www.smalltimes.com/Articles/Article_Display.cfm?ARTICLE_ID=267768&p=109>

Freund Michael S., Lewis Nathan S. A chemically diverse conducting polymer-based “electronic

nose”. Proc. Natl. Acad. Sci. Chemistry, Vol. 92, March 1995: 2652-2656.

Gawel, Richard. “Wine Education Topic: Wine Fermentation – Primary Fementation of Wine.”

<http://www.aromadictionary.com/articles/winefermentation_article.html>

Navratil Marian, Cimander Christian, Mandenius Carl-Fredrik. On-line Multisensor Monitoring

of Yogurt and Filmjolk Fermentations on Production Scale. Journal of Agricultural and Food

Chemistry, Vol. 52, 2004: 415-420.

“Neural Networks & Connectionist Systems.” <http://www.aaai.org/AITopics/html/neural.html>

Ouellette, Jennifer. “Electronic Noses Sniff Out New Markets.” American Institute of Physics:

The Industrial Physicist, February 1999. <http://www.aip.org/tip/INPHFA/vol-5/iss-1/p26.pdf>

Pescovitz, David. “This Nose Knows.” Wired Issue 7.12. Dec 1999.

<http://www.wired.com/wired/archive/7.12/mustread.html?pg=11>

Pinheiro Carmen, Rodrigues Carla M., Schafer Thomas, Crespo, Joao G. Monitoring the Aroma

Production During Wine-Must Fermentation with an Electronic Nose. Biotechnology and

Bioengineering, Vol. 77, No. 6, March 20, 2002: 632-640.

Portfolio: Cyrano Sciences. <http://www.rusticcanyon.com/portfolio/cmp_cyrano.html> “Smiths Detection”. <http://www.sensir.com/> “Serious Winemaking - Fermentation.” < http://freespace.virgin.net/roger.simmonds/dear.htm> “What is Electronic Nose?” <http://science.nasa.gov/headlines/y2004/06oct_enose.htm> “Wine: Primary Fermentation.” <http://www.winedefinitions.com/learningcenter/articles/primaryfermentation.htm> “Wine: Secondary Fermentation.” <http://www.winedefinitions.com/learningcenter/articles/secondaryfermentation.htm>

APPENDIX

% Happiness vs Size of Device% Happiness vs Size of Device% Happiness vs Size of Device% Happiness vs Size of Device

Extremely LargeLargeMediumSmallVery Small0

0.25

0.5

0.75

1

Size of DeviceSize of DeviceSize of DeviceSize of Device

% Happiness

% Happiness

% Happiness

% Happiness

Figure 31 - Percent Happiness versus Size of Device

Size vs. Actual Device SizeSize vs. Actual Device SizeSize vs. Actual Device SizeSize vs. Actual Device Size

Very Small

Small

Medium

Large

Extremely Large

28 32 36 40 44

Actual Device Size (cc)Actual Device Size (cc)Actual Device Size (cc)Actual Device Size (cc)

Size

Size

Size

Size

Figure 32 - Size versus Actual Device Size

% Happiness vs Actual Device Size% Happiness vs Actual Device Size% Happiness vs Actual Device Size% Happiness vs Actual Device Size

0

0.25

0.5

0.75

1

28 32 36 40 44

Actual Device Size (cc)Actual Device Size (cc)Actual Device Size (cc)Actual Device Size (cc)

% Happiness

% Happiness

% Happiness

% Happiness

Figure 33 - Percent Happiness versus Actual Device Size

% Happiness vs Device Weight% Happiness vs Device Weight% Happiness vs Device Weight% Happiness vs Device Weight

Very Light Light Medium Heavy Very Heavy

0

0.2

0.4

0.6

0.8

1

Device WeightDevice WeightDevice WeightDevice Weight

% Happiness

% Happiness

% Happiness

% Happiness

Figure 34 - Percent Happiness versus Device Weight

Weight vs Actual Device WeightWeight vs Actual Device WeightWeight vs Actual Device WeightWeight vs Actual Device Weight

Very Light

Light

Medium

Heavy

Very Heavy

1 3 5 7 9

Actual Device Weight (pounds)Actual Device Weight (pounds)Actual Device Weight (pounds)Actual Device Weight (pounds)

Weight

Weight

Weight

Weight

Figure 35 -Weight versus Actual Device Weight

% Happiness vs Actual Device Weight

0

0.2

0.4

0.6

0.8

1

1 3 5 7 9

Actual Device Weight (pounds)

% Happiness

Figure 36 - Percent Happiness versus Actual Device Weight

TGS 2620

Conc (ppm) Rs/Ro Rs/Ro calc Squ.Diffe. Sum of sq.

100 2 2.02459 0.00060 0.00907 200 1.4 1.31709 0.00687 300 1 1.02421 0.00059 350 0.9 0.93082 0.00095 700 0.6 0.60554 0.00003 1000 0.48 0.48536 0.00003 2000 0.3 0.31575 0.00025 3000 0.21 0.24553 0.00126 4000 0.18 0.20541 0.00065

Rs/Ro=m*(Concentration)n-1 m 35.22801 n 0.37972

Figure 37 - Methodology used to reproduce the output signal versus concentration plot for TGS 2620

TGS 2620 Sensor

y = 35.228x-0.6203

R2 = 1

0.01

0.10

1.00

10.00

100 1000 10000 100000

Concentration of Ethanol (ppm)

Rs/

Ro

Figure 38 -Reproduced output signal versus concentration of ethanol plot for TGS 2060

TGS 4160 gas Conc. , ppm EMF, (mV) EMF, (mV) Squ.Diff. Sum of sq.

350 360 359.92773 0.00522 0.11383 500 350 349.81314 0.03492 700 340 340.27146 0.07369 1000 330 330.15688 0.02461 2000 310 310.50061 0.25061 3000 300 299.00243 0.99515 20000 245 245.20390 0.04158 40000 225 225.54763 0.29990 100000 200 199.56346 0.19057

Emf=m*Ln(concentration)+n m -28.358 n 526.047

Figure 39 -Methodology used to reproduce output signal versus concentration plot for TGS4160

TGS 4160 Sensor

150

200

250

300

350

400

100 1000 10000 100000

Concentration of CO2 (ppm)

EM

F (

mV

)

Figure 40 -Reproduced output signal versus concentration of carbon dioxide for TGS 4160