Elsevier Editorial System(tm) for Neural Networks

Manuscript Draft

Manuscript Number: NEUNET-D-07-00110

Title: An Adaptive Method For Industrial Hydrocarbon Flame Detection

Article Type: IJCNN '07

Section/Category:

Keywords: Artificial neural networks; Signal processing; Flame detection

Corresponding Author: Javid Huseynov,

Corresponding Author's Institution: University of California Irvine

First Author: Javid J Huseynov, PhD Candidate in Computer Science

Order of Authors: Javid J Huseynov, PhD Candidate in Computer Science; Shankar B Baliga, PhD in

Physics; Alan Widmer, B.S. in Physics; Zvi Boger, M.S. in Chemical Engineering

Abstract: An adaptive method for an infrared (IR) hydrocarbon flame detection system is presented. The

model makes use of the joint time-frequency analysis (JTFA) for feature extraction and the artificial neural

networks (ANN) for training and classification. Multiple ANNs are trained independently on a computer,

using the backpropagation conjugate-gradient (CG) method, with input data collected from various flame

and non-flame nuisance signals at four different IR wavelengths. The trained ANN connection weights are

programmed into an embedded system as a part of the voting scheme for distinguishing flames from

nuisance sources. Signal saturation caused by the excessive intensity of some IR sources is resolved by an

adjustable gain control mechanism. The model described herein is employed in an industrial hydrocarbon

flame detector.

An Adaptive Method For Industrial HydrocarbonFlame Detection

Javid J. Huseynov a,b,∗, Shankar B. Baliga b, Alan Widmer b,Zvi Boger c

aSchool of Information and Computer Science, University of California Irvine, Irvine, CA92697, United States

bGeneral Monitors, Inc., 26776 Simpatica Circle, Lake Forest, CA 92630, United StatescOPTIMAL - Industrial Neural Systems, Ltd., Be’er Sheva 84243, Israel

∗ Corresponding author.Email addresses: javid@ics.uci.edu; phone: +1 949 581-4464 (Javid

J. Huseynov), shankarb@generalmonitors.com (Shankar B. Baliga),alanw@generalmonitors.com (Alan Widmer), zviboger@gmail.com (ZviBoger).

Preprint submitted to Elsevier 21 August 2007

* Title Page (With all author details listed)

Abstract

An adaptive method for an infrared (IR) hydrocarbon flame detection system is presented.The model makes use of the joint time-frequency analysis (JTFA) for feature extractionand the artificial neural networks (ANN) for training and classification. Multiple ANNs aretrained independently on a computer, using the backpropagation conjugate-gradient (CG)method, with input data collected from various flame and non-flame nuisance signals atfour different IR wavelengths. The trained ANN connection weights are programmed intoan embedded system as a part of the voting scheme for distinguishing flames from nuisancesources. Signal saturation caused by the excessive intensity of some IR sources is resolvedby an adjustable gain control mechanism. The model described herein is employed in anindustrial hydrocarbon flame detector.

Key words: Artificial neural networks; Signal processing; Flame detection

1 Introduction

The study of fires occupies a unique niche in the world of science and engineeringbecause an unwanted fire is considered a failure in the sense that it is not a desirableoutcome and is to be avoided (Jones, 2004). Therefore, the early detection of hy-drocarbon flames in an industrial facility serves as means ofavoiding such failures,and carries the burden of ensuring both the reliability of the detection and the fastresponse of the system in operation.

Infrared (IR) optical sensors are broadly used in industrial hydrocarbon flame de-tection. Their popularity is dictated by the fixed emission wavelengths of hydro-carbon flames in the IR spectrum, which can be separated from non-flame sourcesand analyzed in various domains. Classical optical hydrocarbon flame detectors arebased on an expert system, where analog signals are collected from the optical sen-sors, converted into digital format, processed, and an output decision is reported onthe presence of flame or lack thereof.

Although simple in appearance, the described model of flame detection becomesmore complex when dealing with IR data from real industrial environments. In ad-dition to the robust identification of an undesired phenomenon, a flame detectionsystem should also be able to distinguish flames from sourcesof environmentalnuisance and ambient noise. IR signals at flame wavelengths can be easily gen-erated by random motion, modulation of heated surfaces, hotair flow, arc weld-ing, sunlight reflecting off a water surface, and other non-flame related environ-mental nuisance. To ensure the adherence of a flame detectionsystem to these re-quirements prior to its deployment, various regulatory compliance standards, suchas FM 3260 (Factory Mutual, 2000), EN 54-10 (European Standard, 2002) and

Preprint submitted to Elsevier 21 August 2007

Manuscript (without any author details listed)

ULC/ORD-C386 (Underwriters Laboratories, 1990) were developed. These stan-dards subject the detection system to a set of reproducible flame and environmentalnuisance sources to test and certify its performance upon deployment.

Strict flame performance regulations have in the past guidedthe designs of a varietyof expert systems based on a limited set of classical time- and frequency-domainsignal processing algorithms. These systems offered simplicity and the definitivereproducibility of performance in standard compliance testing, mostly designed toavoid failures in detection of flames (false negatives). Yet, beyond the regulatorystandards, these systems offer very little flexibility in terms of avoiding the iden-tification of nuisance and ambient noise sources as flames (false positives). Partic-ularly, in IR sensor-based expert systems, the addition of environmental nuisance(heat, hot surface radiation, sunlight, welding, etc.) in the same IR wavelength andelectrical frequency spectrum as hydrocarbon flames makes the rejection of falsepositives very difficult. Some system designers (Baliga et al., 2000; Schuler, 1998;Goldenberg et al., 1994; Det-tronics, 2000; Spectrex, 2005) attempted to ease theproblem with false positives by adding extra sensors at various wavelengths andrestricting the system with a set of more elaborate expert rules. However, in thecurrent market, this solution is extremely pricy in terms ofboth development andunit cost.

As opposed to expert systems, adaptive systems are poised topush the performancein both false positive and false negative cases far beyond the scope of regulatorystandards at no extra cost. The clear advantage of adaptive detection systems basedon fuzzy logic (Thuillard et al., 2000) or neural networks (Huseynov et al. 2005,2007; Chen et al, 1999; Milke, 1995; Milke & McAvoy, 1995; Okayama, 1991a;Okayama, 1991b) is in their reliance on the pattern of the signals rather than onfixed magnitude or phase. For example, an ANN-based detection system can betrained on the data collected from one or more IR sensors observing a variety offalse-positive and false-negative cases. Successful training can automatically derivea set of unique input-output correlation rules that would, otherwise, have to bederived by an expert. Hence, the adaptive approach offers a greater classificationcapability in a much shorter development time.

In this paper, we propose a flame detection system based on JTFA (Zielinski, 1999)and multiple ANNs for distinguishing hydrocarbon flames from environmental nui-sance and background false positives. The design describedhere was implementedin the first industrial flame detector using ANN (Huseynov et al., 2007), which wascertified per FM 3260 (Factory Mutual, 2000), EN-54104 (European Standard, 2002)and ULC/ORD-C3865 (Underwriters Laboratories, 1990), andis currently deployedin the field operation.

In Section 2, we describe some related work on adaptive fire detection systems.Section 3 presents our signal feature extraction scheme andANN design, and Sec-tion 4 summarizes the experimental results.

2

2 Related Work

Over the last two decades, the application of neural networks for the classificationof sources has been a subject of research in fire detection systems. Perhaps, theearliest example was the work by Okayama (1991a) using odor sensors togetherwith a backpropagation neural network to distinguish smoldering fires from envi-ronmental noise such as coffee powder and perfume. In subsequent work, Okayama(1991b) experimented with inputs from carbon monoxide, temperature and smokesensors to estimate a probability of fire.

Expanding on Okayama’s research, Milke & McAvoy (1995) applied neural net-works to a wider range of sensors and fire test cases, such as those from flammableliquid, paper, cotton, cardboard, etc. Their experiments correctly identified 62% ofsmoldering fire sources and 87% of nuisance and ambient non-fire sources.

Ishii et al. (1993) used a time-delayed neural network (TDNN) model to distinguishfires from non-fires in a limited number of tests. Relying on historical information,the TDNN records the transient nature of fire and non-fire conditions. In our ap-proach, for extracting the ANN input features, we also look at the signal over adata window of the past 2.5 seconds. However, in our design each data window isprocessed independently, and no correlations are established between the consecu-tive data windows for the ANN input.

Chen et al. (1999) proposed using the Fourier Transform-based IR spectroscopy(FT-IR) of gas phase products along with the ANN to distinguish flaming and smol-dering fires from environmental nuisance. Their ANN model was formulated usingthe Linear Vector Quantization (LVQ) approach with a hiddenKohonen layer. Theirtraining model implemented with NeuralWorks software (NeuralWare, 1999) pro-duced a 96% success rate in classification and was extensive in analyzing the contri-bution of various gas types to flaming and smoldering fires. But it wasn’t designedfor implementation in a product that must comply with regulatory agency require-ments as in our case, i.e. the false negative error rate in ourcase has to be at 0%.

In our previous work (Huseynov et al., 2005), we presented a model based on asingle large-scale backpropagation ANN and non-restricted feature extraction. Thisdesign had an average 95% classification success rate using asingle network in alltest cases. However, while passing the majority of the compliance tests, some issueswere observed:

(1) The training data from the nuisance sources outside of the flame frequencyband was not useful for classification.

(2) Feature extraction on a longer data window did not improve classification yetdegraded the response time.

(3) The scheme using a single large ANN proved to be overly cumbersome in thetraining process.

3

(4) The placement of sensors at distances less than 40 ft fromthe fire resulted ina signal saturation.

In this work, without changing the training algorithm, we modified the feature ex-traction and the ANN classification schemes. The feature space is brought closer tothe flame flickering frequencies (5 - 15 Hz), and the length of the processing datawindow is reduced to improve the response time by a factor of two. Instead of asingle ANN with a large number of (over 1000) inputs, multiple small ANNs areapplied in series, each trained with the data from only two major types of targetphenomena, i.e. flame vs. sunlight, flame vs. hot surface without light emission,flame vs. heated surface with light emission, etc. We also introduce an adjustablegain control scheme to resolve the issue of signal saturation.

3 Design

The application of ANN in flame detection seeks to distinguish flame conditionsfrom non-flame nuisance and background noise, and to establish numerical rep-resentations of such input-output relationships without apriori knowledge of theconditions. Correct pre-processing of input and target values is essential for goodANN model training (Alfassi et al., 2005). It’s therefore imperative to design anelaborate signal-processing scheme for extracting relevant features from input sig-nal.

3.1 Feature extraction

Typically, signals are analyzed in a time or a frequency domain. Analyzing sensorresponse in a time domain is complex due to the variance of timed signal patternsrelevant to the same environmental phenomenon. For instance, signals producedby the same flame source can vary in amplitude and function depending on dis-tance, angle, presence of obstacles and other non-flame conditions (sunlight, wind,random modulation, bright lights, rain, fog, dust), which may look like a flame toan IR sensor. As opposed to the time domain, the frequency of aflickering flameremains relatively independent of the environmental conditions, so the frequencybecomes an important parameter in analyzing signals. At thesame time, due to thelow-frequency range of a flickering flame, the frequency-only analysis is prone toa low-frequency noise from non-flame sources, so the time domain information isneeded.

To avoid the drawbacks of time-only and frequency-only signal processing methodsand to track the frequencies of non-stationary time-varying signals, we used JTFA(Zielinski, 1999) via the application of the Short-time Fourier Transform (STFT)

4

(Swanson, 2000). The mathematical formulation of STFT is given below:

Xl(k) =N−1∑

n=0

w(n)x(n + lH)e−jωkn,

wheren is the number of time samples,w(n) is a data window,H is the windowshift size,x(n) is the input signal,ωk = 2πfk/fs is the frequencyfk of the kth

Fourier transform bin, normalized by sampling frequencyfs, andl = 0, 1, 2, . . . isa discrete frame.

In an STFT application, an input signal is cut into slices, followed by the applicationof Fast Fourier Transform (FFT) (Cooley & Tukey, 1965) to individual slices. Thefunctions obtained by such segmentation are not periodic. This results in higherFourier coefficients at high frequencies, since FFT interprets jumps between slicesas abrupt changes in signal. Such spectral leakage (Swanson, 2000) is resolved bythe application of data windowing, when an input signal buffer is multiplied by araised cosine wave, as described in Figure 1. Data windowinggradually attenuatesthe amplitude of signal at either end of the input buffer, hence reduces the spectralleakage into adjacent slices and forces the input wave to be more periodic. There areseveral known functions for data windowing such as Hamming,Hanning, Parzen,Gaussian, and others.

In our previous work (Huseynov et al., 2005), we have experimentally identifiedthat the application of the Hamming window results in the best ANN classification.The functional representation of the Hamming window is as follows:

W Hm(n) =1

2

{

1.08 − 0.92 cos(

2πn

N − 1

)}

,

whereN is the size of the window, andn is the variable index.

In our application, the length of data window is set toN = 256 samples and thetime shift is set atH = 25 samples. At the data sampling rate of 10 ms and theresolution of 0.4 Hz, 50 values, in the index range of 3 to 52, are chosen in theoutput of a 256-point STFT. These values contain the frequency information in therange of 1.2 Hz to 20.8 Hz. Thus, combined between four sensors, a 200-inputANN feature vector is formed, as described in Figure 2.

As opposed to our previous approach (Huseynov et al., 2005),where we used a512-point STFT, this approach reduces the response time of the detection systemby a factor of two. Besides this fundamental improvement, the reduction in thesize of STFT also facilitates the efficiency of computation in the embedded systemthereby allowing us to use more than one ANN in the reduced cycle time.

The inputs have to be scaled into common range, so as to give equal importance

5

to values in different ranges. For example, the average magnitude of a signal at 12Hz could be 1000 times larger than the average magnitude of the signal at 5 Hz forthe same time interval. One of the common preprocessing methods isrange scal-ing (Alfassi et al., 2005), in which the range between lowest andhighest expectedvalues is used as a scaling factor,

V scaledi =

Vi − Vmin

Vmax − Vmin

,

which results in 0-1 range of preprocessed inputs and outputs. Another widely usedmethod isauto-scalingof input values (Alfassi et al., 2005), where data columnvalueVi is zero-centered by subtracting the column meanµcol, and unit-normalizedby dividing the result by the column standard deviationσcol.

V scaledi =

Vi − µcol

σcol

In our feature extraction scheme, we make use of the auto-scaling of input values.

3.2 Training Model

The training model used in this design is based on the PCA-CG algorithm (Guterman, 1994)successfully applied in another application (Boger, 2002). The PCA-CG algorithmis based on the conjugate-gradient (CG) descent method for feed-forward networks(Johansson et al., 1992) and can train large-scale ANN models as it starts from non-random initial connection weights derived from a training data set. It also usesPrincipal Component Analysis (PCA) to estimate the number of hidden neuronsfor training. However, due to processing limitations upon validation in an embed-ded system, we use a fixed number of 5 hidden neurons.

To avoid getting the same value if the input sum is zero, a ’bias’ input with aconstant value of one is added to the input layer and to the hidden layer, withadjustable connection weights similar to the other connection weights. To avoidundesired oscillations during the training process, a momentum part proportionalto the most recent connection weight is added to the trainingequation.

The training algorithm consists of the following steps:

1) Form joint input-output data vectorX= xp ∪ yp, makingNp rows of matrixXrepresent the entire data set. The columns ofX are scaled by subtracting the meanof each column from the values in it, and dividing the resultsby standard deviationof each column.

6

2) CalculateΣX[a × a] as

ΣX = E{(X− E{X})T (X− E{X}}

3) Determine the eigenvectors and eigenvalues ofΣX. Select eigenvectorsφ1 · · ·φr

corresponding to the largest eigenvaluesλ1 · · ·λr necessary for reconstructingXwith a chosen information contentξ:

µi =λi

tr(ΣX)=

λi∑a

i=1 λi

Then, assuming thatλi andφi are ordered, the number of neurons in the hiddenlayer,r, would be equal to the number of dimensions necessary to reconstruct theoriginal information with aξ degree of fidelity,

r∑

i=1

µi ≥ ξ

There aren inputs andm outputs, anda = m+n, and ther is the number of nodesin the hidden layer.

4) Compute the initial input to the hidden weights matrixWH as follows (the lastcolumn are the bias values):

WH =

φ11 · · · φn1 h1

φ12 · · · φn2 h2

· · · · · · · · · · · ·

φ1r · · · φnr hr

hi =a

∑

j=n+1

φTijE{Xj}

5) Compute the initial hidden-to-output weights matrixWO (the last column rep-resents the bias values):

WO =

φ(n+1)1 · · · φ(n+1)r un+1

φ(n+2)1 · · · φ(n+2)r un+2

· · · · · · · · · · · ·

φa1 · · · φar ua

7

Ubias =a

∑

i=r+1

φTi E{X}φi = [u1, u2, · · · , ua]

T

6) A conjugate-gradient method (Leonard & Kramer, 1990) is employed to searchfor the optimal weights. The algorithm uses non-random initial connection weights,calculated from characteristics of the training data. The number of hidden neuronsis kept small, usually between 4 and 7, although as mentionedabove, we have useda fixed number of 5 neurons.

One general weakness of ANN modeling compared with the expert systems is thelack of explanation facility for its results (Alfassi et al., 2005). To overcome thisdrawback, a scheme to establish a quasi-quantitative relationship between each in-put and each output, in the form of causal indices (CI) (Baba et al., 1990), is used.The CI describes the magnitude and sign effect of any output when each input valueis changed. It’s calculated from the trained ANN connectionweights as the sum ofthe products of all paths between each input and output:

CI =∑

Wkj × Wji

over allh hidden neurons.Wkj are the connection weights from hidden neuronj tooutputk, Wij are the connection weights from inputi to hidden neuronj.

The advantage of the CI is in the lack of its dependence on a particular input vector,but on the connection weight set that represents all the training input vectors. Thismethod is more reliable that the local sensitivity checks, as it is based on the wholeANN trained on all available states.

In our approach, the CI was an important parameter for choosing inputs only fromthe frequency band (1.2 - 20.8 Hz) that is the most relevant toflame conditions. Asopposed to our previous model (Huseynov et al., 2005), wherethe STFT output (0 -50 Hz) was used entirely as the ANN input, the current model significantly reducesthe training time, complexity, and allows us to use multipleANNs for improvingthe decision confidence.

Once the ANN has been trained, there remains the task of validating the model.The validation is done by presenting the ANN with the validation set of features,extracted by the same method that was used for producing the training and testingsets. If the ANN results of this phase meet the expected errorcriteria, the ANNhas successfully generalized. In our experience, to achieve a generalization, it takesseveral hundred training epochs. Care should be taken not toovertrain the ANN bypresenting a large body of input data and training the ANN, ina brute-force fashion,for an unlimited number of epochs. This may result in a loss ofgeneralizationand will make the resulting ANN connection weights unfit for classification in theembedded system.

8

3.3 Classification Model

The ANN classification model implemented in the embedded system consists of 5hidden neurons and 1 output neuron, indicating either flame or non-flame condi-tion. A unipolar neuron activation (sigmoid) function is used at the output of everyneuron. The model of our implementation with multiple feed-forward ANNs is de-picted in Figure 3, and described in detail below.

Our prior approach (Huseynov et al., 2005), using a single large-scale ANN, hadcomplicated the training process and the flexibility of the classification. It was im-possible to train a single network which would perform well on all the cases in-volving various false positives differing in physical nature but overlapping in theIR domain with certain hydrocarbon flames. In our experiments, we identified threemajor types of false positive sources, some of which overlapin features:

• Heated surfaces with light emission (arc welding, industrial heaters);• Modulated and stationary hot surfaces without light emission (hot plates, hot air

flow, hot air gun);• Direct, modulated, reflected sunlight and bright light surfaces (incandescent and

luminescent lamps, arc welding).

So we have broken down a single large-scale ANN into multiplesmall sized ANNs,each separately trained to distinguish all flames from only agiven set of false pos-itive nuisance sources. During classification, each ANN uses the same set of inputfeatures to make a flame vs. non-flame decision. In order for the observed inputto be classified as flame, all ANNs must produce a flame likelihood value abovecertain predefined threshold. The use of multiple specialized ANNs along with thechanges to the signal processing scheme and the addition of an automatic gain con-trol, described in the next section, results in a more robustperformance.

3.4 Post-Processing

After the application of multiple ANNs, a post-processing scheme is needed to cor-relate the ANN outputs to flame vs. non-flame decision. Several schemes for eval-uation of ANN model outputs exist, including the thresholding of ANN outputs,choosing the highest output, using the Receiver Operator Characteristic (ROC)curve (Alfassi et al., 2005), or examining the outputs of hidden neurons. In thelatter approach, those examples that produce a similar pattern in the outputs of hid-den neurons may be considered belonging to the same class (Boger et al., 1997).

The values of output neurons may also be regarded as the posterior probabilitiesof the class density distribution, which may result in a reduced confidence in theclassification results. As mentioned earlier (Huseynov et al., 2005), it’s very diffi-

9

cult if not impossible to train a single large neural networkwhich would general-ize on all combinations of flame and non-flame conditions and pass all regulatoryperformance tests. Therefore, the application of multipleANNs with further post-processing provides a decision confidence in presence of a real flame condition.

In our application, we use a simple voting scheme based on thresholding of ANNoutputs to accept or reject the existence of flame condition.A threshold of 0.7 (70%likelihood) is applied to each ANN output, and if all four ANNoutputs are abovethis threshold for a period more than 3 seconds, a flame alarm is raised. If two ormore ANN outputs fall below the treshold for a period more than 1 second, thealarm condition is disabled.

3.5 Gain control

An important problem that usually arises in signal processing is the signal satura-tion (Swanson, 2000). An excessive intensity of the analog signal from a sensormay yield an analog signal which is cut off (or saturated) at the range limits ofthe analog-to-digital converter, as shown in Figure 4. The saturated signal lookslike a square wave, and its FFT does not produce valid spectral information, whichultimately invalidates the input to ANN and the whole scheme.

The scaling of the converted signal within the converter’s range is dependent on theelectronic gain of the circuit, which is controlled by the embedded software. Onepotential solution to the saturation problem could be just reducing the electronicgain until the signals stop saturating altogether. However, the reduction in signalmay also shorten the range of detection without contributing to any improvementin the signal-to-noise ratio (SNR). So to alleviate the signal saturation effects, wehave come up with an elaborate expert mechanism, which constantly tracks the rawsignal amplitude between the limits ofVmin + ∆ andVmax − ∆. As soon as theamplitude of the signal falls below or above this range, the system will adjust theelectronic gain and rescale the signal back to the range.

This improvement to the design enables the detection systemto take advantage ofthe ANN classification upon increased IR radiation at closerdistances. For exam-ple, ann-heptane flame burning in a 1 square foot pan at 40 feet causes immediatesignal saturation at the nominal gain, which made it impossible for the ANN to clas-sify this particular phenomenon under the previous design (Huseynov et al., 2005).In the new design, because of the gain control mechanism, thesignal saturation hap-pens at 10 feet or below with minimal adjusted gain, and the ANNs can effectivelydistinguish such an intense flame phenomenon from non-flamesat or greater than10 feet.

10

4 Experimental Results

The design described in this paper has been implemented in the FL4000, an indus-trial flame detector from General Monitors (Huseynov et al.,2007). The followingsubsections describe the details of the ANN training model,details

training implementation, results, and classification performance of the final instru-ment.

4.1 ANN Training

Four independent ANNs have been trained using four data setswith some over-lapping data. As in previous work, the data was collected from the IR sensorsobserving various flame and non-flame nuisance and noise conditions regulatedper FM 3260, EN 54-10 and ULC/ORD-C386 requirements. These conditions in-cluded n-heptane, propane, butane flames at distances from 0to 250 feet, direct,reflected and modulated sunlight, arc welding, modulated heater with light source,modulated hot surface with and without a light source, flashlight, incandescent andluminescent light, vibration and other non-flame nuisance.The training programran in MATLAB 7.1 software on a Windows PC. Each of the four ANNs had inputsof 30,000 samples (30,000 x 200 matrix), out of which 70% wererandomly chosenas a training set, 10% for cross validation and 20% for testing. The target was asingle column indicating either a flame (1) or a non-flame (0) condition.

The structure of the ANN hidden and output layers was the sameas in previousdesign1: 5 hidden neurons and a single output neuron, using aunipolar activationfunction. The training Root-Mean Square (RMS) error rates in nuisance cases foreach ANN are presented in Table 1. The training algorithm converged in an aver-age of 100 - 150 epochs. Obviously, to meet the standard requirements, the flamedetection (false negative) error rate in all cases had to be 0%, which was achievedin the regulatory testing.

4.2 Classification performance

The trained ANN model for classification was implemented in the embedded sys-tem on a Texas Instruments (TI) F2812 Digital Signal Processor (DSP), using Cprogram with a virtual floating point arithmetic library from TI.

Extensive performance tests were conducted on the final instrument to comply withthe regulatory standards. The test results show that, in flame response performance,our ANN-based design is equal or superior to some known expert-systems for IR

11

flame detection (Det-tronics, 2000; Spectrex, 2005). Particularly, our design offersthe on-axis range of 230 feet to detect a nominal 1 square footn-heptane flame witha response time of less than 6 seconds. The highest on-axis range ever offered by anexpert system (Det-tronics, 2000) for the same source is 210feet with a responsetime of over 10 seconds.

In addition, our nuisance testing results show advantages of the ANN-based de-sign over expert systems in eliminating false positives (non-flame sources identifiedas flames). In Table 2, the comparative performance results of rejecting nuisancesources as non-flames are presented for the ANN-based designvs. the best achieve-ments (Det-tronics, 2000; Spectrex, 2005) of the expert systems. The table lists theminimum ”immunity” distances at which the nuisance source is identified as a falsepositive.

As shown in Table 3, our current multiple-ANN design also shows visible improve-ments over our previous design with a single large-scale ANN(Huseynov et al., 2005)in eliminating false positives. The flame response time was improved by two times,while the range remained the same.

5 Conclusion

A design for an industrial IR flame detection using ANN is presented. JTFA(Zielinski, 1999) using STFT is applied to identify relevant signal frequencies asinput features for the ANN. The effects of feature extraction and input reductionhave a dramatic impact on training time and complexity. Other, more advancedmethods, such as the Discrete Wavelet Transform (DWT) (Swanson, 2000) can beapplied to obtain a more relevant set of input features, subsequently resulting inbetter training and classification.

The ANN is trained on a real environmental data using the CG method along withPCA for initializing the connection weights to non-random values. A classificationscheme based on 5 hidden and 1 output neurons is implemented on the DSP. Thebanding of feature frequencies close to the flame flickering range of 1 - 20 Hz re-moves unnecessary noisy input from the ANN, contributes to faster training and animproved classification success rate. The input reduction also enables using morethan one ANN in series for classification, further contributing to lower classifica-tion error rates. Additionally, the signal saturation problem is addressed using again control mechanism, which improves the quality of collected data for input toANN at distances closer to the detected phenomenon.

The detection system described in this paper has been fully implemented in the in-dustrial IR flame detector by General Monitors (Huseynov et al., 2007), which wascertified per North American FM 3260 (Factory Mutual, 2000) and ULC/ORD-

12

C386 (Underwriters Laboratories, 1990), and the European EN 54-10 (European Standard, 2002)regulatory standard for industrial flame detection. Presented results show that theflame detector using our ANN-based classification method achieves longer range(up to 230 ft) of flame detection at shorter response times than those currentlyprovided by the expert systems. At the same time, it providesfor an exceptionaldiscrimination against non-flame sources of environmentalnuisance and electronicnoise. Presented improvements over our previous design (Huseynov et al., 2005)include the two-fold reduction in flame response time and shorter false positive”immunity” ranges.

References

Alfassi, Z. B., Boger, Z., & and Yigal, R. (2005). Statistical Treatment of Analytical Data,CRC Press, 254.

Baba, K., Enbutu, I., & Yoda, M. (1990). ”Explicit representation of knowledge acquiredfrom plant historical data using neural network,”Proceedings of the International JointConference on Neural Networks, 3, 155-160.

Baliga, S. B., Rabe H., & Bleacher, B. (2000). ”Digital Multi-Frequency Flame Detector,”U.S. Patent No. 6,150,659.

Boger, Z., Ratton, L., Kunt, T. A., McAvoy, T. J., Cavicchi, R. E., & Semancik, S. (1997).”Robust classification of ’Artificial Noise’ sensor data by artificial neural networks”,Proceedings of the IFAC ADCHEM’97 Conference, Banff, Canada, 334-338.

Boger, Z. ”Who is afraid of the BIG bad ANN?”Proceedings of the IEEE InternationalJoint Conference on Neural Networks (IJCNN’02), 3, 2000-2005.

Chen, Y., Serio, M. A., & Sathyamoorthy, S. (1999). ”Development of a Fire DetectionSystem Using FT-IR Spectroscopy and Artificial Neural Networks,” Proceedings of theSixth International Symposium of the International Association for Fire Safety Science,791-802.

Cooley, J. W., & Tukey, J.W. (1965). ”An Algorithm for The Machine Calculation of Com-plex Fourier Series”,Math. Comput., 19, 297 - 301.

Det-tronics, Inc. (2000). ”X3301 Multi-Spectrum InfraredFlame Detector”,(http://www.detronics.com).

European Standard EN 54-10. (2002). ”Fire Detection and Fire Alarm Systems”, Part 10,European Committee for Standardization, (http://www.cenorm.be)

Factory Mutual (FM) Global. (2000). ”Radiant Energy-Sensing Fire Detectors for Auto-matic Fire Alarm Signaling”, (http://www.fmglobal.com)

Goldenberg, E., Olami, T., & Arian, J. (1994). ”Method For Detecting A Fire Condition”,U.S. Patent No. 5,373,159.

Guterman, H. (1994). ”Application of Principal Component Analysis to the design of neuralnetworks,”Neural, Parallel and Scientific Computations, 2, 43-54.

Huseynov, J. J., Boger, Z., Shubinsky, G. D., & Baliga, S. B. (2005). ”Optical Flame De-tection Using Large-Scale Artificial Neural Networks”,Proceedings of the IEEE Inter-national Joint Conference on Neural Networks (IJCNN’05), 3, 1959-1964.

Huseynov, J. J., Baliga, S. B., Shubinsky, G. D., & Boger, Z. (2007). ”Flame DetectionSystem,” U.S. Patent No. 7,202,794.

13

Ishii, H., Ono, T., Yamauchi, Y., & Ohtani, S. (1993). ”Fire Detection System by Multi-Layered Neural Network with Delay Circuit,”Proceedings of the Fourth InternationalSymposium of Fire Safety Science, 761-772.

Johansson, E. M., Dowla, F. U., Goodman, D. M. (1992). ”Backpropagation learning formulti-layer feed-forward neural networks using the conjugate gradient method,”Inter-national Journal of Neural Systems, 2(4), 291-301.

Jones, W. W. (2004). ”Development of a Multi-Criteria Algorithm for Fast and ReliableFire Detection”,Proceedings of the Thirteenth International Conference onAutomaticFire Detection (AUBE’04), 184-195.

Leonard, J., & Kramer, M. A. (1990). ”Improvement of the Backpropagation Algorithm fortraining neural networks”,Computers Chemical Engineering, 14(3), 337-341.

Milke, J. A., & McAvoy, T. J. (1995). ”Analysis of Signature Patterns for DiscriminatingFire Detection with Multiple Sensors,”Fire Technology, 31 (2), 120-136.

Milke, J. A. (1995). ”An Application of Neural Networks for Discriminating Fire De-tectors”, Proceedings of the International Conference on Automatic Fire Detection(AUBE’95), 213-222.

NeuralWorks Professional II/PLUS.(1999). NeuralWare Inc., Carnegie, PA,(http://www.neuralware.com)

Okayama, Y. (1991). ”Approach to Detection of Fire in Their Very Early Stage by OdorSensors and Neural Net,”Proceedings of Third International Symposium of Fire SafetyScience, 955-964.

Okayama, Y. (1991). ”A Primitive Study of a Fire Detection Method Controlled by Artifi-cial Neural Net”,Fire Safety Journal, 17, 535-553.

Schuler, F. (1998). ”Dual Wavelength Fire Detection Methodand Apparatus,” U.S. PatentNo. 5,850,182.

Spectrex, Inc. (2005). ”Sharpeye 20/20I Triple IR Flame Detector”’,(http://www.spectrex-inc.com).

Swanson, D. (2000). Signal Processing For Intelligent Sensor Systems, Marcel Dekker, Inc.Thuillard, M. P. (2000). ”Method for Analyzing the Signals of a Danger Alarm System and

Danger Alarm System for Implementing Said Method”, U.S. Patent No. 6,011,464.Underwriters Laboratories Canada (ULC) Standard ULC/ORD-C386. (1990). ”Flame De-

tectors”, (http://www.ulc.ca)Zielinski, T. P. (1999). ”Joint time-frequency resolutionof signal analysis using Gabor

transform,”Proceedings of the 16th IEEE Instrumentation and Measurement TechnologyConference (IMTC’99), 2, 1183-1188.

14

Fig. 1. The Application of STFT with Data Windowing.

Fig. 2. ANN Input Feature Generation.

Fig. 3. The Application of Mutliple ANNs.

Fig. 4. Signal Saturation and Gain Control.

Table 1RMS Error Rates and Number of Epochs for each ANN

ANN (1) (2) (3) (4)

Flame vs. Welding Light Heat Hot Surface

RMS Error 3.2% 2.8% 3.9% 1.8%

No. of Epochs 236 407 364 296

Table 2Adaptive vs. Expert System Performance in False Positive Cases

Nuisance Adaptive Expert

Sources Immunity (ft) Immunity (ft)

Arc Welding @ 190A DC 15 40

Halogen Lamp (500 W) 2 8

Fluorescent Lamp (25 W) 0 3

Incandescent Lamp (60 W) 1 3

Radiant Heater (1,500 W) 1 3

Sunlight 0 0

Table 3Current vs. Previous Adaptive Performance in False Positive Cases

Nuisance Multiple ANN Single ANN

Sources Immunity (ft) Immunity (ft)

Arc Welding @ 190A DC 15 40

Halogen Lamp (500 W) 2 10

Fluorescent Lamp (25 W) 0 4

Incandescent Lamp (60 W) 1 12

Radiant Heater (1,500 W) 1 10

Sunlight 0 > 0 (fail)

1

Table(s)