IEEE TRANSACTION ON INDUSTRIAL ELECTRONICS, VOL. 47, NO.

5, OCTOBER 2000 1161

A Fuzzy Logic Approach to Fire Detection in Aircraft

Dry Bays and Engine Compartments
Simon Y. Foo, Member, IEEE

Abstract—In this paper, a fuzzy logic approach is applied to de- suppressants being released into the environment. Furthermore,
tect hydrocarbon fires in aircraft dry bays and engine compart- this fire alarm system must operate in realtime as well as min-
ments. The inputs to the fuzzy system consist of a set of statistical imize false alarms due to artificial sources such as flashlights
measures derived from the histogram and image subtraction an-
alyzes of successive image frames. Specifically, fuzzy rules based and halogen lights. These requirements make fire detection in
on the median, standard deviation, and normalized first-order mo- aircraft a difficult task.
ment statistical measures of histogram data and the mean statis- It is well known that frame processes using information from
tical measure of image subtraction data of successive frames are two or more frames can determine whether a fire is present in
used to compute the probability of a fire event. This fuzzy logic an image and the extent of the fire. Frame processes or ma-
approach is also tested for false alarms such as those due to flash-
lights and high-power halogen lights. It is shown that image sub- chine vision approach have been used in many industrial ap-
traction analysis can be used to accurately distinguish fires from plications such as security, quality control, etc. [2]. In security
false alarms. applications, frame processes can be used to detect motion and,
Index Terms—Fire detection, fuzzy logic, machine vision. therefore, intruders [4]–[5]. In a similar manner, the ignition and
growth of a fire can be detected by the subtraction of two se-
quential image frames.
I. INTRODUCTION In our previous work [3], we introduced a rule-based ap-
proach for fire detection using a set of heuristics. Specifically
T HE harsh environment of an aircraft dry bay and engine
compartment makes accurate fire detection and response
a difficult task. Fires on board aircraft may be caused by im-
this approach uses inferential statistics, that is, it draws conclu-
sions from the realtime data using a set of heuristics and statis-
pacts of projectiles, electrical sparks, etc. The combination of tical measures derived from the histogram and image subtrac-
heat, fumes, and oil from the hydraulics, fuel lines, etc., inside tion analyzes. However, the disadvantages of this method are: 1)
an aircraft dry bay can be accounted for most false alarms or the system is too dependent on the ad hoc choice of thresholding
failures of conventional smoke/heat sensors. For example, due constants and 2) the system cannot produce a quantitative mea-
to the heat generated from the running engines and the lack of sure (such as probability) of a fire event in the dry bay or engine
proper ventilation, the temperatures of blackbodies inside an en- compartment. This “fuzzy” feature is an important component
gine compartment can soar to over 200 F, making heat sen- of a decision-support and explanation mechanism.
sors such as infrared sensors impossible to distinguish a normal Fuzziness describes event uncertainty such as “slight chance
operating environment from an actual fire. Other conventional of collision,” and impreciseness of linguistic terms such as
methods for fire detection such as smoke sensors rely on chem- “light rain.” Fuzzy logic fits best in applications where the
ical reactions and therefore have too slow response times. variables are continuous and/or mathematical models do not
Recently, visible spectrum machine vision systems have been exist.
developed to detect and characterize hydrocarbon fires [1]–[2]. This paper presents a fuzzy logic approach for fire detec-
Visible characteristics of fires such as brightness, color, spectral tion. Using statistical measures of the histogram and subtraction
texture, spectral flicker, and stationarity are used to discriminate data of successive frames, fuzzy if–then rules are constructed to
them from other visible stimuli and background. The ability to compute the probability of a fire event. This fuzzy system is
detect and categorize a fire in terms of growth patterns is a very also tested for false alarms such as those due to flashlights and
important feature of a fire alarm system [3]. For example, a con- high-power halogen lights.
stant small flickering flame is less likely to be life-threatening
than a fast-growing fire which may engulf the whole dry bay II. HISTOGRAM ANALYSIS
in a few seconds. Subsequently, a proper fire characterization A histogram can provide an immense wealth of information
would enable appropriate response in terms of the right amount about the image. For example, we can distinguish a dark image
of ozone-depleting halon or other chloroflourocarbon (CFC) fire from a bright image and a low-contrast image from a high-con-
trast image based on the distribution of pixel values. We can also
enhance the details of an object by applying histogram equaliza-
Manuscript received May 16, 1999; revised April 9, 2000. Abstract published
on the Internet July 1, 2000. This work was supported in part by Wright-Pat- tion on the original image.
terson Air Force Base and Synergy, Inc., OH, under Contract S2896. High-speed videos of fires in an aircraft dry bay and engine
The author is with the Department of Electrical Engineering, Florida A&M compartments, courtesy of AFRL/FIVS, Wright-Patterson Air
University–Florida State University, Tallahassee, FL 32310 USA (e-mail:
[email protected]). Force Base, OH, were used as our testbed. The fires were ig-
Publisher Item Identifier S 0278-0046(00)08847-X. nited by the impact of 30-mm rounds and other small arms fire.
0278–0046/00$10.00 © 2000 IEEE
1162 IEEE TRANSACTION ON INDUSTRIAL ELECTRONICS, VOL. 47, NO. 5, OCTOBER 2000

Frame 1 Frame 2 Frame 3 Frame 4

Frame 5 Frame 6 Frame 7 Frame 8

Frame 9 Frame 10 Frame 11

Fig. 1. The initial set of 11 greyscale images used in our investigations.

Using a frame grabber, selected frames of video images are dig- taken at times and , respectively, where and are pixel
itized and stored on a computer. The 353 353-pixel images coordinates in a two-dimensional image, is to compare the two
are stored in 8-bit grey-level TGA format. images pixel by pixel [4]. The result is a difference image. This
Fig. 1 shows the set of 11 image frames used in our investiga- approach is applicable only if the two images are registered
tions. (The number 11 is chosen such that there will be a total of and the illumination is relatively constant, as in our application
ten difference images in the image subtraction analysis, see Sec- where the camera is placed at a fixed position. Suppose we
tion III). It is quite clear that these images, generally, are of low have a reference image containing only stationary components.
contrast and sharp “peaks” will be observed in the distributions Comparing this reference image against a subsequent image
of pixel values. Fig. 2 shows the corresponding histograms. For having the same environment but with motion or growth in one
example, Frame 4 is a very bright image of low contrast, and the or more of the components results in the difference of the two
corresponding histogram shows many pixels in the upper 200 images canceling the stationary components. The end result is
greyscale level. It is these “peaks” that hold information whether a difference image where the nonzero entries correspond to the
there is a fire or not. Note that we do not apply histogram equal- nonstationary image components.
ization to the images as this operation would “smooth out” the The difference image between two images taken at times
peaks that we are interested in. and may be defined as

III. IMAGE SUBTRACTION ANALYSIS FOR GROWTH DETECTION

One of the simplest approaches for detecting changes or if
(1)
motion between two image frames and otherwise
FOO: FIRE DETECTION IN AIRCRAFT DRY BAYS AND ENGINE COMPARTMENTS 1163

Fig. 2. Histograms of the set of 11 successive frames used in our investigations.

where is a predetermined threshold. In dynamic image anal- approach is that much information on the difference image
ysis, all pixels in the difference image with value “1” is retained as opposed to the previous approach.
are considered the result of changes in the scenery. Then the Specifically, this function provides valuable information such
1s in the matrix are tallied and compared to another as the change in intensity value of a pixel over a period of time,
predetermined threshold to determine if the growth warrants a which is an indicator of the rate of growth.
response. In practice, there maybe many isolated 1s in Since time constraint is very critical in our application, we
matrix as a result of noise, vibration, and/or misregistration of suggested sampling every 100th pixel of the grey-scale 353
pixels. A suggested solution to the removal of noise is to form 353-pixel - image, thus reducing the difference image to a
four or eight connected regions of 1s in the and then more manageable 1244-pixel matrix. (Ideally, this reduction
ignore any region that has less than a predetermined number in spatial resolution can be achieved by the method of pixel av-
of entries [4]. The major disadvantage of this approach is that it eraging; however, computation constraint would have excluded
may result in ignoring small and/or slow-moving objects, which this routine). Fig. 3 shows the results of image subtraction for
may be of interest in our application. set of 11 test images. For example, we observe that there is sub-
In our approach, we define the difference image as stantial “motion” in Frame 3 from the large number of nonzero
entries in the resulting matrix of image subtraction between
if Frames 3 and 2.

otherwise IV. STATISTICAL ANALYSIS OF SAMPLE DATA

(2) Statistical analysis is a collection of methods to help describe
some aspect or characteristic of a phenomenon under study, and
This equation has the effect of ignoring pixels changing from to draw conclusions using these descriptors [6]. Statistical mea-
bright to dark, and at the same time measuring the gradient sures such as the mean, median, mode, variance, range, skew-
of pixels changing from dark to bright. The advantage in our ness, and kurtosis give us an insight into the underlying physical
1164 IEEE TRANSACTION ON INDUSTRIAL ELECTRONICS, VOL. 47, NO. 5, OCTOBER 2000

Fig. 3. Results of image subtraction for the set of 11 successive frames.

process and probability distribution from which the data was It can also be observed that the mean represents a measure of
drawn. For example, the sample mean for a data set of size is centrality since it indicates where the center of most of the data
given by is likely to be.
The sample median for a data set is the midpoint of the
data set after it has been ordered by ascending value, from the
smallest to the largest data point. Mathematically, the median
of a random variable is the smallest number such that the
(3) distribution function is 0.5.
FOO: FIRE DETECTION IN AIRCRAFT DRY BAYS AND ENGINE COMPARTMENTS 1165

Fig. 5. Statistical analysis on the image subtraction of the set of successive

frames.

first-order moments are calculated in a manner similar to that

of power spectral density (PSD). Similarly, Fig. 5 shows the
statistical data derived from the image subtractions of Fig. 3.
In our previous approach [3], we introduced a set of heuristics
for drawing conclusions from the statistical data. For example,
consider the following.
1) If the standard deviation std dev of the histogram is high,
(exceeding a predetermined threshold) then there is high
contrast in the image, i.e., a good mixture of dark pixels
and bright pixels. Conclusion: presence of bright spots on
dark background, possibly fire.
2) If the std dev is low (less than a predetermined threshold),
then the image either has mostly dark pixels or mostly
bright pixels, i.e., low contrast. Conclusion: either the
Fig. 4. Statistical analysis of the histograms. (a) Mean, median, and std dev. room is very well lit (saturated), or the room is pitch
(b) Normalized first-order moments. black.
3) If the median of the distribution is high (exceeding a pre-
The sample variance for a data set is given by determined threshold), then there is alot of light in the
image. Conclusion: fire activity likely.
4) If the median of the distribution is low (less than a prede-
termined threshold), then there is little light in the image.
(4) Conclusion: fire activity unlikely.
5) If the sum of the first-order moments of the histogram
and it follows that is the standard deviation for the data set. exceeds a predetermined threshold, then there is alot of
A general form of the estimate of the th moment about the light. Conclusion: fire activity likely.
mean for a data set is given by 6) If the difference between two sequential images is greater
than a predetermined threshold, then there is substantial
motion or growth (of fire).
7) If the mean, median, and std dev statistical measures of
(5)
the difference between two sequential images exceed
If , we have the sample variance. If or , we their respective thresholds, then there is substantial
have the sample estimates associated with skewness and kur- motion or growth (of fire).
tosis, respectively. The skewness and kurtosis sample estimates The histogram and image subtraction analyses are performed
were avoided in this speed-critical application for obvious rea- in parallel to determine if a fire is present. Furthermore, the
sons—they are computationally expensive. image subtraction analysis also determines the growth pattern
Fig. 4 shows the mean, median, standard deviation, and the of fire if it exists. It is this set of rules that leads to the develop-
normalized first-order moments of the histograms in Fig. 2. The ment of a fuzzy logic system for fire detection.
1166 IEEE TRANSACTION ON INDUSTRIAL ELECTRONICS, VOL. 47, NO. 5, OCTOBER 2000

V. FUZZY LOGIC AND FUZZY SETS In order to reduce the complexity of the fuzzy rules, we limit
our input variables to the median, std dev, and first-order mo-
Fuzziness describes event uncertainty and impreciseness of
ments of the histogram analysis, and the mean of the subtrac-
linguistic terms. Specifically, fuzziness measures the degree to
tion analysis. The single output is interpreted as the probability
which an event occurs, not whether it occurs or not, according to
of fire event in a particular frame. Both the antecedent and con-
Kosko [7]. We regularly apply fuzziness in our everyday expres-
sequent membership functions are defined as VLOW, LOW,
sions, such as “good chance of light rain,” “not-so-tall men,”
MED, HIGH, and VHIGH, corresponding to “very low,” “low,”
“cool temperature,” “small fire,” and so forth.
“medium,” “high,” and “very high” linguistic terms, respec-
Fuzzy logic has rapidly become one of the most successful
tively.
technologies for developing sophisticated control systems and
Figs. 6 and 7 show the antecedent membership functions for
decision-support expert systems, as shown in [7]–[19]. Fuzzy
the four input variables from the histogram and subtraction sta-
logic fits best in applications where the variables are continuous
tistical data, respectively. Both triangular and trapezoidal mem-
and/or mathematical models are difficult to define or they do
bership functions are used. For example, in the histogram-me-
not exist. In fuzzy logic systems, the imprecision of the input
dian, the linguistic term LOW lies somewhere between 0–100,
and output variables are encoded in “fuzzy” representations, or
whereby a value 50 shows a 1.0 degree of membership.
fuzzy sets, that can be expressed in linguistic terms (e.g., cool,
In the first diagram of Fig. 6, although the maximum reported
warm, and hot).
value from the histogram median is about 190, the universe of
In fuzzy sets, introduced by Zadeh [8]–[10], the imprecise-
discourse is defined up to slightly more than 250 due to the fact
ness of linguistic terms are represented by the degree of mem-
the the grey levels of each pixel ranges from 0 to 255 (totally sat-
bership to a set indicated by a number in the interval [0, 1]. The
urated). Similarly, in the second diagram of Fig. 6, the reported
universe of discourse is the range of values where the fuzzy set
input (std dev of the histogram) ranges from about 40 to about
is defined. The membership of a number to a fuzzy set is
60. Subsequently, a universe of discourse from 0 to slightly more
represented by
than 80 is chosen to accommodate this range. Note that, in the
(6) third diagram of Fig. 6, the membership function “High” has a
slightly different shape from the rest of the membership func-
i.e., the membership function maps every element of tions due to the fact that it has been manually “optimized.” In
the universe of discourse to the interval [0, 1]. Fuzzy mem- general, most of the membership functions have shapes that are
bership functions can have different shapes depending on the chosen ad hoc. More work could have been done on using adap-
applications and the user’s preference. Triangular and trape- tive methods to tune the membership functions but this would
zoidal shapes help simplify computations, therefore, are com- have been outside the scope of this paper.
monly used. Other membership functions include -shaped and Fig. 8 shows the consequent membership functions. Again,
-shaped membership functions. the membership functions in this diagram are not the same as
those in previous diagrams due to the fact that they have been
VI. FUZZY ALGORITHM FOR FIRE DETECTION manually tuned.
In fuzzy systems, fuzzy linguistic descriptions using A total of 300 [i.e., ] rules are constructed to account
if–then–else rules are used to represent the system’s behavior. for all possible combinations of input variables. An example
For example, fuzzy rules associate output fuzzy sets with rule is as follows:
input fuzzy sets of input variables.
The basic fuzzy algorithm, by Tsoukalas et al. [18], con- histogram median is histogram std dev is
sists of multivariate fuzzy if–then rules, each having fuzzy histogram moment is subtraction mean is
input variables on the antecedent side, taking fuzzy
then fire probability is and so forth.
values and one consequent
(output) variable , taking fuzzy values
Given a set of statistical measures (inputs) from the histogram
and subtraction analyses, the crisp output of the fuzzy system is
computed as follows.
First, the inputs are fuzzified based on the membership func-
tions shown in Figs. 6 and 7. Then, the Larsen [15] product im-
plication operator is applied on the fuzzy rules, i.e.,

(7)

In the Larsen implication operator, the fuzzy connective “and”

Specifically, given the fuzzy values for inputs , and corresponds to the “product” of the fuzzy sets. For example, in
the implication relation of the individual rules, and the appro- the fuzzy rule
priate interpretation of the connectives “and” and “else,” gener-
alized modus ponens is used to obtain the output .
FOO: FIRE DETECTION IN AIRCRAFT DRY BAYS AND ENGINE COMPARTMENTS 1167

(a) (b)

(c)
Fig. 6. Histogram-median, histogram-std dev, and normalized 1st-order moments membership functions.

Fig. 7. Subtraction-mean membership functions. Fig. 8. Output fire probability membership functions.

Subsequently, we interpret the “else” connectives as the OR

the membership function of the resulting Larsen product impli- ( ) of all the implication operators , i.e.,
cation relation becomes
DOF DOF
DOF (9)
(8) where we choose the product form of the degree of fulfillment
(DOF) on the antecedent, i.e.,
where is the implication operator. DOF (10)
1168 IEEE TRANSACTION ON INDUSTRIAL ELECTRONICS, VOL. 47, NO. 5, OCTOBER 2000

Fig. 9. Results of applying the fuzzy rules on Frame 1. (a) Consequent membership functions after the Larsen product implication operator. (b) Aggregation of
the consequent fuzzy sets using the “max” operator. (c) Determination of the crisp output using the centroid method.

TABLE I
CRISP OUTPUTS OF THE 11 FRAMES

the geometrical center of the output fuzzy value , where

Fig. 10. Crisp outputs of the fuzzy system based on the set of 11 frames. is formed by taking the union of all the contributions of
rules whose DOF . The centroid of the aggregated out-
and (the number of fuzzy input variables); puts is defined as
(the number of fuzzy if–then rules).
The degree of fulfillment is a measure of the degree of simi-
larity between input and the antecedent of the rule . Before (11)
we defuzzify the output, we also apply the “max” operator to
aggregate the consequent fuzzy sets.
Finally, the centroid or the center of area (COA) defuzzifi-
cation method is used to compute the crisp output of the fuzzy where the summation is carried over discrete values of the uni-
system. In the COA defuzzification, the crisp output value is verse of discourse sampled at points.
FOO: FIRE DETECTION IN AIRCRAFT DRY BAYS AND ENGINE COMPARTMENTS 1169

Frame 1 Frame 2 Frame 3 Frame 4

Frame 5 Frame 6 Frame 7 Frame 8

Fig. 11. Set of additional frames of images for analysis. Frames 1–2: dry bay; frames 3–4: aircraft engine nacelle; frame 5: flashlight; frame 6: 400-W halogen
light; fame 7: 1000-W halogen light; and frame 8: 2000-W halogen light.

VII. COMPUTER ANALYSIS TABLE II

TEST PARAMETERS FOR DISCRIMINATING FIRES FROM ARTIFICIAL
The histogram and image subtraction analyses, along with LIGHT SOURCES
the fuzzy system were implemented using Matlab [20]. In the
fuzzy rules, we chose triangular and trapezoidal-shaped mem-
bership functions for the variables to simplify our computations.
The membership functions are manually tuned by trial and error
to optimize the system’s performance. With sufficient training
data, automatic adaptation of the membership functions is pos-
sible. For example, a gradient descent algorithm can be used
to optimize the antecedent and consequent membership func-
tions. However, this operation would be outside the scope of
this paper. (Future extensions to this paper could include adap-
tive tuning of the fuzzy membership functions).
fires from artificial light sources such as flashlights and halogen
We first applied our fuzzy system on the set of 11 frames.
lights.
Fig. 9 shows the results of applying the fuzzy rules on Frame 1,
Constants are as follows:
where the inputs are the histogram median, std dev, and normal-
ized first-order moments, and the subtraction mean. Fig. 9(a) 1) weather was clear, dry, and at approximately 80 F;
shows the consequent membership functions after the Larsen 2) tests conducted at night;
product implication operation, while Fig. 9(b) shows the ag- 3) camera is positioned at about 76 in from the subject;
gregation of the consequent fuzzy sets using the max operator. 4) camera set at 500 fps shuttered @ 2500.
Finally, the crisp output value of 13.95 is calculated using the Fig. 12 shows the histogram statistical analyses of the test data.
centroid method, as shown in Fig. 9(b). Hence, we interpret the (The image subtractions were deliberately omitted at this time
result as “there is a 13.95% chance of a fire” in Frame 1. to observe the effects of the histograms by themselves). The
The crisp outputs for the entire set of 11 frames are listed in selected inputs were histogram std dev and normalized first-
Table I and plotted in Fig. 10. order moments. The crisp outputs of the fuzzy system for the
From the crisp output values in Table I, we can deduce that miscellaneous frames are listed in Table III.
there is a very low probability (15% or less) of fire in Frames 1, As observed in Table III, we can conclude that there is a very
2, 8, and 11. There is a medium (or good) probability (from 30% low probability of fire in Frames 5, 6, and 7; a good probability
to 70%) of fire in Frames 3, 5, 6, 7, 9, and 10, while there is a of fire in Frames 1 and 8; and very high probability of fire in
very high probability (90% and above) of fire in Frame 4. These Frames 2, 3, and 4. From the visual inspection of the images in
conclusions agree with the visual inspection of the 11 frames in Fig. 11, we noted that the fuzzy system has mistakenly identi-
Fig. 1. fied the false alarms of Frames 1, 3, and 8 as being good or very
We have also tested our fuzzy system using miscellaneous high probability of fire. Frames 1, 3, and 8 can be classified as
frame sequences such as those shown in Fig. 11. Table II low contrasts, with Frame 1 consisting of mostly dark pixels,
shows the test parameters used in discriminating hydrocarbon while Frames 3 and 8 consist of mostly bright pixels.
1170 IEEE TRANSACTION ON INDUSTRIAL ELECTRONICS, VOL. 47, NO. 5, OCTOBER 2000

TABLE IV
CRISP OUTPUTS OF THE SUCCESSIVE FRAMES BASED ON A COMBINATION OF
HISTOGRAM AND IMAGE SUBTRACTION ANALYSES

We observed that the fuzzy system is able to correctly dis-

criminate real fires from false alarms such as flashlights and
halogen lights, as shown in Table IV. The addition of image
subtraction data in our fuzzy system analysis have correctly re-
vealed that there is a very low probability of fire in Frames
1 (dry bay), 5 (engine nacelle), and 7–13 (flashlight, 400-W,
1000-W, and 2000-W halogen lights). We conclude that, while
histograms are able to identify the real fires, histograms alone
are not sufficient to correctly discriminate the real fires from
the false alarms. The reason is that fire is chaotic and, therefore,
exhibits motion, while artificial sources such as flashlights and
halogen lights are not.

VIII. CONCLUSION
We have presented a fuzzy logic approach to fire detection
in aircraft dry bays and engine compartments. The inputs to
the fuzzy system consist of a set of statistical measures de-
rived from the histogram and image subtraction analyses of suc-
cessive image frames. The crisp output of the fuzzy system
yields a probability of a fire event. This fuzzy logic approach
has been tested for false alarms such as those due to flashlights
and high-power halogen lights. Results showed that a combi-
nation of histogram and image subtraction statistical data are
needed for the fuzzy system to correctly discriminate the real
fires from false alarms.
Fig. 12. Statistical analysis of the histograms of dry bay, engine nacelle,
flashlight, and halogen lamps. ACKNOWLEDGMENT
The author would like to express his gratitude to D. Collins
TABLE III for his invaluable input, J. Hodges of Wright-Patterson AFB for
CRISP OUTPUTS OF THE MISCELLANEOUS FRAMES BASED ON
HISTOGRAM ANALYSIS ALONE his kind support of this research, and Lt. Poole and L. Truitt for
providing the high-speed video.

REFERENCES
[1] D. Goedeke, “Machine vision fire detector for millisecond response,”
Photon. Spectra, pp. 18–20, Oct. 1993.
[2] , “Payday for machine vision,” Photon. Spectra, pp. 76–84, Mar.
1994.
[3] S. Y. Foo, “A rule-based machine vision system for fire detection in
aircraft dry bays and engine compartments,” Knowl.-Based Syst., vol.
9, pp. 531–540, 1996.
[4] R. C. Gonzalez and R. E. Woods, Digital Image Processing. Reading,
MA: Addison-Wesley, 1992.
We then tested our fuzzy system using a combination of his- [5] A. K. Jain, Fundamentals of Digital Image Processing. Englewood
togram std dev, histogram normalized first-order moments, and Cliffs, NJ: Prentice-Hall, 1989.
the subtraction mean of successive frame sequences of the dry [6] J. W. Barnes, Statistical Analysis for Engineers and Scientists. New
York: McGraw-Hill, 1994.
bay, engine nacelle, flashlight, and the halogen lights. The crisp [7] B. Kosko, Neural Networks and Fuzzy Systems: A Dynamical Systems
outputs of the fuzzy system are listed in Table IV. Approach. Englewood Cliffs, NJ: Prentice-Hall, 1991.
FOO: FIRE DETECTION IN AIRCRAFT DRY BAYS AND ENGINE COMPARTMENTS 1171

[8] L. A. Zadeh, “Fuzzy sets,” Information Contr., vol. 8, pp. 338–353, [19] E. Cox, “Fuzzy fundamentals,” IEEE Spectrum, vol. 1, pp. 58–61, 1992.
1965. [20] Matlab User’s Guide, The MathWorks, Inc., Natick, MA, Sept. 1993.
[9] L. A. Zadeh, “Outline of a new approach to the analysis of complex
systems and decision processes,” IEEE Trans. Syst., Man, Cybern., vol.
1, pp. 28–44, 1973.
[10] L. A. Zadeh, “Fuzzy logic,” IEEE Computer, vol. 1, pp. 83–93, Apr.
1988.
[11] E. H. Mamdani and S. Assilian, “An experiment in linguistic synthesis
with a fuzzy logic controller,” Int. J. Man–Mach. Stud., vol. 7, no. 1, pp.
1–13, 1975.
[12] M. Sugeno and G. T. Kang, “Structure identification of fuzzy model,” Simon Y. Foo (M’90) received the B.S., M.S., and
Fuzzy Sets Syst., vol. 28, pp. 15–33, 1988. Ph.D. degrees in electrical engineering from the Uni-
[13] T. Takagi and M. Sugeno, “Fuzzy identification of systems and its ap- versity of South Carolina, Columbia, in 1983, 1985,
plications to modeling and control,” IEEE Trans. Syst., Man, Cybern., and 1988, respectively.
vol. 15, pp. 116–132, 1985. He is currently an Associate Professor in the
[14] A. Kandel, Fuzzy Mathematical Techniques with Applica- Department of Electrical Engineering, Florida A&M
tions. Reading, MA: Addison-Wesley, 1986. University–Florida State University, Tallahassee. He
[15] P. M. Larsen, “Industrial applications of fuzzy logic control,” Int. J. specializes in the design and test of CMOS analog
Man-Mach. Stud., vol. 12, no. 1, pp. 3–10, 1980. and digital integrated circuits. He has more than 15
[16] C. C. Lee, “Fuzzy logic in control systems: Fuzzy logic controller—Part years of experience using MOSIS for IC fabrication.
1,” IEEE Trans. Syst., Man, Cybern., vol. 20, pp. 404–418, Mar./Apr. His most recent research activities include fuzzy
1990. logic, neural networks, genetic algorithms, machine vision, computer intrusion
[17] , “Fuzzy logic in control systems: Fuzzy logic controller—Part 2,” detection, mobile agents, and random number generators. He has authored
IEEE Trans. Syst., Man, Cybern., vol. 20, pp. 419–435, Mar./Apr. 1990. more than 30 technical papers and contributed to two books on neural networks.
[18] L. H. Tsoukalas, M. Ishii, and Y. Mi, “A neurofuzzy methodology for Dr. Foo has served as the Faculty Advisor to the IEEE Student Chapter at
impedance-based multiphase flow identification,” Eng. Applicat. Artif. Florida A&M University–Florida State University. He received the Teaching
Intell., vol. 10, no. 6, pp. 545–555, 1997. Incentive Program (TIP) Award in 1995. He is a member of Eta Kappa Nu.