Arid Zone Journal of Engineering, Technology and Environment, June, 2017; Vol.

13(3):356-368
Copyright Faculty of Engineering, University of Maiduguri, Maiduguri, Nigeria.
Print ISSN: 1596-2490, Electronic ISSN: 2545-5818, www.azojete.com.ng

DEVELOPMENT OF A MODEL FIRE DETECTION ALGORITHM

O. S. Ismail1 and C.I. Chukwuemeka2

(1
Department of Mechanical Engineering, Faculty of Technology, University of Ibadan, Nigeria
2
Department of Mechanical Engineering, Faculty of Technology, University of Ibadan, Nigeria)
*Corresponding authors email: os.ismail@ui.edu.ng

Abstract
An adaptive model algorithm for fire (flickering flame, in the infrared region) detection is presented. The
model made use of a Pyro-electric infrared sensor (PIR)/Passive Infrared Detector (PID) for infrared fire
detection. Sample analog signals (around flame flicker region) were generated and simulated within the
framework of the modeled PIR sensor/PID. A Joint Time Frequency Analysis (JTFA) function, specifically
the Discrete Wavelet Transform (DWT) was applied to model the Digital Signal Processing (DSP) of the
generated signals. It was implemented as wavelet analyzer filters for fire and non-fire feature
extraction. A Piecewise Modified Artificial Neural Network (PMANN) and the Intraclass Correlation
Coefficient (ICC) were employed in the decision rule. The PMANN generated polynomials which analyzed
and memorized the signals from DSP. The ICC further categorized cases as 'fire' or 'non-fire' by
comparing data from the PMANN analyzed. The results show that the model algorithm can be implemented
in modern day fire detectors and be used to detect industrial hydrocarbon fires with fewer false alarms than
smoke detectors or ultraviolet detectors.

Keywords: flickering flame detection, passive infrared detector, wavelet filters, signal processing
and artificial neural networks.

1. Introduction
The study of fire and its consequences in industries occupies a special place in the world of science
and engineering because in most cases it results in an unwanted outcome and therefore must be
avoided as much as possible. Therefore, the early detection of hydrocarbon flames in industrial
facilities serves as a means of avoiding such unwanted outcomes. But detection systems must
reliably carry the burden of fast detection and response in system operation, hence the need to
develop as many as possible fast and reliable detection algorithms. Most industrial flames have
hydrocarbons as their main fuel. These flames emit radiations in the optical spectrum, hence most
optical flame detection algorithms are designed based on the hydrocarbon flame spectrum.

Figure 1: Typical Hydrocarbon Flame Emission Spectrum (2-35Hz Characteristic Flickering)

From the spectrum (figure 1), we can have six optical flame detection systems (Nolan, 1996), but
infrared sensors and detectors are the most widely used: Ultraviolet (UV), single frequency

356
Ismail and Chukwuemeka: Development of a Model Fire Detection Algorithm; AZOJETE, 13(3):356-368
ISSN 1596-2490; e-ISSN 2545-5818, www.azojete.com.ng

infrared (IR), dual frequency infrared (IR/IR), ultraviolet/infrared (UV/IR)-simple voting,

ultraviolet/infrared (UV/IR)-ratio measurement and multiband (IR3). However, this work focused
on the infrared (IR) region, since it covers most of the spectrum and covers the CO2 spike region
(the highest constituent of hydrocarbon flames). Although most algorithms for flame detection are
simple in appearance, they become complex when dealing with infrared sources from real
industrial environments. This implies that a good detection algorithm based on IR sensing must be
able to distinguish flames from sources of environmental nuisance and ambient noise (which
includes random motion, modulation of heated surfaces, hot air flow, arc welding, sunlight
reflecting off a water surface, and other non-flame related environmental nuisance).

The use of sensors and neural networks in the detection of fire can be traced to the work carried
out by Okayama (1991). He used odor sensors together with a back propagation neural network to
distinguish smoldering fires from environmental noise such as coffee powder and perfume. In
another subsequent work (Okayama et al, 1991), inputs from carbon monoxide, temperature and
smoke sensors were used to estimate the probability of fire using the back propagation neural
network algorithm. Using the Fourier Transform based IR spectroscopy (FT-IR) of gas phase
products, flaming and smoldering fires were distinguished by an ANN model (Chen et al, 1999).
Their ANN model was formulated using the Linear Vector Quantization (LVQ) approach with a
hidden Kohonen layer. Using the NeuralWorks software, the implementation of their training
model produced a 96% success rate in classification and was extensive in analyzing the
contribution of various gas types to flaming and smoldering fires. Furthermore, a multi-criteria
detection algorithm was developed using a probabilistic neural network (PNN) (Susan et al, 2000).
A large database generated from 20 sensors recording several types of fire and nuisance cases was
analyzed using multivariate methods. Datasets were developed at discrete times corresponding to
the different alarm levels of a conventional photoelectric smoke detector. The datasets were
organized into three classes representing the sensor responses for baseline (non-fire), fires and
nuisance sources. A robust data analysis strategy for use with a sensor array consisting of four to
five sensors for early fire detection and nuisance source rejection was developed using a
probabilistic neural network (PNN). Also, several Advanced Very High Resolution Radiometer
(AVHRR) fire detection algorithms were reviewed by Li Z. et al (2000), with the aim of studying
their principles of operation, limitations, and possible recommendations for improvement.

Other algorithms include the statistical analysis of the apparent source of the heat of fires (Zhu et
al, 2008) at a near infrared zone. After tests and experimentations, it was found out that the
detector functioned well for open flames with very low false alarms while smoldering fires were
hardly detected since there was no direct radiation to the detector, they were only detected when
they had direct radiation. Also an adaptive method for hydrocarbon flame detection was developed
using a Joint Time Frequency Analysis (JTFA) function, namely the Short Time Fourier
Transform (STFT) with the Hamming Window function and the Fast Fourier Transform (FFT) for
fire features extraction from IR sensors and an Artificial Neural Network (ANN) for training and
classification as the decision mechanism (Javid et al, 2008). The ANN was trained using the back
propagation conjugate method (CG). Also, using the Markov Model decision algorithm and
Lagrange wavelet filter banks to extract fire features from signals recorded by pyro-electric
infrared sensors. Fatih et al (2012) modelled a fire detector which can easily detect fire within the
flickering flame frequency. Applying another JTFA function for signal processing, and using the
357
Arid Zone Journal of Engineering, Technology and Environment, June, 2017; Vol. 13(3):356-368
ISSN 1596-2490; e-ISSN 2545-5818; www.azojete.com.ng

PMANN and ICC for decision rule, a new algorithm is proposed in this work. It was simulated
using the MATLAB/SIMULINK software. Figure 2 explains the methodology.

Figure 2: Algorithm Flow Chart

2. Modeling Methodology
2.1 The Pyro-Electric Infrared Sensor (PIR)/Passive Infrared Detector (PID)
A PIR sensor/PID is comprised of the Fresnel lens (which focuses the IR radiation to the sensor),
the PIR sensor which senses the IR radiation and an amplifier/comparator or amplifier/analog to
digital converter (ADC) circuitry depending on the generation of the PID (Emin, 2009). This work
simulated the third generation PID, where the comparator circuitry is replaced by analog to digital
converter (ADC). The ADC gains are then fed into a microprocessor containing the detection
algorithm for further signal processing and decision mechanism (when the model is to be
implemented).
Such a sensor/detector can be modeled as a capacitor with capacitance Cd with a Polyvinylidene
Fluoride (PVDF) film as the dielectric with thickness d and surface area A (Odon, 2010). When IR
radiation of power (t) varying in time is incident on the active surface of the PIR sensor, an
electric charge q(t) is generated. This is transferred as signals with information content either as
voltage V(t) on the detector electrodes or current Ip(t) flowing through the low load resistance of
the detector output. Converting IR radiation into an electric signal is done in 3 stages: converting
radiation power (t) to thermal change on the sensor surface i.e. temperature T(t), the second stage
is the thermal to electric conversion i.e. T(t) to Ip(t), and the last stage is the current to voltage
signal conversion i.e. Ip(t) to V(t). The schematic diagram is given below. The PID detects
infrared radiations from several sources within its range or field of view.

Figure 3: Schematic Diagram for Conversion of IR radiation to Voltage Signal

From figure 3, we have three blocks which represents the earlier mentioned stages. Below are the
model equations for each block (Odon, 2010).
For IR radiation to thermal conversion
358
Ismail and Chukwuemeka: Development of a Model Fire Detection Algorithm; AZOJETE, 13(3):356-368
ISSN 1596-2490; e-ISSN 2545-5818, www.azojete.com.ng

(1)

For thermal to electrical conversion

(2)

For electrical to voltage conversion

(3)

where Cth-thermal capacity of PIR sensor/PID, Gth-thermal conductance of PIR sensor/PID, -

absorption coefficient of radiation, p-pyro-electric coefficient, C-equivalent capacitance (Cd+CL),
R-equivalent resistance (RdRL/(Rd+RL)), Cd & Rd capacitance and resistance of PIR, CL & RL
capacitance and resistance of amplifier.
Using Laplace transfer functions (for the purpose of simulation), the above equations can be
expressed as equation (4), (5) and (6) (Odon, 2010):
For IR radiation to thermal conversion

(4)

For thermal to electrical conversion

(5)

For electrical to voltage conversion

(6)

where is thermal constant, A is the active area and c volume specific heat. and
. e is called the electrical time constant of the equivalent circuit of PIR sensor and
amplifier (e= RC).

From automatic control theory, the procedures for the creation of block diagrams for simulation
involves connecting the block transfer function in series, where the series connection implies
multiplication. For our model, we arranged them in the order GT(s), GTIp(s) and GIpV(s), describing
properties of the appropriate signal conversion stage. Hence the equivalent transfer function is
expressed as: G(s) = GT(s)GTip(s)GIpV(s), giving us the expression in equation (7).

(7)

Equation (7) is the model equation of the PIR sensor/PID. This model was simulated on the
MATLAB/SIMULINK software using standard values for PVDF IR sensor as shown in figure 4.
The IR radiation was represented as an analog sinusoidal wave at a radiation power of =1W. But
before simulation, it was needful to know the best sensor to be used depending on the thickness of
the PVDF film and the thermal time constant. Using Bode plots on MATLAB for three randomly
picked thickness of the PVDF film (d1= 25m, d2= 50m d3= 100m) and three different values of
thermal time constant (th1=0.011, th2=0.022, th3=0.033), it was found out that sensors with
359
Arid Zone Journal of Engineering, Technology and Environment, June, 2017; Vol. 13(3):356-368
ISSN 1596-2490; e-ISSN 2545-5818; www.azojete.com.ng

smaller film thickness had higher voltage responsivity, and the smaller the thermal time constant,
the faster the sensor and also the narrower the frequency range (close to fires) it covered. For the
MATLAB Bode plots, the transfer function G(s) was transformed to form polynomials in
numerator and denominator as expressed in equation (8) (Odon, 2010):

(8)

From the analysis of the MATLAB Bode plots, the PIR sensor/PID with a smaller film thickness
and thermal time constant was used in the simulation. Table 1 shows the parameters for the PVDF
PIR sensor/PID. Random sources of IR radiation at 2Hz, 50Hz and the highest flame flicker
frequency 13Hz (representing the fire) were simulated on MATLAB/SIMULINK and their
corresponding coefficients representing voltage responsitivity fed into the detection algorithm.

Figure 4: MATLAB/SIMULINK block for sample flickering flame signal (13Hz)

Table 1: Parameters for the PVDF PIR sensor/PID

Name Value Unit
Pyro-electric coefficient p 310-6 C/cm2.K
Volume specific heat c 2.4 J/cm3.K
Permittivity 10610-14 F/cm
PIR film thickness d 25 Mm
Thermal conductivity gth 0.00135 W/cmK
Detector active surface A 13210-2 Cm2
PIR detector Capacitance Cd 560 pF
Amplifier input capacitance CL Negligible pF
Amplifier input resistance RL 10 M
Absorption coefficient of radiation 1 -
Electrical time constant of detector- 0.0056 S
amplifier circuit e
Thermal time constant th 0.0110 S

2.2 The Detection Algorithm

a. Digital Signal Processing
The digital signal processing algorithm was developed using the Discrete Wavelet Transform
(DWT) JTFA function. The transform is implemented in real-time as wavelet filters (Schneiders,

360
Ismail and Chukwuemeka: Development of a Model Fire Detection Algorithm; AZOJETE, 13(3):356-368
ISSN 1596-2490; e-ISSN 2545-5818, www.azojete.com.ng

2001). As opposed to the Fourier Transform (FT) or Short Time Fourier Transform (STFT), the
DWT is better in that it makes multi-resolution analysis possible.
The signals are first passed through a window function (the Hamming window) to attenuate the
input signal, thereby reducing spectral leakage and forcing the signal to be more periodic. A
window length of 256 was chosen. This is advantageous since it reduces the response time of the
detection mechanism by two. The window function is expressed as equation (9) (Robert and
Sandra, 2012),
{ ( )} (9)
where N is the window length and n = 1, 2, 3, N. With a 256 window length, a Nyquist
sampling frequency of 50Hz was chosen. The flickering flame frequency of 13Hz (Fatih et al.,
2012) was used as our cutoff frequency. The Nyquist sampling frequency is determined from fs
2(cutoff frequency). Applying the formula, we get 26Hz as our sampling frequency. But a 50Hz
sampling frequency was chosen to widen the frequency range in order to obtain a better sampled
signal.
For real signals, only half of the number of samples (the same as the window length) contains
essential information without redundancy. Hence, using Rayleighs Limit (Robert and Sandra,
2012) our frequency resolution is expressed as equation (10), which becomes equation (11) for our
signals:
(10)
(11)
Real-time wavelet filters is defined by their Quadrature Mirror Filters (QMF) used for DWT.
QMFs are perfect reconstruction filter banks where the calculation of coefficients for the filters at
all levels is seen as making an orthonormal (orthogonal and normalized) projection onto a new
base. These filters banks contain highpass (H) and lowpass (L) filters of length N+1 (Schneiders,
2001), where N is the number of points or filter order which is the same as window length.
Depending on the number of levels denoted as P, we apply the analysis filter to the input signal to
calculate coefficients for the first level. Using N old input points, we calculate the coefficient for
the next level. Hence the total number of old input samples needed for computation of a new
coefficient at a certain decomposition level is defined by equation (12) (Schneiders, 2001):
(12)
where i is the decomposition level and varying between 1 and the maximum P and N the order of
the original filters.

Figure 5 Sample analysis filter bank (2-level)

361
Arid Zone Journal of Engineering, Technology and Environment, June, 2017; Vol. 13(3):356-368
ISSN 1596-2490; e-ISSN 2545-5818; www.azojete.com.ng

From this point, we got new coefficients as the output at each level. This way made on-line
analysis and feature extraction possible with the shortest time delay for each decomposition level.
Such structure was built on the MATLAB software. The code produces a filter matrix A, which
can be implemented as discrete (Finite Impulse Response) FIR filter block on
MATLAB/SIMULINK. Since all coefficients are updated at every sample hit, the time resolution
is increased. Hence, for filter structure as a wavelet analyzer the time duration is expressed as
equation (8) (Schneiders, 2001):
(13)
The time-resolution is equal to the sample time of the system. For DWT perfect reconstruction of
the multi-resolution tree, the frequency resolution (f) is a function of the decomposition level P
expressed as equation (14).
(14)
Using the values for sampling frequency and frequency resolution, we obtained our time resolution
and decomposition level to be expressed as in equations (15) and (16):
(15)

(16)
Using N=256 points, we calculated our filter lengths to be:
Applying the number of points N, filter lengths for the highpass (H) and lowpass (L) filters and
the decomposition level P to our MATLAB code, gave us the needed coefficients for Finite
Impulse Response (FIR) filter block on MATLAB/SIMULINK software, as shown in figure 6.
The coefficients obtained are given in the matrix A.
A = {0.00, 2.82, 0.00, 0.00, 0.00, 0.00}.

Figure 6: MATLAB SIMULINK block diagram for Signal Processing Algorithm

b. The Decision Framework

The decision framework is in two parts. First the PMANN, which analyses and memorizes the
coefficients from the DSP, then the ICC which matches the results from the PMANN for reliability
and resemblance to fire signals. The use of neurons for ANN is very difficult and complex. Hence,
in this model neurons were not used, instead the algorithm made use of polynomial
362
Ismail and Chukwuemeka: Development of a Model Fire Detection Algorithm; AZOJETE, 13(3):356-368
ISSN 1596-2490; e-ISSN 2545-5818, www.azojete.com.ng

approximations or the least square approximation method, whereby polynomial equations were
generated to establish a link between the input and the output (Chukwuka, et al, 2014). For this
algorithm, the input parameters were distributed into the network, rather than lumping them into
the network. Lumping the inputs into the network creates lot of errors while distributing them
reduces the errors. Four distributions were made, which generated four equations resulting in a
Piecewise Modified ANN (PMANN). Figure 7 is the diagram of the algorithm flow chat.

(a) (b)
Figure 7: (a) Flow Chart for proposed PMANN, (b) Inputs and Output of the Neural Network
The polynomial generated is expressed as equation (17) (Chukwuka, et al 2014):
(17)
where k is the number of inputs being considered. For the purpose of simulation k=3. But the
network is trained with the model flame flicker frequency (13Hz) signal, so that it would be able to
differentiate fire cases from non-fire during simulation. Let Wi be the input weights and Sfi their
scaling factor, and ai, bi, ci and di be the coefficients of each input considered. The coefficients are
expressed as in Equation (18) (Chukwuka, et al 2014):

(18)
For our model and simulation, the weight was set to be Wi = 1.0 and scaling factor to be Sfi=0.01.
This simplified the scaling process, making the equations easily evaluated within the framework of
the network. The coefficients obtained could be matched for fire and non-fire cases using
simple statistical correlations like the ICC. The algorithm was designed to receive inputs, sort
them out, adjust their parameters and compute the expected result. The polynomial equations
obtained from the above analysis were logged into a MATLAB m-file. The ensuing program ran
for different data sheets containing signal coefficients from the DSP. Using the Intraclass
Correlation Coefficient (ICC), data from the PMANN analysis are differentiated into fire and
non-fire cases. The ICC is used to quantify the degree to which measurements with a fixed
degree of relatedness match each other in terms of quantitative trait. Also this statistical analysis
can be applied to assess the consistency (or agreement) of quantitative measurements made by
different observers measuring the same quantity. All these are classified as reliability analysis.
363
Arid Zone Journal of Engineering, Technology and Environment, June, 2017; Vol. 13(3):356-368
ISSN 1596-2490; e-ISSN 2545-5818; www.azojete.com.ng

Hence, this method was applied to analyze the data gotten from PMANN. The coefficients
obtained from the data of sample signals of frequency 2Hz, 50Hz and 13Hz (fire signal) were
matched against those of the 13Hz (fire signal) used for training. For a perfect match, the scenario
was recorded as fire otherwise it was recorded as non-fire. This analysis was carried out on
SPSS 16.0 software. The results obtained were for the class 2 or two-way random single and
average measures (consistency/absolute agreement) ICC with a 95% confidence interval. Here, the
measurement raters are chosen at random. The reliability of the analysis is interpreted between the
lower and upper bound of the confidence interval (Figure 8).

Figure 8: SPSS ICC Reliability Analysis

3.0 Results and Discussion
For the selection of the model PIR sensor/PID, we have the following MATLAB bode diagrams.

(a) (b)
Figure 9: (a)Voltage Responsivity (Rv) versus Piezo (PVDF) film thickness (b) Voltage
Responsivity (Rv) versus thermal Time Constant
From the figure 9 it can be deduced that voltage responsivity of the PIR sensor/PID increases with
decrease in PVDF thickness. The voltage responsivity of the sensor increases with decrease
thermal time constant. From both plots, it can be deduced that the best sensor to be used for our
model is an infrared sensor with a thin PVDF film and a small thermal constant.

Figures 10: (a), (b) and (c) were obtained after infrared signals modeled at the flickering frequency
(13Hz), and signals other than the flickering fire frequency, i.e. 2Hz and 50Hz were simulated
using the MATLAB/SIMULINK block for the PIR sensor/PID.
364
Ismail and Chukwuemeka: Development of a Model Fire Detection Algorithm; AZOJETE, 13(3):356-368
ISSN 1596-2490; e-ISSN 2545-5818, www.azojete.com.ng

a b c
Figure 10 (a) PIR Sensor/PID Output for fire at 13Hz, (b) PIR Sensor/PID Output for radiation at
2Hz (c) PIR Sensor/PID Output for radiation at 50Hz

From figure 10, the difference among the various radiations can easily be seen. A step function,
sequence interpolator and pulse generator (Figure 4) were added together to generate the IR analog
signals. This accounted for the oscillatory nature of the graphs. The fire radiation produced higher
values (from -3 and 3 V) for voltage responsivity. If the model is to be implemented, the voltages
are fed into a microcontroller with the Digital Signal Processing algorithm (DSP). Therefore, a
similar case was achieved when the voltages were fed into the MATLAB/SIMULINK interface
containing the corresponding DSP blocks discussed earlier (figure 6). Below are the results of the
signal processing for a simulation time of 10s.

a b c
Figure 11: (a) Radiation at 2Hz after DSP, (b) Radiation at 50Hz after DSP, (c) Fire Radiation after DSP.

For a hamming window length of N = 256, and wavelet filter implemented on SIMULINK as a
discrete Finite Impulse Response (FIR) filter block, the 256 coefficients (or samples) considered
by the window function were filtered and the recorded coefficients narrowed down to 52 samples.
This further fine tunes the detection process. This can be seen from Figure 15, 16 and 17,
respectively, for the results of the signal processing for a simulation time of 10 seconds. After
DSP, each signal produces a set of coefficients which are fed into the PMANN for the decision
mechanism. Out of the 52 coefficients produced by the DSP algorithm, 18 were used in the
training of the network (these contained recorded information, i.e. non-zero coefficients as seen
from figures 11 (a), (b) and (c)). The other 34 coefficients could be rounded up to zero; hence, they
were not needed for the training. The DSP output from the fire signal (13Hz) is our measured and
expected output; therefore, it was used to train our network. After training and classification, the
other signals (2Hz and 50Hz) were then passed through the network for analysis. Below are the
results for the training and analysis.

365
Arid Zone Journal of Engineering, Technology and Environment, June, 2017; Vol. 13(3):356-368
ISSN 1596-2490; e-ISSN 2545-5818; www.azojete.com.ng

a b c
Figure 12: (a) PMANN training of signal at 13Hz (ideal, hence the perfect match of both graphs)
(b) PMANN analysis of signal at 2Hz, (c) PMANN training of signal at 50Hz
The blue graph is the expected (measured) output of PMANN, while the green is graph for the
signal under analysis (predicted). Figure 12 (a) shows the training of the network, hence, perfect
match between the predicted and measured. While Figures 12 (b) and 12 (c) are the results of the
analysis.
Using the ICC, the samples (or coefficients) were analyzed and decisions were made for fire and
non-fire scenarios. Tables 1, 2 and 3 explain the decision rule. Our estimated reliability between
the 13Hz sample signal and fire training signal gave 1, with 95% CI (1.00, 1.00), which matches
exactly. Hence, such a scenario is recorded fire (Table 1). As indicated in Table 2, our estimated
reliability between the 2Hz sample signal and fire training signal gave an average of 0.004, with
95% CI (-1.660, 0.627), which is a miss-match. Hence, such a scenario is recorded non-fire.
Moreover, Table 3 shows that our estimated reliability between the 50Hz sample signal and fire
training signal is an average of 0.144, with 95% CI (-1.222, 0.670), which is a miss-match. Hence,
such a scenario is recorded non-fire.
Table 2: Intraclass Correlation Coefficient for 13Hz Sample Signals (e.g. open fires)
13Hz Sample Intraclass 95% Confidence Interval (CI) F Test with True Value 0
Signal Correlation Lower Bound Upper Bound Value df1 df2 Sig
Single Measures 1.000 1.000 1.000 1.946E19 18 18 0.000
Average 1.000 1.000 1.000 1.946E19 18 18 0.000
Measures

Table 3: Intraclass Correlation Coefficient for 2Hz Sample Signals (radiations from sources like
Humans)
2Hz Sample Intraclass 95% Confidence Interval (CI) F Test with True Value 0
Signal Correlation Lower Bound Upper Bound Value df1 df2 Sig
Single Measures 0.002 -0.454 0.457 1.004 17 17 0.497
Average 0.004 -1.660 0.627 1.004 17 17 0.497
Measures

Table 4: Intraclass Correlation Coefficient for 50Hz Sample Signals (background radiations)
50Hz Sample Intraclass 95% Confidence Interval (CI) F Test with True Value 0
Signal Correlation
Lower Bound Upper Bound Value df1 df2 Sig
Single Measures 0.078 -.379 0.504 1.168 18 18 0.373
Average 0.144 -1.222 0.670 1.168 18 18 0.373
Measures

366
Ismail and Chukwuemeka: Development of a Model Fire Detection Algorithm; AZOJETE, 13(3):356-368
ISSN 1596-2490; e-ISSN 2545-5818, www.azojete.com.ng

With these results, our model can be real life using appropriate electrical equipment to detect only
real FIRE cases.

4.0 Conclusion
An adaptive model for fire detection has been discussed with emphasis on automatic fire detection.
Using various equations, new algorithms for infrared fire detection have been developed. The
sensor was modeled using Laplace transforms and the fire signal detection mechanism was
modeled using the Hamming window function and discrete wavelet transforms. Using PMANN
and ICC as the decision rule, the detector is able to differentiate between fire and non-fire
radiations.

References
Chen,Y., Serio, MA., Sathyamoorthy, S. 1999. Development of a fire detection system using FT-
IR spectroscopy and artificial neural networks. Fire Safety Science - Proceedings of the sixth
International Symposium, 791802.

Chukwuka, GM., Aderemi, OA., and Obolo, MO. 2014. Oil Well Characterization and Artificial
Gas Lift Optimization using Neural Networks combined with Genetic Algorithm. Discrete
Dynamics in Nature and Society, 2014: 1 11. DOI: http://dx.doi.org/10.1155/2014/289239

Emin BS. 2009. Pyro electric Infrared (PIR) Sensor Based Event Detection. A Thesis Submitted to
the Department of Electrical and Electronics Engineering and the Institute of Engineering and
Sciences of Bilkent University in Partial Fulfillment of the Requirements for the Degree of Master
of Science.

Fatih E., Toreyin BU., Soyer EB., Inac I, Gunay O., Kose K., Cetin E.A. 2012. Wavelet based
flickering flame detector using differential PIR sensors. Fire Safety Journal, 53: 13 - 18.

Javid JH., Shankar BB, Alan W., Boger Zvi. 2008. An adaptive method for Hydrocarbon flame
Detection. Journal of Neural Networks, 21: 398405.

Li Z., Kaufman YJ., Ichoku C., Fraser R., Trishchenko A., Giglio L., Jin J., Yu X. 2000. A Review
of AVHRR-based Active Fire Detection Algorithms: Principles, Limitations, and
Recommendations. Canada Centre for Remote Sensing, Ottawa, Canada, and NASA Goddard
Space Flight Center, Greenbelt, MD.

Nolan, DP, 1996. Handbook of fire and explosion protection engineering principles for oil, gas,
chemical, and related facilities. Noyes Publications, USA.

Odon, A. 2010. Modeling and Simulation of the Pyro-electric Detector using MATLAB/Simulink.
Measurement Science Review, 10( 6): 195 - 199.

Okayama Y., Takashi I., Toru S. 1991. Design of Neural Net to Detect Early Stage of Fire and
Evaluation by Using Real Sensors' Data. Fire Safety Science-Proceedings of the Fourth
International Symposium, 751-759

367
Arid Zone Journal of Engineering, Technology and Environment, June, 2017; Vol. 13(3):356-368
ISSN 1596-2490; e-ISSN 2545-5818; www.azojete.com.ng

Okayama, Y., 1991. Approach to detection of fire in their very early stage by odor sensors and
neural net. Fire Safety Science - Proceedings of third international symposium of fire safety
science, 955964.

Robert JS, Sandra LH. 2012. Fundamentals of Digital Signal Processing Using MATLAB,
Cengage Learning, First Stamford Place, Stamford, USA.

Schneiders MGE. 2001. Wavelets in Control Engineering. Masters Thesis submitted to

Eindhoven University of Technology Faculty of Mechanical Engineering, Dynamics and Control
Technology.

Susan LR., Ronald ES., Sean JH., Frederick WW., Daniel TG., Brooke DS., Scott AH., 2000,
Multi-criteria fire detection systems using a probabilistic neural network. Sensors and Actuators B:
Chemical, 69 (3): 325335.

368