Proceedings of the 2002 IEEE

-
International Conference on Robotics 8 Automation
Washington, DC May 2002

Fire detection and Isolation for Intelligent Building System

Using Adaptive Sensory Fusion Method

Ren C. Luo, Kuo L. Su, and Kuo H. Tsai

Intelligent Automation Laboratory
Department of Electrical Engineering, National C h u g Cheng University
160 Shang-Shing, Ming-Hsiung, Chia-Yi, Taiwan 621
Email: [email protected], [email protected]. [email protected]

Abstract In the last two decades, various methods for detection

and isolation of sensory errors in process data have been
This paper describes an algorithm and experimental proposed including generalized likelihood ratio method[ lo],
resultsfor detecting and isolating sensory failures for afire measurement test method[ 111, parity space method[ 121,
detection system. In the fire system, we use two smoke pseudo nodes method[l3] and extended Kalman filtering
sensors, two flame sensors and two temperature sensors to method[l4, 151. The aim of this paper is to extend the
detect fire event. These sensors can be classified into two adaptive fusion algorithm to detect and isolate sensor faults
groups, and each group has one smoke sensol; one flame for the fire detection system.
sensor and one temperature sensor. The sensory failure and The paper is organized as follows. In the second
isolation techniques described in this paper are based on section, we present the principles of fire detection sensors.
weight variation of the sensory adaptive firsion method. The third section is dedicated to adaptive fusion method.
From the simulation and experimental implementation How to use the adaptive fusion method to detect sensor
results, it demonstrates that the method can exactly find out errors is be described in the section four. The fifth section is
which sensor to befaulty and isolate. That is to say, when a the simulation and experimental results for the proposed
sensory failure occurs, the system can exactly locate the methods. The final section presents brief concluding
sensoryfailure. remarks.

I. Introduction 11. Fire detection system

Intelligent buildings can provide safety, convenience,
Most commonly used sensor for fire detection is smoke
and efficiency for society in the 2lStcentury. An intelligent
sensor. However, smoke sensor cannot provide accurate fire
building system (IBS) is integrated by many services and
detection information. For example, when a smoker stands
subsystems. One of the most important subsystems is the
close to the smoke sensor, it may give a false alarm.
fire detection system in an intelligent building[lJ].
Therefore, we need to use multiple variable sensors to
We can use redundant and complementary information
detect the fire accident and fuse the sensory data to
results from multisensor fusion to enhance system
generate a reliable fire detected signal. In this paper, we
reliability and efficiency of fire detection system. Wang et
employ smoke sensors, flame sensors, and temperature
a1[6] developed a multisensor fire detection algorithm using
sensors for fire detection.
neural network. One temperature and one smoke density
There are two techniques for detection of smoke:
sensor signal were fused for ship fire alarm system. Healey
photo-electric smoke detectors and ionization smoke
et al. [7] presented a real-time fire detection system using
detection. The principle of ionization smoke detection is
color video input. The spectral, spatial, and temporal
that the air cannot conduct electricity (no current flows),
properties of fire were used to derive the fire-detection
when no smoke happens. When smoke happens, an
algorithm. The experimental results demonstrated system
ionizing radioactive source is brought close to the plates
performance on a buming jet fuel. Neubauer [SI applied
and the air itself is ionized. In other words, it will generate
genetic algorithms to an automatic fire detection system.
a tiny current. In this research, we use the ionization smoke
The on-line identification of stochastic signal models for
sensor to detect fire.
measured fire signals were presented. Ruser and Magori [9]
Flame sensors look for characteristic emissions of either
described the fire detection with a combination of
infra-red or ultra-violet light from the flames. We use the
ultrasonic and microwave Doppler sensor.
ultra-violet sensor (R2868) to detect the flame. It’s peak
These sensors may contain errors that can influence the
wavelength is 220 um and sensing wavelength is 185-260
decision output in the fire detection system. Accidents may
um.
happen in this dangerous condition. So that we are
Temperature is a fundamental physical parameter and is
interested in using fault detection and isolation techniques
easy to measure. When the fire is happening, the
to validate the data provided by sensors before using them
temperature may increase on the circumambiency. The
for fire detection system.
temperature sensors include thermocouples, thermistors,
semiconductor junction effects, RTD's and thermo switches. H, : Signal is absent.
In this research, we use semiconductor sensor (AD590). H , : Signal is present.
This sensor has the positive temperature coefficient about
0.7, and its linearity is withii 0.5% for temperature range
between -65°C and 150°C. The standard output of AD590 We set a priori probabilities of the two hypotheses
is 1pA/ OK. represented b y P ( H 0 ) = Po and P(H,) = e.
The fire
The hardware structure of the fire detection system is detection system contains 6 sensors: two smoke sensors,
shown in Figure 1. The decision center is a personal two flame sensors and two temperature sensors. The
computer to perform computation of adaptive fusion sensors can be classified into two groups, and each group
algorithms. In general, these sensory signals were tiny has one smoke sensor, one flame sensor and one
voltage output, and we must process these signals in order temperature sensor. The principles of these sensors are not
to standardize voltage output(OV-4-1OV) by amplifier the same and the weight vales are time-varying in real
circuit. Finally, these sensory signals that become to binary conditions. Hence the weight values are not equal at all
digital signals output transmit to the interface card by time and we can use the adaptive decision to obtain the
comparison circuit. The data from local decision center are optimal values.
transmitted to the decision center. The transmission According to the above assume, we can know that every
interface is a 16-bit specifically designed interface card. observant values at the individual detectors are statistically
We use 6 sensors to detect fire: two smoke sensors, two independent, and the condition probability is defined by
flame sensors and two temperature sensors. These sensors P(XilHb, i = l , 2,3;j=1,2. Each sensor can make a decision
are classified into two groups. Each sensor group contains Vi, i=l, 2,3.
one smoke sensor, one flame sensor, and one temperature If H, is declared
sensor. These complementary sensors can enhance the vi= If HI is declared
(1)
reliability of the detection.
We have following conditions in the fire detection
system.
'Dlj = 'D2j
j = 1,2,3 (2)
'Flj = 'F2j
where j presents sensor type (smoke sensor, flame
sensor or temperature sensor). The reinforcement update
rule of adaptive decision fusion is as follows:

1
If Ui = +1 and H I
1
If Ui = -1 and H O
Figure 1. The hardware structure of fire detection system

111. Adaptive Fusion Method II m1i

1
If U, = -1 and H I
The decision architecture of the fire detection system is
shown in Figure 2 using adaptive fusion method. The
adaptive fusion method is proposed by Ansari[l6, 17, 181.
Assume that the operation of each sensor is independent in CL
When H I occurs.
the fire detection system. The measurable probability is PD AWo= (4)
and the probability of false alarm is PF. In general, PD

*+*-
and PF are assumed to be constant. Therefore, the When H O occurs.
weighting functions can be obtained as
W = log(PDI PF). PD and PF are determined by = + A R . , i=O, 1, 2,...where and
confidence level of the measurement signal. We can q-
represent the weight value after and before each
conduct experiments to determine PD and PF for update, and
different sensors. In reality, PD and PF under operating m is the number of H , .
conditions are time-varying. Therefore, the weight h c t i o n mli is the number of vi = +1 ,and U: = 1.
should be made time-varying [19].
Consider a binary hypothesis testing case with the moi is the number of Ui = -1, and U#:= -1.
following two hypotheses:
 where ut*is real experimental output by data fusion. Therefore, we can define the weight ratio WR,jg as

We can set a threshold q, when the Q,(i) is bigger than q,

we can say that has one sensor is faulty, then we can isolate
it. So that we can integrate them as follows:

Figure 2. Adaptive fusion architecture

1' 7
Qj<i>

S , ( i ) = 1 Q j ( i )2 17 and W , is smal
0 Q j ( i )2 7 and W , is big
Next, we can get more exact decision output for the fire
detection system. The more exact decision output Y(i) is
(9)

defined as
IV. Fault detection algorithm
The paper uses 6 sensors to detect fire occurrence, and k=l j=1
these sensors are classified into two groups. Each group has When Y(i) is bigger than another fire detection threshold
one smoke sensor, one flame sensor and one temperature h, we can say that fire is happened. Finally, we can draw
sensor. We use the adaptive fusion method in each group the architecture of fire detection from equation (9) and (10).
for fire detection, and compare the weight variance for each It is shown in Figure 4.
group for sensor fault detection and isolation. The adaptive
multisensor fusion architecture is shown in Figure 3.
In reality, the weight h c t i o n W,,(i) is time-variable for
adaptive multisensor fusion method. We define the weight
variable function Q,(i) to be
Q,0 )=I Y,(9- wzJ(9 I j = 1,2,3 (5)

Figure 4. The architecture of fire detection system

V. Experimental results
We use 6 sensors to detect fire happened, and these fire
sensors were classified two group. We call them as group I
(K=l) and group I1 (K=2). We define weight functions
verse to sensors to be shown table I.
Table I. The weight function verse to sensors
Figure 3. Adaptive multisensor fusion architecture
If the fue is occured, we can get the equation ( 6 ) by
equation (3)

In another condition, the fire is not occured, we can get

the equation (7), too.
, . .I I-

(7)
n + n

Where m oi, m l i , w0,W O are time-variable functions.

1779
 Next, we use the proposed method to detect sensory
failure and isolate it. In the experiment, we assume that
three types of sensors can detect fire signal simultaneously.
In the Figure 5, the weight variable function is almost the
same, when the fire detection system has not any sensor
failure. If one smoke sensor is error in group I, the weight
variable function is variable for one sensor error, and the
function WRl(i) is become bigger than others(see WR1 of 0.8E ~ ......; . . . . . . . '. \ :

Figure 6). The training number of the sensor failure 0.8; ~

. , . ,

accident location was from 300 to 600. So that we can say ............i.........i ........... :. ......... i............,:..... '.*".i
**i ................i................:.................. :................. i ................................... j
that the fire system has one smoke sensor error. But we 0 " s o a r s m R m
can't know which group smoke sensor to be faulty. We can "1 nmDx

see the Figure 7 and Figure 8. We compare Wll(i)and Wlz(i) Figure 7. The weight of group I for one smoke sensor error
function. The weight function Wll(i) of faulty sensor is
become small. We can detect the sensor fault detection and
isolation.
If one flame sensor is error in the another experiment, we
can get the weight variable function W&(i) to become
bigger than others. It is shown in Figure 9, and we can
know the flame sensor failure in the group I1 by Figure 10
and Figure 11. Because that W&) is bigger than WJ$).
The proposed method can provide more exact results in
sensor fault detection and isolation. We can use the result to
apply in intelligent building and security robot, and the
method is easy to design the hardware circuits. ................;..................: ..............:..................:................. i................ .i
o 1 m m s a m s o a a
h!A$nQnnbr

Figure 8. The weight of group I1 for one smoke sensor error

Tmww-

Figure 5. The weight different function for no sensor error

,m
,c.1 ....... i :.
....... ...... i. ..... Figure 9. The weight different function
for flame sensor error

rniiinp nsnb*

Figure 6. The weight different function

for smoke sensor error

Figure 10. The weight of group I for one flame sensor error
mmq -
1780
 [8] Neubauer, A., “Genetic algorithms in automatic fire
22
detection technology,” Second Intemational
Conference On Genetic Algorithms in Engineering
Systems: Innovations and Applications, 1997, pp.
180-185.
[9] Ruser, H.; Magori, V., “Fire detection with a
combined ultrasonic-microwave Doppler sensor,” in
Proceedings of IEEE Ultrasonics Symposium, Vol.],
1998, pp. 489-492.
[lo] S. Narasimhan, R. S. H. Mah, “Generalized
Likelihood Ratio Method for Large Error
Identification”, AIChE Journal. Vo1.33 No.9,pp.
1514-1521,1987
Figure 11. The weight of group I1 for [Ill R. S. h. Mah, A. C. Jamhane, “Detection of Large
Errors in Process Data”, AIChE Journal, Vo1.28 No.5,
one flame sensor error
pp. 828-831, 1982.
[12] J. Ragot, A. Aitouche, F. Kratz and D. Maquin,
VI. Conclusion “Detection and Localization of Gross Errors in
The sensor fault detection and isolation method for fire Instrument using Parity Space Technique”,
detection system is presented. This paper proposed the International Journal of Mineral Processing, Vo1.3 1,
sensor fault detection and isolation method via adaptive pp. 281-299, 1991.
multisensor fusion algorithm, and the fire detected sensors [I31 J. Ragot, D. Maquin and Darouach, “Analysis of
include smoke sensors, flame sensors and temperature Generalized Bilinear Systems Application to
sensors. We compare the weight-variable on each group by Diagnosis”, Proceeding of IMACS-IFAC Symposium,
using adaptive multisensor fusion scheme. Testing results Lille, France, 1991.
indicated that the proposed method is a suitable solution for [14] D. Del Gobbo, “Sensor Failure Detection and
sensor fault detection and isolation. In the future, we will Identification using Extended Kalman Filtering”,
implement the proposed method on embedded system for Master thesis, Dept’: Aerospace Mech. Eng., West
intelligent building system. Virginia Univ., 1998.
[15] D. Del Gobbo, M. Napolitano, P. Famouri and M.
References Innocenti, “Experimental Application of Extended
C. W. Wang and A. T. P. So, “BuildingAutomation In Kalman Filtering for sensor Validation”, IEEE
The Century,” in Proceedings of the 4-th International Transactions on Control System Technology, Vo1.9,
Conference on Advance on Advances in Power N0.2, March 2001, pp. 376-380.
System Control, Operation Management, APCOM-97, [16] G Chen, N. Ansari and Z. Siveski, “Performance
Hong Kong, November 1997, pp.819-824. Analysis and Realization of Decision Fusion for
M. Azegami and H. Fujixoshi, “A Systematic Macroscopic Diversity in Cellular Wireless Systems,”
Approach to Intelligent Building Design,” IEEE IEEE 1997, pp.424-428.
Communications Magazine, October 1993, pp.46-48. [17] N. Anasri, J. -G Chen and Y. -Z. Zhang Adaptive
Kujuro and H. Yasuda, “Systems Evolution in decision fusion for unequiprobable sources ‘I, IEE
Intelligent Building,” IEEE Communication Proceedings Radar, Sonar and Navigation 1997,
Magazine, October 1993, pp.22-26. pp.105-Ill.
Marion R. Finley, Jr. Ancilla Karakura and Raphael [I81 N. Anasri, E. S . H. Hou, B. 0. Zhu and J. G Chen,
Nbogni, “Survey of Intelligent Building Concepts,” “Adaptive Fusion by Reinforcement Learning for
IEEE Communication Magazine, April 1991, Distributed Detection Systems,” IEEE Transactions
pp.18-20. on Aerospace and Electronic System, Vo1.32, No.2,
M. Fiax, “Intelligent Building,” IEEE April 1996, pp.524-53 1.
Communications Magazine April 1991, pp.24-27.
Wang Xihuai; Xiao Jianmei; Bao Minzhong, “A ship
fire alarm system based on fuzzy neural network,”in
Proceedings of the 3rd World Congress on Intelligent
Control and Automation, Vol. 3, 2000, pp. 1734
-1 736.
Healey, G.; Slater, D.; Lin, T.; Drda, B.; Goedeke,
A.D., “A system for real-time fire detection,” in
Proceedings of IEEE Computer Society Conference
on Computer Vision and Pattern Recognition, 1993,
pp. 605-606.

1781