- Email: [email protected]
Fire Alarm System Based on Multi-Sensor Bayes Network
Available online at www.sciencedirect.com Available online at www.sciencedirect.com
Procedia Engineering
ProcediaProcedia Engineering 00 (2011) Engineering 29 000–000 (2012) 2551 – 2555 www.elsevier.com/locate/procedia
2012 International Workshop on Information and Electronics Engineering (IWIEE)
Fire Alarm System Based on Multi-Sensor Bayes Network Chen Jinga,b*, Fu Jingqia b
a School of Mechatronics Engineering and Automation, Shanghai University, Shanghai, 200072, China School of Electrical and Information Engineering, Anhui University of Science and Technology, Huainan, 232001,China
Abstract Aimed at the problem of false alarm and missing alarm caused by information uncertainty existing in the fire alarm system, Bayesian network (BN) is proposed to analyze fire alarm system. The paper elaborates the internal logic relationship between the fire alarm and the physical-chemical characteristics generated in the process of fire burning by analyzing fire mechanism. Based on defining node variables in BN, multi-sensor Bayesian network model for the fire alarm system is established in Netica. Probabilistic inference and sensitivity analysis of finding node verified that analyzing fire probability through the multi-sensor Bayesian network model is feasible and effective.
© 2012 Published by Elsevier Ltd. Selection and/or peer-review under responsibility of Harbin University of Science and Technology Keywords: Fire alarm system; Multi-sensor; Bayesian network; Probabilistic inference; Sensitivity Analysis
1. Introduction Fire (as a disaster) is one of the most damaging disasters in modern society, how to prevent people's lives and society's wealth from harm caused by fire has become a major issue facing on currently. Fire alarm system is an important means of fire prevention, its function is to sample the signal from fire scene process and determine whether the fire occurred. At present, the main research areas are intelligent fire alarm system based on multi-sensor. Multi-sensor data fusion, neural network, image processing technology and fuzzy logic decision have been used for fire alarm system [1], [2], [3], [4]. In engineering practice, the system structures, control objects and working conditions in different systems are mutative, there are many uncertainties, we hope that fire alarm system can cope with uncertainty problems and has functions of adaptive and self-learning. BN applied in the paper is an
* Chen Jing. Tel.: 086-18955435543; fax: +0-000-000-0000 . E-mail address: [email protected].
1877-7058 © 2011 Published by Elsevier Ltd. doi:10.1016/j.proeng.2012.01.349
22552
andJingqi/ Fu Jingqi / Procedia Engineering 29 (2012) 2551 – 2555 ChenChen Jing Jing and Fu Procedia Engineering 00 (2012) 000–000
effective tool to express and infer uncertain knowledge. Bayesian network model for fire alarm system is built based on mechanism analysis of fire burning. We carry out related probability inferences and the analysis of sensitivity to findings through the Bayesian network model. 2. Bayesian Network Theory Bayesian network [5][6] is a graphical network based on probabilistic inference, which encode a causal relationship between variables visually. It is one of the most effective models in the field of uncertain knowledge representation and inference. 2.1. Definition of Bayesian Network A Bayesian network consist of two parts: the network structure and network parameters, corresponding to qualitative description and quantitative description of the research field respectively [7],[8]. Network structure is a directed acyclic graph structure S, which is composed of the node variables set V(V={V1,V2,…Vn})and directed edges set L(L={ViVj|Vi,Vj∈V}). Node variable Vi is abstraction of the problem; Directed edge L indicates dependence or causal relationship between variables. Network structure is expressed as: S=(V,L). Network parameters P are local probabilities distribution set to reflect the relationship between variables that is the conditional probability table (CPT), CPT lists all possible conditional probabilities of each node relative to its parent nodes: P={P(Vi|V1,V2,…Vi-1),Vi∈V}. Thus, the mathematical model of a Bayesian network is: B=(S,P)=(V,L,P). If Vpi is the parent nodes set of variables Vi, the joint probability distribution of V is: n (1) P( V ) = P( V ,V ,LV , ) = P(V | V ) 1
2
n
∏ i −1
i
pi
2.2. Construction of Bayesian Network Building Bayesian networks is base to solve practical problems. The steps to build Bayesian network include: define node variables; build a network structure; determine network parameters. Constructing network is not simple sequence of above steps, but carries out alternately. First, define the node variables. Premise to create BN firstly determine node of network and their states on behalf of the object properties, the task relies on expert knowledge in related fields to complete mainly. The node can be divided into three categories: input node, output node, and transitional node. The variables are two kinds of variables: discrete variable and continuous variables. It is very simple to define the states of discrete variable. In line with the precision of the research object, the continuous variable C is discretized into finite number states, conversion process is shown as formula (2) [9][10]: ⎧ ST1 ,C0 ≤ C < C1 ⎪ ST ,C ≤ C < C ⎪ 2 C=⎨ 2 1 L ⎪ ⎪⎩STk ,C k −1 ≤ C < C k
(2)
Second, Establish network structure. There are two ways to establish the network structure: expert knowledge method and learning method. The former is based on expert knowledge to determine the structure of Bayesian network; the latter is the so-called learning structure. Third, determine network parameters. There are three ways to determine the network parameters: expert knowledge, the formula, learning parameters method.
Chen – 2555 ChenJing Jingand andFuFuJingqi Jingq// Procedia Procedia Engineering Engineering29 00 (2012) (2012) 2551 000–000
2.3. Probabilistic Inference of Bayesian Network Probabilistic inference is to calculate probability. Let us suppose that all random variables’ set is V, the given node variables’ set E is a subset of V, where, E’s value is expressed as e (e is assumed to be True or False), i.e. E=e. the given node variables are obtained usually by the sensors, we call them findings. The query node variables’ set is assumed as Q, its value is expressed as qi. Probabilistic inference is to calculate the conditional probability of query node variable when finding E=e has been given [9]: p( Vi = qi , E = e ) (3) p( Qi = qi | E = e ) = p( E = e ) Probabilistic inference of BN includes: causal inference, diagnostic inference, Support inference. 3. Multi-sensor Bayesian network of fire alarm system 3.1.
Mechanism analysis of Fire
Fire is the disaster caused by burning out of control. Fire releases heat, accompanied by flame, burning sound, and generates new materials such as smoke, CO and so on in the process of burning. The fire detection means used commonly in engineering is extracting the physical-chemical characteristics of burning processes, including flame, combustion products, burning sound. The physical quantities that can express characteristics are called as characteristic parameters of fire detection [11] [12]. 1) Flame. Fire flame has obvious physical characteristics: flame radiation, flame spectroscopy, flame shape and flame flicker. Characteristic parameters include infrared light, ultraviolet light and flame shape. Sensors involved include infrared sensors, ultraviolet sensors, image sensors and so on. 2) Combustion products. Combustion products include gas and solid product, accompanying with temperature change. The main components of gas products are CO and CO2, Corresponding sensor is gas sensor. Solid product is smoke, sensors are smoke sensor, electrostatic sensor and image sensor. Characteristic parameter of temperature change is temperature, sensors are temperature sensors. 3) Burning sound. Characteristic parameter is infrasound. Corresponding sensor is sound sensor [13]. 3.2. Build fire alarm system based on multi-sensor Bayesian network First, define node variables for Bayesian network. Based on the mechanism analysis of the fire, the characteristic parameters of the fire are treated as the input node variables of BN, the flame, combustion products and burning sound as the intermediate node variables, and the fire as the output node variable. All nodes are the natural nodes. Second, establish BN structure. Netica provides two ways to build BN structure, one is expert knowledge; the other is learning samples from cases. This article integrates expert knowledge and learning samples to create BN structure. Third, determine CPT. Netica provides three methods to define CPT of the nodes: first is expert knowledge; second is using formula, editing language of formula follows language used commonly now such as C, C + + and Java; third is to obtain CPT based on leaning data samples from cases. At the same time of defining CPT, the network structure is built as well. The article integrates two ways of expert knowledge and editing formula to define CPT, shown in figure 1. Each node embodies implicitly CPT of the node variable. Figure.2 is the CPT of the fire node. There are 3 parent nodes of the fire node and the states combination of corresponding nodes at the left of the figure.2, the right is probability distribution of fire node relative to states combination of its parent nodes. Note that the probabilities of each row in the table must sum exactly to 1.0.
2553 3
42554
andJingqi/ Fu Jingqi / Procedia Engineering 29 (2012) 2551 – 2555 ChenChen Jing Jing and Fu Procedia Engineering 00 (2012) 000–000 Ultraviolet Present 50.0 Absent 50.0
Present Absent
Infrared 50.0 50.0
CO2 High Medium Low
CO
33.3 33.3 33.3
High Medium Low
Burning Picture Present 50.0 Absent 50.0
Electrostatic Present 50.0 Absent 50.0 Burning Gas
Present Absent
Present Absent
Flame 50.0 50.0
Smoke Picture Present 50.0 Absent 50.0
33.3 33.3 33.3
Temperature
50.0 50.0
High Low
Burning Sound Present 50.0 Absent 50.0
Present Absent
Solid Product
50.0 50.0
Present Absent
50.0 50.0
Burning Product Present 50.0 Absent 50.0
Fire 50.0 50.0
Figure.1 Compiled Bayesian network model of fire alarm system
Figure.2 Conditional probability table of fire node
4. Probabilistic Inference of Bayesian Network 4.1. Forecast Analysis of Fire In BN shown in figure 3, assume that the CO concentration is finding variable. If the current CO concentration exceeds standard level, finding node status is high. Forecast analysis is to calculate the posterior probability of other node variables based on the findings. In Netica, CO variable state is set to high (High = 100%), indicating that state of finding variable is known, the probability of the entire network is updated by automatic updating function. The results shown in figure 3, where, probability of burning gases (Present) roses up to 70% from 50%, burning products (Present) from 50% to 57.5%, fire (Fire) from 50% to 53.7%, showing that overrun of CO concentration increases the fire probability. Besides CO concentration, let us assume that CO2 exceeds the standard also. The results is shown in figure 4, where, probability of fire roses up to 59.4%from53.7%, showing that the fire probability is higher. The above analysis is consistent with the engineering practice. Ultraviolet Present 50.0 Absent 50.0 Infrared Present 50.0 Absent 50.0
CO2 High Medium Low
100 0 0
33.3 33.3 33.3
Burning Picture Present 50.0 Absent 50.0
Present Absent
Flame 50.0 50.0
70.0 30.0
Burning Sound Present Absent
50.0 50.0
Temperature High Low
50.0 50.0
Ultraviolet
Smoke Picture Present 50.0 Absent 50.0
Present Absent
Present Absent
50.0 50.0
Present Absent
50.0 50.0
57.5 42.5
Flame Present Absent
50.0 50.0
53.7 46.3
Figure.3 Fire forecast based CO finding variable
CO 100 0 0
50.0 50.0
Smoke Picture Present Absent
Present Absent
Burning Sound Present Absent
50.0 50.0
50.0 50.0
Temperature High 50.0 Low 50.0
50.0 50.0 Solid Product Present 50.0 Absent 50.0
Burning Product Present Absent
68.8 31.3
Fire
Fire Present Absent
High Medium Low
Electrostatic
Burning Gas Present 100 Absent 0
Burning Product Present Absent
CO2 100 0 0
Burning Picture
Solid Product Present Absent
High Medium Low
50.0 50.0
Infrared
Electrostatic Present 50.0 Absent 50.0 Burning Gas
Present Absent
CO High Medium Low
Present Absent
59.4 40.6
Figure.4 Fire forecast based both CO and CO2 finding variables
4.2. Sensitivity Analysis of Finding Sometimes it is useful to know how much our belief in a particular node is influenced by findings at other nodes. If it is very sensitive, it is important for us to know the state of that node. Netica can compute a node's sensitivity to findings easily. Suppose that we want to know which nodes can most influence our knowledge of "Flame". Click on the top of the title region of the "Flame" box to select it, and then choose "Network->Sensitivity to Findings". A detailed report will be produced in the "Netica - Messages" window. For brevity, the sensitivity of "Flame" to other findings is only shown in table 1. We can find that the most influential nodes of "Flame" are its parents and children from table 1. Tab.1 Sensitivity of "Flame" to a finding at another node
Chen – 2555 ChenJing Jingand andFuFuJingqi Jingq// Procedia Procedia Engineering Engineering 29 00 (2012) (2012) 2551 000–000 Node Flame Burning-picture Fire Ultraviolet Infrared Other nodes
Mutual Info 1.00000 0.18872 0.11871 0.04557 0.04557 0.00000
Percent 100 18.9 11.9 4.56 4.56 0.00
Variance of Beliefs 0.2500000 0.0625000 0.0400000 0.0156250 0.0156250 0.0000000
Conclusions The paper has applied Bayesian network for analyzing fire alarm system, made good use of the powerful graphical knowledge expression and effective uncertain findings processing ability, which is an attempt to study fire alarm system. Probabilistic inference show that fire alarm system based on BN model is feasible and effective. In practical applications, there are some works to do, including collecting samples of fire cases, establishing database, and gradually modify and improve and update BN model. Acknowledgments This paper is supported by the National “863” Foundation of China (2007AA04Z174), the Scientific and Technologic Fund of Huainan city (2009A05011). References [1] Zhang Zhuan-cheng. A Fire-alarm prediction system based on multi-sensor information fusion technique [J]. Transducer and Microsystem Technologies. 2006.09:22-25. [2] Ran Hai-chao, Liu Li-hui. The Fire Detection System on Neural Network[J]. FIRE SAFETY SCIENCE. 2000.01:34-38. [3] LI Xiao-qin. Design and Implementation of Fire Detection and Alarm System based on Image Processing [J]. Network & Computer Security. 2010.11:35-37. [4] Yao Jia-fei, Li Bei-hai. Atrium Fire Detecting System Based on Fuzzy-neural Network [J]. China Instrumentation. 2007.11:36-38. [5] JENSEN F V, NIELSEN T D. Bayesian Networks and Decision Graph [M]. Berlin: Springer, 2007. [6] KORB K B, Nicholson A E. Bayesian Artificial Intelligence [M]. London, UK: Chapman & Hall/CRC Press, 2004. [7] Pearl J. Probabilistic reasoning in intelligent system: Networks of plausible inference [M]. San Mateo: Morgan Kaufman Publishers, 1988. [8] Pollino. etal. Parameterization and evaluation of a Bayesian belief network for use in an ecological risk assessment [J]. Environmental Modelling and Software 2007，22 (8):1140-1152. [9] Zhang Lian-wen, Guo Hai-peng. Introduction to Bayesian networks[M]. Science Press, 2006,11. [10] Wang Guang-yan, etal. The Study of Simulation Metamodel Based on Bayesiab Network and Its Application in Equipment Battlefield Damage Simulation[J]. MATHEMATICS IN PRACTICE THEORY, 2010(4):59-69. [11] Cheng Xiao-fang. Principle and Method for Fire Detection [J]. China Safety Science Journal. 1999,2:24-29; 1999,4:1-5. [12] CHENG Xin, WANG Da-chuan, YIN Dong-liang. Image Type Fire Flame Detecting Principle [J]. FIRE SAFETY SCIENCE. 2005,10:239-246. [13] Fan Heng. Research on Acoustic Emission Monitoring Technology Application in Fire Detection [J]. Fire Science. 2009(1):108-111.
2555 5
Procedia Engineering
ProcediaProcedia Engineering 00 (2011) Engineering 29 000–000 (2012) 2551 – 2555 www.elsevier.com/locate/procedia
2012 International Workshop on Information and Electronics Engineering (IWIEE)
Fire Alarm System Based on Multi-Sensor Bayes Network Chen Jinga,b*, Fu Jingqia b
a School of Mechatronics Engineering and Automation, Shanghai University, Shanghai, 200072, China School of Electrical and Information Engineering, Anhui University of Science and Technology, Huainan, 232001,China
Abstract Aimed at the problem of false alarm and missing alarm caused by information uncertainty existing in the fire alarm system, Bayesian network (BN) is proposed to analyze fire alarm system. The paper elaborates the internal logic relationship between the fire alarm and the physical-chemical characteristics generated in the process of fire burning by analyzing fire mechanism. Based on defining node variables in BN, multi-sensor Bayesian network model for the fire alarm system is established in Netica. Probabilistic inference and sensitivity analysis of finding node verified that analyzing fire probability through the multi-sensor Bayesian network model is feasible and effective.
© 2012 Published by Elsevier Ltd. Selection and/or peer-review under responsibility of Harbin University of Science and Technology Keywords: Fire alarm system; Multi-sensor; Bayesian network; Probabilistic inference; Sensitivity Analysis
1. Introduction Fire (as a disaster) is one of the most damaging disasters in modern society, how to prevent people's lives and society's wealth from harm caused by fire has become a major issue facing on currently. Fire alarm system is an important means of fire prevention, its function is to sample the signal from fire scene process and determine whether the fire occurred. At present, the main research areas are intelligent fire alarm system based on multi-sensor. Multi-sensor data fusion, neural network, image processing technology and fuzzy logic decision have been used for fire alarm system [1], [2], [3], [4]. In engineering practice, the system structures, control objects and working conditions in different systems are mutative, there are many uncertainties, we hope that fire alarm system can cope with uncertainty problems and has functions of adaptive and self-learning. BN applied in the paper is an
* Chen Jing. Tel.: 086-18955435543; fax: +0-000-000-0000 . E-mail address: [email protected].
1877-7058 © 2011 Published by Elsevier Ltd. doi:10.1016/j.proeng.2012.01.349
22552
andJingqi/ Fu Jingqi / Procedia Engineering 29 (2012) 2551 – 2555 ChenChen Jing Jing and Fu Procedia Engineering 00 (2012) 000–000
effective tool to express and infer uncertain knowledge. Bayesian network model for fire alarm system is built based on mechanism analysis of fire burning. We carry out related probability inferences and the analysis of sensitivity to findings through the Bayesian network model. 2. Bayesian Network Theory Bayesian network [5][6] is a graphical network based on probabilistic inference, which encode a causal relationship between variables visually. It is one of the most effective models in the field of uncertain knowledge representation and inference. 2.1. Definition of Bayesian Network A Bayesian network consist of two parts: the network structure and network parameters, corresponding to qualitative description and quantitative description of the research field respectively [7],[8]. Network structure is a directed acyclic graph structure S, which is composed of the node variables set V(V={V1,V2,…Vn})and directed edges set L(L={ViVj|Vi,Vj∈V}). Node variable Vi is abstraction of the problem; Directed edge L indicates dependence or causal relationship between variables. Network structure is expressed as: S=(V,L). Network parameters P are local probabilities distribution set to reflect the relationship between variables that is the conditional probability table (CPT), CPT lists all possible conditional probabilities of each node relative to its parent nodes: P={P(Vi|V1,V2,…Vi-1),Vi∈V}. Thus, the mathematical model of a Bayesian network is: B=(S,P)=(V,L,P). If Vpi is the parent nodes set of variables Vi, the joint probability distribution of V is: n (1) P( V ) = P( V ,V ,LV , ) = P(V | V ) 1
2
n
∏ i −1
i
pi
2.2. Construction of Bayesian Network Building Bayesian networks is base to solve practical problems. The steps to build Bayesian network include: define node variables; build a network structure; determine network parameters. Constructing network is not simple sequence of above steps, but carries out alternately. First, define the node variables. Premise to create BN firstly determine node of network and their states on behalf of the object properties, the task relies on expert knowledge in related fields to complete mainly. The node can be divided into three categories: input node, output node, and transitional node. The variables are two kinds of variables: discrete variable and continuous variables. It is very simple to define the states of discrete variable. In line with the precision of the research object, the continuous variable C is discretized into finite number states, conversion process is shown as formula (2) [9][10]: ⎧ ST1 ,C0 ≤ C < C1 ⎪ ST ,C ≤ C < C ⎪ 2 C=⎨ 2 1 L ⎪ ⎪⎩STk ,C k −1 ≤ C < C k
(2)
Second, Establish network structure. There are two ways to establish the network structure: expert knowledge method and learning method. The former is based on expert knowledge to determine the structure of Bayesian network; the latter is the so-called learning structure. Third, determine network parameters. There are three ways to determine the network parameters: expert knowledge, the formula, learning parameters method.
Chen – 2555 ChenJing Jingand andFuFuJingqi Jingq// Procedia Procedia Engineering Engineering29 00 (2012) (2012) 2551 000–000
2.3. Probabilistic Inference of Bayesian Network Probabilistic inference is to calculate probability. Let us suppose that all random variables’ set is V, the given node variables’ set E is a subset of V, where, E’s value is expressed as e (e is assumed to be True or False), i.e. E=e. the given node variables are obtained usually by the sensors, we call them findings. The query node variables’ set is assumed as Q, its value is expressed as qi. Probabilistic inference is to calculate the conditional probability of query node variable when finding E=e has been given [9]: p( Vi = qi , E = e ) (3) p( Qi = qi | E = e ) = p( E = e ) Probabilistic inference of BN includes: causal inference, diagnostic inference, Support inference. 3. Multi-sensor Bayesian network of fire alarm system 3.1.
Mechanism analysis of Fire
Fire is the disaster caused by burning out of control. Fire releases heat, accompanied by flame, burning sound, and generates new materials such as smoke, CO and so on in the process of burning. The fire detection means used commonly in engineering is extracting the physical-chemical characteristics of burning processes, including flame, combustion products, burning sound. The physical quantities that can express characteristics are called as characteristic parameters of fire detection [11] [12]. 1) Flame. Fire flame has obvious physical characteristics: flame radiation, flame spectroscopy, flame shape and flame flicker. Characteristic parameters include infrared light, ultraviolet light and flame shape. Sensors involved include infrared sensors, ultraviolet sensors, image sensors and so on. 2) Combustion products. Combustion products include gas and solid product, accompanying with temperature change. The main components of gas products are CO and CO2, Corresponding sensor is gas sensor. Solid product is smoke, sensors are smoke sensor, electrostatic sensor and image sensor. Characteristic parameter of temperature change is temperature, sensors are temperature sensors. 3) Burning sound. Characteristic parameter is infrasound. Corresponding sensor is sound sensor [13]. 3.2. Build fire alarm system based on multi-sensor Bayesian network First, define node variables for Bayesian network. Based on the mechanism analysis of the fire, the characteristic parameters of the fire are treated as the input node variables of BN, the flame, combustion products and burning sound as the intermediate node variables, and the fire as the output node variable. All nodes are the natural nodes. Second, establish BN structure. Netica provides two ways to build BN structure, one is expert knowledge; the other is learning samples from cases. This article integrates expert knowledge and learning samples to create BN structure. Third, determine CPT. Netica provides three methods to define CPT of the nodes: first is expert knowledge; second is using formula, editing language of formula follows language used commonly now such as C, C + + and Java; third is to obtain CPT based on leaning data samples from cases. At the same time of defining CPT, the network structure is built as well. The article integrates two ways of expert knowledge and editing formula to define CPT, shown in figure 1. Each node embodies implicitly CPT of the node variable. Figure.2 is the CPT of the fire node. There are 3 parent nodes of the fire node and the states combination of corresponding nodes at the left of the figure.2, the right is probability distribution of fire node relative to states combination of its parent nodes. Note that the probabilities of each row in the table must sum exactly to 1.0.
2553 3
42554
andJingqi/ Fu Jingqi / Procedia Engineering 29 (2012) 2551 – 2555 ChenChen Jing Jing and Fu Procedia Engineering 00 (2012) 000–000 Ultraviolet Present 50.0 Absent 50.0
Present Absent
Infrared 50.0 50.0
CO2 High Medium Low
CO
33.3 33.3 33.3
High Medium Low
Burning Picture Present 50.0 Absent 50.0
Electrostatic Present 50.0 Absent 50.0 Burning Gas
Present Absent
Present Absent
Flame 50.0 50.0
Smoke Picture Present 50.0 Absent 50.0
33.3 33.3 33.3
Temperature
50.0 50.0
High Low
Burning Sound Present 50.0 Absent 50.0
Present Absent
Solid Product
50.0 50.0
Present Absent
50.0 50.0
Burning Product Present 50.0 Absent 50.0
Fire 50.0 50.0
Figure.1 Compiled Bayesian network model of fire alarm system
Figure.2 Conditional probability table of fire node
4. Probabilistic Inference of Bayesian Network 4.1. Forecast Analysis of Fire In BN shown in figure 3, assume that the CO concentration is finding variable. If the current CO concentration exceeds standard level, finding node status is high. Forecast analysis is to calculate the posterior probability of other node variables based on the findings. In Netica, CO variable state is set to high (High = 100%), indicating that state of finding variable is known, the probability of the entire network is updated by automatic updating function. The results shown in figure 3, where, probability of burning gases (Present) roses up to 70% from 50%, burning products (Present) from 50% to 57.5%, fire (Fire) from 50% to 53.7%, showing that overrun of CO concentration increases the fire probability. Besides CO concentration, let us assume that CO2 exceeds the standard also. The results is shown in figure 4, where, probability of fire roses up to 59.4%from53.7%, showing that the fire probability is higher. The above analysis is consistent with the engineering practice. Ultraviolet Present 50.0 Absent 50.0 Infrared Present 50.0 Absent 50.0
CO2 High Medium Low
100 0 0
33.3 33.3 33.3
Burning Picture Present 50.0 Absent 50.0
Present Absent
Flame 50.0 50.0
70.0 30.0
Burning Sound Present Absent
50.0 50.0
Temperature High Low
50.0 50.0
Ultraviolet
Smoke Picture Present 50.0 Absent 50.0
Present Absent
Present Absent
50.0 50.0
Present Absent
50.0 50.0
57.5 42.5
Flame Present Absent
50.0 50.0
53.7 46.3
Figure.3 Fire forecast based CO finding variable
CO 100 0 0
50.0 50.0
Smoke Picture Present Absent
Present Absent
Burning Sound Present Absent
50.0 50.0
50.0 50.0
Temperature High 50.0 Low 50.0
50.0 50.0 Solid Product Present 50.0 Absent 50.0
Burning Product Present Absent
68.8 31.3
Fire
Fire Present Absent
High Medium Low
Electrostatic
Burning Gas Present 100 Absent 0
Burning Product Present Absent
CO2 100 0 0
Burning Picture
Solid Product Present Absent
High Medium Low
50.0 50.0
Infrared
Electrostatic Present 50.0 Absent 50.0 Burning Gas
Present Absent
CO High Medium Low
Present Absent
59.4 40.6
Figure.4 Fire forecast based both CO and CO2 finding variables
4.2. Sensitivity Analysis of Finding Sometimes it is useful to know how much our belief in a particular node is influenced by findings at other nodes. If it is very sensitive, it is important for us to know the state of that node. Netica can compute a node's sensitivity to findings easily. Suppose that we want to know which nodes can most influence our knowledge of "Flame". Click on the top of the title region of the "Flame" box to select it, and then choose "Network->Sensitivity to Findings". A detailed report will be produced in the "Netica - Messages" window. For brevity, the sensitivity of "Flame" to other findings is only shown in table 1. We can find that the most influential nodes of "Flame" are its parents and children from table 1. Tab.1 Sensitivity of "Flame" to a finding at another node
Chen – 2555 ChenJing Jingand andFuFuJingqi Jingq// Procedia Procedia Engineering Engineering 29 00 (2012) (2012) 2551 000–000 Node Flame Burning-picture Fire Ultraviolet Infrared Other nodes
Mutual Info 1.00000 0.18872 0.11871 0.04557 0.04557 0.00000
Percent 100 18.9 11.9 4.56 4.56 0.00
Variance of Beliefs 0.2500000 0.0625000 0.0400000 0.0156250 0.0156250 0.0000000
Conclusions The paper has applied Bayesian network for analyzing fire alarm system, made good use of the powerful graphical knowledge expression and effective uncertain findings processing ability, which is an attempt to study fire alarm system. Probabilistic inference show that fire alarm system based on BN model is feasible and effective. In practical applications, there are some works to do, including collecting samples of fire cases, establishing database, and gradually modify and improve and update BN model. Acknowledgments This paper is supported by the National “863” Foundation of China (2007AA04Z174), the Scientific and Technologic Fund of Huainan city (2009A05011). References [1] Zhang Zhuan-cheng. A Fire-alarm prediction system based on multi-sensor information fusion technique [J]. Transducer and Microsystem Technologies. 2006.09:22-25. [2] Ran Hai-chao, Liu Li-hui. The Fire Detection System on Neural Network[J]. FIRE SAFETY SCIENCE. 2000.01:34-38. [3] LI Xiao-qin. Design and Implementation of Fire Detection and Alarm System based on Image Processing [J]. Network & Computer Security. 2010.11:35-37. [4] Yao Jia-fei, Li Bei-hai. Atrium Fire Detecting System Based on Fuzzy-neural Network [J]. China Instrumentation. 2007.11:36-38. [5] JENSEN F V, NIELSEN T D. Bayesian Networks and Decision Graph [M]. Berlin: Springer, 2007. [6] KORB K B, Nicholson A E. Bayesian Artificial Intelligence [M]. London, UK: Chapman & Hall/CRC Press, 2004. [7] Pearl J. Probabilistic reasoning in intelligent system: Networks of plausible inference [M]. San Mateo: Morgan Kaufman Publishers, 1988. [8] Pollino. etal. Parameterization and evaluation of a Bayesian belief network for use in an ecological risk assessment [J]. Environmental Modelling and Software 2007，22 (8):1140-1152. [9] Zhang Lian-wen, Guo Hai-peng. Introduction to Bayesian networks[M]. Science Press, 2006,11. [10] Wang Guang-yan, etal. The Study of Simulation Metamodel Based on Bayesiab Network and Its Application in Equipment Battlefield Damage Simulation[J]. MATHEMATICS IN PRACTICE THEORY, 2010(4):59-69. [11] Cheng Xiao-fang. Principle and Method for Fire Detection [J]. China Safety Science Journal. 1999,2:24-29; 1999,4:1-5. [12] CHENG Xin, WANG Da-chuan, YIN Dong-liang. Image Type Fire Flame Detecting Principle [J]. FIRE SAFETY SCIENCE. 2005,10:239-246. [13] Fan Heng. Research on Acoustic Emission Monitoring Technology Application in Fire Detection [J]. Fire Science. 2009(1):108-111.
2555 5