论文部分内容阅读
Abstract
A forest fire is a severe threat to forest resources and human life. In this paper， we propose a forest?fire detection system that has an artificial neural network algorithm implemented in a wireless sensor network （WSN）. The proposed detection system mitigates the threat of forest fires by provide accurate fire alarm with low maintenance cost. The accuracy is increased by the novel multi?criteria detection， referred to as an alarm decision depends on multiple attributes of a forest fire. The multi?criteria detection is implemented by the artificial neural network algorithm. Meanwhile， we have developed a prototype of the proposed system consisting of the solar batter module， the fire detection module and the user interface module.
Keywords
forest fire detection； artificial neural network； wireless sensor network
W1 Introduction
ireless sensor networks （WSNs） have been the focus of research over the past few years because of their potential in environmental monitoring， target tracking， and object detection [1]. WSNs have also been studied in the context of detecting forest fires， which threaten forest resources and human life. WSNs are not costly and can detect forest fires in real time， unlike current detection methods based on human observation and unlike spot weather forecasts or even satellite monitoring. WSNs can also provide information about environmental conditions within the forest， which is useful for predicting forest fires [2]. Moreover， forest fire detection and prediction is associated with specific location information provided by individual sensor nodes.
Although some practical experiments have been conducted using WSNs to collect sensed data from a forest fire [3]-[5]， there are still some challenges to using WSNs for this purpose. A fire detector may sound an alarm based on a simple threshold， which gives rise to false alarms even though the sensing unit of the fire detector may be highly sensitive. False alarms occur for two main reasons：
·A photoelectric smoke sensing unit is sensitive to white aerosol particles from a smoldering fire but also to dust [6].
·Environmental conditions in the forest often severely disturb the normal behavior of the sensing unit. Sunlight and artificial light are primary sources of interference with the flame?sensing unit.
Limited power supply to sensor nodes makes it difficult to detect fires over a long period of time. The potential energy sources for sensor nodes can be classified according to whether they store energy within the sensor nodes （e.g.， in a battery）， distribute power to the sensor node through a wire， or scavenge available ambient power （e.g， using a solar battery on the sensor node）. Considering the volume of the sensor node， manner of deployment， and forest conditions， the solar battery is one of the most promising sources of energy for detecting forest fires over a long period of time. However， existing works on solar batteries for sensor nodes， e.g.， [8]-[13]， overlook the problem of intermittent sunlight in the forest. In this paper， we propose a forest fire detection system that includes an artificial neural network algorithm implemented in a WSN. Overall， the main contributions of this paper are as follows：
·The multi?criteria detection depends on multiple attributes of a forest fire and is introduced into WSNs to increase the accuracy of detecting a forest fire.
·An artificial neural network algorithm is used to fuse sensing data that corresponds to multiple attributes of a forest fire into an alarm decision.
·We introduce the principle of the proposed system as well as a prototype comprising TelosB sensor nodes and a solar battery to power the WSN.
2 System Description
For the sake of clarity， we consider a WSN with only one base station and hundreds of sensor nodes. Because a WSN with multiple base stations can be regarded as multiple WSNs （each comprising one base station and corresponding sensor nodes）， the proposed system can also be implemented if the WSN has multiple base stations. Therefore， there are [n] sensor nodes in the WSN， each denoted [sj]， [1≤j≤n]. A forest fire [f] has [l] attributes， each denoted [rfi]， [1≤i≤l]. Attribute [rfi] can be sensed by the sensing unit [ui]. A [ui] on an [sj] is denoted [uji]. The output sensing data of [uji] is denoted [oji]. For simplicity， we assume that [sj] has [l] types of sensing units covering [l] attributes of the forest fire. We use a multilayer back?propagation artificial neural network to fuse sensing data [oji]. The total number of layers in the artificial neural network is denoted [m]. The input vector of the [k]th layer， [1≤k≤m]， is denoted [Ak-1]. The output vector of the [k]th layer is denoted [Ak]. Therefore， [A0] and [Am] represent the input and output of the artificial neural network， respectively. The alarm decision is denoted[ad].
3 Proposed Forest Fire Detection Method
In our proposed system， detection is made more accurate by using multiple criteria， which means the[ad] is based on multiple criteria of the forest fire. Multi?criteria detection is implemented by the artificial neural network algorithm. Because of the artificial neural network， the proposed system has low overhead and has self?learning capabilities； that is， it trains itself to build up the relations between sensing data and correct [ad].
3.1 MultiCriteria Detection
In a system that depends on one attribute of a forest fire to raise alarms， there is a high probability of false alarms because of inherent system drawbacks or external disturbances. To overcome such drawbacks and counter external disturbances， the system must take into account the multiple attributes of a forest fire. This is referred to as multi?criteria detection （Definition 1）. With multi?criteria detection， multiple attributes of a forest fire are sensed by different types of sensing unit. Therefore， a sensing unit that has been interfered with cannot raise a false alarm. Together， multiple sensing units confirm an alarm. Multi?criteria detection increases the accuracy of detecting a forest fire. Definition 1 （Multi?criteria detection）. Multi?criteria detection is represented as a function with multiple arguments [rf1，rf2，...，rfl] ， which refer to the attributes of forest fire [f] ， and one [ad]， given by：
[ad=f（rf1，rf2，...，rfl）] （1）
The attributes [rf1，rf2，...，rfl] could be any combination of the attributes of a forest fire. The directly sensed attributes of a forest fire are flame and heat， which are sensed by the flame sensing unit and heat sensing unit. The flame emits visible light， but the forest fire also emits a lot of radiation， the spectral distribution of which is the radiation intensity with respect to different wavelengths and is not uniform. In theory， the radiation intensity is determined by the temperature of the fire. The radiation intensity from a blackbody with respect to the wavelength and temperature is described by Planck’s radiation law （2）， where [h] is Planck’s constant， [c] is the speed of light， [λ] is the wavelength， [k] is Boltzmann’s constant， and [T] is the temperature [14].
[I=2hc2λ5exp[hc/λkT]-1] （2）
Therefore， radiation intensity can be the basis for detecting a forest fire given that the typical temperature of a forest fire is [600 °C-1000 °C] [15]. The ultraviolet sensing unit and infrared flame sensing unit work by detecting radiation intensity. Other attributes that can be used to identify a forest fire include combustion products. It is well known that a forest fire gives off bursts of carbon dioxide， carbon monoxide， water vapor， and dust.
3.2 Artificial Neural Network Algorithm
We use the multi?layer back?propagation artificial neural network for multi?criteria detection. Although data fusion in WSNs has been covered in much of the literature [16]-[18]， the topic has not been considered in the context of forest fires. A multi?layer back?propagation artificial neural network is widely used to emulate the non?linear relationship between its input and output. However， computation in this kind of network is not complex because the network is a combination of neurons dealing with simple functions. Moreover， multi?layer back?propagation artificial neural network is capale of self?learning， which means it can train itself to build up relations between the inputs and desired targets.
3.2.1 Making an Alarm Decision
Without loss of generality， we assume that the multi?layer back?propagation artificial neural network is implemented on [sj] with [l] types of sensing units that cover [l] attributes of the forest fire. Sensing data [oji] of [uji] on [sj] corresponds to [rfi] of the forest fire. [A0=[a01，...，a0l]T] （3）
For clarity， let all [oji]， [1≤i≤l] comprise a column vector [A0] （3）， where [a0i=oji]. Vector [A0] is the input to the multiple layer artificial neural network.
Specifically， [A0] is the input to the first layer of the multi?layer artificial neural network. In the first layer， [A0] is multiplied by weight matrix [W1] with dimension [s1×l] and bias vector [B1] ， including [s1] neurons in the first layer. The intermediate computation result of the first layer is denoted [N1] and is given by：
[N1=W1A0+B1] （4）
Then， [N1] is sent to transfer function [F1]， which may be a linear or nonlinear. That is， [F1] may be a hard?limit function or sigmoid function depending on the specific problem it needs to solve. In general， transfer functions in the multi?layer artificial neural network are easy to compute. Transfer function [F1] operates on every element of [N1]. The result of transfer function[F1]， denoted [A1]， is the output of the first layer：
[A1=F1（N1）=F1（n11）?F1（n1s1）] （5）
The fusion of sensing data proceeds in the second layer of the multi?layer artificial neural network. The output [A1] of the first layer becomes the input of the second layer. The calculation process of the second layer is similar to that of the first layer except the second layer has its own [W2]， [B2]， and [F2]. In general， the calculation of the [i]th layer is given by：
[Ai=Fi（WiAi-1+Bi）， 1≤i≤m-1 ] （6）
where [m] is the number of layers in the artificial network.
For a decision to be made on whether there is a forest fire or not， the output of the [m]th layer， [Am] （7）， is confined to one element.
[ad=Am=Fm（WmAm-1+Bm）] （7）
This is done by letting the m th layer contain only one neuron. If the alarm decision is confined to a Boolean value， we need to choose the transfer function， the output of which is a Boolean value， for the m th layer， such as the hard limit function.
An alarm decision is made by inputting the attributes of a forest fire into the multi?layer back?propagation artificial neural network， as shown in Theorem 1.
Theorem 1. The [ad]of the multi?criteria detection is computed by the recursive （8） given sensing data [A0=[a01，...，a0l]T] corresponding to [l] attributes of forest fire[f].
[ad=Fm（WmAm-1+Bm）Ai=Fi（WiAi-1+Bi），1≤i≤m-1] （8）
3.2.2 Self?Learning Capability
Given sensing data that corresponds to multiple attributes of a forest fire and given correct alarm decisions， the multi?layer back?propagation artificial neural network trains itself to build relationships between the sensing data and correct alarm decisions. However complex the relationship， it is easy for the multi?layer back?propagation artificial neural network to fulfil the task. Essentially， self?learning means having the output of the multi?layer back?propagation artificial neural network approximate the target output by adjusting the weight matrixes and biases. This adjustment is made in order to minimize the mean?square error （MSE） between the output and target output. Suppose [q] inputs are denoted[A0i]， [1≤i≤q]. Corresponding to [A0i]， the output is denoted [Ami]， and the target output is denoted [Tmi]. Thus， the MSE for the [i]th iterated adjustment is：
[MSE（i）=E（Tmi-Ami）T（Tmi-Ami）] （9）
For clarity， the adjustment of the weight matrixes and biases is expressed by Theorem 2.
Theorem 2. In the self?learning， the [i]th iterated adjustment of the weight matrixes and biases is conducted according to （10） and （11）， where [α] is a constant for the learning rate， and [?（j）] is computed by the recursive （12） where [D（j）=diag（Fj）′（nj1），...，（Fj）′（njsj）].
[Wj（i）=Wj（i-1）-α?（j）（Aj-1）T， 1≤j≤m] （10）
[Bj（i）=Bj（i-1）-α?（j）， 1≤j≤m] （11）
[?（j）=D（j）（Wj+1）T?（j+1），1≤j≤m-1?（m）=-2D（m）（Tmi-Ami）] （12）
After self?learning， the multi?layer back?propagation artificial neural network builds up a mathematical relationship between the sensing data and correct alarm decisions. Then， the artificial neural network can make an accurate alarm decision.
4 Implemented Prototype
We have developed a prototype of the forest fire detection system using an artificial neural network in a WSN. The system mainly comprises three parts： the solar battery， fire?detection module， and user interface.
4.1 Solar Battery
To consistently power the unattended sensor nodes deployed in a forest where only intermittent sunlight is available， we develop a solar battery （Fig. 1）. The energy from the solar panel is buffered by the super capacitor. When the energy in the supper capacitor reaches a threshold， the super capacitor starts to recharge the Li?ion battery. Because of the intermittent sunlight in the forest， the energy produced by the solar panel is not enough to recharge the battery. If not buffered in the super capacitor， this energy is wasted. On the other hand， the charge?discharge cycles of the Li?ion battery are limited. It is better to charge a Li?ion battery until it is full； otherwise， the life of the battery decreases. On the contrary， the super capacitor has almost infinite charge?discharge cycles and is ideal for frequently pulsing applications.
Here we discuss implementation of the solar battery in detail. The solar panel of the battery is 110 × 95 mm and comprises eight cells connected in parallel and generating 550 mA at 2 V. Theoretically， the maximum energy generated in one hour can sustain a sensor node for 26 days， i.e.， 1100 mAh/（0.53 mA × 3.3 V × 24 h）， provided that the sensor nodes work on a 10% duty cycle with an average current of 0.53 mA. Energy from the solar panel is buffered by two 150 F 2.5 V super capacitors wired in parallel. A 3.7 V 700 mAh Li?ion battery is used to continually save energy. The fully charged Li?ion battery can power a sensor node working on a 10% duty cycle for 55 days， i.e.， [700 mAh/（0.53 mA×24 h）] . We chose MAX1674 and ISL6292 integrated circuits as the DC?DC converters， which have a conversion efficiency of around 90%. 4.2 Fire Detection Module
The fire detection module is responsible for multi?criteria detection. The module comprises five TelosB sensor nodes， four of which monitor the forest fire. That is， they convert the attributes of a forest fire into sensing data. The multi?layer back?propagation artificial neural network is implemented on each individual sensor node because the sensor node is endowed with four types of sensing units. However， for the purpose of analysis， raw sensing data besides the fire alarm are transmitted to users. The last sensor node acts as the base station， collecting sensing data and the fire alarm from the other four sensor nodes. For simplicity， four sensor nodes communicate with the base station directly in one?hop communication. Each TelosB sensor node has a 16 bit 8 MHz mirocontroller， an RF transceiver compliant with IEEE 802.15.4， and four sensing units. These sensing units sense temperature （[-40 °C-123.8 °C]）， relative humidity （RH）， infrared light （320 nm-1100 nm）， and visible light （320 nm-730 nm）. Hence， each sensor node can monitor the four attributes of a forest fire.
The architecture of the artificial neural network is shown in Fig. 2.
The back?propagation artificial neural network in the fire detection module is a two?layer network. There are four neurons in the first layer， because of the four sensing units in a TelosB sensor node， and one neuron in the second layer. The transfer function for the first layer is log?sigmoid function （f1 in Fig. 2） and is given by：
[f（x）=1/（1+e-x）] （13）
The transfer function for the second layer is a linear function （f2 in Fig. 2） and is given by：
[f（x）=x] （14）
4.3 User Interface Module
The user interface module is responsible for displaying the raw sensing data to the user. First， the sensing data and fire alarm are transmitted from the base station to the user. The data flow is shown in Fig. 3. Sensing data from sensor nodes are transmitted to the base station by wireless communication. The base station is a gateway between WSNs and the Internet and forwards the sensing data to a user client. The medium between the base station and user client is the Internet. Therefore， the user may be located far away from the fire?detection system. Socket communication is facilitated by Java. Next， the user interface module displays the sensing data to the user. The graphical interface draws curves for each typed of sensing data over time. The graphical interface is refershed accourding to the arrival of new sensing data. Therefore， the curves of the graphical interface are synchronous with the sensing units on sensor nodes. Each type of sensing data is displayed by a tab in the graphical interface. 5 Conclusion
A forest fire can threaten forest resources and human life. This threat can be mitigated by timely and accurate alarms. WSNs are widely used for environmental monitoring； therefore， we use a WSN for forest fire detection. To increase the accuracy of the detection system， we propose multi?criteria detection for forest fires. In this paper， multi?criteria detection is implemented by the artificial neural network algorithm. To power the sensor nodes in the forest where only intermittent sunlight is available， we develop a solar battery module. We developed a prototype of the proposed system comprising solar batter module， fire detection module， and user interface module.
References
[1] I. Akyildiz， W. Su， Y. Sankarasubramaniam， and E. Cayirci， “A survey on sensor networks，” IEEE Communications Magazine， vol. 40， no. 8， pp. 102-114， 2002.
[2] B. Son， Y. Her， and J. Kim， “A design and implementation of forest?fire surveillance system based on wireless sensor network for south korea mountains，” International Journal of Computer Science and Network Security， vol. 6， no. 9， pp. 124-130， 2006.
[3] D. Doolin and N. Sitar， “Wireless sensors for wildfire monitoring，” in Proceedings of SPIE， vol. 5765， no. 2， 2005， pp. 477-484.
[4] C. Hartung， R. Han， C. Seielstad， and S. Holbrook， “Firewxnet： A multitiered portable wireless system for monitoring weather conditions in wildland fire environments，” in Proceedings of the 4th international conference on Mobile systems， applications and services， Uppsala， Sweden， 2006.
[5] M. Hefeeda and M. Bagheri， “Wireless sensor networks for early detection of forest fires，” in MASS ’07， Pisa， Italy， 2007， pp. 1-6.
[6] G. Pfister， “Multisensor/multicriteria fire detection： a new trend rapidly becomes state of the art，” Fire Technology， vol. 33， pp. 115-139， 1997.
[7] S. Roundy， M. Strasser， and P. Wright， Powering ambient intelligent network， Germany： Springer， 2005， pp. 271-299.
[8] M. Minami， T. Morito， and H. Morikawa， “Solar biscuit： A batteryless wireless sensor network system for environmental monitoring applications，” in the 2nd international Workshop on Networked Sensing Systems， San Diego， USA， 2005.
[9] P. Dutta， J. Hui， and J. Jeong， “Trio： enabling sustainable and scalable outdoor wireless sensor network deployments，” in Information Processing in Sensor Networks， Nashville， TN， USA， April 19-21， 2006， pp. 407-415.
[10] V. Raghunathan， A. Kansal， J. Hsu， J. Friedman， and M. Srivastava， “Design considerations for solar energy harvesting wireless embedded systems，” in Information Processing in Sensor Networks， UCLA， Los Angeles， USA， April 25?27， 2005. [11] X. Jiang， J. Polastre， and D. Culler， “Perpetual environmentally powered sensor networks，” in Information Processing in Sensor Networks， UCLA， Los Angeles， USA， April 25?27， 2005.
[12] P. Corke， P. Valencia， P. Sikka， T. Wark， and L. Overs， “Long?duration solar?powered wireless sensor networks，” in Proceedings of the 4th Workshop on Embedded Networked Sensors， Cork， Ireland， June 25-26， 2007， pp. 33-37. doi： DOI： 10.1145/1278972.1278980.
[13] J. Taneja， J. Jeong， and D. Culler， “Design， modeling， and capacity planning for micro?solar power sensor networks，” in Information Processing in Sensor Networks， St. Louis， USA， April 22?24， 2008， pp. 407-418.
[14] S. S. Gyltner and K. Stetter， “A fire detection system and method for early detection of fire，” International Publication Number WO 2009/080581 A1， World Intellectual Property Organization， 2009.
[15] E. Valendik and I. Kosov， “Effect of thermal radiation of forest fire on the environment，” Contemporary problems of Ecology， vol. 1， no. 4， pp. 399-433， 2008. doi： 10.1134/S1995425508040012.
[16] G. Xing， R. Tan， B. Liu， J. Wang， X. Jia， and C.?W. Yi， “Data fusion improves the coverage of wireless sensor networks，” in Proceedings of the 15th annual international conference on Mobile computing and networking， Beijing， China， 2009， pp. 157-168.
[17] L. Freitas， A. Coimbra， V. Sacramento， S. Rosseto， and F. Costa， “Sol： a data fusion protocol in wireless sensor networks for controlled environment，” in IEEE INFOCOM Workshops， Rio de Janeiro， Brazil， 2009， pp. 1-2. doi： 10.1109/INFCOMW.2009.5072191.
[18] H. Luo， H. Tao， H. Ma， and S. Das， “Data fusion with desired reliability in wireless sensor networks，” IEEE Transactions on Parallel and Distributed Systems， vol. 22， no. 3， pp. 501-513， 2011. doi： 10.1109/TPDS.2010.93.
Manuscript received： 2015?04?22
A forest fire is a severe threat to forest resources and human life. In this paper， we propose a forest?fire detection system that has an artificial neural network algorithm implemented in a wireless sensor network （WSN）. The proposed detection system mitigates the threat of forest fires by provide accurate fire alarm with low maintenance cost. The accuracy is increased by the novel multi?criteria detection， referred to as an alarm decision depends on multiple attributes of a forest fire. The multi?criteria detection is implemented by the artificial neural network algorithm. Meanwhile， we have developed a prototype of the proposed system consisting of the solar batter module， the fire detection module and the user interface module.
Keywords
forest fire detection； artificial neural network； wireless sensor network
W1 Introduction
ireless sensor networks （WSNs） have been the focus of research over the past few years because of their potential in environmental monitoring， target tracking， and object detection [1]. WSNs have also been studied in the context of detecting forest fires， which threaten forest resources and human life. WSNs are not costly and can detect forest fires in real time， unlike current detection methods based on human observation and unlike spot weather forecasts or even satellite monitoring. WSNs can also provide information about environmental conditions within the forest， which is useful for predicting forest fires [2]. Moreover， forest fire detection and prediction is associated with specific location information provided by individual sensor nodes.
Although some practical experiments have been conducted using WSNs to collect sensed data from a forest fire [3]-[5]， there are still some challenges to using WSNs for this purpose. A fire detector may sound an alarm based on a simple threshold， which gives rise to false alarms even though the sensing unit of the fire detector may be highly sensitive. False alarms occur for two main reasons：
·A photoelectric smoke sensing unit is sensitive to white aerosol particles from a smoldering fire but also to dust [6].
·Environmental conditions in the forest often severely disturb the normal behavior of the sensing unit. Sunlight and artificial light are primary sources of interference with the flame?sensing unit.
Limited power supply to sensor nodes makes it difficult to detect fires over a long period of time. The potential energy sources for sensor nodes can be classified according to whether they store energy within the sensor nodes （e.g.， in a battery）， distribute power to the sensor node through a wire， or scavenge available ambient power （e.g， using a solar battery on the sensor node）. Considering the volume of the sensor node， manner of deployment， and forest conditions， the solar battery is one of the most promising sources of energy for detecting forest fires over a long period of time. However， existing works on solar batteries for sensor nodes， e.g.， [8]-[13]， overlook the problem of intermittent sunlight in the forest. In this paper， we propose a forest fire detection system that includes an artificial neural network algorithm implemented in a WSN. Overall， the main contributions of this paper are as follows：
·The multi?criteria detection depends on multiple attributes of a forest fire and is introduced into WSNs to increase the accuracy of detecting a forest fire.
·An artificial neural network algorithm is used to fuse sensing data that corresponds to multiple attributes of a forest fire into an alarm decision.
·We introduce the principle of the proposed system as well as a prototype comprising TelosB sensor nodes and a solar battery to power the WSN.
2 System Description
For the sake of clarity， we consider a WSN with only one base station and hundreds of sensor nodes. Because a WSN with multiple base stations can be regarded as multiple WSNs （each comprising one base station and corresponding sensor nodes）， the proposed system can also be implemented if the WSN has multiple base stations. Therefore， there are [n] sensor nodes in the WSN， each denoted [sj]， [1≤j≤n]. A forest fire [f] has [l] attributes， each denoted [rfi]， [1≤i≤l]. Attribute [rfi] can be sensed by the sensing unit [ui]. A [ui] on an [sj] is denoted [uji]. The output sensing data of [uji] is denoted [oji]. For simplicity， we assume that [sj] has [l] types of sensing units covering [l] attributes of the forest fire. We use a multilayer back?propagation artificial neural network to fuse sensing data [oji]. The total number of layers in the artificial neural network is denoted [m]. The input vector of the [k]th layer， [1≤k≤m]， is denoted [Ak-1]. The output vector of the [k]th layer is denoted [Ak]. Therefore， [A0] and [Am] represent the input and output of the artificial neural network， respectively. The alarm decision is denoted[ad].
3 Proposed Forest Fire Detection Method
In our proposed system， detection is made more accurate by using multiple criteria， which means the[ad] is based on multiple criteria of the forest fire. Multi?criteria detection is implemented by the artificial neural network algorithm. Because of the artificial neural network， the proposed system has low overhead and has self?learning capabilities； that is， it trains itself to build up the relations between sensing data and correct [ad].
3.1 MultiCriteria Detection
In a system that depends on one attribute of a forest fire to raise alarms， there is a high probability of false alarms because of inherent system drawbacks or external disturbances. To overcome such drawbacks and counter external disturbances， the system must take into account the multiple attributes of a forest fire. This is referred to as multi?criteria detection （Definition 1）. With multi?criteria detection， multiple attributes of a forest fire are sensed by different types of sensing unit. Therefore， a sensing unit that has been interfered with cannot raise a false alarm. Together， multiple sensing units confirm an alarm. Multi?criteria detection increases the accuracy of detecting a forest fire. Definition 1 （Multi?criteria detection）. Multi?criteria detection is represented as a function with multiple arguments [rf1，rf2，...，rfl] ， which refer to the attributes of forest fire [f] ， and one [ad]， given by：
[ad=f（rf1，rf2，...，rfl）] （1）
The attributes [rf1，rf2，...，rfl] could be any combination of the attributes of a forest fire. The directly sensed attributes of a forest fire are flame and heat， which are sensed by the flame sensing unit and heat sensing unit. The flame emits visible light， but the forest fire also emits a lot of radiation， the spectral distribution of which is the radiation intensity with respect to different wavelengths and is not uniform. In theory， the radiation intensity is determined by the temperature of the fire. The radiation intensity from a blackbody with respect to the wavelength and temperature is described by Planck’s radiation law （2）， where [h] is Planck’s constant， [c] is the speed of light， [λ] is the wavelength， [k] is Boltzmann’s constant， and [T] is the temperature [14].
[I=2hc2λ5exp[hc/λkT]-1] （2）
Therefore， radiation intensity can be the basis for detecting a forest fire given that the typical temperature of a forest fire is [600 °C-1000 °C] [15]. The ultraviolet sensing unit and infrared flame sensing unit work by detecting radiation intensity. Other attributes that can be used to identify a forest fire include combustion products. It is well known that a forest fire gives off bursts of carbon dioxide， carbon monoxide， water vapor， and dust.
3.2 Artificial Neural Network Algorithm
We use the multi?layer back?propagation artificial neural network for multi?criteria detection. Although data fusion in WSNs has been covered in much of the literature [16]-[18]， the topic has not been considered in the context of forest fires. A multi?layer back?propagation artificial neural network is widely used to emulate the non?linear relationship between its input and output. However， computation in this kind of network is not complex because the network is a combination of neurons dealing with simple functions. Moreover， multi?layer back?propagation artificial neural network is capale of self?learning， which means it can train itself to build up relations between the inputs and desired targets.
3.2.1 Making an Alarm Decision
Without loss of generality， we assume that the multi?layer back?propagation artificial neural network is implemented on [sj] with [l] types of sensing units that cover [l] attributes of the forest fire. Sensing data [oji] of [uji] on [sj] corresponds to [rfi] of the forest fire. [A0=[a01，...，a0l]T] （3）
For clarity， let all [oji]， [1≤i≤l] comprise a column vector [A0] （3）， where [a0i=oji]. Vector [A0] is the input to the multiple layer artificial neural network.
Specifically， [A0] is the input to the first layer of the multi?layer artificial neural network. In the first layer， [A0] is multiplied by weight matrix [W1] with dimension [s1×l] and bias vector [B1] ， including [s1] neurons in the first layer. The intermediate computation result of the first layer is denoted [N1] and is given by：
[N1=W1A0+B1] （4）
Then， [N1] is sent to transfer function [F1]， which may be a linear or nonlinear. That is， [F1] may be a hard?limit function or sigmoid function depending on the specific problem it needs to solve. In general， transfer functions in the multi?layer artificial neural network are easy to compute. Transfer function [F1] operates on every element of [N1]. The result of transfer function[F1]， denoted [A1]， is the output of the first layer：
[A1=F1（N1）=F1（n11）?F1（n1s1）] （5）
The fusion of sensing data proceeds in the second layer of the multi?layer artificial neural network. The output [A1] of the first layer becomes the input of the second layer. The calculation process of the second layer is similar to that of the first layer except the second layer has its own [W2]， [B2]， and [F2]. In general， the calculation of the [i]th layer is given by：
[Ai=Fi（WiAi-1+Bi）， 1≤i≤m-1 ] （6）
where [m] is the number of layers in the artificial network.
For a decision to be made on whether there is a forest fire or not， the output of the [m]th layer， [Am] （7）， is confined to one element.
[ad=Am=Fm（WmAm-1+Bm）] （7）
This is done by letting the m th layer contain only one neuron. If the alarm decision is confined to a Boolean value， we need to choose the transfer function， the output of which is a Boolean value， for the m th layer， such as the hard limit function.
An alarm decision is made by inputting the attributes of a forest fire into the multi?layer back?propagation artificial neural network， as shown in Theorem 1.
Theorem 1. The [ad]of the multi?criteria detection is computed by the recursive （8） given sensing data [A0=[a01，...，a0l]T] corresponding to [l] attributes of forest fire[f].
[ad=Fm（WmAm-1+Bm）Ai=Fi（WiAi-1+Bi），1≤i≤m-1] （8）
3.2.2 Self?Learning Capability
Given sensing data that corresponds to multiple attributes of a forest fire and given correct alarm decisions， the multi?layer back?propagation artificial neural network trains itself to build relationships between the sensing data and correct alarm decisions. However complex the relationship， it is easy for the multi?layer back?propagation artificial neural network to fulfil the task. Essentially， self?learning means having the output of the multi?layer back?propagation artificial neural network approximate the target output by adjusting the weight matrixes and biases. This adjustment is made in order to minimize the mean?square error （MSE） between the output and target output. Suppose [q] inputs are denoted[A0i]， [1≤i≤q]. Corresponding to [A0i]， the output is denoted [Ami]， and the target output is denoted [Tmi]. Thus， the MSE for the [i]th iterated adjustment is：
[MSE（i）=E（Tmi-Ami）T（Tmi-Ami）] （9）
For clarity， the adjustment of the weight matrixes and biases is expressed by Theorem 2.
Theorem 2. In the self?learning， the [i]th iterated adjustment of the weight matrixes and biases is conducted according to （10） and （11）， where [α] is a constant for the learning rate， and [?（j）] is computed by the recursive （12） where [D（j）=diag（Fj）′（nj1），...，（Fj）′（njsj）].
[Wj（i）=Wj（i-1）-α?（j）（Aj-1）T， 1≤j≤m] （10）
[Bj（i）=Bj（i-1）-α?（j）， 1≤j≤m] （11）
[?（j）=D（j）（Wj+1）T?（j+1），1≤j≤m-1?（m）=-2D（m）（Tmi-Ami）] （12）
After self?learning， the multi?layer back?propagation artificial neural network builds up a mathematical relationship between the sensing data and correct alarm decisions. Then， the artificial neural network can make an accurate alarm decision.
4 Implemented Prototype
We have developed a prototype of the forest fire detection system using an artificial neural network in a WSN. The system mainly comprises three parts： the solar battery， fire?detection module， and user interface.
4.1 Solar Battery
To consistently power the unattended sensor nodes deployed in a forest where only intermittent sunlight is available， we develop a solar battery （Fig. 1）. The energy from the solar panel is buffered by the super capacitor. When the energy in the supper capacitor reaches a threshold， the super capacitor starts to recharge the Li?ion battery. Because of the intermittent sunlight in the forest， the energy produced by the solar panel is not enough to recharge the battery. If not buffered in the super capacitor， this energy is wasted. On the other hand， the charge?discharge cycles of the Li?ion battery are limited. It is better to charge a Li?ion battery until it is full； otherwise， the life of the battery decreases. On the contrary， the super capacitor has almost infinite charge?discharge cycles and is ideal for frequently pulsing applications.
Here we discuss implementation of the solar battery in detail. The solar panel of the battery is 110 × 95 mm and comprises eight cells connected in parallel and generating 550 mA at 2 V. Theoretically， the maximum energy generated in one hour can sustain a sensor node for 26 days， i.e.， 1100 mAh/（0.53 mA × 3.3 V × 24 h）， provided that the sensor nodes work on a 10% duty cycle with an average current of 0.53 mA. Energy from the solar panel is buffered by two 150 F 2.5 V super capacitors wired in parallel. A 3.7 V 700 mAh Li?ion battery is used to continually save energy. The fully charged Li?ion battery can power a sensor node working on a 10% duty cycle for 55 days， i.e.， [700 mAh/（0.53 mA×24 h）] . We chose MAX1674 and ISL6292 integrated circuits as the DC?DC converters， which have a conversion efficiency of around 90%. 4.2 Fire Detection Module
The fire detection module is responsible for multi?criteria detection. The module comprises five TelosB sensor nodes， four of which monitor the forest fire. That is， they convert the attributes of a forest fire into sensing data. The multi?layer back?propagation artificial neural network is implemented on each individual sensor node because the sensor node is endowed with four types of sensing units. However， for the purpose of analysis， raw sensing data besides the fire alarm are transmitted to users. The last sensor node acts as the base station， collecting sensing data and the fire alarm from the other four sensor nodes. For simplicity， four sensor nodes communicate with the base station directly in one?hop communication. Each TelosB sensor node has a 16 bit 8 MHz mirocontroller， an RF transceiver compliant with IEEE 802.15.4， and four sensing units. These sensing units sense temperature （[-40 °C-123.8 °C]）， relative humidity （RH）， infrared light （320 nm-1100 nm）， and visible light （320 nm-730 nm）. Hence， each sensor node can monitor the four attributes of a forest fire.
The architecture of the artificial neural network is shown in Fig. 2.
The back?propagation artificial neural network in the fire detection module is a two?layer network. There are four neurons in the first layer， because of the four sensing units in a TelosB sensor node， and one neuron in the second layer. The transfer function for the first layer is log?sigmoid function （f1 in Fig. 2） and is given by：
[f（x）=1/（1+e-x）] （13）
The transfer function for the second layer is a linear function （f2 in Fig. 2） and is given by：
[f（x）=x] （14）
4.3 User Interface Module
The user interface module is responsible for displaying the raw sensing data to the user. First， the sensing data and fire alarm are transmitted from the base station to the user. The data flow is shown in Fig. 3. Sensing data from sensor nodes are transmitted to the base station by wireless communication. The base station is a gateway between WSNs and the Internet and forwards the sensing data to a user client. The medium between the base station and user client is the Internet. Therefore， the user may be located far away from the fire?detection system. Socket communication is facilitated by Java. Next， the user interface module displays the sensing data to the user. The graphical interface draws curves for each typed of sensing data over time. The graphical interface is refershed accourding to the arrival of new sensing data. Therefore， the curves of the graphical interface are synchronous with the sensing units on sensor nodes. Each type of sensing data is displayed by a tab in the graphical interface. 5 Conclusion
A forest fire can threaten forest resources and human life. This threat can be mitigated by timely and accurate alarms. WSNs are widely used for environmental monitoring； therefore， we use a WSN for forest fire detection. To increase the accuracy of the detection system， we propose multi?criteria detection for forest fires. In this paper， multi?criteria detection is implemented by the artificial neural network algorithm. To power the sensor nodes in the forest where only intermittent sunlight is available， we develop a solar battery module. We developed a prototype of the proposed system comprising solar batter module， fire detection module， and user interface module.
References
[1] I. Akyildiz， W. Su， Y. Sankarasubramaniam， and E. Cayirci， “A survey on sensor networks，” IEEE Communications Magazine， vol. 40， no. 8， pp. 102-114， 2002.
[2] B. Son， Y. Her， and J. Kim， “A design and implementation of forest?fire surveillance system based on wireless sensor network for south korea mountains，” International Journal of Computer Science and Network Security， vol. 6， no. 9， pp. 124-130， 2006.
[3] D. Doolin and N. Sitar， “Wireless sensors for wildfire monitoring，” in Proceedings of SPIE， vol. 5765， no. 2， 2005， pp. 477-484.
[4] C. Hartung， R. Han， C. Seielstad， and S. Holbrook， “Firewxnet： A multitiered portable wireless system for monitoring weather conditions in wildland fire environments，” in Proceedings of the 4th international conference on Mobile systems， applications and services， Uppsala， Sweden， 2006.
[5] M. Hefeeda and M. Bagheri， “Wireless sensor networks for early detection of forest fires，” in MASS ’07， Pisa， Italy， 2007， pp. 1-6.
[6] G. Pfister， “Multisensor/multicriteria fire detection： a new trend rapidly becomes state of the art，” Fire Technology， vol. 33， pp. 115-139， 1997.
[7] S. Roundy， M. Strasser， and P. Wright， Powering ambient intelligent network， Germany： Springer， 2005， pp. 271-299.
[8] M. Minami， T. Morito， and H. Morikawa， “Solar biscuit： A batteryless wireless sensor network system for environmental monitoring applications，” in the 2nd international Workshop on Networked Sensing Systems， San Diego， USA， 2005.
[9] P. Dutta， J. Hui， and J. Jeong， “Trio： enabling sustainable and scalable outdoor wireless sensor network deployments，” in Information Processing in Sensor Networks， Nashville， TN， USA， April 19-21， 2006， pp. 407-415.
[10] V. Raghunathan， A. Kansal， J. Hsu， J. Friedman， and M. Srivastava， “Design considerations for solar energy harvesting wireless embedded systems，” in Information Processing in Sensor Networks， UCLA， Los Angeles， USA， April 25?27， 2005. [11] X. Jiang， J. Polastre， and D. Culler， “Perpetual environmentally powered sensor networks，” in Information Processing in Sensor Networks， UCLA， Los Angeles， USA， April 25?27， 2005.
[12] P. Corke， P. Valencia， P. Sikka， T. Wark， and L. Overs， “Long?duration solar?powered wireless sensor networks，” in Proceedings of the 4th Workshop on Embedded Networked Sensors， Cork， Ireland， June 25-26， 2007， pp. 33-37. doi： DOI： 10.1145/1278972.1278980.
[13] J. Taneja， J. Jeong， and D. Culler， “Design， modeling， and capacity planning for micro?solar power sensor networks，” in Information Processing in Sensor Networks， St. Louis， USA， April 22?24， 2008， pp. 407-418.
[14] S. S. Gyltner and K. Stetter， “A fire detection system and method for early detection of fire，” International Publication Number WO 2009/080581 A1， World Intellectual Property Organization， 2009.
[15] E. Valendik and I. Kosov， “Effect of thermal radiation of forest fire on the environment，” Contemporary problems of Ecology， vol. 1， no. 4， pp. 399-433， 2008. doi： 10.1134/S1995425508040012.
[16] G. Xing， R. Tan， B. Liu， J. Wang， X. Jia， and C.?W. Yi， “Data fusion improves the coverage of wireless sensor networks，” in Proceedings of the 15th annual international conference on Mobile computing and networking， Beijing， China， 2009， pp. 157-168.
[17] L. Freitas， A. Coimbra， V. Sacramento， S. Rosseto， and F. Costa， “Sol： a data fusion protocol in wireless sensor networks for controlled environment，” in IEEE INFOCOM Workshops， Rio de Janeiro， Brazil， 2009， pp. 1-2. doi： 10.1109/INFCOMW.2009.5072191.
[18] H. Luo， H. Tao， H. Ma， and S. Das， “Data fusion with desired reliability in wireless sensor networks，” IEEE Transactions on Parallel and Distributed Systems， vol. 22， no. 3， pp. 501-513， 2011. doi： 10.1109/TPDS.2010.93.
Manuscript received： 2015?04?22