Threshold Based Data Aggregation Algorithm To Detect Rainfall Induced Landslides

Transcription

1 Threshol Base Data Aggregation Algorithm To Detect Rainfall Inuce Lanslies Maneesha V. Ramesh P. V. Ushakumari Department of Computer Science Department of Mathematics Amrita School of Engineering Amrita School of Engineering Amrita University Amrita University Kerala, Inia Kerala, Inia Abstract Lanslies are one of the environmental isasters that cause massive estruction of human life an infrastructure. Real time monitoring of a lanslie prone areas are necessary to issue fore warning. To accomplish real time monitoring, massive amount of ata have to be collecte an analyze within a short span of time. This work has evelope a metho for effective ata collection an aggregation by implementing threshol alert levels. The sampling rates of threshol alert levels will etermine amount of ata collecte an aggregate which will reuce the power consumption by each wireless sensor noes. This work also helps to etermine the appropriate sampling rates for each threshol level, an the expecte number of ata packets in the queue. The time elay in receiving the ata packet at the analysis station can be etermine by using the value of expecte number of ata packets in the queue. 1 Introuction Environmental isasters are largely unpreictable an occur within very short spans of time. Fore warning of environmental isasters is more challenging than other applications ue to the fact that they occur within a very short span. Therefore technology has to be evelope to capture relevant signals in a very short span of time. Wireless sensors are one of the cutting ege technologies that can respon to rapi changes of ata. The fore warning of an environmental risk can be one if an only if the relevant ata within very short time spans are quickly capture with high time resolution, processe an transmitte to the analysis station. Wireless sensor network technology has all the above mentione capabilities. However, it has its own limitations such as relatively low amounts of battery power an low memory availability compare to many existing technologies. It oes, though, have the avantage of eploying sensors in hostile environments with a bare minimum of maintenance. This fulfills a very important nee for any real time monitoring. The scenario uner consieration for this work is lanslies. We aim to eploy wireless sensor networks in a lanslie prone area to etect lanslies 1. One of the main concerns for the successful implementation of a lanslie etection application is how to hanle the ata packets receive from each of the wireless sensor noes such that an accurate etection of the scenario or event is mae 1 This work has been partially fune by the WINSOC project, a Specific Targete Research Project (Contact Number ) co-fune by the INFSO DG of the European Commission within the RTD activities of the Thematic Priority Information Society Technologies.

2 possible. This paper provies information about a ata aggregation technique that can be use for a wireless sensor network eploye to etect rainfall inuce lanslies. The ata packets collecte from the eployment site have to be efficiently transmitte an aggregate for goo preiction results. This paper introuces one metho for ata aggregation which uses a threshol alert level to aggregate the ata packets. For the case of a wireless sensor network implemente using threshol alert levels, it fins the probability of ata packets being in the queue an also the expecte number of ata packets that will be in the queue at any ranom time. 2 Lanslie Detection Using Wireless Sensor Network Lanslie is a short live an suenly occurring phenomena that has catastrophic effects all aroun the worl. Inia faces lanslie every year with an annual loss of about US $400 million. Although there are ifferent types of lanslies, this work mainly concentrates on rainfall inuce lanslies which occur commonly in Inia. The wireless sensor noes will be eploye in a lanslie prone area. Each of the wireless sensor noes will be connecte to a sensor column (see Figure 1) containing various geophysical sensors. The ata from the geophysical sensors will be sample an collecte by the wireless sensor noes in the lowest layer of the wireless sensor network which has two layer hierarchy with low level noes an high level noes. The ata packets collecte by the low level noes will be sent to the higher layers for further processing. Each ata packet will consist of heterogeneous ata receive from ifferent geophysical sensors connecte to the low level noes. The information receive from basic ata analysis of the geo physical sensors in the sensor column inclues the time stamp, sequence number of the ata packet, sensor ata, sensor sensitivity an location. The network contains a two layer hierarchy consisting of a lowest layer which carries out the functions of ata collection an transmission to the Figure 1: Sensor Column higher layers. The upper layer will aggregate the ata an forwar it to the sink noe. Using a Wi-Fi connection the sink noe will then transmit the ata to a network server connecte to a VSAT network near the eployment site at Iukki, Kerala, South Inia. The satellite network will transmit the ata to a remote ata analysis center for further processing. The remote analysis station will be situate at at AMRITA University. 3 Data Aggregation In A Lanslie Scenario Real time applications require continuous monitoring, which in turn generates large amount of ata packets for analysis. The ata packets generate from the low level noes have to be efficiently aggregate an analyze by higher layer noes to etect or preict any emergence of a critical situation such as a lanslie. Threshol alert levels have been use for efficient ata aggregation with necessary ata.

3 3.1 Threshol Alert Levels Threshol base alert levels for the whole network have been evise for minimizing energy consumption an also for efficient ata hanling. In each of the threshol levels the ata packets collecte will be sample at ifferent rates. As the rate of ata packet collection an transmission changes the amount of energy rain will change accoringly. Three threshol alert levels have been evise for lanslie scenario. They are: Low alert level: During the low alert level, the external perceive risk is very low. Each sensor noe will transit between sleep an monitor states over a relatively longer time interval, T low. The rate of ata packet transmission from the cluster hea to the sink (the α% of total ata receive by the cluster hea from the low level noes) will be low (approximately 1 10%) in this alert level. Meium alert level: During the meium alert level, the external perceive risk is average. Each sensor noe will transit between sleep an monitor states over a meium time interval, T meium. The rate of ata packet transmission from the cluster hea to the sink (the α% of total ata receive by the cluster hea from the low level noes) will be meium (approximately 50 60%) in this alert level. High alert level: During the high alert level, the external perceive risk is very high. Each sensor noe will transit between sleep an monitor states over a shorter time interval, T high. The rate of ata packet transmission from the cluster hea to the sink (the α% of total ata receive by the cluster hea from the low level noes) will be high (approximately %) in this alert level. 3.2 Data Aggregation Metho As the network initiates each low level wireless noe will sample the geophysical sensors attache to the wireless sensor noe. The collecte ata will be converte as a ata packet consisting of a maximum of eight heterogeneous ata. The ata in the ata packet will be analyze using multivariate analysis. The result of the analysis will be compare with a preetermine threshol of the network alert level which is apriori known to the wireless noe (The preetermine threshol values will be foun using our laboratory experiments an fiel ata.). If the ata analysis result provies parameter value more than the preetermine threshol value, then the noe will change its state from monitor to active accoring to the state transition (as shown in Figure 2)evelope by Maneesha et.al[1]. Otherwise it will rop the packet an return to sleep state for a perio of time etermine by the alert level (such as T low, T meium, ort high ). The noe which has change its state to active will sen a long pulse to wake up all its group members in the same cluster. The cluster members on receiving the wake up call will change their state to active an their ata packet will be sent to the cluster hea. So continuous sampling of geophysical sensors may not be followe by transmission of the ata packets, if the parameter value oes not cross the preetermine threshol value. The cluster hea will wait to receive the ata packets from its cluster members. The cluster hea can process each packet arriving from the flow an aggregate them incrementally until they are all processe, as in [3]. However, this will consume more processing power. To avoi that situation an save power, we are employing bulk processing of ata. As the ata packets from the cluster members are receive, the cluster hea will perform ata aggregation algorithm base on multivariate analysis. The aggregate values along with an α% of ata ( where α% is some fixe constant percentage for each alert level) receive from the low level noes will be forware by the cluster heas to the sink. The sink will not perform complicate ata aggregation algorithms. However, the network alert level will etermine by the sink noe. The amount of ata receive at the sink will iffer since the constant value of α% of ata packet will iffer with respect to the alert levels an also with respect to the number of chil noes attache

4 Figure 2: Low Level Noe State Transition to each cluster hea. The sink noe will etermine the ratio of the ata packets which is efine as the number of ata packets recommening an alert change ivie by the number of ata packets total. If this ratio excees certain threshols, the alert level can change. The state of the network will thereby alter accoring to the ratio of the ata packets that have recommene an alert level change. The entire wireless network can therefore change its threshol alert level from one state to another. The alert level change of the whole network from high to low is a very rare change. As a result, the change from high to low alert level for the whole network will occur if an only if the maximum number of ata packets show such a result is necessary. The alert level change from meium to high has to be ecie very rapily ue to the fact that lanslies may occur in a short span. As a result, a minimum number of ata packets will etermine the transition from meium alert to high alert. The exact values of the ratio of ata packets that recommen for an alert change from one alert level to another will be etermine through simultion, laboratory experiments, an the collection of fiel ata. The alert level of the network will change if the ratio of ata packets recommening the change of alert level is greater than or equal to the pre etermine ratio etermine from continuous simulation an laboratory tests. Otherwise it will remain in the same alert level. The alert level change means all wireless sensor noes in the whole wireless network will change their alert level. Due to an alert level transition, low level noes have to change the sampling rate of the geophysical sensors an also the existing preetermine threshol value to the new alert level s preetermine threshol value. The frequency of ata transmitte in each alert level is ifferent. In the high alert level, the arrival rate at the cluster hea an sink noe will be a maximum while the arrival rate at the low alert level will be a minimum. That of the meium alert level will be an intermeiate value between the arrival rates of the high an low alert levels. This makes the whole arrival pattern non-homogeneous an also these ifferent rates of transmission of ata packets will contribute towars minimizing the whole network s energy consumption, which will in turn reuce the number of ata packets to be processe by the whole network for a ranom amount of time. The ata packets to be processe for a ranom amount of time will be etermine mainly by the threshol levels. Therefore the energy use by the whole network will be much less compare to that use by a network that is active through out the eployment. The ata aggregation for the whole network consists of a three level aggregation algorithm. The three levels has been iscusse earlier are at low level noes, cluster heas, an sink noe. The three level ata aggregation algorithm will reuce the amount of multivariate analysis to be performe for the whole network. Initially, in low level noes, the multivariate analysis will be performe an the

5 ata, along with the result will be save in a buffer. The multivariate analysis will not be performe for the next ata set. It only etermines the ifference with the previous ata. As a result, at each point in time, the buffer will have one ata packet (initial or fresh ata from just after a whole network alert level transition), the noe s result of multivariate analysis an the latest ata. Once the ifference reaches the threshol, the noe will wake up all its cluster members. All of the cluster members will then sen the ata packet to the cluster hea an multivariate analysis will be performe for the group of ata. Thus, it is not necessary to perform multivariate analysis for each an every ata arrival at low level noes. This will reuce the energy consume for ata processing which will in turn reuce the network level energy consumption.at the sink noe, multivariate analysis is not performe. It etermines the ratio of ata packets that recommen transition to a new alert level. This is a simple operation that consumes very little processing power. 4 Data Aggregation in a Subtree A subtree consiere for our stuy consists of chil noes connecte to a cluster hea, i.e., the subtree will have only single hop ata packet transmissions. The arrival pattern of ata packets to the cluster hea are ifferent with respect to each alert level. The amount of ata packets waiting for service will change accoring to the arrival rate transition. This will become a critical parameter for any real time application because if the queue size excees the buffer size then ata loss will occur. The arrival rate which causes ata loss is not an efficient one since the purpose of ata aggregation is not attaine an also energy is lost collecting an transmitting the ata. Energy consumption minimization an efficient ata aggregation can be performe if the monitoring application is aware of the optimal arrival rate, queue length, waiting time, etc. This will help to avoi ata packet an energy loss. Data arrival to the cluster hea can be moele as a non-homogeneous Poisson process with arrival rate λ(t), at time t. The network of chil noes connecte to a cluster hea can be treate as a M/M/1 queue with N-Policy. The service time neee by a cluster hea is exponential with rate of service µ. The cluster hea starts servicing whenever ata from all or a percentage of the chil noes (N g ) arrive at time t or whenever N g (> 1) units accumulate in the queue. Otherwise the cluster hea will be in a sleep state. (A cluster hea uses only three state transitions. It oes not use the monitor state). As soon as the service of one batch is over, the cluster hea will continue the ata aggregation if there exists N g waiting ata. Otherwise the cluster hea will change its status to sleep state an starts the processing of ata when N g ata accumulate in the queue. Any new arrival after N g waiting ata will be lost to the system with probability 1. Let X(t) enote the number of ata packets in the queue at time t. Let Y (t) enote the state of the cluster hea. Then X(t) has the state space {0, 1, 2,..., N g 1, N g } an Y (t) can be efine as Y (t) = { Then the process 1, if theclusterheaisbusy 0, if theclusterheaisileorsleep Z(t) = {(X(t), Y (t)) ; t 0} (1) is a non-homogeneous birth/eath process over the state space E = {0, 1, 2,..., N g 1} {0, 1} (N g, 1). To stuy the process Z(t), efine P ij (t) = P r {(X(t), Y (t)) = (i, j)}, (i, j) E (2) The Forwar Kolmogorov Differential Difference Equations associate with Z(t) are the following: t P 00 (t) = λ(t)p 00 (t) + µp 01 (t) (3) t P i0 (t) = λ(t)p i0 (t)+λ(t)p i 10 (t)+µp i1 (t) (4) where i = 1, 2,..., N g t P01 (t) = (λ(t) + µ) P01(t)+λ(t)PNg 10(t)+µPNg1(t) (5)

6 t P i1 (t) = (λ(t) + µ) P i1 (t) + λ(t)p i 11 (t) (6) where i = 1, 2,..., N g 1 t P N g1 (t) = µp Ng1(t) + λ(t)p Ng 11(t) (7) In matrix notation, Equations 3 to 7 can be written as P (t) = T P (t) (8) with P (0) = P 0 where P (t) = [ The P 00 (t),, P Ng 10 (t), P 01 (t),, P Ng1 (t) ] expecte number of ata packets in the queue when the server is busy is given by: is a matrix of orer (2N g + 1) 1 with its erivative [ P (t) = t P 00 (t), t P 10 (t),..., ] t P N g1 (t) an P 0 is a vector with initial conition P 0 = (1, 0, 0,..., 0), which means initially there are none in the system. T is a (2N g + 1) (2N g + 1) rate matrix. The rate matrix T (5 x 5)is given by λ (t) λ (t) λ (t) λ (t) 0 0 T = µ 0 µ λ (t) λ (t) 0 0 µ 0 µ λ (t) λ (t) 0 0 µ 0 µ (9) has the unique solution. P (t) = T P (t) P (t) = P 0 e tt (10) The solution to the equation 11 will provie the probability at a time t. Using the probability, the expecte number of ata packets in the queue can be foun as follows: Expecte number of ata packets in queue when the cluster hea is in the sleep state is given by: E [x/y (t) = 0] = E [x/y (t) = 1] = N g 1 i=1 N g 1 i=1 ip i0 (12) ip i0 (13) Thus, the expecte number of ata packets in the queue at any ranom time is given by: E [x] = 1 E [x/y (t) = j] P (Y (t) = j) j=0 E [x] = E [x/y (t) = 0] P.0 + E [x/y (t) = 1] P.1 (14) where P.0 is the probability of the cluster hea being in the sleep state,p.1 is the probability of the cluster hea being in the busy state.from the above results, we will be able to fin the ifferent arrival rates, the waiting time istribution, etc. 5 Data Aggregation from the Cluster heas to a Sink Noe which can be compute as P (t) = Ce tt C 1 P 0 (11) where A = C 1 T C is the Joran Canonical form of T. To fin C 1 T C, the characteristics of T can be compute by iagonalizing the T matrix. Thus we get the characteristic roots as λ 1, λ 2,..., λ 2Ng an 0. Using these roots, compute the invertible matrix C such that C 1 T C is the Joran Canonical form of T. Let the sink noe be connecte to N c cluster heas. The ata packets from the cluster heas have to be collecte, analyze an forware to the satellite network by the sink noe. Each cluster hea, after completing its multivariate analysis for a group of ata at time t, will transmit an α% of ata packets receive from the low level noes an the result of the analysis to the sink noe. The number of ata packets sent from each cluster hea iffers with respect to the alert level an also with respect to the number of chil noes attache to the

7 cluster hea. Therefore, the number of ata packets arriving at the sink noe from each cluster hea will iffer. The sink will maintain one queue for each of the cluster heas since it has to hanle more than one ata packet from each of the queues at a time t. As a result, if you have N c cluster heas, the sink will maintain the same number of queues. The sink has to receive an analyze the ata in bulk, as a α% = n of the number of ata packets, from each queue an then etermine the ratio of ata packets recommen for state transition. If the ratio of ata packets recommene for the transition is greater than the ratio shown in Table??, then the whole network will change the alert level, otherwise the alert level remains the same. 6 Conclusion an Future Work This paper iscusses ata aggregation in a wireless sensor network, in particular the network use for the lanslie scenario. It iscusses the ifferent steps neee for ata aggregation, the expecte number of ata packets in the queue, an their arrival rates. The avantage of this algorithm is that it reuces the amount of energy consume by the whole network as well as helps in fining the arrival rate of ata packets. This work will be teste in our laboratory set up, through wireless network simulators, an through the stuy of fiel ata. In the future, this work will be extene to preict the arrival rates to be aopte for a lanslie scenario at the cluster hea level an at the sink noe level. This work can be extene to etermine bouns on the number of intermeiate noes if the number of low level noes is known. It can also be extene to stuy the elay incurre for ata analysis with respect to each arrival rate, with respect to changes in the ensity of the network, an with respect to other factors that have a irect effect on preicting any real time scenario. Venkat.,Factors an Approaches for Energy Optimize Wireless Sensor Network to Detect Rainfall Inuce Lanslies, International Conference on Wireless Networks, WORLDCOM-07, Las Vegas, USA, Jun [2] Terzis. Anreas., Ananarajah. Annalingam., Moore. Kevin., Wang. I-Jeng.,Slip Surface Localization in Wireless Sensor Networks for Lanslie Preiction, IPSN 06, USA, April 19-21, [3] Akkaya. Kemal., Younis. Mohame., Youssef. Moustafa.,Efficient Aggregation of Delay-Constraine Data in Wireless Sensor Networks, Department of Computer Science an Electrical Engineering, University of Marylan. [4] Zhang. Honghai., Hou. C. Jennifer.,On the Upper Boun of α-lifetime for Large Sensor Networks,ACM Transactions on Sensor Networks, Vol. 1, No. 2, Pages , November [5] Linsey. Stephanie., Raghavenra. Cauligi.,Sivalingam. Krishna.,Data Gathering in Sensor Networks using the Energy Delay Metric,Proceeings of the 15th International Parallel an Distribute Processing Symposium, IEEE, 2001 [6] Mehi, J., Stochastic Moels in Queuing Theory, Acaemic Press, An Imprint of Elsevier., 2n eition, References [1] Ramesh. Maneesha., Raj. Rehna., Freeman. Joshua., Kumar. Sangeeth., Rangan.