Research On Algorithms For Anomalous Event Detection And Anomaly Interpretation In Wireless Sensor Networks

Wireless sensor network(WSN)is a self-organized wireless network composed by a large number of cheap,small sensor nodes.Wireless sensor network can detect anomalous events that occur in the monitoring area by sensing,collecting and analyzing monitoring data.The existing event detection techniques recognize anomalous events based on setting a decision threshold for a single observation attribute,or based on a changing trend of data,which are not extensible in the applications involving multiple types of sensors.In addition,there are few studies related to abnormal event interpretation in the field of anomalous event detection in WSN.The existing algorithms can only find anomalous events,lacking the explanations for the causes of event,the abnormal features,the correlations among features and so on,and it is difficult for decision makers to make use of it.Therefore,it has theoretical significance and practical value to study the mechanism of anomaly interpretation in WSN.To improve the quality of event detection and make up for the lack of anomaly interpretation in existing algorithms,we propose an algorithm for anomalous event detection based on multi-attribute correlations,and an algorithm for anomalous event interpretation based on correlated subspaces.Besides,we design and implement a prototype system for abnormal event detection and anomaly interpretation of WSN.The main contributions of the dissertation are summarized as follows.(1)Considering that the existing algorithms detect anomalous events based on spatio-temporal correlations,and ignore the impacts of intrinsic correlations of non spatio-temporal attributes on detection results,which lead to a problem of high false alarm rate.This paper proposes an algorithm for anomalous event detection based on multi-attribute correlation(MACAED).First,a dependency model of non spatio-temporal attributes is constructed based on Bayesian network.In this model,the dependency structure of attributes is obtained by structure learning process.And by parameter learning,the conditional probability table of each node can be calculated,which indicates the probability dependencies of different non spatio-temporal attributes.Then,we propose a concept named attribute correlation confidence to measure the similarity between the current sensor reading and the sample data attribute pattern.Finally,the abnormal events can be detected on the basis ofspatio-temporal correlations and attribute correlations.Experimental results on real dataset show that the MACAED algorithm improves the accuracy from 2% to 6.2%compared with existing algorithms,with a decrease of 44.5% in the false alarm rate.And it suppresses the impact of interference events effectively.(2)Aiming at the problem that most event detection algorithms do not explain the anomalous events well.This paper proposes an algorithm for anomalous event interpretation based on correlated subspaces(CSAEI).First,we calculate the anomaly score of the subspaces based on the separability,and the anomaly scores are weighted to filter out the low dimension correlated subspaces,which are more interesting to decision makers.Then,we propose a concept named subspace correlation degree based on the principle of conditional independence,and design an algorithm for anomalous subspace search(SCDASS)based on this measure.Experimental results on both synthetic dataset and real dataset show that the performance of SCDASS algorithm is significantly improved compared with traditional algorithms based on scoring and searching,and it has strong scalability.Finally,on the basis of SCDASS algorithm,an anomaly event interpretation algorithm CSAEI is proposed.Experimental results on real dataset show that,this algorithm can output the interpretation of anomalous events,which provides a favorable basis for the decision makers to analyze the causes of anomalous event.(3)In order to verify the feasibility of the algorithms proposed in this paper,we design and implement an anomalous event detection and interpretation system for WSN.This system can not only detect abnormal events effectively,reducing the impact of noise data and interference factors,but also provide an interface showing the region where abnormal events occur.Besides,this system interprets the reasons for the occurrence of abnormal events.