Distributed Outlier Detection And Analysis For Wireless Sensor Network

Outlier detection and analysis is one of the key problems in the research field of wireless sensor networks.There are two main sources of the sensor data being abnormal.One is that events can trigger the data becoming outliers.Another is that the soft failure of the sensor nodes can also lead to the data becoming outliers.Outliers can affect the users' decision,so that the research of outliers detection is the necessary measure in security of wireless sensor networks.This paper set up a perfect system of outlier detection and analysis,which consists of three detection algorithms.The algorithms are shown as follows.Distributed outlier detection algorithm based on credibility feedback.First,the algorithm uses spatial correlations of nodes' observations,evaluate the initial credibility by using the comparison results between the nodes and their neighbours.Then,the final credibility is evaluated based on credibility feedback and Bayesian theorem,and it can determine whether the nodes' observations are outliers or not.Finally,the accurate detection results are obtained after adjusting for the outlier set.Simulation results show that the algorithm have high detection accuracy and low false alarm rate.Event detection algorithm based on modified low-rank representation clustering algorithm and modified random forest classification algorithm.First,based on the temporal correlations of nodes' observations,sliding window data streams are clustered by using modified low-rank representation clustering algorithm,have event detection,and save the characteristics of event data.Finally,using modified random forest classification algorithm,the outliers can be classified in real time,and determine whether themselves are event data or not.Simulation results show that the algorithm can determine whether an event has happened or not at express speed without training set.Fault detection algorithm based on simulated annealing and least squares method in event area.In the event nodes set,there may be some faulty nodes which are mistaken for event nodes.On the based of the principle of simulated annealing,the cluster head performs generalized least square fitting by changing fitting nodes set.Within the event nodes set,the algorithm can find out a group of nodes which have better fitting effect.After the comparison between other event nodes and their fitting values,the algorithm can determine whether the nodes' observations are fault data or not.