Research On Outlier Detection Of Wireless Sensor Networks Based On Support Vector Data Description

As the core technology of IoT industry,wireless sensor network has been applied in many fields.WSN is often deployed in the complex and interferential environment,which may cause the generation of abnormal data or lower the data quality,and have negative impact on the decisions.So,it is very important in both theorectical and practical fields to study the outlier detection method of WSNs.This paper focuses on studying the outlier detection method using support vector data description(SVDD),and the main contributions of this paper are as follows:(1)An outlier detection algorithm based on Toeplitz matrix random feature mapping SVDD for wireless sensor networks(TRFF)is proposed.By applying the approximate kernel function of random Fourier feature mapping to the traditional SVDD algorithm,the time complexity of SVDD is reduced.The cyclic characteristics of Toeplitz matrix can reduce the memory consumption caused by the storage random feature matrix.The experimental results showed that TRFF algorithm achieved better detection rate,lower false alarm rate and less time than traditional algorithm.(2)An outlier detection algorithm based on model selection SVDD for wireless sensor networks(TSRFF)is proposed.Because of high real-time demand,it is necessary to design a random feature mapping algorithm which can realize the mapping in the lower dimension.The traditional random feature mapping algorithm has poor stability in low dimension feature space.Thus,over-fitting or under-fitting often occurs in the decision model.By applying the model selection strategy to the SVDD algorithm of Toeplitz matrix random feature mapping,the model selection of TSRFF is realized under the low feature dimension.The over-fitting error and the under-fitting error are calculated based on support vector method.The experiments showed that the TSRFF algorithm had good stability under the low feature dimension.(3)An outlier detection algorithm based on adaptive SVDD for wireless sensor networks(ATRFF)is proposed.Since the change of sensor data stream is random,the generalization ability of fixed decision model will deteriorate over time.In ATRFF algorithm,the density of data distribution is used as the criteria for updating the decision model efficiently.The kernel Euclidean distance method is used to simplify the training dataset and improve the real-time performance.The experiments showed that ATRFF algorithm can update the decision model efficiently and realize outlier detection with high detection rate and low false alarm rate.