Research Of Outlier Detection Algorithm Of Wireless Sensor Network In A Non-Stationary Environment

Wireless sensor network integrating multi-disciplinary is an emerging technology. It makes the communication between people and the world enter into a new level, and is considered as the second largest network technology after the Internet. Recently, promoted by the market demands of increasing efficiency and cutting costs, it has been widely used in industrial, household and other fields. In the engineering applications, the factors of complex work environment, interference and invasion can lead to outliers, which would deteriorate the quality of data and cause error in judgment and response. The resources, energy and communication bandwidth within wireless sensor network are limited. Faced with the common challenges of unstable environment and outlier detection in multi-dimensional datasets, the existing filtering technologies and outlier detection technologies can not accurately and efficiently detect the outlier of wireless sensor network in engineering applications.To solve these problems, the current outlier detection algorithms based on the support vector machine with good detection accuracy and the distribution density with good detection efficiency have been improved to propose some new comprehensive online outlier detection algorithms. With synthetic pseudo-random datasets, the simulation analysis of the outlier detection performance of new algorithms has been introduced. To verify the accuracy and efficiency of improved algorithms for outlier detection in engineering applications, a new engineering prototype for health monitoring of large metal structure has been developed based wireless sensor network technology, and the comparison test on quay cranes and tensioner of pipelay vessel have been introduced. Funded by the Shanghai science and technology committee foundation (Grant 11DZ2290500), the research work in this paper are as follows:(1) Comprehensive online outlier detection algorithms have been proposed based on the hypersphere one-class support vector machine. The existing outlier detection algorithms based on support vector machine have some limitations in model optimization and iterative update, and need to repeatedly solve quadratic programming problem causing poor detection efficiency. To solve these problems, comprehensive online outlier detection strategy has been introduced, and a fast model optimization algorithm based on the average Euclidean distance has been proposed. By combining these stage detection methods, a comprehensive online fast outlier detection algorithm has been proposed. Adopting Linear kernel and Gaussian kernel functions, the simulation analysis, based on the synthetic pseudo-random datasets and real wireless sensor network datasets, showed that the proposed algorithm can improve the accuracy and efficiency of outlier detection.(2) Comprehensive online outlier detection algorithms have been proposed based on density. The existing outlier detection algorithms based on density have limitation in model optimization and change detection, which result in low outlier detection accuracy and frequently model update. To solve these problems, a fast model optimization algorithm and a change detection algorithm based on the multi-granularity deviation factor have been proposed. By combining these stage detection methods, two comprehensive online fast outlier detection algorithm has been proposed. The simulation analysis based on the synthetic pseudo-random datasets and real datasets showed that the proposed algorithms can improve the outlier detection accuracy and efficiency. In addition, the reasonable range of the counting field coefficient has been studied to complement the existing work in this field.(3) A portable wireless strain monitoring system for the health monitoring of large metal structure has been developed, and the effectiveness of the improved algorithms in engineering applications has been verified. To verify the improved algorithms and solve the issues of decentralized measuring points, limited power supply and difficult implementation of wired systems, a portable wireless strain monitoring system based on wireless sensor network has been developed. Engineering tests of pipeline stress measurement in tensioner of pipelay vessel and drawbar stress measurement in quay cranes have been introduced, in which appear outlier. To improve the data quality and analysis accuracy, all improved outlier detection algorithms have been integrated into the system for engineering verification.With the simulation analysis and engineering validation, the proposed new outlier detection algorithms have improved the detection efficiency and accuracy significantly, and can meet the demand of outlier detection of wireless sensor network in engineering applications. With these algorithms, the portable wireless strain monitoring system also can be used in the health monitoring of large metal structure.