Research Of Outlier Detection And Data Completion Based On Wireless Sensor Network

In recent years, with the rapid extensive application of wireless sensor network, human life andsocial development has improved a lot. However, the limitations as well as the distribution ofenvironmental condition of wireless sensor networks cause missing, error and other problems of datasamples which seriously affect the application of wireless sensor networks. So how to find a goodsolution to this issue is still urgent problem for people.This thesis proposes solutions in view of anomalies and missing data and validates specificalgorithms by experiment. The main work and innovation are as follows:(1) Analysis and research on outlier detection and data completion in wireless sensor network,pointing out the existing problems, then improving new algorithm and verifying them by a largenumber of experiments.(2) Support Vector Machine (SVM) technology can solve abnormal data and can avoid thedimension disaster problem, but for large scale data, the kernel function mapping process of SVMtechnology need large overhead. To solve this problem, this thesis proposes KNN-SVM algorithm.Since the vast majority of data collected by sensor nodes are normal data and the target of anomalydetection is to identify abnormal data, KNN-SVM algorithm first cut data samples and most of thecropped data samples are normal samples; the detect remaining small part of the data samples so thatthe number of data samples are greatly reduced, thereby reducing computational overhead.(3) The existing data completion methods in WSN mainly include linear interpolation algorithmand spatial correlation based missing data completion algorithm which rely single on time correlationor spatial correlation, and estimate missing data on multiple attributes. Aiming at this problem, thisthesis improves OptSpace algorithm and puts forward Ioptspace algorithm which can be applied incompletion of missing data in WSN. Experiments show that Ioptspace algorithm is more effective andowns more accurate estimation compared with the existing linear interpolation algorithm and datacompletion based on the spatial correlation in WSN.