Research Of Outlier Detection Based On Quarter-Sphere SVM And Principal Component Analysis In WSN

Wireless Sensor Network(WSN)is widely used in a variety of specific environments thanks to its small size and low cost of sensor nodes.WSN integrates multiple functions such as sensing,processing and short-range wireless communication.Nevertheless,the resource constraints and the harsh environment in which the sensor nodes are subjected also make the generated data vulnerable to pollution of noise,errors,data loss,duplicated values and conflicting information.In the context of WSN,these unreliable data are called outliers.Outlier detection is essential to ensure data quality,security monitoring and reliable detection of critical events.Following are the methods proposed in order to solve the problem of outlier detection in WSN in this thesis.Aiming at the local outliers of the single sensor node,a method based on Quarter-Sphere Support Vector Machine(QS-SVM)is proposed.A quarter-sphere SVM model is established using the data collected by the nodes,combined with the particle swarm optimization(PSO)algorithm to find opimum parameters suitable for the model.In the way,the model can achieve the best detection effect.The experimental results show that the proposed technique can reduce the false positive rate while ensuring a high detection rate.A method of detecting local outliers based on incremental principal components analysis(PCA)of Euclidean distance is proposed for the data in a nonstationary distribution in WSN.This incremental manner is used instead of the batch approach to update the detection model.The iterative formulas avoid the dumplicated computational on the sensor nodes.At the same time,an exponential forgetting factor is also introduced to the iterative formula to increase the ability to deal with non-stationary data distribution for the nodes.The experimental results show that this method can effectively deal with outlier detection in a non-stationary data distribution.For the detection of outliers in a distributed environment,a PCA formulation based on Spatio-Temporal-Attribute correlations(STA-PCA)is proposed.By combining the temporal correlation of the same node in sequence time,the spatial correlation of adjacent nodes in space and the attribute correlation of different dimensions from the same measurement value,a new PCA calculation is used for detection.In addition,the proposed method uses spatial correlation to further distinguish outliers detected as errors or events.The experimental results show that the STA-PCA method has obvious performance improvement compared with the method without considering attribute correlation.Although the three methods proposed are effective but they still have deficiencies.Therefore,further research work is needed.