Data-driven anomaly detection using OCSVM with Boundary optimzation

With large amount of industrial data available, data-driven anomaly detection methods have been widely used for ensuring industrial safety and preventing economic losses. Among them, one-class support vector machine (OCSVM) is an effective unsupervised method for detecting abnormal points by establishing a decision boundary in the kernel space and has obtained much attention in the research field in recent years. Through solving the convex quadratic programming problem, OCSVM obtains a classifier with maximum distance to the origin and satisfying maximum quantity of training samples, without utilizing anomaly data points. However, the outliers in the training set, which are more likely to be selected as support vectors, will lead to the degradation in the performance of OCSVM. Because the outliers are away from the center of the dataset and cannot represent the real characteristics for the data. To address this drawback, the clustering-based and LOF outlier detection method (CLOF), which has high computational efficiency and modeling accuracy, is adopted for eliminating the outliers and optimizing the final acquired boundary of OCSVM. Experimental results on shuttle flight data have illustrated the superior performance of the proposed approach.