Robust one-class SVM for fault detection

• A new method is proposed to reduce outliers' influence on OCSVM model.
• Some suspected outliers are identified and removed first.
• The decision boundary enclosing the cluster core of the target class is obtained.
• The outliers are identified and removed based on this boundary.
• The final OCSVM model is obtained on the remaining target samples.

One-class SVM (OCSVM) has been widely adopted in many one-class classification (OCC) application fields. However, when there are outliers in OCC training samples, the OCSVM performance will degrade. In order to solve this problem, a new method is proposed in this paper. This method first identifies some “suspected outliers” and removes them so as to obtain the decision boundary enclosing the “cluster core”. Then outliers are identified by this boundary and are removed from OCSVM training. The effectiveness of this proposed method is verified by experiments on UCI benchmark data sets and Tennessee Eastman Process data sets.