Fault Diagnosis Based on Self-Tuning Support Vector Machine in Sample Unbalance Condition

The unbalance of sample quantities between faulty samples and normal samples leads to a deviation of classifying hyperplane while using support vector machine (SVM), hence decreases the accuracy of SVM-based fault diagnosis. A new self-tuning support vector machine (St-SVM), which could automatically adjust the penalty factors for risk function, is proposed for this issue. This method selects informative samples by booststrapping approach, amplifies their risk penalty factors and decreases the deviation of classification hyperplane brought by sample unbalance, and hence improves the accuracy of fault diagnosis. The St-SVM has been applied to the diagnosis of transformer faults. During the experiment, the positive and negative samples yield equal loss risks, and the diagnostic performance with unbalanced samples is significantly improved. It demonstrates the effectiveness of the proposed approach.