A Hybrid Least-square Support Vector Machine Approach to Incipient Fault Detection for Oil-immersed Power Transformer

In this article, incipient fault detection methods using a novel hybrid classifier are developed for dissolved gas analysis of oil-immersed power transformers. New fault features are derived by analyzing various industry standards of dissolved gas analysis. Two effective data pre-processing methods are employed for improving diagnosis accuracies. Bootstrap is first utilized to equalize sample numbers of different fault types, and then the logarithmic transform is applied to generate additional classification features. In experiments, a least-square support vector machine, support vector machine, and support vector data description are developed as fault classifiers, and the optimal parameters of the three classifiers are obtained using particle swarm optimization. A comprehensive comparison is made regarding the performance of the three support vector machine based classifiers for the first time in the area of dissolved gas analysis. Moreover, classification boundaries are illustrated to provide an in-depth understanding upon the performance of each classifier with clear visualization figures. The results indicate that least-square support vector machine can significantly improve the diagnosis accuracy of dissolved gas analysis along with the proposed pre-processing methods.