Optimal Feature Selection for SVM Based Fault Diagnosis in Power Transformers

Power transformer is one of the most vital equipment in an electrical system and its failure results in huge economic losses. Amongst the various data driven techniques available in literature for diagnosing faults in a power transformer, Support Vector Machine (SVM) is one of the most promising. In this context, SVMs have typically been implemented using all the gaseous species available from dissolved gas analysis (DGA). In this work, we propose to enhance the diagnostic performance of SVMs by using them with an optimally identified subset of gaseous species available from DGA. We propose to use mutual information to identify these optimal species (features). The approach is applied on industrial datasets corresponding to various commonly encountered faults in power transformers. The results show that better diagnostic performance is obtained when CO2 concentration measurement is not used.