New transformer fault diagnosis method based on multi-class relevance vector machine

The transformer fault diagnosis is naturally a multi-classification problem with few sample data and a lot of uncertainties. Among the existing transformer fault diagnosis methods, a large number of sample data and amount of computation are needed for Bayesian Network (BN), and the adjustment of the coefficient is difficult for support vector machine (SVM). So a new method of transformer fault diagnosis based on multi-class relevance vector machine (M-RVM) is proposed. The method takes ratios of feature gases as inputs and Fast Type-II ML and expectation maximization (EM) are adopted. Diagnostic outputs are probability for each fault category and fault type with the highest probability is taken as diagnosis result. Experimental results show that the diagnosis speed is sufficient for project needs and M-RVM shows higher diagnosis accuracy compared with BN and SVM.