CS-RVM and its application in fault diagnosis of power transformers

Since different severities of damage may be induced by misclassification of transformer faults, the classification correctness alone may not be practically meaningful, for which, the CS-RVM (Cost-Sensitive Relevance Vector Machine) is proposed. It takes the minimum misclassification cost as its objective and applies Bayesian risk theory to predict the class of new sample. A typical case is studied to verify its cost sensitivity, based on which, it is adopted to the transformer fault diagnosis. The transformer fault diagnosis based on DGA (Dissolved Gas Analysis) shows that, the global diagnosis correctness of CS-RVM is slightly higher than that of BPNN (BP Neural Network) or SVM (Support Vector Machine) and slightly lower than that of M-RVM, while the diagnosis correctness of CS-RVM for the fault class with higher misdiagnosis cost is higher, showing its cost sensitivity. The diagnosis speed of CS-RVM meets the requirement of projects for transformer fault diagnosis.