A New Classification Method for Radial Faults of Transformer Winding by FRA and PSO-RVM

Accurately diagnosing winding faults of transformers is of great importance to the stable operation of the electric grid. In this paper, by combining frequency response analysis (FRA) with relevance vector machine (RVM) optimized by particle swarm optimization (PSO), a new method, namely PSO-RVM, is proposed to classify typical radial faults of transformer winding including forced buckling (FDB) and free buckling (FB). First, a series of FB and FDB faults are artificially set on a transformer winding model and corresponding FRA measurements are conducted to collect the FRA data at the same time. Second, RVM optimized by particle swarm optimization (PSO) is introduced to set up the PSO-RVM classifier. Then, as inputs of the PSO-RVM, the features of the FRA dataset are extracted by using different numerical indices. After training and testing, the classification results of PSO-RVM for FB and FDB faults are obtained by the leave-one-out (LOO) and 10-fold cross-validation methods. Finally, the feasibility of the PSO-RVM classifier are validated by comparison with PSO-SVM. The comparison result indicates that the PSO-RVM has a very high classification accuracy rate which is similar to the PSO-SVM, but only a very small number of relevance vectors are used which is much less than the number of the support vectors used in the PSO-SVM.