Detection and classification of partial discharge using a feature decomposition-based modular neural network

This paper develops a feature decomposition-based modular neural network (MNN) for the recognition of partial discharge (PD) sources. The original statistical analysis-based feature set is naturally partitioned into three disjointed feature subsets. These subsets are independently fed into three neural subnetworks. The aggregation of the sub-networks, by an integrating unit using a majority vote strategy, provides the final assignment of PD patterns to a particular PD source. Compared with a single neural network (SNN) with the same feature vector, the training of MNN is faster, the network is more robust, and the success rate of classifying "unseen" patterns is higher.