Design of Ensemble Fuzzy-RBF Neural Networks Based on Feature Extraction and Multi-feature Fusion for GIS Partial Discharge Recognition and Classification

A new topology of ensemble fuzzy-radial basis function neural networks (EFRBFNN) based on a multi-feature fusion strategy is proposed to recognize and classify a pattern of reliable on-site partial discharge (PD). This study is concerned with the design of an ensemble neural networks based on fuzzy rules and the enhancement of its recognition capability with the aid of preprocessing technologies and multi-feature fusion strategy. The key points are summarized as follows: (1) principal component analysis (PCA) and linear discriminant analysis (LDA) algorithm are utilized to reduce the dimensionality of input space as well as extracting features. (2) statistical characteristics (SC) are obtained as the complementary characteristics of the PD. (3) the proposed network architecture consists of two-branch radial basis function neural networks (RBFNN) based on fuzzy rules, which can effectively reflect the distribution of the input data. Two types of RBFNN are designed which are based on hard c-means (HCM) and fuzzy c-means (FCM) clustering respectively. To fuse the learned features by PCA and LDA, we design a multi-feature fusion strategy that not only adjusts the contribution of different features to the networks but also enhances the recognition ability for PD. The performance of the proposed networks is evaluated using PD data obtained from four types of defects in the laboratory environment, and noise that might occur in power grids is also concerned. The experimental results of the proposed EFRBFNN show the satisfied recognition requirement for PD datasets.