High-frequency Voltage Signal Fault Diagnosis Based on Multi-channel Hierarchical Structure Capsule-Attention Network

An important research area for covered conductors is partial discharge detection. However, the complex distribution of high-frequency signals with particular noise and the long tail problem create a considerable obstacle to accurate detection. This paper proposes a Capsule-Attention model with a Multi-channel Hierarchical structure, which separately utilize ResNet and BiGRU to extract short- and long-term high-frequency and pre-filter existing irrelevant information. The capsule layer is good at multidimensional and spatial information representation, and we utilize it to recognize and classify fault conditions. The attention mechanism's weight distribution quantifies the importance of features in the entire signal. Considering the class imbalance of the long-tail distribution of samples, We propose a Focal-MCC (FMCC) loss function. The proposed loss reshapes the typical cross-entropy loss, hence reducing the weights allocated to well-classified instances. Matthews Correlation Coefficient penalizes false positives/false negatives to prevent parameters from falling into local minimums. The experiments indicate that the model proposed is useful in detecting partial discharge faults and has application value. The code is available at https://github.com/qiumia030/capsule_attention.