Pattern Recognition of GIS Partial Discharge Optical Signal Based on GWO-CNN Algorithm

During the manufacturing, transportation, and assembly processes of Gas Insulated Switchgear (GIS), internal insulation defects can lead to partial discharge (PD), posing risks to operational reliability. This study proposes a pattern recognition method for GIS PD optical signals using a Grey Wolf Optimization-Convolutional Neural Network (GWO-CNN) algorithm. Features including Gray-Level Co-occurrence Matrix (GLCM), gray difference statistics, and Hu invariant moments are extracted from two-dimensional Phase-Resolved Partial Discharge (PRPD) maps of four defect types (particle, floating, surface, and needle-tip discharges), forming a 19-dimensional feature vector. The CNN model enhances recognition accuracy through deep learning, while the GWO algorithm optimizes hyperparameters (e.g., learning rate) to prevent overfitting and improve convergence. Experimental results demonstrate that the GWO-CNN model achieves a recognition accuracy of 92.79%, outperforming the standalone CNN (89.21%). Additionally, GWO optimizes the learning rate from 0.001 to 0.01, enhancing global optimization capabilities. This hybrid approach provides an effective solution for accurate and robust PD pattern recognition in GIS, contributing to improved insulation defect diagnosis in power systems.