Recognition of multiple defects and signal intensities from superimposed partial discharge sources in gas-insulated switchgear

This study addresses the problem of poor recognition accuracy when a partial discharge pattern contains two defective discharge sources. First, a representative pattern for each defect was constructed using a gas-insulated switchgear (GIS) from three different defects including oil stain, metal particles, and metal protrusion. Then, superimposed defect patterns with two different defects and varying intensities were generated. Second, the feature extraction capabilities of the VGG16, ResNet50, MobileNet, and Inception V3 deep learning models were applied to produce the feature maps. Finally, the decision strategy was proposed to determine the combination of defect types and discharge intensity ratios based on cosine similarity analysis. The results showed that when the MobileNet model utilized traditional learning networks, the recognition accuracy was only 43%. By adopting the proposed method, the accuracy improved to 85.3%. Moreover, combining MobileNet, VGG16, and Inception V3 in a PSO-weighted ensemble increased the defect-combination accuracy to 89.3%. In addition, using a grayscale colormap slightly improved the accuracy compared with that achieved using a jet colormap. The results demonstrate the model’s capability to deconstruct mixed signals into specific intensity ratios.