Toward Real-Time Monitoring of High-Voltage Insulators: Progressive Flashover Classification Using Quantized Deep Learning

Abstract

High-voltage insulators play a critical role in ensuring the reliability of power transmission systems by preventing flashover under severe environmental conditions. Traditional monitoring techniques rely on visual inspection or static classification schemes, which often fail to capture the progressive nature of surface discharge activity leading to flashover. This article presents a novel machine learning framework that addresses this limitation by classifying leakage current signals into five distinct operational stages: negligible leakage current, leakage current starting, leading to flashover, preflashover, and flashover. This multistage classification approach enables more accurate early warning of impending flashover by identifying subtle changes in leakage current behavior that precede catastrophic insulation failure. Controlled contamination experiments were conducted using porcelain insulators under varying environmental stressors and leakage current data was systematically acquired, segmented, and labeled based on amplitude variations, harmonic distortion, and dry band arcing characteristics. The proposed model, based on inception modules with residual connections, effectively captures multiscale temporal patterns in leakage current signals. Furthermore, posttraining quantization was applied to compress the model for edge deployment, achieving a 91.4% reduction in model size and a 90% decrease in inference time with negligible accuracy loss. Comparative evaluation against conventional neural networks and state-of-the-art ML architectures demonstrated the superior classification accuracy, robustness, and computational efficiency of the proposed framework. This architecture not only facilitates early detection of flashover stages but also enables low-latency, low-power deployment on resource-constrained devices, such as embedded systems, in remote substations.