Efficient Identification of Cable Partial Discharge: A Spiking Neural Network Approach

Cable partial discharge (PD) identification is a critical technology for ensuring insulation reliability in power systems. To address the high computational resource demands of traditional deep learning methods in edge device deployment, this article proposes a low-power event-driven spiking neural network (SNN)-based approach for cable PD identification. By reconstructing the temporal feature extraction mechanism of AlexNet and leveraging the sparse computation and biologically inspired event-driven properties of SNNs, the model replaces traditional activation functions with integrate-and-fire (IF) neurons. A three-layer convolutional encoder is designed to extract spatiotemporal dynamic features from phase-resolved partial discharge (PRPD) patterns, combined with differentiable surrogate gradients for end-to-end training. Experiments on a real-world cable PD dataset demonstrate that, compared to traditional AlexNet, the proposed low-power SNN-based method effectively reduces computational resource requirements while maintaining high identification accuracy.