A Partial Discharge Classification Method Based on Enhanced Spiking Neural Network

This paper addresses the limitations of traditional spiking neural networks (SNNs) in partial discharge (PD) classification tasks, namely insufficient accuracy and high model complexity. It proposes a lightweight and high-accuracy PD classification framework based on enhanced spiking neural networks (ESNN). The framework consists of three stages: 1) PD Data Acquisition and Preprocessing: PD signals are captured through sensors and signal acquisition systems, and the training and testing datasets are constructed. 2) Design and Training of the ESNN Model: The ESNN model leverages the temporal dynamics of traditional SNNs while introducing several innovative components, including a lightweight block-based tokenization module, spiking self-attention (SSA) mechanism, spiking relative position encoding (RPE), and multi-spike neuron layers. These enhancements significantly improve feature extraction and classification accuracy. Additionally, the model adopts a residual connection structure to ease the training of deep networks and employs an optimized lightweight architecture to effectively reduce computational complexity. 3) PD Classification with the Trained ESNN Model: The trained ESNN model is utilized to classify test data, achieving automated identification of PD types. Experimental results demonstrate that the proposed method achieves substantial improvements in classification accuracy and model complexity compared to traditional SNN methods. This study provides a novel solution to address the challenges of low accuracy and high computational complexity in PD classification using existing SNN approaches.