Partial Discharge Small Sample Pattern Recognition for Generator Stator based on SP-Attention and Triplet-DenseNet

A novel partial discharge (PD) pattern recognition method based on SP-Attention and Triplet-DenseNet is proposed. Generator stators face a multitude of aging stresses that can compromise their insulation systems, making reliable PD detection crucial for assessing their condition. Traditional methods face challenges due to the scarcity of comprehensive and well-labeled datasets. Our proposed method integrates SP-Attention to enhance feature extraction throughout Phase-Resolved Partial Discharge (PRPD) spectra, establishing critical long-range dependencies within the dataset. Furthermore, we employ Triplet-DenseNet, a contrastive learning framework designed to efficiently process features in datasets with limited samples. This system significantly improves diagnostic capabilities, especially in distinguishing between highly similar PD patterns. In laboratory conditions, common insulation defects were simulated and subsequent tests were conducted. These tests were shown to confirm that our method surpasses traditional deep learning models in terms of accuracy and robustness. The proposed method achieves an accuracy exceeding 97% in scenarios characterized by limited and similar sample sets, significantly enhancing the analysis of generator stator PD.