Partial Discharge Pattern Recognition Method Based on SIFT-BOVF and Siamese-VGG with Limited Training Samples for Generator Stator

Accurate pattern recognition under conditions of limited training samples is a significant challenge in partial discharge (PD) detection for generators. To address this issue, a novel method based on SIFT-BOVF and siamese visual geometry group (Siamese-VGG) is proposed for PD pattern recognition under limited training samples conditions. Firstly, phase-resolved partial discharge (PRPD) samples of different defects are collected from laboratory experiments to construct an imbalanced sample set. The dataset includes both abundant and limited samples, used for training the Siamese-VGG model. Next, scale-invariant feature transform (SIFT) features are extracted from the imbalanced samples to construct a bag of visual features (BOVF) model. A sample retrieval process is then applied to identify samples similar to the target. Finally, the similarity between samples is computed using a shared loss function within the Siamese-VGG framework, enabling accurate PD pattern recognition. Experimental results demonstrate that the proposed SIFT-BOVF and Siamese-VGG approach achieves superior classification accuracy compared to other models under limited training sample conditions. When the limited training sample set contains only five training samples, the proposed method achieves a recognition accuracy of 93.20%, significantly enhancing the accuracy of generator PD pattern recognition under imbalanced sample conditions.