KRMNet: learning core representations for partial discharge pattern recognition via masked autoencoders and mixed position coding

Abstract

Partial discharge pattern recognition (PDPR) is a crucial cornerstone for condition monitoring and safe operation of electrical devices. It has become an important hotspot in the field of energy systems. However, it faces several challenges, including noise interference, signal complexity, and difficulty in data labeling. This study proposes an efficient multi-scale masked autoencoder (MAE) network (KRMNet) to effectively address these challenges. KRMNet learns common and important multi-scale features and long-range semantic dependencies of partial discharge signals. Furthermore, by using the MAE structure and the transformer as the backbone, the model extracts key distinguishing features from phase-resolved partial discharge (PRPD) signals with background noise, interference, and labeling issues in an efficient and self-supervised manner. In addition, the efficient multi-scale module uses an efficient multi-scale attention mechanism to aggregate key information from multiple feature dimensions. The integration of the efficient multi-scale attention mechanism and contrastive learning methods improves the model’s ability to distinguish key information and resist interference. Experiments on two challenging PRPD datasets show that our proposed KRMNet achieves detection accuracies of 88.5% and 90.2% on noisy (ECPD) and clean (PUPD) datasets, respectively. This finding suggests that the method faces challenges in effectively managing noise interferences and missing labels.