Knowledge Distillation and Multi-task Feature Learning for Partial Discharge Recognition

To achieve accurate detection and recognition of partial discharge (PD) in switchgear, developing an intelligent PD diagnosis system has garnered significant attention in recent years. Due to inevitable noise interference and high similarity of different PD signals, detecting and identifying PDs using a portable PD detector poses significant challenges. In this study, we aim to transfer the knowledge acquired by the large-scale network to a lightweight network for precise PD recognition. To achieve this, we employ a k-means clustering model to effectively separate signals originating from different sources, thereby obtaining Phase Resolved Partial Discharge (PRPD) patterns. Then, we introduce knowledge distillation and a multi-task feature learning framework to extract discriminative features from PRPD patterns. We conduct experiments and compare the proposed method against some state-of-the-art methods on our constructed PD recognition dataset to evaluate the superiority of the proposed method.