Pattern Recognition Methods of Multi-source Partial Discharge Based on the Improved Deformable DETR Model and its Application

Pattern recognition methods of partial discharge (PD) utilizing images are efficient for the single PD source, yet they face challenges in recognizing the multi-source PD. An object detection model is proposed for the recognition of multi-source PD according to Deformable detection with transformers (Deformable DETR). Typical single-source PD and multi-source PD signals are collected by experiment. Two types of PD spectra, namely phase-resolved partial discharge spectrum and polar coordinate phase-resolved spectrum, are used to generate the data set. The denoising training task and Bayesian optimization algorithm are introduced to optimize the performance of the Deformable DETR model. Single-source and multi-source PD spectra are identified by the optimized PD Deformable DETR model. Results show that the proposed model can effectively recognize the source of single- and multi-PD patterns. In addition, compared with common types of object detection models, the performance of the PD Deformable DETR model can be evidently improved at the cost of losing a few efficiencies. Finally, the PD spectra of real motors with insulation defects are identified by the PD Deformable DETR model. The recognition accuracy reaches 91%, which shows the validity of this proposed method. Additionally, the acquisition and recognition program of PD spectrum is developed. The paper provides novel perspectives for identifying multi-source PD. ©2024 Chin.Soc.for Elec.Eng.