A Visual Diagnosis Method of Power Transformer Multi-Source Partial Discharge Based on YOLO-URLKS

Effective monitoring and early warning of multi-source partial discharge (PD) are critical to ensuring the reliable operation of power transformers. However, the scarcity of multi-source Phase Resolved Partial Discharge (PRPD) patterns persists, limiting model generalization and diagnostic performance. Current PRPD-based methods lack deep feature information extraction, hindering quantitative evaluation and risk prediction. To address these limitations, a multi-source PRPD visual diagnosis method for power transformers using YOLO-URLKS is proposed, which leverages deep feature information of PRPD patterns to diagnose the insulation status of power transformers. Initially, five PD models of power transformers are developed, and PRPD patterns are obtained through the voltage experiments, constructing a single- and multi-source hybrid PRPD dataset. Subsequently, the backbone and neck modules of YOLO11 are optimized using UniRepLKNet and StarNet modules. Furthermore, SIoU is adopted in place of CIoU to enhance training stability. Additionally, a visual diagnosis algorithm is developed to analyze PRPD features and identify power transformer fault conditions. Experimental results demonstrate that the proposed method achieves high diagnostic accuracy for multi-source PRPD patterns with relatively low computational complexity.