The safe and stable functioning of power grids is fundamentally dependent on the insulation integrity of transformers, which are vital assets within the system. Partial discharge (PD) detection serves as a vital method for assessing transformer insulation integrity. Addressing the limitations of traditional PD identification methods—which rely on manual feature extraction, exhibit low recognition efficiency, and demonstrate poor generalization capabilities—this paper proposes the MDW-YOLO algorithm for identifying PD defect waveforms in transformers. The MDW-YOLO algorithm is an extension of YOLOv8 that incorporates a Multi-Scale Large Kernel Attention (MLKA) mechanism within its backbone network. It enhances the model’s perception of detailed waveform features through parallel large-kernel convolutions and a gated fusion strategy. Furthermore, to enhance the integration of multi-scale features, the model reconstructs the fusion pathway into a BiFPN, employing weighted fusion for adaptive feature combination. This significantly improves the unified detection capability for both minute discharge pulses and macro discharge patterns. To refine localization, the method incorporates the MPDIoU loss function, achieving precise bounding box regression through direct optimization of diagonal point distances. Experimental results report a mean average precision of 0.935 for MDW-YOLO, meeting the practical need for transformer partial discharge waveform detection. The collective improvements are validated, and the algorithm’s superiority is established, through detailed ablation and comparative testing.
Liu et al. (Mon,) studied this question.