Multi-scale feature interaction-driven fault detection in photovoltaic modules

Some fine-grained faults in photovoltaic (PV) module images are prone to missed detection, while the overall detection accuracy remains insufficient. To address these issues, a PV module fault detection method based on multi-scale feature interaction is proposed in this study. In this method, an improved C3K2 module (C3K2-HL) integrating global context modeling with local feature enhancement is first designed. It replaces the shallow C3K2 module in the backbone to preserve the local structural information of low-contrast, fine-grained faults. Second, a Multi-Scale Feature Transmission (MSFT) neck is constructed to improve cross-scale feature alignment and reduce information loss through multi-branch feature interaction and adaptive feature transmission. Finally, the DySample dynamic upsampling module, which performs upsampling by learning sampling locations, is introduced to preserve the fine-grained features of small faults and reduce interference from background gridline textures. Experimental results show that the improved model achieves a mean average precision of 87.4% at an intersection-over-union (IoU) threshold of 0.5 (mAP50). This value is 2.5 percentage points higher than that of the baseline YOLOv11s. The model also achieves an mAP of 57.5% averaged over IoU thresholds ranging from 0.5 to 0.95 (mAP50-95). The proposed improvements enhance the overall detection performance for PV module faults and provide a feasible technical solution for multi-scale PV module fault detection in complex backgrounds.