Adaptive multi-scale YOLOv8s framework for defect detection in photovoltaic facilities

Photovoltaic (PV) modules are widely used in building-integrated photovoltaic systems and renewable energy infrastructure. Accurate defect detection is essential for improving the reliability, energy efficiency, and sustainable operation of PV systems. However, existing deep learning-based inspection methods still suffer from missed detections, imprecise localisation of small defects, and limited robustness under complex operating conditions due to variations in defect scales and background interference. This study proposes an adaptive multi-scale attention-enhanced YOLOv8s framework (AMA-YOLOv8s) for intelligent PV module defect detection. Compared with the baseline YOLOv8s, the proposed framework incorporates three newly designed modules specifically developed for PV defect detection: the Adaptive Bottleneck Attention Mechanism (ABAM), Multi-scale Adaptive Fusion Module (MSAFM), and Adaptive Pooling Fusion Module (APFM). ABAM is introduced to enhance defect-related feature representation and suppress irrelevant background information, whereas MSAFM and APFM are designed to strengthen multi-scale feature fusion and improve the detection of small and irregular defects. In addition, a comprehensive PV defect dataset was constructed by integrating publicly available images, field-collected samples, and data augmentation strategies. Experimental results show that AMA-YOLOv8s achieves a precision of 93.2%, an F1-score of 92.7%, and an mAP@0.5 of 93.5% on the current evaluation dataset. Compared with the YOLOv8s baseline, the AMA-YOLOv8s provides incremental improvements in detection performance while maintaining a comparable real-time inference speed. Its overall performance is competitive with the evaluated detectors, although recent lightweight YOLO variants retain advantages in some individual metrics and computational complexity. These results demonstrate the practical effectiveness of the proposed coordinated attention and multi-scale feature-fusion design for PV defect detection.