Real-time defect localisation in photovoltaic cells using electroluminescence images on an edge device

Photovoltaic (PV) cells convert solar energy into electricity. Defects from manufacturing or environmental factors can reduce efficiency and system stability, making timely detection crucial. Electroluminescence (EL) imaging is essential for detecting defects, and deep learning (DL) has achieved notable success in analysing EL images. However, detection efficiency remains underexplored in existing methods. To address this limitation, this paper proposes an innovative method for real-time PV cell defect localisation (PVCDL), based on the YOLOv10 architecture. Two key enhancements are introduced to improve detection performance. First, the original partial self-attention module (PSAM) was replaced with coordinate attention module (CAM), which has fewer parameters and provides faster inference speed. Second, the original CIoU was replaced with WIoU, delivering more accurate prediction results. Subsequently, the proposed method is validated using a publicly available PV cell defect detection dataset. The experimental results confirm the method’s effectiveness, achieving an mAP of 89.5%, with inference speeds of 118.4 FPS on GPUs and 33.5 FPS on an edge device, demonstrating real-time detection performance even on resource-constrained hardware. Compared to existing methods, our approach exhibits the smallest parameter size, the fastest inference speed and an excellent balance between detection accuracy and efficiency.