Efficient photovoltaic infrared image fault detection via channel compression and model optimization

Abstract This study addresses the challenges in photovoltaic infrared image fault detection related to large computational and parameter demands, as well as the scarcity of fault localization models. We propose VCE-YOLOv8n, a high-precision and lightweight fault detection model based on an improved YOLOv8n architecture. The model replaces the original YOLOv8n backbone and C2f module with the VanillaNet backbone and the lightweight C3Ghost module, enhancing feature representation while reducing computational load and parameters, thus improving detection speed. Additionally, the Efficient Channel Attention (ECA) mechanism is incorporated to strengthen multi-scale feature representation with minimal additional computational overhead. Comparative experiments between Slim pruning and heuristic channel compression demonstrate that the heuristic method is optimal for secondary compression, reducing the model parameters, computational cost, and memory usage by 64%, 59.2%, and 61.9%, respectively. The optimized model achieves an accuracy of 91.2%, recall of 88.9%, and mean Average Precision (mAP @ 0.5) of 91.7%. It operates at a detection speed of 166.6 frames per second, with an average inference time of 6.0 milliseconds per frame. These results demonstrate that VCE-YOLOv8n effectively balances low network complexity and high detection accuracy, providing a solid foundation for deployment on mobile and embedded platforms.