KD-PH-YOLO: Low-Power Infrared Defect Detection for Sustainable Photovoltaic Edge Inspection

Abstract

Infrared defects in photovoltaic (PV) modules captured during unmanned aerial vehicle (UAV) inspection are typically small and low-contrast, while onboard edge platforms are constrained by memory, logic resources, and power consumption. To address these challenges, this paper proposes KD-PH-YOLO for PV infrared defect detection and deployment on the Zynq-7020 platform. Based on YOLOv8, PH-YOLO removes redundant deep-layer computation and introduces a P2 detection head to preserve fine-grained information for small defects. Hardware-friendly Weighted Feature Fusion (HWFF) and Lightweight Attention-CBAM (LA-CBAM) are incorporated to enhance multiscale feature fusion and defect responses. Soft-label and multiscale feature distillation are further employed to improve the lightweight student model without increasing inference complexity. For edge deployment, INT8 quantization and hardware-aware acceleration are applied to map the model onto the Zynq-7020. Experimental results show that KD-PH-YOLO achieves an mAP@0.5 of 92.4% with only 1.48 M parameters and 6.9 GFLOPs. After hardware deployment, the model retains an mAP@0.5 of 91.8%. The Zynq-7020 implementation achieves an average latency of 184.3 ms per frame and a throughput of 5.43 FPS, with power consumption of 3.2 W and an energy efficiency of 1.7 FPS/W. The proposed method therefore provides a favorable accuracy–complexity–energy-efficiency trade-off for resource-constrained PV edge inspection.