KDBiDet: A Bi-Branch Collaborative Training Algorithm Based on Knowledge Distillation for Photovoltaic Hot-Spot Detection Systems

Abstract

Photovoltaic (PV) power generation technology is an effective approach for alleviating the current energy crisis. However, PV modules are prone to hot-spot faults and shorter service lives because of harsh long-term operating conditions. Therefore, in this study, we propose a bi-branch collaborative training algorithm based on knowledge distillation for PV hot-spot detection systems, focusing on the accuracy and computational efficiency of the algorithm. The proposed approach uses a knowledge distillation model based on a “teacher and student” collaborative training structure to improve detection accuracy while ensuring inference speed. The anchor-free YOLOX detection algorithm is used as the student network to reduce the number of network parameters. This includes a decoupled prediction head designed to improve detection performance. The teacher network first uses a HorNet-based cross-stage partial network (CSPHN)-based backbone network to enhance feature representation ability by capturing higher order spatial interactions. Second, a bi-branch multilevel feature adaptive fusion (BiMAF) module is designed to fuse multiscale features from both global and local perspectives in a parallel manner. Third, a shift contextual transformer (SCT)-based prediction head is adopted to enhance the long-range interaction between cross-scale features, thereby significantly improving the accuracy of the detection algorithm in dense scenes. Finally, comparative experiments are conducted on 13 classical detection algorithms to verify the effectiveness of the proposed detection network. The experimental results demonstrated that the proposed system can achieve fast and accurate detection of multiscale hot-spot faults under various harsh conditions, achieving an AP50 metric of 82.2%.