基于最优传输理论的光伏电池异常检测精确分类

Abstract Solar energy, particularly photovoltaic (PV) systems, plays a crucial role in combating climate change . However, PV cell anomalies such as black cores and cracks, caused by environmental factors, significantly degrade their performance. Traditional detection methods are often inefficient and risky, while existing YOLO models like YOLOv9 face challenges in accurately detecting anomalies with irregular shapes or sizes. These anomalies lead to low confidence in predictions and inaccurate classification results . In this paper, a precision classification framework for anomaly detection in PV cells is introduced, leveraging optimal transport (OT) theory. The framework operates in two stages. In the first stage, an anomaly prototype pool is constructed by clustering features within ground-truth boxes using k-means. Anomaly prototypes are selected based on their cosine similarity to normal prototypes, with those exhibiting lower similarity to normal regions being chosen. To ensure diversity among the prototypes, an orthogonal loss is applied during this stage. In the second stage, OT theory is utilized to match YOLO-predicted bounding boxes with the prototypes. A cosine similarity matrix is first created between the bounding box features and the prototypes. The Sinkhorn-Knopp algorithm then generates an OT transport plan based on this matrix, refining the classification scores. This process enhances the accuracy of both anomaly classification and localization. Experiments conducted on the PVEL-AD dataset demonstrate that the proposed framework, when integrated with YOLOv9, achieves a 95.8% mAP@0.5, marking a 2.6% improvement over the baseline method . Additionally, the True Positive Rate (TPR) increases by 1.6%, while the False Positive Rate (FPR) decreases from 2.5% to 1.1%. Visualizations further confirm a reduction in false negatives and improved localization accuracy . The paper also discusses the framework’s scalability and computational trade-offs, validating its effectiveness in enhancing the precision of PV anomaly detection .