As the global energy transition accelerates and climate challenges intensify, photovoltaic (PV) power generation plays an increasingly critical role, highlighting the importance of efficient and accurate fault detection in PV panels. Traditional detection methods are often inefficient, imprecise, and labor-intensive. Although thermal imaging offers advantages, it faces limitations such as low contrast between targets and background, and the high computational complexity of existing algorithms. To address these issues, this study proposes a thermal imaging-based fault detection method using the YOLOv12 algorithm. YOLOv12 incorporates a regional attention mechanism and the R-ELAN structure to enhance detection precision and inference speed. Experiments are conducted using an augmented dataset collected from a PV plant in northern China. The results demonstrate that YOLOv12 achieves a mAP @ 0.5 of 92.6%, GFLOPs of 6.3, a recall of 81.2%, and the fastest inference speed of 0.7 ms among the evaluated models. The proposed method effectively balances accuracy, efficiency, and real-time performance, offering a reliable and practical solution for the operation and maintenance of PV systems.