Deep learning-driven fault detection in photovoltaic modules using UAV assistance

Automated inspection of photovoltaic (PV) modules is increasingly important for large-scale solar farms, where manual thermal inspection is time-consuming and susceptible to missed faults. This study presents a UAV-assisted Photovoltaic Fault Detector (PVFD) that combines radiometric thermal imaging with a lightweight YOLOv8s-based detection framework. A dataset of 1,251 radiometric thermal images was collected over one year from three PV installations using a FLIR One Pro thermal camera mounted on a DJI Phantom 4 UAV. The framework detects three fault categories—hotspots, diode faults, and affected cells—using a fault-specific One-vs-Rest (OvR) strategy in which three independently trained YOLOv8 detectors are integrated for multi-fault inspection. On the held-out test set, the proposed framework achieved an accuracy of 0.85, precision of 0.89, recall of 0.86, and mAP@0.5 of 0.77. The corresponding 95% Wilson confidence interval for accuracy was 0.764–0.908. A controlled comparison with a conventional multi-class YOLOv8 model showed that the OvR formulation increased overall accuracy from 0.82 to 0.85 and average recall from 0.83 to 0.86. Class-wise mAP improved from 0.70 to 0.73 for hotspots, 0.66 to 0.71 for diode faults, and 0.62 to 0.66 for affected cells, while inference throughput decreased from 48 FPS to 43 FPS. These results demonstrate that the OvR formulation provides a measurable sensitivity–efficiency trade-off and improves detection of less represented PV fault categories, supporting its use in UAV-assisted thermal inspection.