Automated defect detection in solar module electroluminescence images using YOLOv9 variants

Defective photovoltaic (PV) panels cause a reduction in energy generation. To maximise green energy production, close monitoring and improvement of the photovoltaic health index are essential. Current methods rely on visual inspection of electroluminescence (EL) images by experts which is time-consuming and requires highly trained personnel. This work presents an automated defect detection approach on PV panels using the state-of-the-art object detection model YOLOv9. Four variants of YOLOv9 (gelan-e, gelan-c, yolov9-c, and yolov9-e) are trained on three datasets (ELDDS, ELDDS1400C5, and PVELAD). Two similar datasets (ELDDS, ELDDS1400C5) are merged to make a larger dataset. The proposed YOLOv9-e model on merged data achieved the mAP@0.5 of 0.815 outperforming other approaches by 3.8%. Notably, on the ELDDS1400c5 dataset commonly used for comparison, the proposed YOLOv9-e variant achieves a competitive mAP@0.5 of 0.766 without architectural modifications compared to 0.777 for YOLOv5s.