Performance Analysis of YOLO11 and YOLO26 for Electroluminescence-Based Photovoltaic Defect Detection Under Class-Imbalanced Conditions

Solar energy is rapidly being adopted across the world and many photovoltaic systems (PV) are being installed. The increased demand for solar PVs has brought along the need for better inspection processes that can accurately detect defects like electrical and structural anomalies. Electrical and structural defects can cause solar PVs to operate at reduced power efficiency and lower lifetime expectancy. Inspection methods should maximize time efficiency and return accurate diagnosis, therefore out of many viable techniques electroluminescence (EL) imaging has proven to be effective in identifying solar PV anomalies. In this study, we provide an in-depth comparison of YOLO11 and YOLO26 object detection architectures for the task of detecting solar PV defects from EL images. We train YOLO11 and YOLO26 across different scale models namely nano(n), small(s), and medium(m). However, to analyze the influence of class imbalance, the experiments are conducted on a dataset comprising of 12 defect classes and its subsequent reduced subsets generated by exempting defect categories with limited samples based on predefined instance thresholds. The experimental results obtained demonstrate that defect detection performance improves as underrepresented classes are removed. YOLO26s was able to achieve the highest performance with a mAP@50: 95 of 0.6224 and an F1 score of 0.8366 on a 6-class subset. YOLO11n also exhibited strong performance with only 2.58M parameters and less training time. These findings clearly provide valuable insights on model selection for scalable and real time PV health monitoring systems.