Advancing photovoltaic cells defect detection in electroluminescence images through exploring multiple object detectors

Automated methods can provide accurate, time-efficient and cost-effective solutions for monitoring of photovoltaic (PV) modules in order to deal with underperformance and unreliability issues. However, these methods are not yet widespread at commercial level and still under study at laboratory scale due to many practical limitations. The present study deals with deep learning based enhanced classification and detection of multi-defects in electroluminescence (EL) images of PV cells, with a focus on practical application in field particularly on unseen data in multiple unknown imaging conditions. It explores potential of multiple state-of-the-art deep learning object detectors i.e. Detection Transformer, EfficientDet, FasterRCNN, YOLOv7, YOLOv8, and YOLOv9 with different variants, formations, techniques and experimentation, aiming enhancement in defect detection and understandings into trade-offs. These detectors were trained on polycrystalline cells with cracks, finger interruptions, black cores, and thick line defects. The proposed YOLOv9 GELAN-e with PGI and GELAN architectures achieves promising results of 94.30 % [email protected]. Moreover, testing experimentation is carried out on unseen data obtained in unknown imaging conditions (taken from multiple sources like a PV inspection center and public sources) and having wider diversity to generalize findings for providing insights into real challenges at practical level and elucidating possible solutions/directions. The proposed network also performed well on this data, which includes half-cut and full poly/mono crystalline cell images as well as Generative (Artificial Intelligence generated) images, making it promising for a wider range of practical applications.