Defect detection in photovoltaic cells with electroluminescence images using deep learning algorithms

The rapidly increasing global installed capacity of solar power plants has made the sustainability of the operational lifetime and efficiency of PV modules a critical engineering challenge. Microcracks and structural deformations occurring on cell surfaces during manufacturing, transportation, and on-site installation processes lead to significant energy losses over time. Although EL imaging is an industrial standard for detecting such defects, the manual analysis of EL images requires substantial expertise and is highly time-consuming. In this thesis, a hybrid deep learning architecture that combines object detection and segmentation approaches is proposed for the autonomous, fast, and pixel-accurate detection of PV cell defects. Within the scope of the study, the state-of-the-art single-stage detection architecture YOLOv11 is integrated with the foundational segmentation model SAM 2. The proposed method first localizes defects using bounding boxes generated by YOLOv11 and subsequently transfers these coordinates to SAM 2 as box prompts, enabling precise masking of the morphological boundaries of defects. Model training and evaluation were conducted on a comprehensive dataset comprising a total of 13,631 images, constructed by combining the open-source EL2021 v10 dataset with original field-acquired EL images. Experimental results demonstrate that the proposed hybrid model achieves a mAP@0.5 score of 0.885 across the defect classes black_core, crack, finger, horizontal_dislocation, short_circuit, star_crack, and thick_line. Class-wise analysis reveals that visually prominent defects such as black core and short circuit are detected with exceptionally high average precision values of up to 99.5%, while competitive performance is also achieved for the crack class, which is known to be the most challenging to detect. By combining the real-time detection capability of YOLOv11 with the zero-shot segmentation ability of SAM 2, this study presents a high-accuracy, industrially applicable autonomous inspection system that significantly reduces manual annotation costs and enhances quality control processes in photovoltaic module inspection.