Automatic anomaly classification of photovoltaic cells with deep semantic segmentation of electroluminescence images

The need for automated monitoring of the performance of solar cells grows as solar energy is being utilized on a greater scale. Electroluminescence (EL) imaging presents itself as an effective artificial intelligence solution for detecting internal defects in images of photovoltaic (PV) cells, such as micro-cracks, which are not visible to the human eye. This paper proposes a detection model to investigate the effectiveness of deep learning models in automatically detecting and classifying PV faults with semantic segmentation of EL images. To extract and learn a set of discriminating image features, the proposed deep architecture is initialized by the residual network 50 (ResNet50) as a backbone encoder and then fine-tuned on EL images that are segmented and classified semantically by several U-networks (U-Net). Additionally, an accurate masking procedure is applied to the image ground-truth annotations to enhance the detector’s accuracy. Augmentation techniques are also performed on EL images to compensate for the limited diversity and quantity of data, enhancing the classifier’s generalization capacity. Extensive experiments show the effectiveness of the semantic U-Net-based deep architecture in classifying PV cells. ResNet50 combined with Attention-U-Net demonstrated high classification accuracy, achieving a weighted F1-score of 98%, a precision of 98%, and a recall of 98%. It also achieves a mean intersection of union of 97% for EL image segmentation. The proposed semantic deep encoder–decoder provides an effective means of automatically detecting and classifying defective PV cells.