Deep Learning System for Defect Classification of Solar Panel Cells

Solar photovoltaic technology can be regarded as a safe energy generation system with relatively less pollution, noiseless, and abundant solar source. The operation and maintenance costs for solar panels are almost negligible as compared to costs of other renewable energy systems. However, due to the exposure to different weather conditions like extreme heat, humidity, dust storms and rain, the panel modules are liable to kind of defects which lead to power degradation, shading, bridging, power loss and fire hazard. Visual inspection of solar panel cells by experts is stressful, time consuming and unreliable. The Electroluminescence (EL) method with the use of infra-red cameras make deep learning algorithms promising in solving the problem. In this paper, we applied several deep learning networks such as AlexNet, SENet, ResNet18, ResNet34, ResNet50, ResNet101, ResNet152, GoogleNet (Inception V1), Xception, Vision Transformer (Vit), YOLOv3, and SqueezeNet to classify solar PV cell defects. We applied the models on the 2,624 elpv benchmark images using both binary and four classifications. But due to limited defect classifications with elpv benchmark dataset, we extracted EL images from publicly available datasets of a total of 18,347 Photovoltaic (PV) cells images with 11 types of defects in addition to the non-defective PV cells. We compared the results of the elpv benchmark with the extracted elpv images using accuracy, F1 score, precision and recall metrics. The results show that the xception model has the highest accuracy from 56.296% on elpv benchmark to 91.399% on the extracted elpv datasets.