Classification of clean and faulty images in solar panels using transfer learning based deep learning models

Solar energy is a renewable, inexhaustible and clean energy source. In order to reduce carbon emissions and dependence on fossil fuels, solar energy has been widely used, especially in recent years. Photovoltaic panels play an important role in obtaining solar energy efficiently. Failures may occur in photovoltaic panels due to environmental factors or electrical irregularities. Early and fast detection of faults in panels is of great importance in terms of increasing efficiency and obtaining energy without loss. However, manual detection of these faults is both costly and time consuming. For this reason, many deep learning models are currently used to detect these faults. In this thesis, a model combining ResNet and EfficientNet, two transfer learning based deep learning models, is proposed to classify the defects in solar panel images. A six-class solar panel dataset is used for the experimental studies. Consisting of 885 images, the dataset consists of dirty, clean, electrically damaged, physical damage, snow covered and bird drops classes. In the study, individual measurements were performed using transfer learning based architectures (VGG16, VGG19, MobileNet, DenseNet) and the results were compared with the proposed model. When the results are compared, it is observed that the proposed model shows superior performance. The ResNet architecture used in the proposed model reduces the gradient problem and allows deep networks to be trained in a healthier way thanks to their residual connections. EfficientNet architecture, on the other hand, optimizes the depth, resolution and width of the network to achieve more efficient results. When all architectures of ResNet and EfficientNet were used, the highest accuracy was calculated with ResNet101 and EfficientNet B1. The combination of ResNet101 and EfficientNetB1 achieved the best results with 87.55% accuracy, 87.92% precision, 88.75% recall and 88.13% F1-score.