Solar Panel Fault Detection and Classification Using Transfer Learning and Deep Neural Networks

The maintenance of solar panels is critical for ensuring their efficiency and effectiveness in reducing greenhouse gas emissions. However, traditional monitoring of faults in solar panels, such as dust and physical damage, can be labor-intensive and impractical since solar panels are installed in large land areas. The study explores different deep-learning models to automate fault detection and classification in solar panels, focusing on three classes: clean, dirty, and broken. A dataset was constructed by combining images from multiple open-source datasets, and advanced preprocessing techniques were applied to enhance the diversity of the training data. Four convolutional neural network (CNN) architectures that utilize transfer learning were implemented and evaluated on this dataset: custom-made CNN, ResNet50, VGG19, and ConvNeXT. The VGG19 model demonstrated the best performance, achieving a test accuracy of 96.3%. It also excelled in key metrics such as F1-score, precision, and recall. Cross-validation further confirmed the robustness of the VGG19 model.