A deep learning based framework for solar panel segmentation and fault classification enhanced with explainable AI

With the increasing adoption of solar photovoltaic (PV) systems worldwide, efficient monitoring and fault detection have become essential to ensure optimal performance and longevity. Traditional PV panel inspection methods are often time-consuming, labor-intensive, and lack precision. To address these challenges, this study presents a deep learning-based framework for solar panel segmentation along with faults and abnormalities classification enhanced with Lime based Explainable AI (XAI). The proposed approach employs a hybrid segmentation model that combines the strengths of U-Net and RefineNet, leveraging DenseNet as the backbone. The segmentation module is designed to perform binary segmentation to ensure the precise localization of solar panels before the fault classification stage. For fault classification, the MobileNetV3Small-based transfer learning model has been utilized, which can classify diverse real-world solar panel fault and abnormality conditions, including physical and electrical damage, dust, bird droppings, and snow cover. Furthermore, LIME-based XAI has been incorporated to improve interpretability, providing insight into the decision-making process of the model. The segmentation model achieved an Intersection over Union (IoU) of 93.63% and 91.24% for two publicly available datasets. The classification model attained 97.53% accuracy. The proposed method outperforms existing approaches in both segmentation and classification while maintaining computational efficiency. By combining segmentation and fault classification as a sequential model, this framework enables automated solar panel monitoring with high reliability and transparency.