Intelligent fault detection in photovoltaic (PV) systems plays a critical role in ensuring sustainable energy production and system reliability. This can be done using infrared thermal imagery from the Infrared Solar Modules dataset. In realworld, imbalanced dataset of low-resolution images may exist. This work applies artificial intelligence and image processing techniques to classify anomalies in solar panels. To address the class imbalance challenge, a convolutional neural network (CNN)-based model was developed alongside a comprehensive preprocessing pipeline, including data augmentation, class weighting, and undersampling of dominant classes. Model performance was evaluated using loss, accuracy, AUC-PR, F1-score, and G-Mean to ensure a robust and fair assessment. The results demonstrate that targeted preprocessing significantly improved generalization and minority class recognition, reducing false positives and negatives and enabling more balanced predictions. Nonetheless, persistent misclassifications highlight the need for further dataset enrichment, advanced architectures, and real-time integration. The findings reinforce the value of AI-driven diagnostics for PV monitoring while underscoring the importance of data quality and tailored pre processing strategies.
Dekhandji et al. (Wed,) studied this question.