Comparative analysis of methods used to solve class imbalance problem in dataset for CNN-based fault diagnosis of solar photovoltaic panels

This study presents a comprehensive approach for improving fault diagnosis in solar Photovoltaic (PV) panels by addressing the critical challenge of class imbalance in image-based datasets. A novel RGB image dataset containing three critical and real-world PV fault categories—including glass breakage, dust and bird droppings is developed to support deep learning-based classification. The research systematically evaluates four class balancing strategies: undersampling, oversampling, transformation-based augmentation, and generative adversarial network (GAN) based augmentation, using the InceptionV3 architecture. Among the methods tested, GAN-based augmentation achieves the best classification performance, with 86.02% accuracy and an F1 score of 86.00%, demonstrating its effectiveness in generating realistic minority class samples. To validate model robustness, the GAN augmented dataset is further evaluated using the YOLOv8 classifier, which achieves a superior accuracy of 90.1%. These results highlight the capability of GAN to enrich training data while mitigating class imbalance, and the strength of YOLOv8 in high-accuracy fault detection. The novelty of this study lies in the creation of a domain-specific RGB dataset and the systematic benchmarking of multiple balancing techniques in the context of solar PV fault detection. The proposed framework enhances classification performance under imbalanced conditions and contributes toward the development of automated, scalable PV inspection systems suitable for large-scale deployment.