Automated fault detection in solar panels using customized EfficientNetB0 with explainable AI solution for real-time monitoring

Solar energy offers a sustainable solution for power generation, reducing dependence on fossil fuels. However, surface-level anomalies such as dust, snow, bird droppings, and structural damage significantly impact the operational efficiency of solar panels. This study presents a deep learning-based fault detection framework utilizing a customized EfficientNetB0 architecture for the classification of common panel defects. The architecture incorporates global average pooling, batch normalization, and dropout layers to enhance fault sensitivity and ensure better generalization, while maintaining computational efficiency. The proposed model is trained on a diverse image dataset covering six fault types under varying environmental conditions, achieving an accuracy of 96.42%, with precision, recall, and F1 scores of 98%, 93%, and 95%, respectively, indicating competitive performance in the evaluated dataset. Grad-CAM-based visual explanations are employed to enhance model transparency by highlighting critical decision-making regions. The system demonstrates consistent performance under different lighting and weather conditions, providing a robust solution for real-time solar panel condition monitoring and maintenance.