Deep learning approach for photovoltaic fault diagnosis using DINOv2 and infrared imaging

Abstract

This paper presents a deep learning-based methodology for the non-invasive detection and classification of faults in photovoltaic (PV) modules using DINOv2, a Vision Transformer model trained with self-supervised learning. Leveraging a dataset of 20,000 real-world thermal images collected from PV systems across 25 countries, the proposed system enhances defect visibility through image sharpening and addresses class imbalance using the SMOTE technique. DINOv2's global attention mechanism enables it to outperform traditional convolutional neural networks (CNNs) in identifying both localized and subtle defects such as cracks, hot spots, and soiling. The model is trained to detect anomalies and classify them into 12 categories with a strong generalization capability, achieving a detection accuracy of 98.23% and a classification accuracy of 96.19%. Comprehensive experiments, including comparisons with ResNet50 and EfficientNetB0, validate the effectiveness of the proposed approach. The system offers a scalable and accurate solution for automated PV monitoring, with potential applications in predictive maintenance, real-time inspection, and energy efficiency optimization in solar power systems.