Abstract
Advancing renewable energy solutions requires efficient and durable solar Photovoltaic (PV) modules. A novel mechanism based on Deep Learning (DL) and Residual Network (ResNet) for accurate cracking detection using Electroluminescence (EL) images of PV panels is proposed in this paper. Different kinds of ResNet architectures, where ResNet34, ResNet50, and ResNet152 were tested, came out with an F1-Score of 86.63%, 87.37%, and 88.89%, respectively. Although the accuracy for ResNet152 is slightly higher, ResNet34 was chosen as the best model since it gives us a trade-off between detection performance and computational performance. The main contribution in this research is the design of an efficient crack detection system trained on a large PV power dataset composed of 2000 EL images collected from different polycrystalline and monocrystalline cells. Although the dataset has some imperfections, to guarantee the presence of many cell states in each subset, it was split into training (70%), validating (20%), and testing (10%). This research demonstrates the application of advanced DL frameworks for early defect diagnosis from raw data to enhance PV panel maintenance, thereby bolstering the sustainability of solar systems. This research also has a significant impact on the academic industry, offering practical solutions for the renewable energy sector during periods of sustainable energy instability, particularly when new materials supplement PV panel usage. The technology preserves the efficiency of solar modules and encourages clean energy solutions by accurately identifying PV panel faults. The study lays a foundation for the further development of image-based defect detection methods in PV systems.