EfficientNet-driven deep learning for accurate detection of faults in photovoltaic cells

Abstract

The global transition to renewable energy resources has driven the need for potent international practice attention on optimized Photovoltaic (PV) system performance, particularly increased PV cell efficiency. Due to environmental conditions and manufacturing defects, the process of energy conversion through these cells may regress substantially. Early identification of these issues can be the key to improving maintenance procedures and extending the useful lives of solar panels. In this study, a robust Deep Learning (DL) framework based on the EfficientNetV2 architecture is presented to increase the accuracy of fault identification in PV cells. Three EfficientNetV2 variants—EfficientNetV2B0, EfficientNetV2B2, and EfficientNetV2M—were evaluated to identify the most effective model for this paper. This solution combines sophisticated techniques for image preprocessing and augmentation by employing a dataset of 2,500 images, including both defective and non-defective cells. This approach makes it possible for us to tailor the model, especially for the difficult process of anomaly detection. Among the assessed models, EfficientNetV2M exhibited the highest performance, achieving an overall accuracy of 89.6%, precision of 88.6%, recall of 77.5%, and F1-score of 82.7%, signifying superior generalization and learning capability. EfficientNetV2B2 demonstrated the highest validation accuracy of 82.4% in the feature extraction phase, indicating strong consistency throughout training. The ability to accurately detect both nuanced and overt issues underscores the suitability of the EfficientNetV2-based approach for the preventive maintenance of PV systems.