Automated Fault Classification in Solar Panels Using Transfer Learning with EfficientNet and ResNet Models

Authors : Rojbin Akınca, Hüseyin Fırat, Mehmet Emin Asker

Pages : 164-173

Doi:10.36222/ejt.1533783

View : 18 | Download : 35

Publication Date : 2024-12-24

Article Type : Research Paper

Abstract :Classifying and detecting faults in solar panels using deep learning methods is crucial to ensuring their efficiency and longevity. In this study, we propose a model that concatenates ResNet and EfficientNet to classify faults in solar panel images. ResNet\\\'s advantage lies in its residual connections, which help mitigate the vanishing gradient problem and improve training of deep networks. EfficientNet is known for its scalability and efficiency, providing a balanced trade-off between accuracy and computational cost by optimizing network depth, width, and resolution. Together, these models enhance fault classification accuracy while maintaining efficiency. To evaluate the performance of the proposed model, experimental studies were conducted using a solar panel dataset with six classes: bird-drops, covered snow, dusty, clean, electrical and physical damage on the surfaces of solar panels. The results demonstrated that the ResNet101 + EfficientNetB1 concatenation achieved superior performance, with an accuracy of 87.55%, F1-score of 88.13%, recall of 88.75%, and precision of 87.92%. This concatenation provided significant improvements in fault classification metrics compared to individual models.
Keywords : Solar panel, Transfer learning, EfficientNet, ResNet, Fault classification