A Deep Learning Approach for Fault Detection in Photovoltaic Systems Using MobileNetV3

Abstract

This study investigates the use of the MobileNetV3 deep learning architecture for fault detection in Photovoltaic (PV) systems. The research developed a model capable of classifying solar panels under six different conditions: clean, physically damaged, electrically damaged, snow covered, bird droppings covered, and dusty panels. Using a dataset obtained from Kaggle, pre-processed and divided into training (70%) and test (30%) sets, the MobileNetV3 model achieved a validation accuracy of 95%. Confusion matrix analysis showed high classification accuracy, in particular 100% accuracy for snow-covered and bird droppings-covered panels, with F1 scores as high as 98.73% for certain classes. Training and validation curves confirmed stable learning with low loss values. Compared to models such as EfficientB0 + SVM and InceptionV3-Net + U-Net, MobileNetV3 demonstrated competitive accuracy and computational efficiency, making it suitable for resource-constrained devices. This approach improves energy efficiency, reduces manual inspection, and promotes sustainable energy production. Future work will expand the dataset to include different climatic conditions and fault scenarios, improving the robustness and real-world applicability of the model.