DEEP LEARNING-BASED AUTOMATIC DETECTION OF SOILING AND FAULTS ON PHOTOVOLTAIC PANELS

Dust accumulation on the surfaces of photovoltaic (PV) panels, as well as mechanical and electrical defects, are among the main factors contributing to reduced energy generation at solar power plants. This problem is particularly pronounced in arid and dusty regions such as Central Asia. Manual inspection of large PV arrays is labour-intensive, time-consuming and prone to errors. This paper proposes a lightweight and explainable deep learning system for automatically classifying PV panel surface conditions into six classes: clean, dusty, contaminated with bird droppings, electrically damaged, mechanically damaged and snow-covered. Transfer learning is applied using the MobileNetV3-Large neural network pretrained on the ImageNet dataset. To mitigate class imbalance, a class-weighted loss function and data augmentation techniques are employed. The Grad-CAM method is integrated to localise damaged regions. For each detected condition, the system estimates a potential power loss range based on scientific literature and presents the results to users through a FastAPI web application with an Uzbek-language interface. On a publicly available dataset containing 3,065 images, the model achieved a validation accuracy of 78.81 %, a macro-averaged recall of 0.86 and a weighted F1-score of 0.79. All computations were performed exclusively on a central processing unit without a graphics accelerator. Rare but operationally important classes — snow-covered panels (0.98), panels contaminated with bird droppings (0.90) and electrically damaged panels (0.90) — were identified with high recall. The proposed approach provides a low-cost and practical decision-support tool for maintenance operations in regions experiencing rapid development of solar energy.

Hali tarjima qilinmagan