Аннотация
Solar panel performance and longevity depend heavily on reliable condition monitoring, yet classification approaches often struggle with class imbalance and limited expressivity in their final layers. This paper explores the application of the Kolmogorov-Arnold Network (KAN) framework to solar panel condition classification using a dataset of 804 images across six distinct categories. The proposed EfficientNet-KAN architecture integrates a pre-trained EfficientNetB0 backbone with a KANLinear classifier that uses B-spline basis functions to replace traditional linear mappings. Our implementation addresses the inherent class imbalance through weighted sampling, class-weighted loss functions, and targeted augmentation strategies. Experimental results show a weighted F1-score of 0.85, with varying performance across different conditions—clean, dusty, bird-drop, snow-covered, electrical-damage, and physical-damage. This study provides insights into how KANLinear modules might capture visual patterns associated with different panel conditions, contributing to the ongoing development of condition monitoring approaches for renewable energy systems.