Evidential multi-model CNN integration for visual fault detection in solar panels

This study proposes an evidential fusion framework for classifying visual defects in solar panels using convolutional neural networks (CNNs) and Dempster-Shafer theory (DST). Three pretrained CNN models: ResNet50, MobileNetV2, and EfficientNetB0 are fine-tuned to detect multiple defect types, and their outputs are fused at the logit level using DST. Unlike conventional ensemble strategies such as majority voting, the proposed method explicitly accounts for uncertainty and conflict among predictions by assigning belief masses to sets of hypotheses. Experimental evaluations conducted on a multiclass solar panel dataset demonstrate that DST fusion consistently outperforms individual models and majority voting across all macro-averaged metrics, particularly in underrepresented or visually ambiguous classes such as Physical-Damage and Dusty. These findings underscore the potential of uncertainty-sensitive model fusion to enhance the robustness and interpretability of automated photovoltaic inspection systems.