PVEL-ViT: Adaptive token selective vision transformer for photovoltaic cell defect classification in electroluminescence imaging

Abstract

The rapid expansion of photovoltaic (PV) technology has increased the need for reliable and automated defect detection to ensure the long-term efficiency and durability of PV modules. Electroluminescence (EL) imaging provides detailed visualization of internal defects and is a key modality for data-driven classification systems. This study proposes a vision transformer (ViT)-based framework, named PVEL-ViT, for accurate classification of PV cell defects from EL images. The architecture is constructed on a MetaFormer-based backbone and integrates an embedded multi-scale feature fusion module with an adaptive token selective attention mechanism. The designed PVEL-ViT captures fine-grained local defect patterns and global context while dynamically emphasizing defect-relevant tokens and suppressing redundant background information, thereby improving discriminative feature learning and robustness. The proposed framework is evaluated on an EL imaging dataset and compared against recent convolutional and transformer-based models. Experiments show that EfficientNetV2-m and EfficientViT-b2 achieve accuracy rates of 0.9386 and 0.9359, respectively, whereas PVEL-ViT attains a higher accuracy of 0.9584, corresponding to accuracy gains of 2.11% and 2.40% over these models. The obtained results indicate that PVEL-ViT provides a robust and scalable defect classification performance on EL imagery, enhancing the reliability of automated PV inspection pipelines and supporting monitoring in PV modules.