Lightweight vision architecture with mutual distillation for robust photovoltaic defect detection in complex environments

Haoran Zhang, Gong Boao, BohanMa, Zhiyong Tao, Shi Wang

Abstract:

With the rapid growth of solar photovoltaic installations, defect detection in PV power stations has become
crucial for ensuring operational safety and economic efficiency, as undetected defects can lead to significant
performance degradation and potential hazards. Unmanned Aerial Vehicle (UAV)-based Electroluminescence
(EL) imaging offers an efficient solution for large-scale inspection. However, the harsh environmental condi
tions and complex imaging scenarios pose significant challenges to detection models, while edge computing
deployment demands strict resource constraints. This study introduces SCRViT, a lightweight deep learning
model that substantially improves detection performance on low-quality EL images through a spatial-channel
reconstruction mechanism and a peer network co-learning strategy. Experimental results demonstrate that the
proposed method achieves 88.19% detection accuracy on simulated outdoor environment datasets, surpassing
state-of-the-art approaches by 4.77% while reducing model parameters by 55.6%. Through multi-dimensional
interpretability studies – including Shapley value feature attribution, GradCAM attention pattern analysis, and
information-theoretic mechanism analysis – this research systematically elucidates the model’s environmental
adaptation mechanisms. This lightweight yet robust solution enables real-time defect detection on edge devices,
improving inspection efficiency and reducing operational costs while providing reliable decision support for
practical applications in complex outdoor environments.