DASFormer-Net: A Diffusion-Augmented Self-Supervised Dual-Attention Transformer Framework for Trustworthy Photovoltaic Defect Analysis

Abstract

To guarantee the quality, safety, and durability of solar modules, reliable PV defect inspection is crucial. However, the current deep learning methods are faced with challenges of few annotated samples, serious class imbalance for rare defects, and not enough integration of local and global contextual features, and inadequacy of interpretability for the model, limiting their applications in industrial quality control systems. To overcome these difficulties, in this paper, we present DASFormer-Net, a diffusion-augmented self-supervised dual-attention transformer network for reliable defect analysis of PVs. It is proposed that the framework is based on a physics-inspired conditional latent diffusion model, which is used to create realistic synthetic samples of defects that are not common, but still maintain important PV structural properties. To avoid reliance on large annotated datasets, a self-supervised masked autoencoder is used to extract robust feature representations from unlabeled EL images. It adopts a novel dual-attention fusion paradigm, in which a ConvNeXt V2 branch is introduced for local feature learning, while a Swin Transformer V2 branch is introduced for the global context modeling. Moreover, a multi-task learning strategy is coupled with the hierarchical context pyramid for defect classification, semantic segmentation, severity grading and confidence prediction. Inference pipeline is complemented with explainability tools derived from Grad-CAM++ to enhance model trustworthiness and uncertainty quantification from Monte Carlo dropout. DASFormer-Net is evaluated through extensive experiments on both public and industrial photovoltaic datasets, outperforming the current state-of-the-art methods in terms of macro F1-score (89.7%), mean Intersection-over-Union (83.5%), Dice Similarity Coefficient (86.2%), and Expected Calibration Error (0.064). The proposed framework accurately, interpretably and reliably solves the problem of automatic inspection of photovoltaic defect, and also has a good adaptability for industrial quality assurance in actual production process.