Robust PV fault diagnosis under data constraints using a hybrid vision transformer with cross-attention fusion and uncertainty-aware explainability.

Solar photovoltaic (PV) systems are crucial in modern energy infrastructures for the shift to sustainable renewable energy, making accurate and reliable fault diagnosis essential for maximizing energy yield, operational efficiency, and system availability. However, reliable automated fault diagnosis faces challenges like data scarcity, class imbalance, and lack of calibrated uncertainty in deep learning models. This paper introduces HyViX-PV, a hybrid deep learning framework that integrates EfficientNet-B3 and Vision Transformer (ViT-B/16) through a novel cross-attention fusion mechanism to enhance renewable PV monitoring and its associated energy efficiency. Unlike conventional static fusion approaches, the proposed HyViX-PV module dynamically combines local texture features and global contextual representations via learnable query-key interactions, enabling adaptive feature prioritization based on input characteristics. The presented framework incorporates Monte Carlo dropout for uncertainty quantification, facilitating confidence-aware decision-making and selective prediction abstention. In addition, a multi-granularity explainability pipeline, combining Grad-CAM, attention roll-out, and fusion weight analysis, provides transparent and interpretable diagnostic insights aligned with domain expert reasoning. The proposed model is evaluated using a five-fold cross-validation, a locked test set, and hash-based data leakage prevention. Experimental results demonstrate that HyViX-PV achieves 92.22% accuracy and 0.9230 macro F1-score, outperforming state-of-the-art CNN, ViT, and hybrid baselines while maintaining efficient inference. Cross-dataset validation shows strong generalization with only a 1.05% performance drop. Furthermore, it attains 92.66% accuracy and 0.9364 macro F1-score robustness under severe class imbalance realistic scenarios. By enabling dependable, interpretable, and uncertainty-aware fault diagnosis, the proposed framework supports improved reliability, reduced maintenance costs, and enhanced operational performance of solar PV installations, contributing to the broader adoption and effective utilization of renewable energy systems.