Benchmarking Vision Transformers and CNNs for Thermal Photovoltaic Fault Detection with Explainable AI Validation

AI models can be validated against real-world physics: The study demonstrates that explainable AI techniques (XRAI saliency analysis) can verify whether deep learning models are actually learning the correct reasons for their predictions—in this case, confirming that models identified physically meaningful thermal signatures like hotspots and linear thermal paths rather than spurious patterns.

Vision Transformers outperformed traditional deep learning for this task: Swin Transformer achieved 94% accuracy in binary fault detection compared to CNN approaches, suggesting that newer AI architectures may be more suitable for specialized infrastructure monitoring applications, though this finding may be specific to thermal imaging.

Not all faults are equally detectable: The research revealed significant performance gaps across different fault types, with electrical faults detected reliably (F1-scores >0.90) but environmental factors like soiling remaining challenging (F1-scores 0.20-0.33), highlighting that AI systems have real limitations that must be understood before deployment in critical infrastructure.

Interpretability is essential for infrastructure adoption: The paper emphasizes that high accuracy alone is insufficient for deploying AI in energy systems—stakeholders require understanding of why the model makes decisions, and this study shows that explainability validation can build justified confidence in automated monitoring systems.

Systematic validation frameworks are needed before deployment: This work represents the first systematic comparison combining performance benchmarking with physics-based interpretability validation for this application, suggesting that responsible AI deployment in critical infrastructure requires multi-layered evaluation beyond accuracy metrics.