Spatial Attention Mechanism Derived from Diffusion Model for Photovoltaic Cell Fault Diagnosis

Artificial intelligence is integral to modern energy systems, with a particular focus on renewable energy, where photovoltaic play a central role. Artificial intelligence-driven photovoltaic fault diagnosis has emerged as a critical research topic. However, existing photovoltaic fault diagnosis methods exhibit poor cross-domain generalization, limiting their realworld applicability. This work introduces a diffusion model-based spatial attention mechanism for fault diagnosis. The proposed approach reconstructs fault images into their health counterparts, deriving spatial attention by disentangling fault textures from photovoltaic cell background semantics. To enhance computational efficiency, an approximation strategy is designed. Experiments on three public electroluminescence datasets demonstrate state-of-the-art performance, achieving an average accuracy of $\mathbf{9 5. 9 \%}$ for single-domain PV fault detection, $\mathbf{9 7. 0 \%}$ for singledomain classification, $75.5 \%$ for cross-domain fault detection, and $53.1 \%$ for cross-domain classification.