Fault Detection for Induction Motors Using Diffusion-Enhanced Deep Learning Models With Infrared Thermography

Infrared fault recognition for induction motors remains challenging because thermal patterns are strongly influenced by operating load, ambient temperature, surface emissivity, and viewing conditions, while different fault modes may produce subtle yet highly correlated temperature distributions. These factors lead to large intra-class variability and small inter-class margins, making conventional handcrafted descriptors and standard image-level augmentation insufficient for robust classification. To address these challenges, infrared thermography is adopted as a non-contact sensing modality that captures spatially distributed heat signatures linked to mechanical and electrical anomalies, and a diffusion-enhanced classification framework is developed to improve generalization under realistic thermal variations. Specifically, Stable Diffusion is employed as an offline image-to-image augmentation module and integrated with an EfficientNetV2 backbone to generate semantically consistent thermal samples with diversified appearance. The generation process is guided by component-state textual prompts, where each prompt describes both the monitored motor component and its corresponding operating or fault state. These generated samples expand the effective training distribution beyond limited geometric or photometric transforms. The resulting DM-EfficientNetV2 improves feature learning robustness and reduces confusion among visually similar fault states. Experimental validation on a multi-component motor thermal dataset demonstrates that the proposed method achieves an overall accuracy of 98.2% $\pm ~0.3$ % and consistently outperforms the conventional EfficientNetV2 training scheme without diffusion augmentation, confirming the effectiveness of Stable Diffusion-based offline distribution expansion for infrared motor fault recognition.