Improving explainability of attention-based deep learning models for thermal analysis systems

To improve the explainability and reliability of deep learning models for thermal monitoring and industrial fault diagnosis, an Entropy-Guided Attention Filtering (EGAF) framework is presented. Conventional explainability techniques often produce noisy, spatially inconsistent visual explanations, making them difficult to interpret from an engineering perspective, even though attention-based neural networks can achieve acceptable forecasting performance. The proposed framework is an explainability pipeline that integrates transformer attention components, spatial attention modules, entropy-guided filtering, and multi-layer attention aggregation to address this challenge. It can be used with various deep learning architectures, such as Vision Transformer (ViT), attention-based ResNet50, and ConvNeXtV2 with focal self-attention, to generate informative thermal attention maps for extraction and refinement. Experiments were run on CIFAR-10, ISIC 2019, and SST-2 benchmark datasets and compared to Grad-CAM, LIME, SHAP, guided backpropagation, and traditional attention visualization approaches. The EGAF framework showed the highest explainability, fidelity, and localization with the lowest DAUC (0.281), highest IAUC (0.824), IoU (0.801), and localization accuracy of 91.7%. Noise suppression of noisy activations and concentration of attention were further improved by entropy-guided filtering, leading to greater similarity in correspondence with the predefined reference target regions. Comparative and robustness analyses showed that the framework consistently provided stable, semantically meaningful explanations in complex thermal environments and under noisy conditions. These results demonstrate EGAF’s effectiveness in the attention localization, explanation fidelity, and noise suppression aspects in heterogeneous test datasets. While the results are encouraging for thermal imaging and industrial fault analysis, further validation using diverse industrial thermal images is required to ensure applicability in real-world scenarios. To sum up, the proposed EGAF framework provides a computation-effective and highly interpretable solution to enhance transparency and trustworthiness in thermal engineering applications with AI assistance.