Explainable Artificial Intelligence for Thermal Condition Assessment of High-Voltage Power Equipment Using Infrared...

Area of Science:

• Electrical Engineering
• Materials Science
• Artificial Intelligence

Background:

• Infrared thermography is crucial for non-contact thermal monitoring of high-voltage power equipment.
• Abnormal temperature patterns in thermographic images can signal developing faults or insulation degradation.
• Purely data-driven deep learning models may yield temperature predictions inconsistent with heat transfer physics.

Purpose of the Study:

• To investigate a physics-constrained convolutional neural network (CNN) framework for accurate spatial temperature field estimation from thermographic images.
• To enhance the physical consistency and interpretability of deep learning models for thermal monitoring applications.

Main Methods:

• Developed a physics-constrained CNN framework incorporating a diffusion-based Laplacian residual loss derived from the heat equation.
• Evaluated model performance against a baseline CNN using experimental thermographic data.
• Explored learnable thermal diffusivity and source-term augmentation, and applied Monte Carlo dropout for uncertainty estimation.

Main Results:

• The physics-constrained model demonstrated improved performance over the baseline CNN, with the best configuration achieving an RMSE of 12.013 °C and a CAP R-squared of 0.4646.
• Learnable thermal diffusivity and source-term augmentation did not yield significant performance improvements.
• Uncertainty estimation revealed higher predictive variance near hotspot boundaries and steep thermal gradients.

Conclusions:

• Physics-based regularization and uncertainty estimation enhance the physical coherence and interpretability of thermographic prediction models.
• The proposed framework shows potential for improving the reliability of thermal monitoring in power equipment.
• Further validation on diverse datasets and settings is needed to confirm generalizability.