Distilled Probability Regression Model for Heat Reliability Analysis

Facing the complex thermal environment in space, the heat reliability analysis of the satellite is important for the operation. To tackle the resource consumption and overconfidence of traditional methods for heat reliability analysis, a deep learning-based surrogate model is developed in this paper. By distilling from an ensemble, the model maps from a heat layout to a temperature field as an image-to-image probabilistic regression task with uncertainty. With the output obtained by the distilled model, heat reliability based on the failure rate of the circuit board in a satellite is analyzed. The results evaluate the performance of the model for prediction and heat reliability analysis.