Research On Uncertainty Quantification Method And Application Of Deep Surrogate Models For Satellite Thermal Field Analysis

The rapid development of space technology and the increasing demand for applica-tions requirements have brought about more stringent requirements on the high-quality and swift system design of satellites.As a significant part of the satellite system design,the efficiency and accuracy of satellite thermal field analysis directly affect the overall satellite performance,development cost and design level.For the past few years,the use of deep surrogate model to replace physical simulation achieves near real-time thermal field prediction,which has become the primary solution to such problems.However,the large number of parameters and complex structure of the deep surrogate model used for satellite thermal field analysis tasks lead to the high complexity of uncertainty quan-tification,and the two common uncertainty quantification methods based on ensmeble and sampling face the challenges of inefficiency and inaccuracy in this task.As to this problem,this dissertation improves the two methods above and proposes efficient and accurate methods for quantifying the uncertainty of deep surrogate model,fully exploits the information contained in the uncertainty of the deep surrogate model,and applies it to the improvement of satellite thermal field prediction accuracy and satellite thermal field reliability analysis.In summary,the main research contents of this dissertation include the following four aspects:First of all,a transfer learning-based uncertainty quantification method for deep sur-rogate model is proposed to address the problem of high training cost of ensemble-based deep surrogate model uncertainty quantification methods.This method uses the seed net-work already trained to continuously widen so as to obtain the hatching network.Since the widened hatching network inherits part of the parameters from the seed network,the hatching network can converge quickly.Taking advantage of multiple hatching networks to quantify the uncertainty of the deep surrogate model can greatly improve the quantifi-cation efficiency.Experimental results show that the proposed method can significantly reduce the time cost required to train multiple ensemble models,and thus can quickly quantify the deep surrogate model uncertainty for satellite thermal field analysis.In the second place,in allusion to the problem of low quantification accuracy of sampling-based deep surrogate model uncertainty quantification methods,a random ReLU-based deep surrogate model uncertainty quantification method is proposed.This method embeds randomness into the ReLU module of the deep surrogate model,and amplifies this random effect through a large number of ReLU modules to obtain more diverse out-puts,and thus quantifies model uncertainty with higher accuracy.Experimental results show that the proposed method cannot only improve the prediction accuracy of the deep surrogate model for the thermal field,but also effectively capture the relationship between the uncertainty of the deep surrogate model and the error of the thermal field prediction.Then,to address the few-shot problem of difficult construction of high-precision sur-rogate models in satellite thermal field analysis due to the less amount of labeled data,a surrogate model construction method for satellite thermal field prediction based on model uncertainty and semi-supervised learning is proposed.In combination with the relation-ship between the uncertainty of the deep surrogate model and the error of the thermal field prediction,the component layout data without thermal field labels are pseudo-labeled to form a new dataset.and then the new data set is used to train the model again to further improve the prediction accuracy of the model for the thermal field.Experimental results show that the method can effectively improve the thermal field prediction accuracy,and the more unlabeled data,the more obvious the improvement is.Finally,a satellite thermal field reliability analysis method based on the interval pre-diction of the surrogate model is proposed for the problem of satellite thermal field re-liability analysis considering the uncertainty of the deep surrogate model.The interval prediction results of satellite thermal field can be obtained by combining the information of the standard deviation of thermal field prediction provided by the uncertainty of the deep surrogate model.Compared with the previous single-value prediction results of the satellite thermal field,the interval prediction results include the effects brought by the uncertainty of the deep surrogate model.From the interval prediction results of the ther-mal field,more reasonable reliability interval estimates of the thermal field can be further obtained,thus providing a more scientific reliability ranking for the preferential selection of the component layout.The dissertation conducts a comprehensive and in-depth study on the quantification and application of uncertainty in deep surrogate models.Efficient and accurate uncer-tainty quantification can assist in the construction of high-precision surrogate models with few-shot and achieve more reasonable thermal field reliability analysis,which lays a solid foundation for further accurate and reliable satellite system design.