Research On The Key Technology Of Intelligent Thermal Control For Space Telescope Based On Deep Learning

As one of the key technical support systems for spacecraft,the thermal control system has the important task of providing a reliable thermal environment for all space mission units,such as electromechanical equipment and science payloads,throughout the mission cycle of the spacecraft.Traditional thermal control technologies are very practical for spacecraft operating in fixed orbits because of their high reliability and simple control.However,with the continuous introduction of various new space payloads,the temperature control requirements for space telescopes,which are very sensitive to temperature,have become increasingly demanding.When facing complex space missions such as reorbiting and rapid maneuvering,the thermal control system is required to be able to adjust the thermal control strategy autonomously according to the current environmental changes,which poses a great challenge to the traditional thermal control technology: on the one hand,it is required to improve the efficiency of traditional thermal analysis modeling and calculation,and to realize adaptive modeling for multiple different working conditions;on the other hand,it is required to get rid of the over-reliance on the experience of thermal engineers in the thermal design optimization process,so as to improve the optimization efficiency and realize the automation and intelligence of the optimization process;and on the other hand,it is required that the active thermal control system is capable of adaptively adjusting the system parameters online according to the environment and system changes under unsupervised conditions to achieve intelligent autonomous operation.For this reason,this paper gives an overview of the development status of space telescope and its thermal control technology,and proposes an intelligent thermal control technology based on deep learning for space telescope with the advantages of the application of artificial intelligence technologies such as deep learning in surrogate modeling,multi-objective optimization and intelligent control,and around thermal analysis modeling,thermal design optimization,and online real-time tuning and control of active thermal control systems etc.,and carried out research on key technologies such as thermal analysis surrogate modeling for space telescopes based on deep learning,sensitivity analysis of thermal design parameters based on machine learning,thermal design optimization based on statistical machine learning algorithms and intelligent thermal control based on deep reinforcement learning.First,the application of deep learning in thermal analysis modeling for space telescopes was studied.In order to improve the speed and accuracy of generating thermal analysis data sets for surrogate modeling,an intelligent batch processing system for thermal analysis was developed by combining the macro operation development tools embedded in NX/SST with MATLAB and Windows batch scripting language to realize the whole automatic operation of ”parameter sampling>parameter input>simulation calculation>result extraction”;Based on this,the performance of the deep learning-based surrogate model on fitting the finite element method-based thermal analysis model of the space telescope under different working conditions was investigated,and an optimal structured deep neural network(DNN)was determined as the surrogate model for the thermal analysis model.The adaptive fitting of surrogate models under different working conditions can be achieved by transfer learning on the basis of only a small amount of training data,which not only improves the efficiency of thermal analysis modeling and computation,but also improves the poor generalizability of traditional surrogate models,showing a promising future for thermal analysis surrogate modeling of large and complex space telescopes.Then,on the basis of constructing the thermal analysis surrogate model using neural networks,a machine learning-based sensitivity analysis framework for thermal design parameters was proposed and extended to a general framework for sensitivity analysis of thermal design parameters applicable to various types of space payloads.The computational efficiency of the intelligent batch processing system for thermal analysis was further improved by introducing parallel computing.Based on this,the concept of multi-fidelity metamodel was introduced,and the performance of the radial basis function neural network(RBF neural network)surrogate model based on multi-fidelity model on fitting the thermal analysis model of space telescope based on finite element method was investigated,and the RBF neural network was optimized to further improve its prediction accuracy by an improved mind evolution algorithm.In the process of sensitivity analysis,the thermal analysis surrogate model was used to replace the thermal analysis model based on the finite element method to speed up the calculation process,while the density-based sensitivity analysis method was introduced to characterize the density-based sensitivity index by the cumulative function of the corresponding output distribution of each input parameter of the surrogate model.The sensitivity analysis framework was used to analyze the thermal design parameters on the main heat transfer paths of the space telescope and compared with other traditional sensitivity analysis methods to verify the superiority of this general framework and to lay a theoretical foundation for further optimization of the thermal control system design.After fully studying the application of machine learning in thermal analysis surrogate modeling and sensitivity analysis of thermal design parameters for space telescopes,an optimization design method for intelligent and autonomous optimization of thermal control system parameters by Bayesian optimization algorithm was proposed.This method gets rid of the over-reliance on engineers' experience during the traditional thermal design process,and it reduces the computational cost of thermal analysis by means of the DNN-based surrogate model,and adopts a Gaussian process to consider the a priori parameter information,and sets up a acquisition function to heuristically evaluate the current model so as to compare the uncertainty region with the known region with better target values at the upper level and make trade-offs,and then realize the continuous optimization and updating of thermal design parameters.In comparison with various traditional optimization methods,the theoretical and experimental analysis shows that the proposed optimization method in this paper has fewer iterations and faster convergence,which verifies the feasibility and superiority of the application of machine learning in the optimization of the thermal design of space telescopes.Finally,an intelligent thermal control strategy based on deep reinforcement learning for space telescopes was proposed.The application of two classical reinforcement learning algorithms,the Actor-Critic and Deep Deterministic Policy Gradient,in the adaptive tuning of PID thermal controller parameters were studied respectively to achieve the intelligent and autonomous operation of the thermal control system under unsupervised and m K-level precision thermal control.By introducing the ideas of deep learning and transfer learning,the temperature control effect and adaptive adjustment capability of the PID thermal controller before and after the switch of the controlled object were analyzed without re-modeling and training.Finally,the effectiveness and practicality of the intelligent thermal control strategy proposed in this paper were verified by building a ground test system.