Data Driven Attitude Takeover Control For Combined Spacecraft

With the rapid development of space technology,on-orbit service missions are increasingly diversified and complicated.Effective attitude takeover control of targets through service spacecraft is the foundation stone for subsequent on-orbit refueling,onorbit maintenance,debris removal,and other missions.The takeover control methods for cooperative targets are mature enough and have been implemented in several on-orbit applications.However,when it comes to the takeover control of non-cooperative targets with complex structure,unknown inertia matrix,and unknown attitude maneuverability,there often exist multiple challenges such as incomplete target information and difficulties in accurate on-orbit identification.This dissertation is dedicated to solving the problem of attitude takeover control of combined spacecraft after capturing by the servicer,exploring the mechanism of input response of the combined spacecraft under complicated on-orbit scenarios,breaking through the traditional model-based attitude control theory,and carrying out data-driven modeling and control methods for the attitude takeover task of combined spacecraft.Based on the background of on-orbit missions,this dissertation closely focuses on the data-driven modeling and control problem of combined spacecraft,aiming to build a novel on-orbit maneuvering system with a high value of engineering practicability,enhancing the versatility,autonomy,robustness,and reliability of the implementation of on-orbit mission,and improving the space intelligent perception and learning control capability.The details are as follows:To address the problem of attitude takeover control of combined spacecraft with completely unknown dynamics,a convolutional neural network-based controllability prediction strategy is proposed to build a general and universal takeover risk assessment mechanism by considering the multiple mission requirements and constraints.Furthermore,a novel online policy iteration Q-learning based attitude tracking control algorithm is proposed,which directly learns the control policy from the system input/output data without identification and any assumptions on the model form.The monotonicity,convergence,and optimality of the iterative sequence are analyzed theoretically.Meanwhile,the off-policy learning is used to perform the policy iteration and an optimal Q-value function approximator is constructed employing neural network structure,which ensures the safety of the on-orbit learning of combined spacecraft.Taking the priori model into consideration,a Gaussian process-based data-driven attitude takeover control strategy is first proposed for combined spacecraft,and Lyapunov stability and boundedness of the closed-loop system in the sense of probability are demonstrated in detail.Furthermore,considering the real-time implementation requirements,the strong maneuverability,and the limited onboard computational resources,a sparse online Gaussian process-based attitude takeover control strategy for the combined spacecraft is presented,which makes full use of the on-orbit data to implement the incremental update of the data-driven model and improves the performance and steady-state accuracy of the closed-loop system.What’s more,the strategy significantly reduces the computational load during the learning process.The algorithm is robust to measurement noise,external disturbances,and active attitude maneuvers of the target,and can effectively support fast and efficient online adaptive processing of new missions.To address the problem of attitude takeover control of combined spacecraft with safety constraints in actual on-orbit missions,firstly,model predictive control with precisely known models for the combined attitude takeover are given and analyzed in detail,respectively.Secondly,a variational inference-based sparse online Gaussian process model predictive control strategy is introduced to achieve attitude takeover tasks under multiple safety constraints and target attitude maneuvers.Wherein,a novel sparse online Gaussian process regression update routine is designed to adaptively handle time-varying and abrupt change scenarios in on-orbit missions.Furthermore,inspired by the biological concept of "memory and learning",a long/short term dual Gaussian process-based model predictive control algorithm is presented.In this way,the system can "remember" task data to obtain a better dynamic response and steady-state accuracy in historical missions.Also,the system is enabled to deal with unknown and new tasks quickly and efficiently,which can meet the requirements of future complex space missions.Based on the simulation platform,the dissertation provides an extensive numerical simulation analysis of the combined spacecraft attitude takeover control tasks,and verifies the effectiveness and practical value of the control strategies in this dissertation.