Machine Learning Method In Spaceflight Control

With the gradual deepening of human exploration into space,deep space exploration and asteroid exploration have attracted more and more attention recently.Low-thrust spacecraft has been widely used in deep space exploration due to its high specific impulse.Without a doubt,it is difficult to real-time control the spacecraft in such a complicated environment of deep space since the control is continous.The most remarkable thing about the space exploration mission is how to design an optimal trajectory in the case of limited fuel,explore more targets and be robust to the unknown disturbance.It is of great significance to complete these missions.Therefore,this thesis focuses on solving these problems with state-of-the-art machine learning algorithms.In the first place,this thesis studies on searching the preliminary range of multi?target low-thrust mission.The main belt asteroids are clustered using a density-based spatial clustering of applications with noise algorithms and select the core asteroid as the first target to explore.An orbit indicator was proposed to surrogate the transfer value to represent the difficulty of transferring between asteroids.It's a new approximator and it avoids the complexity of computing the transfer value.The orbit indicator defines a new distance and then finishes the clustering.Secondly,a new ensemble learning technique is used in fast estimating the optimal arrival mass for low-thrust spacecraft between near earth objects.As the low-thrust spacecraft transferring orbits are complex in dynamics,and its difficulty in optimization,it is time-consuming to design an actual orbit,so it's meaningful to estimate the transfer mass in the preliminary phases of designing an interplanetary trajectory.Furthermore,a good estimation of mass cost will also help to get better results in the vast combinatorial part.Finally,the deep neural network is used as an on-board representation of the optimal guidance profile,we use feedforward deep neural network architectures trained in a supervised manner.Traditional optimal trajectories for lunar soft landing cannot be calculated in real time if there's error when spacecraft entering orbit or during the descending stage,the deep neural networks here trained with state variables as sample and control variables as label achieve optimal control capabilities so that it can replan the trajectory in real time.