Search

While deep learning has shown its promises in solving spacecraft optimal control problems, it still requires a large amount of training samples and is lack of the interpretability. In this article, we introduce the nonparametric flexibility of kernel approaches into the deep learning architecture, and develop a deep kernel learning model to provide an interpretable control prediction in various aerospace missions. In our framework, the core is a Gaussian process function with RBF kernel that can transform the latent features into infinite-dimensional latent space to obtain interpretable similarities. It hence intuitively brings the positive or negative correlations between the predicted result and all training samples. The parameters of both the deep architecture and the kernel are jointly trained in an end-to-end manner. Simulation results in two aerospace missions show that the proposed model outperforms the stand-alone deep neural network model and the Gaussian process model in control prediction accuracy and interpretability under insufficient training samples. [ABSTRACT FROM AUTHOR]

Spacecraft relative trajectory planning is central to many space missions like on-orbit service and debris removal. While early attempts of machine-learning-based trajectory planning have been witnessed, they still face the open problem of how to acquire sufficient training samples to conduct robust training. In this article, we have introduced a meta-learning framework to improve the adaptation ability of the planner when facing new initial conditions. To achieve this goal, we divided the training trajectories as sub-training samples and fake testing samples. Then, the meta planner is trained by repeatedly conducting a trail training-and-testing process. To this end, the gradient information of the meta learner is initially obtained on the subtraining sets and is further adjusted by looking at its testing performance on those fake testing data. Therefore, the meta planer explicitly take account of the potential testing performance into consideration and, hence, alleviate the overfitting phenomena with few training trajectories. Simulation results substantiate the effectiveness of our approach, as well as the advantages of quick adaptation to new initial conditions without overfitting. [ABSTRACT FROM AUTHOR]