An on-line deep learning framework for low-thrust trajectory optimisation

Abstract In the preliminary interplanetary mission design stage, a fast low-thrust (LT) transfer cost approximator will improve the mission design efficiency and enable us to design more complex missions. In this study, we propose a Deep Neural Network (DNN) based on-line framework for training approximators for fast low-thrust trajectory optimisation. The online characteristic refers to the ability to continuously adjust the trained DNNs using new LT transfer data from newly found asteroids, which avoids repeated and costly re-training and is particularly useful for mission scenarios where new data are obtained regularly. The framework contains a DNN-classifier and an online-DNN regressor. The Bayesian optimisation (BO) technique is adapted to determine the network structure as well as the feature selection and processing methods. The proposed DNN-classifier significantly improves the LT optimisation convergence rate from 7% to over 98%. The proposed on-line DNN regressor proves to have better generalization ability and scalability comparing to series network structure, giving a mean relative error (MRE) of approximately 0.6% in the test Near-Earth Asteroid (NEA) rendezvous mission scenario. The proposed on-line DNN framework can also be extended to solve other trajectory optimisation problems in other mission scenarios, such as Main-Belt Asteroid (MBA) missions, active debris removal mission (ADR), etc.