Online Tuning of Pre-trained Deep Neural Network for Guidance Against Non-cooperative Pursuer

In the astronautics, it is widely acknowledged as a hard problem that using neural network to handle the intelligent observation and guidance when dealing with non-cooperative pursuing vehicles. Due to the difficulty of obtaining training data set from non-cooperative vehicles, the inconsistency between training data and utilization data, application of neural network controller is by far very limited. In order to tackle this problem, an online tuning scheme of neural network is proposed in this paper aiming to boost the observation and counter action abilities against non-cooperative pursuer. In the proposed methodology, the deep neural network controller is pre-trained through dataset from generative network before deployment, and fine-tuned online through real time observations data set. In this way, the calculation pressure is alleviated for onboard computer and the contribution of real time observation data set is utilized more correctly, thus guaranteeing the efficiency and effectiveness of online tuning. Simulation results suggested that, both in the observation ability and counteractive guidance to avoid the pursuer, the proposed method all trumped traditional networks without online tuning, which resulting a higher chance to survive the pursuit.