Satellite racemization method based on deep reinforcement learning

Abstract

The invention discloses a satellite racemization method based on deep reinforcement learning, and the method is characterized in that the method comprises the following steps: marking data samples ofa known satellite, and establishing a sample data set of the known satellite; training the sample data set by using a full convolutional neural network, so that the terminal can understand and identify known satellites in the images or videos to obtain confidence maps of key points of the known satellites in the images or videos; tracking the motion trails of the key points in the video, and estimating the pose of a known satellite through a PNP algorithm; training optimal racemization through a DDPG algorithm, and completing satellite racemization by brushing the side edge of the spacecraft sailboard through racemization of the space manipulator. According to the method, by means of deep reinforcement learning, racemization of a high-speed spinning out-of-control satellite is achieved, meanwhile, a computer makes contact with data and a model environment in combination with visual information, the optimal grabbing pose is trained, and the target capturing accuracy of the space mechanical arm is improved.

Description

technical field [0001] The invention relates to a satellite derotation method based on deep reinforcement learning, and belongs to the technical field of satellite derotation methods. Background technique [0002] With the increase in the number of spacecraft in orbit and their wide application, real life is increasingly inseparable from the various application functions provided by spacecraft in orbit. Due to the limitation of the space on-orbit working organization's own conditions and the influence of the space environment, without any supply and maintenance, the operation is often forced to stop due to limited fuel, outdated equipment or module failure, and a new system has to be remanufactured and launched to replace it. Replacement, resulting in unnecessary losses and waste. GEO is the geosynchronous orbit. Carrying out GEO on-orbit maintenance and service and research on related technologies can effectively prolong the service life of the on-orbit system, and at the ...

AI Technical Summary

Problems solved by technology

Q-learning makes it possible to find the optimal action strategy without the knowledge of immediate reward function and state transition function. In other words, Q-learning makes reinforcement learning no longer dependent on the problem model, but still needs to know the final reward or goal state

Images