Deep Learning Based Cooperative Scheduling with Distributed Non-linear Predictive Coding for Small Spacecraft Swarms

A deep learning based non-linear predictive coding (NLPC) source compression and cognitive cooperative relay scheduling approach is developed for small-spacecraft swarms used in imaging based remote sensing missions. A LEO satellite swarm complemented by a number of earth stations and GEO satellite relays is specifically considered, although the developed approach is applicable for lunar and other planetary and space destinations as well. Each swarm node employs a recurrent convolutional neural network (RCNN) to predict its future observations based on immediate past observations and encodes the prediction error as its payload. A novel image reconstruction method that ensures maintaining an exact replica of the RCNN at the earth receiver without explicitly sharing the network weights is also proposed. Proposed cognitive cooperative relay scheduling algorithm also makes use of the same RCNN to predict future payload size based on which nodes are categorized as source and relay nodes. A multi-objective optimization algorithm is formulated that allows relay capacity allocation by taking into account various desired performance objectives and priority handling. Simulation results verify that the proposed NLPC and cooperative scheduling can significantly increase the amount of image data delivered to the earth while also drastically cutting down image delay and improving fairness among data delivered from each satellite node.