A Back Propagation Based Spiking Neural Network Approach for Intelligent Link Decisions In Satellite Communication
MetadataShow full item record
A Spiking Neural Network (SNN) with neuromorphic architecture for optimal link decisions is put forward in this paper. SNNs can adapt to the various changes in the working environment quickly, for maintenance or advancement of the selected performance metrics. Such results can be appealing for satellite networks with orbital operations involving either stationary or manned aids, which would provide directions for autonomy in CN decisions. The satellite on-board processing capabilities, traditionally, have been a limiting factor for advanced satellite communication strategies. Additionally, with deep space explorations rising, the demand for bandwidth is increasing, which can be achieved by making communication systems more efficient. Manual updating procedures for satellite operations gives rise to chances of configuration errors. Since AI has been showing continuous improvements and glorious performances, when applying it to convert manual operations to intelligent ones, some errors can be avoided. In scenarios where the delay time of an operator responding is considerable, the spacecraft must be able to autonomously make decisions. Intelligent systems can help improve spacecraft reliability by being trained to react to unexpected situations and guide the spacecraft to safer operational states with autonomous decision-making. This serves as an apt area to apply an SNN model for a lighter space network on a first-hop level. This literature will be focused on enabling flexible routing for link selection with the help of Spiking Neural Networks. This path selection problem is approached by applying Spike Neural Network (SNN) to classify satellite downlinks based on link cost to improve learning, and later analyze the classification and link decision capabilities of the network with respect to a traditional neural network. The spiking network has inbuilt Back Propagation (BP) implemented in the framework, Nengo. The system managed to achieve better accuracy even when activation was provided in hidden layer instead of output layers. Tweaking the firing rates, epochs and batch size of the data might yield better results. For the LEO scenario, a maximum accuracy of 86% was obtained for synthetic data using SNN and for the GEO scenario, a maximum of 98.5% was obtained.