Machine Learning Models for Optimisation of Satellite Laser Communication Terminals and Optical Space Networks

Machine Learning Models for Optimisation of Satellite Laser Communication Terminals and Optical Space Networks

Utah State University, Logan, UT

Laser communication systems in optical SatCom systems are mainly used in point-to-point networks, possibly deploying standard routing solutions which perform sub-optimally when deployed in transport and access satellite networks. Optical signal acquisition is also affected by the changing satellite environment and atmospheric conditions.

In response, Craft Prospect and partners focused on upgrading optical systems by using machine learning (ML) methods to improve laser communication terminals and to simplify interconnectivity of networked laser SatCom systems. In this context we target the European OPS-SAT Versatile Optical Laboratory for Telecoms (VOLT) mission led by Craft Prospect and the European High Throughput Optical Network (HydRON) project which benefits from using machine learning systems.

Based on initial research and development work on enhancement of satellite free space laser communication systems with machine learning, two use cases and their requirements were identified. Firstly, leveraging autonomic networking principles to develop distributed ML agents for space network nodes which monitor local network state and can autonomously make rerouting decisions on impaired links in a localised way to improve the average network throughput. This enables links between optical ground stations (OGS) and the space network segments to quickly switch in a smooth and responsive way without having to have multiple paths open; searches carried out for preset alternate paths in long/overloaded flow table lists; or an over-reliance on a software defined networking (SDN) controller. Secondly, for space segment laser terminals, development of a ML model to boost detection accuracy in coarse acquisition in reduced signal-to-noise ratio (SNR) cases with strong background light conditions and varying beacon intensity. This is relevant for the scenario where a LEO satellite receives an uplink signal from OGS with strong atmospheric fluctuations or background Earth reflections and the detector is saturated. Similarly, in inter-satellite links (ISL) where satellites are undergoing relative movement and the signal intensity rate of change is rapid, the acquisition sensor quickly arrives in a region where SNR is low, or the sensor is saturated.

In this contribution, we present the developed ML models for these use cases and describe training techniques and datasets. We discuss performance results from tests carried out with a network simulator with the results compared to an SDN controller solution to demonstrate the benefits. For small network topologies, the ML solution resulted in the throughput at the network endpoint being 16.7% higher following link degradation than the SDN controller solution. The spatial acquisition detection errors and accuracy under different SNR conditions using the ML solution was compared to a benchmark Centre of Gravity detection method typically used in laser terminal beacon acquisition. Results showed that the beacon detections from the ML solution were closer in distance to the true spot than the conventional method in low SNR conditions by over an order of magnitude, in addition to having an overall accuracy of 97.77%.

With this we show how to enable autonomic routing in optical data, and laser terminals with improved acquisition rates in networks for OPS-SAT VOLT and for multitudes of future interconnected laser SatCom missions.