This research was done by scientists from the Catalan Centre for Telecommunications technology and a combination of several companies and organisations such as Eutelsat, GMV Aerospace, and the European Centre for Space Applications and Telecommunications.
The problem that they were presented with was that for satellite communications operations human operators are still very widely relied on. This results in human error and latency affecting issue resolution and operation adaptation. More specifically they focused on:
- Interference detection
- Flexible payload configuration
- Congestion prediction
Their proposed solution involves a combination of machine learning and data science. The innovation in this case is triggered by the availability of relatively deep datasets for satellite operations and advancements in machine learning. Machine learning specifically is good for unusual pattern recognition which is usually the area where humans excel and data science is used to interpret the complex results.
Interference detection is done by using an Autoencoding Neural Network based on CNN for its pattern recognition. It was trained on no-interference data and exploits the rigidity of task specific models. ANNs very simply put try to reconstruct their input, so the authors of the paper exploit this fact, by training the model to reconstruct no-interference data, meaning that when data with interference is input, there will be a detectable non-insignificant error in the output that can be interpreted as interference in the input data. This solution detects only 19% false positives versus the 31% of the existing energy detectors. It also works in a wider signal range.
Flexible payload configuration means allocating satellite resources dynamically. The ones mentioned in this paper are beam pattern, transmit power, and specific frequency. For the solution they use a DNN and Genetic Algorithm combination which proposes solution configurations to human operators, so that they don't have to come up with solutions completely from scratch. This application was evaluated by measuring the reduction of unmet user capacity for satellite connection requests, and 32% reduction was reached compared to only human operators.
User demand congestion prediction is traditionally done by analysing recent usage data which generally means only short term predictions are available and the prediction models are not adaptive to unplanned usage spikes. a long-short-term-memory NN was applied and compared with conventional data science algorithms used for this, and resulted in an ~10% prediction error decrease.
Long-short-term-memory networks: https://machinelearningmastery.com/gentle-introduction-long-short-term-memory-networks-experts/#:~:text=Long%20Short%2DTerm%20Memory%20
Satellite beam patterns: https://web.archive.org/web/20200402075612/http://www.satsig.net/satellite-internet-access.htm
Autoencoders: https://en.wikipedia.org/wiki/Autoencoder
Slides: