Machine Learning for Interference Detection and Mitigation on Deep Space Telecom Signals

Detecting anomalies and interference is an essential capability of contemporary flight radios and ground support equipment. The advent of software-defined radios (SDRs) has made this task feasible. These radios allow real-time monitoring and potential mitigation of interference by adjusting frequency bands. This feature is useful as spacecraft components degrade and interference from other instruments may disrupt communications. The implementation of automated anomaly detection systems on SDRs can enhance both spacecraft and ground station testing by identifying abnormal waveforms and potential interference in communication channels. First, this work investigates how machine learning compares against classical signal processing approaches for the anomaly detection task on SDR signals, when no assumptions can be made about the interference signal. By adding more realistic assumptions to the modeled signals, we show cases where classical signal processing methods start to fail in comparison to machine learning approaches. In the second part, we explore the more general problem of mitigating the anomaly by removing it from the received signal. A classical approach (independent component analysis) is shown to be effective if the problem has more receivers than signal sources. This work considers an under-determined setting, i.e., the number of receivers is less than number of sources, where deep learning has been shown to be effective in separating multiple sources from one receiver, as illustrated in the literature for audio source separation.