Anomaly Detection for Spacecraft Radios Based on Open-Loop Recording Data

Anomaly and interference detection is a very valuable feature of modern flight radios and ground support equipment. With the rise of software-defined-radios (SDRs), this has now become a possibility. In the case of flight radios, the ability to monitor the channel for any possible interference gives rise to the possibility of mitigating the interference by potentially shifting frequency bands. This has previously occurred in flight for the Mars Reconnaissance Orbiter when an interference tone from its CRISM instrument caused its Electra relay radio receiver processing to fail. However, the detection of this interference tone was manual and was facilitated by the Electra radio’s ability to capture short duration, narrowband open loop captures. Considering that spacecraft components fail gradually, and interference patterns of other instruments may eventually begin to impact the radio, it is crucial to enable automatic and periodic anomaly detection on the spacecraft. Additionally, the spacecraft radio undergoes considerable testing during its integration and test campaign. Considering SDRs are used during this testing on the ground station emulation side, it is possible to implement the anomaly detection algorithms on these SDRs to support testing and identifying when a flight radio might be producing abnormal waveforms or interference might be seeping into its downlink channel from neighboring equipment. Finally, these anomaly detection techniques implemented on the ground support equipment can also monitor the uplink channel for interference if it can sample from the uplink frequency. In this work, we evaluate a collection of relatively simple algorithms for anomaly detection schemes from the PyOD (Python Outlier Detection) library among others for time-series anomaly detection. The input comprises of open-loop captured I/Q data (which assumes that the spacecraft radio supports open-loop recording like the Electra radio). The training is semi-supervised in the sense that the training data only contains nominal data, and no assumptions are made as to the nature of the interference. These algorithms were executed on synthetic uplink direct carrier modulation data and receiver-operating-curves (ROC) were derived for various interference types, including tone, chirp, and filtered additive white Gaussian noise. The data was swept over desired signal to noise ratio and interference power to desired signal power ratio. The results show that as the interference power increases, the anomaly detection schemes tend to perform better, and tone interference tends to be easiest to detect relative to the white Gaussian noise interference. We summarize the best performing algorithms and discuss their effectiveness in detail.