RFI Novelty Detection using Machine Learning Techniques

Abstract

In order to ensure that the Dominion Radio Astrophysical Observatory (DRAO) continues to be a great asset to the Canadian astronomical community we must work to actively protect the RF cleanliness of the site. One aspect of this much larger effort is the site monitor project. This is currently realized by an omnidirectional monitoring station mounted on the roof of the main building.A pitfall of previous RFI monitoring projects on site has been the volume of data produced, combined with the time limitations of personnel. Occupancy plots have been produced, but this tool has very limited value for day-to-day maintenance of the site. Simply, no eyes have been available to look at all of the data.Our aim is to deal with the data first: to build a rich description of the RF scene at the site in order to automatically separate “normal” events from “novel” events. To do this we use features extracted from both the spectrogram and the complex baseband waveform. This includes center frequency, bandwidth, received power, transmission duration, time of day, high-order cumulants, and more. We use unsupervised learning techniques to cluster events in this multidimensional space into hierarchical groups. The clustering results allow us to study populations of events and their relationships, rather than individual or small sets of events as in a spectrogram. This feature space also allows us to relate waveforms with similar modulations across frequency, and to reveal temporal patterns. Work is ongoing to bring this analysis into a realtime observing state, in order to provide up-to-date notifications about novel RF events occurring at the DRAO site.