Speaker
Description
Radio SETI searches utilising the Green Bank Telescope over many years have produced a vast database of observations. With the large volume of this data, there still exists potential for the discovery of interesting SETI candidates—as well as non-SETI related anomalies—through re-analysis with newly developed detection methods. However, radio frequency interference is prevalent within this data, and the absence of a known reference signal makes extracting potential signals challenging. To confront these issues, we explore the application of machine learning methods designed to enhance signal detection capabilities.
Specifically, the methods explored within this work utilise a deep autoencoder trained on observational data and injected signals. By examining the latent space of the autoencoder, we explore two approaches for anomaly detection. The first employs a simple reconstruction loss method, identifying sections of the dataset with the highest reconstruction loss as anomalies. The second approach leverages the “on-off” observation pattern of the Green Bank Telescope, wherein a primary target system A is observed alternately with nearby systems B, C, or D, producing a sequence of observations in the order ABACAD. Anomalous data segments are identified if observations of system A occupy a distinct region of the latent space compared to observations of systems B, C, or D, suggesting unique signals present only during observations of the primary target.
These methods are evaluated by injecting previously unseen anomalous classes of broadband complex signals in a proof-of-concept test.
