Detecting Dispersed Radio Transients in Real Time Using Convolutional Neural Networks

We present a methodology for automated real-time analysis of a radio image data stream with the goal to find transient sources. Contrary to previous works, the transients we are interested in occur on a time-scale where dispersion starts to play a role, so we must search a higher-dimensional data space and yet work fast enough to keep up with the data stream in real time. The approach consists of five main steps: quality control, source detection, association, flux measurement, and physical parameter inference. We present parallelized methods based on convolutions and filters that can be accelerated on a GPU, allowing the pipeline to run in real-time. In the parameter inference step, we apply a convolutional neural network to dynamic spectra that were obtained from the preceding steps. It infers physical parameters, among which the dispersion measure of the transient candidate. Based on critical values of these parameters, an alert can be sent out and data will be saved for further investigation. Experimentally, the pipeline is applied to simulated data and images from AARTFAAC (Amsterdam Astron Radio Transients Facility And Analysis Centre), a transients facility based on the Low-Frequency Array (LOFAR). Results on simulated data show the efficacy of the pipeline, and from real data it discovered dispersed pulses. The current work targets transients on time scales that are longer than the fast transients of beam-formed search, but shorter than slow transients in which dispersion matters less. This fills a methodological gap that is relevant for the upcoming Square-Kilometer Array (SKA). Additionally, since real-time analysis can be performed, only data with promising detections can be saved to disk, providing a solution to the big-data problem that modern astronomy is dealing with.