I found his paper: iopscience.iop.org/article/10.3847/…/ad7fe6 (no paywall 😃)

From the intro:

VARnet leverages a one-dimensional wavelet decomposition in order to minimize the impact of spurious data on the analysis, and a novel modification to the discrete Fourier transform (DFT) to quickly detect periodicity and extract features of the time series. VARnet integrates these analyses into a type prediction for the source by leveraging machine learning, primarily CNN.

So they start with some good old fashioned signal processing, before feeding the result into a convolution neutral net. The CNN was trained on synthetic data.

Image

FC = Fully Connected layer, so I assume they mix FC and convolution layers in their NN. I need to read the whole paper, which won’t happen right now.