Exploiting the Latent Space of Deep AutoEncoders for the Identification of Signal Pulses in Noisy Time-Series

We propose a data-driven procedure, based on convolutional variational autoencoders, to identify the presence of signal pulses in long time series. The dataset consists of synthetic waveforms, each composed of non-Gaussian noise and a log-normal-shaped signal of variable intensity, with a length of 10,000 samples. The model heavily compresses the input waveforms, allowing a direct study of such a reduced representation. After training for 150 epochs on 7500 waveforms, a region in the latent space where the network encodes time-series presenting only background noise emerges, allowing, in turn, to tag as candidates for containing a signal those falling outside. When applied to a test dataset of freshly generated waveforms, 100% of events with signal amplitudes well above the baseline noise are correctly labelled, and this fraction only decreases for amplitudes comparable with accidental noise pulses. This approach was designed to fully exploit the measurements in dual-phase Liquid Argon Time Projection Chambers, as the one of the Recoil Directionality experiment, built in the context of the Darkside project.