Denoising 3D Time Projection Chamber data using convolutional neural networks

AbstraktAbstract

Spurious signals caused by microdischarges are a known effect inherent to all gaseous detectors. During the reconstruction in imaging and tracking detectors, such as time projection chambers, these signals are added to the actual track-generated signal as extra pixels or clusters, compromising the performance of the detector. The usual approach to remove these noise patterns is by hardware-dependent heuristics and conditions. In this work, we study the usage of denoising convolutional neural networks (NN) to clean the signals from a Time Projection Chamber (TPC) prototype. We show that this denoising provides also a tool for the selection and rejection of detector events that do not contain any track. The output provided by the neural network is compared with the results obtained using a conventional algorithm. The Physics of the events measured by the detector (such as the shape of the tracks) is used to assess and compare the quality of the two algorithms and how much they improve the existing data set.

Spurious signals caused by microdischarges are a known effect inherent to all gaseous detectors. During the reconstruction in imaging and tracking detectors, such as time projection chambers, these signals are added to the actual track-generated signal as extra pixels or clusters, compromising the performance of the detector. The usual approach to remove these noise patterns is by hardware-dependent heuristics and conditions. In this work, we study the usage of denoising convolutional neural networks (NN) to clean the signals from a Time Projection Chamber (TPC) prototype. We show that this denoising provides also a tool for the selection and rejection of detector events that do not contain any track. The output provided by the neural network is compared with the results obtained using a conventional algorithm. The Physics of the events measured by the detector (such as the shape of the tracks) is used to assess and compare the quality of the two algorithms and how much they improve the existing data set.