Removing Noise From Simulated Events at The Main Drift Chamber of BESIII Using Convolutional Neural Networks

Abstract

BESIII is the particle detector of the Beijing Electron-Positron Collider, which is a {\tau} -charm factory working at energies around 4 GeV. The first part of the detector, around the collision site, is called the Main Drift Chamber, MDC. The events recorded at MDC are mixed with the background noise of various origins. On average, about 10% of the hits of an event are noises. Still, the noise level differs event by event, and some of the events might even get more noise hits than signal hits, making the analysis less efficient. The standard algorithms of the offline software system of BESIII reconstruct signal tracks using the polluted data. This reduces the reconstruction efficiency of high noise tracks. In this article, we test the idea of using supervised deep learning techniques to remove this noise beforehand. We generate Monte Carlo events, then mix them with noise hits coming from real data. At first, we use deep learning techniques to classify the hits based on their individual features. Then, we simplify every event to a 40 by 43 picture and use image recognition tools to remove the noise. The average noise level for these events with only two signal tracks is about 30%. On average, the techniques presented in this article can purify Bhabha events to nearly 99% while preserving about 99% of the signal tracks.