The high performance parameterization for deep learning in pulse shaping

Hui Wang; Dong Wang; Fan Zhang; Ni Fang; Yanwei Kui; Zhuo Zhou; Shiqiang Zhou

doi:10.1016/j.apradiso.2023.110787

The high performance parameterization for deep learning in pulse shaping

Appl Radiat Isot. 2023 Jun:196:110787. doi: 10.1016/j.apradiso.2023.110787. Epub 2023 Mar 24.

Authors

Hui Wang¹, Dong Wang², Fan Zhang³, Ni Fang¹, Yanwei Kui¹, Zhuo Zhou¹, Shiqiang Zhou¹

Affiliations

¹ PLAC, Key Laboratory of Quark & Lepton Physics (MOE), Central China Normal University, Wuhan, 430079, China.
² PLAC, Key Laboratory of Quark & Lepton Physics (MOE), Central China Normal University, Wuhan, 430079, China. Electronic address: dongwang@mail.ccnu.edu.cn.
³ Hubei Collaborative Innovation Center for High-efficiency Utilization of Solar Energy, Hubei University of Technology, Wuhan, 430068, China.

PMID: 36996532
DOI: 10.1016/j.apradiso.2023.110787

Abstract

In high energy physics, front-end data acquisition systems based on analog-to-digital converters (ADC) can provide multiple aspects (time, energy, position) of information when an incident particle is detected. To process the shaped semi-Gaussian pulses from ADCs, multi-layer neural networks (aka. deep learning recently) show excellent accuracy and promising real-time capability. However, several factors, such as sampling rate and precision, neural network quantization bits, and intrinsic noise, complicate the problem and make it hard to find a cost-effective solution with high performance. In this article, we analyze above factors in a systematic way to study the effect of each one on the performance of the network individually when other factors are controlled. Moreover, the proposed network architecture can provide both the time and energy information from a single pulse. When the sampling rate is 2.5 MHz, sampling precision is 5-bit, the network tested in this work with an 8-bit encoder and a 16-bit decoder (designated as N2) achieved the best comprehensive performance in all conditions.

Keywords: Data reduction methods; Deep learning; Network quantize; Pulse shaping.