Recovery of saturated signal waveform acquired from high-energy particles with artificial neural networks
Yu Liu,Jing-Jun Zhu,Neil Roberts,Ke-Ming Chen,Yu-Lu Yan,Shuang-Rong Mo,Peng Gu,Hao-Yang Xing
DOI: https://doi.org/10.1007/s41365-019-0677-0
2019-09-30
Nuclear Science and Techniques
Abstract:Artificial neural networks (ANNs) are a core component of artificial intelligence and are frequently used in machine learning. In this report, we investigate the use of ANNs to recover the saturated signals acquired in high-energy particle and nuclear physics experiments. The inherent properties of the detector and hardware imply that particles with relatively high energies probably often generate saturated signals. Usually, these saturated signals are discarded during data processing, and therefore, some useful information is lost. Thus, it is worth restoring the saturated signals to their normal form. The mapping from a saturated signal waveform to a normal signal waveform constitutes a regression problem. Given that the scintillator and collection usually do not form a linear system, typical regression methods such as multi-parameter fitting are not immediately applicable. One important advantage of ANNs is their capability to process nonlinear regression problems. To recover the saturated signal, three typical ANNs were tested including backpropagation (BP), simple recurrent (Elman), and generalized radial basis function (GRBF) neural networks (NNs). They represent a basic network structure, a network structure with feedback, and a network structure with a kernel function, respectively. The saturated waveforms were produced mainly by the environmental gamma in a liquid scintillation detector for the China Dark Matter Detection Experiment (CDEX). The training and test data sets consisted of 6000 and 3000 recordings of background radiation, respectively, in which saturation was simulated by truncating each waveform at 40% of the maximum signal. The results show that the GBRF-NN performed best as measured using a Chi-squared test to compare the original and reconstructed signals in the region in which saturation was simulated. A comparison of the original and reconstructed signals in this region shows that the GBRF neural network produced the best performance. This ANN demonstrates a powerful efficacy in terms of solving the saturation recovery problem. The proposed method outlines new ideas and possibilities for the recovery of saturated signals in high-energy particle and nuclear physics experiments. This study also illustrates an innovative application of machine learning in the analysis of experimental data in particle physics.
physics, nuclear,nuclear science & technology