Low-resolution gamma-ray spectrum analysis using comprehensive training set and deep ResNet architecture

Abstract

Considering limitations of training sets and architectures in existing methods for gamma-ray spectrum analysis based on artificial neural networks, we propose a method for constructing a comprehensive training set and a modified ResNet architecture to perform low-resolution gamma-ray spectrum analysis. The constructed training set reproduces the shape feature diversity of spectra acquired from real detection scenarios by using various parameter settings in Monte Carlo simulations. The proposed ResNet constitutes the largest and deepest architecture ever applied to gamma-ray spectrum analysis, containing 51 layers and more than 107 parameters. Results from tests on simulated and measured spectra show that the key performance values of average precision and F1 score of the proposed network are substantially better than those of a convolutional neural network and a fully connected network. Moreover, our approach provides weak ray identification, fake peak discrimination, and overlapping peak resolution. This study demonstrates the feasibility of applying deep learning to gamma-ray spectrum analysis and introduces an approach to achieve general, accurate, sensitive, and reliable gamma-ray spectrum analysis.