Advanced Spectroscopy Time-Domain Signal Simulator for the Development of Machine and Deep Learning Algorithms

Abstract

Machinelearning methods, particularly deep learning (DL), have become essential for advanced signal processing. These methods often depend on annotated datasets, which can be limited or even unavailable in many cases. One area significantly affected is nuclear spectroscopy, where the lack of annotated datasets is due to the challenges of manually labeling signals recorded in the time domain. To address this issue, it is necessary to use simulators to generate annotated signals, ensuring that the generated time signals are as realistic as possible. This letter introduces a novel simulator designed to generate time-domain signals for gamma spectroscopy. Unlike traditional energy-spectrum simulators, our approach simulates raw sensor output for training advanced DL models. The simulator is analytically trackable, highly customizable, and lightweight, enabling researchers to tackle challenges, such as pile-up events and noise suppression. Case studies demonstrate its practical application in high-activity measurement scenarios.