Smart Data Analytic-Based Transform of Low-Count Gamma-Ray Spectra for Enhancing Isotope Detection by Using a Self-Learning Window of Relevance Vector Regression

Abstract

Identification of isotope signatures in low-count gamma-ray spectra measured with a portable NaI detector is of great importance with concern to nuclear security and non-proliferation. Specifically, in hidden radioactive source search scenarios, where a mobile detector acquires consecutive measurements in a very small interval of time (for instance every second), the measured spectra may exhibit a low number of counts and high fluctuation. These two factors challenge the radioisotope identifier algorithm both in detection accuracy and processing time by providing a high number of spectral peaks that need to be checked against known signatures. In this article, a new smart data analytics-driven transform for enhancing isotope detection is presented. In particular, a machine learning (ML) tool, i.e., the relevance vector regression (RVR), is utilized to transform the obtained signal into a form where the isotopic peaks are resolved with higher efficiency while several spectral peaks resulting from fluctuation are removed. Initially, the transform uses the data within a step-size window as its training set—self-learning (i.e., no external datasets are used) window of RVR—and subsequently replaces the window values with the RVR outputs. The presented data analytics-based transform is tested on a set of real-world gamma-ray spectra measured with a low-resolution NaI detector. Results exhibit that the presented transform improves detection certainty as obtained by two radioisotope identifier algorithms, and concurrently reduces the number of random peaks by providing a smoothed signal with fewer spectral peaks to be checked.