Classification of radioxenon spectra with deep learning algorithm

Highlights

• Deep learning for Beta-Gamma coincidence radioxenon spectra classification.

• Screening samples that are not interesting in the CTBT context.

• Without making use of the screening threshold values and background spectra.

• Noble gas classification by CNN technique as prescreening for CTBT relevant samples.

Abstract

In this study, we propose for the first time a model of classification for Beta-Gamma coincidence radioxenon spectra using a deep learning approach through the convolution neural network (CNN) technique. We utilize the entire spectrum of actual data from a noble gas system in Charlottesville (USX75 station) between 2012 and 2019. This study shows that the deep learning categorization can be done as an important pre-screening method without directly involving critical limits and abnormal thresholds. Our results demonstrate that the proposed approach of combining nuclear engineering and deep learning is a promising tool for assisting experts in accelerating and optimizing the review process of clean background and CTBT-relevant samples with high classification average accuracies of 92% and 98%, respectively.

Introduction

International Monitoring Systems (IMS) can detect the amount of radioxenon in the atmosphere by sampling and measuring activity concentrations of ^131mXe, ¹³³Xe, ^133mXe, and ¹³⁵Xe to identify and prevent any nuclear test explosion. Radioxenon can also be detected from civilian sources such as commercial nuclear power plants and medical isotopes at some locations worldwide (Kalinowski et al., 2010). Observed characteristics of radioxenon concentration time series are exceedingly different from one another (Plastino et al., 2010; Schöppner et al., 2013). Therefore, categorization experts need to review and screen samples based on the environmental background of total concentration at each station to distinguish subsequent civilian sources from a nuclear explosion. Before categorization, each sample should meet the minimum criteria to be considered for analysis, the so-called state of health criteria (SOH). Most samples can pass this step without requiring any consideration, although some samples may warrant further investigation (Saey and De Geer, 2005; Saey et al., 2010).

Since significant environmental background sources are unevenly distributed globally, the IMS noble gas network has been spread over regions to obtain global coverage as environmental background level differs at different stations (Stoehlker et al., 2011; Ringbom et al., 2014). Concentrations of approximately 60+ samples are measured every day based on the collection time of three types of radioxenon systems in 12-or 24-h air samples at 40 operational IMS noble gas components, including 25 certified systems, six installed, and nine under construction. Further system development will also increase the number of samples to be analyzed every day. Noble gas systems going through the acceptance process at the time of this study are designed for 6-h and 8-h sampling periods. SAUNA III and Xenon International have 6- of sample time (Haas et al., 2017), while SPALAX NG has 8-h of sampling time (Topin et al., 2020). When these new systems are rolled out to replace existing systems, the daily number of samples will be at least doubled. Thus, analysts may employ a fast pre-screening method to highlight relevant samples prior to their interactive analysis to prioritize work assignments.

Past studies have considered two main approaches to analyze radioxenon concentrations and effectively categorize noble gas data (Ringbom, 2005; Kalinowski et al., 2008; Burnett et al., 2018). The first approach is based on the distinction between radioxenon signals from reactors and nuclear signals. This method considers the absolute concentration of radioxenon and accounts for the activity concentration ratios among four isotopes when identifying sources of samples (Kalinowski et al., 2008, 2010; Haas, 2008; Postelt, 2014). The second approach involves anomaly observations concerning the history of absolute activity concentration spotted at the respective site (Keller, 2012; Ely et al., 2013; Postelt, 2014; Schoeppner and Plastino, 2014).

A categorization scheme is needed to screen out samples below some values to assist the National Data Centers (NDCs) in the day-to-day and manual review of uninteresting samples in the CTBT context. However, there might be the risk of samples that are not categorized as anomalous by some automatic process, while if additional information exists, they might be intertwined with a nuclear explosion (Plastino et al., 2010). Radioxenon is categorized into three levels based on absolute activity concentrations. For each sample, the threshold for the anomalous concentration is updated. It is obtained with a special formula from absolute activity concentrations of all samples of the previous 365 days (Postelt, 2014). Accordingly, level one represents no detection of radioxenon, while level two shows that radioxenon is detected at a normal concentration level (see Fig. 1). The abnormal concentration for each isotope is defined as the median plus the abscissa (default value three) times interquartile ranges (Q75 - Q25) using all spectra taken at a particular station in the previous year. Level three indicates that radioxenon is detected and at least one concentration is above the abnormal threshold (Postelt, 2014).

Fig. 2 shows a noble gas categorization scheme (three levels) for Xe-133 concentrations during one year for the USX75 station. The Lc value (Curie's critical level) indicates at what concentration a signal is accepted as a real presence of the isotope. Every signal below Lc is categorized as Level 1(A) and characterized by a green marker. Measurements above Lc but below the anomalous threshold can be screened out and categorized as Level 2(B) (yellow markers) unless other information warrants further analysis. These measurements are at normal concentrations for that station. In contrast, samples with measured concentrations above the abnormal threshold for the respective station and current sample are categorized as Level 3(C) (red marker) to raise attention for possible further investigation. It should be noted that the abnormal threshold value (dotted red line) is not a constant.

Ringbom et al. (2014) and Saey et al. (2010) have discussed the distribution of activity concentrations of all measurements from two or four radioxenon isotopes at various stations with a long time series. They also analyzed the mean, median, and abnormal concentration values of the data set. Various categorization schemes, including isotopic activity ratios, are proposed for further investigation as additional flags are added to the categorization level in the screening process. Such methods mainly focus on events that possess strong characteristics of quick releases from nuclear explosions and events inconsistent with known environmental background sources (Kalinowski et al., 2008; Postelt, 2014). In some cases, samples may be miscatagorized as ‘normal’ while they come from a nuclear explosion. Therefore, one may treat the categorization scheme with caution as additional help in the sample review process.

The analysis method for categorization currently employed by the International Data Center (IDC), known as the “Net-Count Calculation” (NCC) method, was developed around 15 years ago (Ringbom and Axelsson, 2020). To determine total counts for each ROI, gas background, detector background, quality control spectra, interference corrections, and calibration are necessary for analyzing every individual sample. The net counts discriminate between events attributed to the isotope of interest versus counts attributed to other isotopes or environmental backgrounds. Precise energy calibration of the detector is needed for the accuracy of the ROI analysis (Cooper et al., 2019; Liu et al., 2020). Using computer algorithms can assist experts in handling a massive amount of data by focusing merely on the most significant samples without necessarily calculating activity concentrations, critical limits, and abnormal thresholds for each isotope in each year (Postelt, 2014; Foltz Biegalski and Biegalski, 2001).

Over the past several decades, artificial intelligence (AI) has been growing fast with numerous practical applications and active research topics (Pannu, 2015; Duan et al., 2019). Specifically, Machine Learning (ML) methods analyze data and automatically extract rules for classifying data into predefined categories. These methods have been applied over the past three decades to different fields such as mechanical diagnostics, satellite image analysis, credit card fraud identification, medical diagnosis, flood susceptibility mapping, and so on (Do Koo et al., 2018; Kumar and Iqbal, 2019; Bui et al., 2020). Deep learning can enhance the performance of several problems. Deep learning can often solve issues easier by automatically learning all features in one move instead of manually creating good layers of representations for their results, known as feature engineering. It also helps a model to learn all layers of representation simultaneously rather than in a sequence. However, deep learning methods need a lot of data which can be prohibitive for some problems (Raghavendra et al., 2018; Azarkhalili et al., 2019).

The application of machine learning can facilitate the study of patterns in data derived from IMS stations (i.e., seismic, infrasound, and radionuclide stations) to optimize processing and automatic response to a possible nuclear explosion. Some studies have presented methods using Principal Component Analysis (PCA) and Self-Organized Maps (SOM) on seismic data (Sick et al., 2015). In addition, Sinambela et al. (2019) have proposed a machine learning wavelet-based approach to classify and detect nuclear explosion waveform signals from Indonesia's tele-seismic recording stations. The most advanced implementation of a machine learning algorithm in IDC operations is NET-VISA, a method for building events by network processing of waveform detections (Sereno Jr and Patnaik, 1993; Coyne et al., 2009; Russell et al., 2010; Schneider et al., 2010). Other studies have shown that the application of a DNN standard model, a self-normalizing neural network (SNN), a fully convolutional neural network (FCN), and a long short-term memory (LSTM) network for the classification of infrasonic events (Solomon et al., 2018).

Furthermore, to detect nuclear explosions by monitoring radionuclide and measuring the concentration of radioactive particles and noble gases, Bellinger and Oommen (2012) have employed pattern recognition of One-Class classifiers with data from the industrial emitter. They found that the performance of models depends on the distance between possible explosion regions and the respective monitoring station. Besides, they argued that atmospheric conditions such as fluctuation of wind speed and its direction should be considered when calculating the environmental background level. However, very few studies have used machine learning in categorizing radioxenon. For instance, Stocki et al. (2010) have investigated several machine learning algorithms, including Naıve Bayes, Neural Networks, Decision Trees, k-Nearest Neighbors, and Support Vector Machines using simulated data sets. They showed that nonlinear induction algorithms outperformed a simple linear discriminator in nonlinear domains.

This study presents a machine learning perspective for the first time in categorizing the Beta-Gamma coincidence of radioxenon spectra based on real data. Using raw data of the USX75 station during seven years, we show that analysis can be conducted without background information, interference correction, or mathematical equations for calculating the activity concentration of the sample and other 365 moving day's data. In detail, we apply convolutional neural networks (CNN) as a deep learning algorithm ‌on the entire two-dimensional Beta-Gamma radioxenon spectra in the format of IMS sample histograms. In our proposed approach, we utilize actual raw data without applying interactive re-calibration. In other words, CNN does not rely on specific ROIs and information corresponding to each detection system. As discussed in section 1.2, first, we classify three levels (A, B, and C) of radioxenon spectra labeled as level one, level two, and level three, respectively. We then combine level two and level three to categorize radioxenon according to the presence or absence of radioxenon in each spectrum to find clean background¹ spectra. Finally, we classify level three, which contains atypical concentrations of radioxenon isotopes from level two and level one to predict the existence of xenon with high activity concentration.