Identification and quantification of anomalies in environmental gamma dose rate time series using artificial intelligence

Abstract

Gamma dose rate (GDR) monitors are the most widely used tool for continuous monitoring of environmental radioactivity. They are inexpensive to procure and operate, and generally require little maintenance. However, since no spectral information is available, the detection limit for irregularities is correspondingly high; A value around 20 nSv/h is often called out. By adding weather data to the GDR measurement and a sequence of machine learning algorithms, the anomaly detection sensitivity can be significantly increased while simultaneously decreasing the number of false positives. The algorithms were designed such that an integrated safety net prevents false negatives. First, the precipitation-induced GDR peaks from washed-out Radon progeny are removed by means of regression, provided that a check of the regression parameters shows sufficient agreement with past data at the measurement site. A neural network then calculates the expected value of the remaining GDR baseline for the prevailing conditions. Finally, an anomaly detection is carried out on the remainder between the expected and actual GDR baseline value. Extreme value theory is used to detect point anomalies, and hierarchical clustering of subsequences for slower processes. By combining the two detection methods, the full spectrum of irregularities is covered. The algorithms were implemented in Python and trained with real measurement data from the German GDR monitoring network. For verification, the data were enriched with results from JRODOS simulations of a nuclear power plant accident. Altogether, the presented methodology can lower the detection limit of irregularities to about 4 nSv/h, i. e. about a factor of 5 below the previous consensus value. The algorithm detects as well as quantifies the anomaly in the GDR, allowing for additional conclusions like potentially involved isotopes. Most important, it allows to refrain from the current practice of defining fixed alarming thresholds between the two contradicting goals of high sensitivity and low false alarm rate. Instead, it allows to transition to the more natural alarming on deviations from the expectation.

Keywords: Anomaly detection; Artificial intelligence; Environmental gamma dose rate; Machine learning.