Anomaly detection in xenon time series data for nuclear explosion monitoring

In the context of nuclear explosion monitoring, the Preparatory Commission for the Comprehensive-Nuclear-Test-Ban Treaty Organization (CTBTO) is building the International Monitoring System (IMS) as part of the verification regime. One important component of the IMS are 40 noble gas stations that are globally distributed to periodically sample the atmosphere and measure activity concentrations of four radioxenon isotopes commonly used as tracers for nuclear explosions. Currently, CTBTO applies a median-based abnormal threshold in order to detect anomalous occurrences in the collected xenon time series data. However, many civilian xenon sources like medical isotope production facilities (MIPFs) and nuclear power plants (NPPs) produce xenon signals that frequently exceed the abnormal threshold and are capable of masking the signals of real nuclear explosions. Additionally, the global xenon background is not static and changes with time and location. Although proven effective at detecting major events, the current median-based technique produces a high false positive rate. Modern machine learning-based anomaly detection (ML based AD) techniques can adaptively learn the normal behavior of a particular noble gas station while taking into account local influences such as nearby MIPFs or meteorological conditions. This Master’s thesis explores the application of modern ML-based AD techniques on xenon time series data collected by SAUNA-type noble gas stations of the IMS. The hypothesis for this dissertation states that ML-based AD techniques have improved detection capabilities compared to the current threshold based approach due to their ability to learn different models depending on the specific noble gas station. Four modern anomaly detection algorithms are implemented and tested, namely an Autoencoder, Gaussian Mixture, Isolation Forest and One-class SVM. Furthermore, two main model types are tested. First, the cross-isotope approach enables algorithms to learn from all four isotopes simultaneously. These isotopes are 135Xe, 133Xe, 133mXe and 131mXe. This allows models to find relevant relationships between isotopes and use this knowledge to their advantage. Second, the isotope-specific approach aims at modelling the normal behavior of a singular isotope without considering concurrent isotopic behavior. This is in line with the current technique. Both models are either trained on point-based or sequence-based input data to determine whether the temporal nature of the data is exploitable. Artificially augmented data sets are used to validate model performance. Four contributions are made. First, we show that ML-based AD techniques significantly outperform CTBTO’s current median-based technique. Second, ML algorithms are shown to generally take advantage of learning from all isotopes simultaneously and thereby outperform the isotope-specific models. Third, models trained with point-based input data tend to outperform their sequence-based counterparts suggesting temporal dependencies are difficult to exploit and that the ability to consider concurrent isotopic behavior is more relevant than the values of temporally neighboring observations. Finally, we propose the overall best performing ML-based AD technique as a better alternative to the current median-based technique.