Probabilistic Ensemble of Statistical and Machine Learning Techniques for the Detection of Abnormal Conditions in Safety-Critical Systems: Application to the Simulator of an Advanced Nuclear Reactor

The paper proposes a framework for the robust identification of anomalies (hereafter also called outliers) in the behavior of dynamic safety-critical systems, which is based on an ensemble of statistical and machine learning techniques, relying on three diverse principles: (i) proximity-based algorithms, defining outliers as those data points lying in low-density regions or considerably far from their neighbors; (ii) statistical models, capturing the underlying probabilistic structure of the data and identify the outliers as those observations characterized by the smallest likelihood; (iii) (deep) learning-based techniques, trained to reconstruct multivariate data and spot out abnormal samples as those characterized by the largest reconstruction error. Each algorithm of the ensemble is tailored to compute anomaly scores for all the available observations and quantify their "degree of outlyingness". These sets of anomaly scores are finally combined within an original probabilistic framework (based on cumulative distribution functions) to get a unique, robust and reliable ranking. The developed approach is tested on transients representing the time evolution of several physical parameters (fluid temperatures) taken from the simulator of a Generation-IV nuclear reactor. The results show that the ensemble improves the anomaly detection performance with respect to the individual detectors, with accuracy, precision and recall values of 97.24 %, 96.94 % and 98.77 %, respectively.

Keywords: Anomaly detection, Advanced Nuclear Reactor, Machine Learning, Probabilistic Ensemble, Outlyingness scores, Cumulative distribution, Proximity-based algorithms, Deep learning, Statistical models.