Development and implementation of an ensemble of intelligent techniques for the detection of abnormal conditions in safety-critical applications

An ensemble of statistical techniques and artificial intelligence methods is proposed for the detection of abnormal conditions (anomalies) in safety-critical applications. Stacked Autoencoders (SAEs), the Local Outlier Factor (LOF) method and Gaussian Mixture Models (GMMs) are considered in the ensemble. The outputs of these algorithms (i.e., properly defined anomaly scores) are normalized within an original probabilistic framework and combined for the robust and conservative identification of anomalies in datasets: different combination functions are employed, including "averaging", "maximization" and "AKPV", which averages the scores of the top three anomaly detectors. SAEs are also used as efficient tools for dimensionality reduction in the overall framework. The developed approach is tested on two functional datasets made of time series: 1) the ECG5000 dataset; and 2) a set of transients representing the time evolution of several physical parameters (e.g., fluid temperatures) taken from the simulator of a Generation-IV nuclear reactor, i.e., a Molten Salt Fast Reactor (MSFR). The global accuracy and the Area Under The Curve (AUC)-Receiver Operating Characteristics (ROC) scores are used for assessing the performance of the individual methods and of the ensemble. The results show that the use of an ensemble improves the performance in the detection of anomalies with respect to the individual detectors.