Deep Ensemble Novelty Detection—Novelty Detection and Fault Identification in Multivariate Data

Safety-critical systems require rapid detection of abnormal behavior to prevent damage and reduce downtime. Monitoring only individual data inputs above a threshold may be insufficient, as this approach may be too slow to respond effectively to rapidly evolving situations and fails to consider the complex interactions between various system components. Thus, advanced fault detection methods are required. Novelty detection methods address this problem by identifying previously unknown states based on knowledge of known states. We present a novel approach called Deep Ensemble Novelty Detection (DEND) for accurate and stable novelty detection performance across multivariate data such as stationary time series, non-stationary time series, tabular data, and vectorized images. In addition, we demonstrate how to identify the fault causes of the detected unknown states by the utilization of a decision tree (DT) to learn the hierarchical fault causes and the use of a multilayer perceptron (MLP) to explain the fault causes.