A multivariate method for detecting and characterizing the changes in responses of sensors when extreme outliers arise

In the machine learning area, when rare events occur, the problem is treated simply as an outlier or anomaly detection; however, all strategies are applied when the events have already occurred. In this work, rare events correspond to events with a very low probability of occurrence, significant impact after they occur, difficult detection, and little predictability. The detection and alarm of these events are only possible if previous and predictable changes have been observed in the measurable variables, still with low probability but may herald the appearance of a rare event. In this work, these previous changes are called extreme outliers. We propose a multivariate method for detecting and describing the changes in the sensor system due to extreme events that may precede a rare event. We use long-tail probability density distributions to observe the low-probability events and to define the magnitudes of the sensor signal that can determine the presence of extreme outliers. After detection, the method uses unsupervised and supervised learning to describe the changes in the monitored system. The results show the method’s effectiveness in detecting and describing changes in time series for different scenarios and criticality levels that precede extreme outliers or rare events. The approach allows the development of alerts according to the level of criticality the sensor system achieves. The proposed method can help in decision-making to reduce the impact of this class of events in critical industrial environments. •We detect extreme outliers in systems monitored by time series.•We use the Theory of Extreme Values to predict the occurrence of extreme outliers.•We describe the changes that occur in the system after extreme outliers occur.•We apply supervised and unsupervised learning to interpret changes that occurred.•Proposal of a workflow for detecting and describing anomalies that occurred.