Aberrant measurements: Detection, localization, suppression, acceptance and robustness

The detection of outliers in a series of measurements, but even more so their location, is a necessity when these measurements are to be used in a monitoring system. This detection/localization can only be done if redundant information is available, which may be based on the model of the system on which the measurements were collected. In some cases, however, it is not necessary to detect and locate outliers. Instead, a robust approach to their use may be preferred, one that minimizes the influence of these outliers, such as using a median rather than a mean. In this paper, the focus will be on the notion of robustness through a few examples and notably by proposing extensions to two well-known data processing techniques (data reconciliation and principal component analysis). The numerical examples proposed clearly show how to implement these two techniques and how to use them in a system monitoring procedure.

Automated summary

Detecting and locating outliers in measurements used for monitoring systems is crucial. Redundant information is needed for this. Sometimes, a robust approach that minimizes the impact of outliers is preferred.