Abstract

Statistical change detection plays a key role in prognosis and diagnosis of faults. Design of change detection algorithms is well known in theory, but practice often violates the idealised perquisites that theory requires. This talk provides tutorial insight in existing methods for change detection in the presence of Gaussian or Non-Gaussian noise with focus on reliable diagnosis / prognosis where accurate knowledge of detection and false alarm probabilities is required. As these features are determined by properties of the test statistic of the particular problem, the talk discusses the consequences of correlation in real life, and compares theoretical results with those obtained in actual applications. A methodology for change detection is suggested that adapts to real-life conditions. Using estimation of distribution parameters for the actual test statistic, threshold for hypothesis testing and test sequence length are shown to be parameters in an optimization problem that can guarantee prescribed properties. Detection with multiple detectors, based on different indicators of change, is then discussed, and it is shown how a Copula description of joint probability can be employed to assess properties of combined detectors. Selected industrial applications from unmanned aircraft and drilling illustrate the methods.