A unified framework for statistical change detection

The authors address the problem of detecting abrupt changes in the statistical properties of signals and dynamical systems, which is of key interest for failure detection and monitoring. They present a unified parametric statistical framework for both the design and the performance analysis of change detection algorithms. This framework is introduced in the simplest case of change in the scalar parameter of the distribution of an independent random sequence. Situations of increasing complexity are investigated, namely additive changes in linear regression and state-space models and spectral changes in both scalar and multidimensional cases. The link with the analytical redundancy and detection filter approaches is also stressed. A general tool for performance evaluation is introduced: the average run length function, which can be used for estimating both the mean time between false alarms and the mean delay for detection, and for tuning the algorithms