Fault detection in the presence of undermodelling

It is shown that the application of the functional control limit theorem (FCLT) reduces a difficult analysis of the pre-limiting stage to an asymptotically accurate analysis of the properties of Gaussian martingales. The FCLT approach allows mainly a local (small change) detection analysis. The FCLT based design provides schemes which are uniformly most powerful for small changes and long data records. The resulting simple detection schemes allow various forms of a priori information about the change model to be incorporated. The sensitivity of the alarm function is adjusted are in such a manner that certain deviations from the nominal model are emphasized while others are deemphasized. This approach relaxes the requirements towards accuracy of a nominal model and in particular can be used to increase the robustness of detection algorithms with respect to the undermodelling errors