Abstract:

An important objective in monitoring mechanical and other systems is the detection of changes in the underlying dynamics governing system behavior. Statistical change detection (SCD), or change-point detection, can play a crucial role in the early detection of small changes in mechanical systems. The primary challenge is to detect changes manifest in very small deviations in the statistical behavior of the observations. Statistical change detection algorithms act as novelty detectors, obviating the need to collect examples of all possible fault signature patterns for the purpose of training pattern classifiers. With respect to SCD, the use of nonlinear models can provide useful benefits. In particular, detector performance may be enhanced through the use of more appropriate model structures, such as when processing nonlinear and/or non-Gaussian data. Second-order statistics are, in general, adequate when data being fitted are Gaussian. When data are non-Gaussian, or represent the output of a nonlinear process, higher-order statistics (HOS) (e.g., bispectrum, trispectrum, etc.) may be more useful in characterizing a process. An attractive aspect of the SCD approach is its ready extensibility to include higher-order statistical moments, where necessary, to exploit additional information in the data. Since HOS are polynomial functions of the observed data, polynomial neural networks are well suited for the implementation of signal detectors based on HOS. In this paper, the performance benefits arising from the use of nonlinear models in SCD algorithms are considered. [See NOISE-CON Proceedings for full paper.]