Stochastic change detection in uncertain nonlinear systems using reduced-order models:classification

Abstract

A reliable structural health monitoring methodology (SHM) is proposed to detect relativelysmall changes in uncertain nonlinear systems. A total of 4000 physical tests were performedusing a complex nonlinear magneto-rheological (MR) damper. With the effective (or‘genuine’) changes and uncertainties in the system characteristics of the semi-active MRdamper, which were precisely controlled with known means and standard deviation of theinput current, the tested MR damper was identified with the restoring force method(RFM), a non-parametric system identification method involving two-dimensionalorthogonal polynomials. Using the identified RFM coefficients, both supervised andunsupervised pattern recognition techniques (including support vector classification andk-means clustering) were employed to detect system changes in the MR damper. Theclassification results showed that the identified coefficients with orthogonal basis functioncan be used as reliable indicators for detecting (small) changes, interpreting the physicalmeaning of the detected changes without a priori knowledge of the monitoredsystem and quantifying the uncertainty bounds of the detected changes. Theclassification errors were analyzed using the standard detection theory to evaluate theperformance of the developed SHM methodology. An optimal classifier designprocedure was also proposed and evaluated to minimize type II (or ‘missed’) errors.