Parametric Fault Detection in Nonlinear Systems-A Recursive Subspace-based Approach

This paper deals with the problem of detecting nolinear systems’ parametric faults modeled as changes in the eigenvalues of a local linear state-space model. The linear state-space model approximations are obtained by recursive subspace system identification techniques, from which the eigenvalues are extracted at each sampling time. Residuals are generated by comparing the eigenvalues against those associated with a local nominal model derived from a neural network predictor describing the nonlinear plant dynamics in free fault conditions. Parametric fault symptoms are generated from the eigenvalues residuals, whenever a given predefined threshold is exceeded. The feasibility and effectiveness of the proposed framework is demonstrated in a practical case study.