Fault Diagnosis in Dynamic Systems Using a Robust Output Zeroing Design Method

Abstract

A new method of analysing faults in the m measurements of an n^th order dynamic system is presented. The estimation error space of an observer is used for fault diagnosis. The design uses a robust observer in which the weighted linear combination of m estimation errors is equivalent to the output zeroing problem. In order to design a fault monitor which is insensitive to r (multivariable) input and parameter-induced disturbances, n-1 eigenvectors are assigned to the null space N(WC_O) where W is a weighting matrix and C_O is the nominal (no-fault) measurement matrix. Motions away from the null space regime are due to parameter variations, disturbances or sensor faults. The problem is to distinguish the variations due to sensor faults from all other uncertain signals. By assigning the eigenstructure of a dual problem and exploiting an orthogonality relation between the eigenstructures of the dual systems a robust design can be achieved. The effects of parameter variations are nulled and the fault monitor is rendered sensitive to sensor faults defined in a given spectral bandwidth.