Detection of Sensor Faults by Means of Multivariable Calculation Methods

Abstract

This paper deals with the detection and identification of abrupt changes (such as failures) in a multidimensional stochastic dynamic system which is representable by a vector autoregressive (AR) process, without any a priori knowledge about the process parameters. Two approaches are presented: In the first model the multidimensional system is considered an arrangement of multiple linearly coupled stochastic scalar autoregressive processes, and the residual is determined by inverse filtering. Alternatively, but using the AR approach, the system is modelled by a single stochastic process, and the analytical redundancy comprised in the multiple sensor signals can further contribute to residual generation. In both models the decision on an abrupt change in signal statistics is based on the generated residual signals and performed via the generalized likelihood ratio (GLR) approach, discovered by Willsky and Jones (1974). Finally some variations of the basic approach yielding a reduced number of numerical operations are presented, assuming either stationary state and/or using a bivariate calculation method.