Fault Diagnosis With Process Uncertainty

Abstract

A nonparametric pattern classification method is introduced for fault diagnosis of complex systems. This method represents the fault signatures by the columns of a multi-valued influence matrix (MVIM), and uses adaptation to cope with fault signature variability. In this method, the measurements are monitored on-line and flagged upon the detection of an abnormality. Fault diagnosis is performed by matching this vector of flagged measurements against the columns of the influence matrix. The MVIM method has the capability to assess the diagnosability of the system, and use that as the basis for sensor selection and optimization. It also uses diagnostic error feedback for adaptation, which enables it to estimate its diagnostic model based upon a small number of measurement-fault data.