Fault detection in multimode nonlinear systems

Abstract

Modern complex industrial plants are commonly working on multiple operating conditions because of different product specifications and some external restrictions. As a consequence, the plant characteristics change from one operating condition to another due to nonlinearity and set-point changes in the system. Hence, the statistical model obtained from traditional MSPM techniques for an operating mode is not valid anymore for the others and will induce false alarms. This is because the basic assumption in MSPM methods that the data should follow uni-modal Gaussian distribution. Under the assumption that the data corresponding to each operating point follows a multivariate Gaussian distribution with various statistical properties, the available historical data can be seen as a mixture of Gaussian components with different mean vectors and covariance matrices.