Process monitoring and fault detection based on multivariate statistical projection analysis

Multivariate statistical process control (MSPC) has been applied to performance monitoring for chemical process. However, conventional methods of MSPC are based on the premise that the extracted latent variables must be subjected to normal distribution, which often can´t be satisfied. In this paper, a new method based on independent component analysis (ICA) and principal component analysis (PCA) is presented for process performance monitoring, in which a two-step procedure is employed. At first step, process operating information with non-normal distribution is extracted by means of ICA, and the density function of this part of information is estimated by means of parzen density estimator for calculating the confidence limits. At the second step, the information with normal distribution is extracted from the underlying residual data sets by PCA, and the confidence limits are determined on Q and hotelling T² statistics. With the primary advantage of which no assumption of normal distribution on process data sets is needed, the proposed method is applied to a double-effect evaporator, and the simulation results verify it effective.