Chemical process monitoring and fault diagnosis based on independent component analysis

Multivariate statistical process control (MSPC) has been successfully applied to performance monitoring and fault diagnosis for chemical processes. However, classical methods of MSPC are based on the premise that the separated latent variable must be subjected to normal distribution, which sometimes can´t be satisfied. In this paper, a new method based on independent component analysis OCA) whose goal is to find a line representation of nomgaussian data to depict the chemical process and improve the monitoring performance of the system is presented. Due to the uncertainty of the probability distribution of the independent component, the paper devises a kind of classifier with Parzen density estimation for classifjring the normal data and fault data. Then the nonisothemal CSTR is monitored and diagnosed by the present method, the simulation result verifies the effectiveness of ICA-based monitoring method.