Multivariate statistical monitoring of continuous processes.

Abstract

Meeting product specifications and process safety have been major concerns in the chemical industry. Increasing number of process variables has made monitoring and analysis of any process deviations difficult. Hence, principal component analysis (PCA), a method to reduce dimensionality, can be used to simplify the process analysis, monitoring and optimization of chemical processes. Fault detection indices, such as squared prediction error (SPE), also known as Q statistic, and Hotelling's T^2 statistics, are used to detect and diagnose process faults. The results from these statistics consider the correlation between the variables with high contribution and the process faults. Hence, they give plant operators a better grasp of the process changes and steps can be taken to rectify the faults. However, the limitation is that it does not imply causality of the process variables to the fault. A state-of-the-art method, wavelet analysis combined with PCA, is also discussed. Wavelet transform is used to de-noise the process signals before carrying out dimensional reduction using PCA. This process helps to reduce the time taken to identify the fault so that any rectification actions can be taken promptly. These two methods are applied to the Tennessee Eastman problem and the results are rather promising. Fault detection and diagnosis are successful using both techniques. Wavelet analysis combined with PCA is found to perform slightly better for fault detection compared to conventional PCA alone but both techniques exhibit similar performance in fault diagnosis.