Process monitoring using non-linear statistical techniques

Abstract

A large number of process variables are usually measured and stored in computer data base during process operation. These variables are usually highly correlated and the real dimensionality of the monitored process is considerably less than that represented by the number of process variables collected. Successful process performance monitoring requires the efficient handling of large amounts of monitored plant data. Principal component analysis reduces the dimensionality of the process by creating a new set of variables, principal components, which attempt to reflect the true underlying system dimension. Process performance can then be monitored in a low dimensional principal component space. Linear process performance monitoring is based upon plots of scores and squared prediction errors from a principal component model. However, for highly non-linear processes, this form of monitoring may not be efficient since the process dimensionality cannot be represented by a small number of linear principal components. Non-linearly correlated process variables can be reduced to a set of non-linear principal components, through the application of non-linear principal component analysis. Efficient process monitoring can then be performed in a low dimensional non-linear principal component space. In parallel with the conventional multivariate plots, the use of accumulated scores provides a significant breakthrough in the separation of different operating conditions/faults, leading to robust early warning of potential plant malfunctions. An application to the condition monitoring of a polymerisation reactor demonstrates the effectiveness of the non-linear monitoring approach.