Multivariate statistical process control and property inference applied to low density polyethylene reactors

Low density polyethylene (LDPE) is produced in high pressure tubular reactors. All the fundamental polymer properties are extremely difficult to measure and are usually unavailable. However, many on-line measurements such as the temperature profile down the reactor and the solvent flowrate are available on a frequent basis. This paper investigates, via a simulation study, the use of the multivariate statistical method, PLS (Projection to Latent Structures or Partial Least Squares), to develop inferential prediction models for these properties using the available process data. The dimensionality reduction aspects of PLS are also exploited to develop a multivariate statistical control plot for monitoring the operating performance of the reactors. The ability of the plot to detect some common reactor operating problems is investigated via the simulation. (A)