The chemical industry often involves batch processes in which online monitoring plays an important role for batch quality assurance. Often times, process conditions are electronically measured very frequently (i.e. every minute) over the entire time-span of the processing steps for each produced batch. After multiple batches have been produced, a rather large data set of the processing conditions is available to help correlate the changing of process conditions with resulting batch quality. In addition, the trajectories of the process variable levels between good and defect batches can be compared to examine if abnormal behavior exists in defect batches.
With a real Dow polymerization process example, this presentation discusses how the rather large process data set (many processing measurements per batch) is combined with the single batch quality assessment to enable meaningful analyses. In addition, several advanced statistical techniques used for root cause investigation on defect batches including partial least squares, principal component analysis, discriminant analysis, partition, logistic regression, multivariate SPC charts and batch-level modeling will be described. By using multiple statistical techniques, we were successfully able to eliminate a variety of process variables as potential root causes for why defect batches kept occurring. This conclusion enables the business to focus on the identification of alternative potential root causes and the resulting data required for assessment for ultimately solving the problem.