Multivariate monitoring of startups, restarts and grade transitions using projection methods

Process transition policies are traditionally developed using detailed first principles models in conjunction with optimization algorithms. Mechanistic models are, however, often not readily available or difficult and time consuming to develop. In these situations, a data base approach to improve transition policies and product quality would be very helpful. Data buses collected during past transitions can be used for this purpose. The successful use of multivariate projection methods, namely multiway principal component analysis (PCA) and multiway projection to latent structures (PLS), to improve both process transition performance and product quality using historical records of transition data is demonstrated in this work. The nature of the transition data in the general case, where all variables are not present for the entire duration of the transition is examined. Multivariate SPC approaches are proposed to assess the successful completion of a transition ("production readiness for the new grade"). Finally, analysis tools are suggested for diagnosing the reasons for past transition problems and for monitoring new transitions to ensure repeatable high quality transitions. These latter methods are aimed at reducing the amount of off-specification materials and reducing transition time.