Choose Language Hide Translation Bar

New Insights into Process Deviations Using Multivariate Control Charts

Stephen Pearson, PhD, Chemical Process Statistician, Syngenta


In this presentation we will capture multivariate batch data in the form of letters of the alphabet, using a LEGO® MINDSTORMS® kit. With a known training letter, unknown letters can be identified based on multivariate properties. The manufacture of chemical active ingredients is a multivariate batch process. It can take experienced scientists years to understand how the various inputs to the process interact. Persistent problems are often multivariate in nature (such as an interaction between temperature, pressure and an impurity), which can make them difficult to solve. While the problem remains, significant losses in productivity can occur. By utilising domain experts in conjunction with the multivariate control charts in JMP, it is often possible to troubleshoot the process deviation. Unfortunately, the output is encoded in eigenvalues and eigenvectors, which can be non-trivial to understand. An application has been built in JSL to reformat the output of this platform into simple graphs with descriptive text of the principal components and interactive filters. The key steps in preparing, analysing and visualising the data will be demonstrated.

A paper detailing the process including the code for the LEGO® robot are available to download.