Modeling Dataset Selection in Multivariate Statistical Process Control

Using methods from multivariate statistical analysis, statistical process control has found wide applications in different industrial processes. Owing to the data-based nature of statistical process monitoring, it is relatively easy to apply to real processes with nonlinear and time varying behavior, in comparison with other methods based on systems theory or rigorous process methods. In a typical industrial application of statistical process control, the purpose is diagnose an evolving batch as normal or not, and to assist the trouble shooting during batch process. For the application of batch process in biochemical process, especially process with limited number of total runs, knowing of minimum number of historical data is really important. In this paper, a simple approach to calculating the minimum number modeling dataset for a validate model is developed. In addition, the proposed approach effectively estimates the extent of capturing meaningful variations and analyzes the stability of the partial least square (PLS) model. The method is illustrated by several industrial applications.