Partial least squares confidence interval calculation for industrial end-of-batch quality prediction

The interest in partial least squares (PLS) for prediction of industrial batch quality has increased with the introduction of the US/FDA Process Analytical Technology initiative. A batch process adds complexity to modeling in part because of the evolving product composition and operating conditions. This paper outlines methods for PLS modeling and online prediction calculations for a batch process. The application of the multiple regression confidence interval calculation is examined and deficiencies are identified with respect to PLS batch monitoring. Modifications to the PLS prediction confidence interval calculations are proposed for the application of batch monitoring. Field trial data from a Lubrizol specialty chemical plant is used to support conclusions that the modifications provide a more accurate assessment of the prediction confidence based on the amount of over-determined data available during online processing. The modified confidence interval also adds information and supports an intuitive understanding for the operator interface.