Stage-Based Process Analysis and Quality Prediction for Batch Processes

Abstract

A process analysis and quality prediction scheme is proposed based on a stage-based PLS modeling for batch processes. Without any requirement of prior process knowledge, the scheme first divides a batch process into stages of different process characteristics. Subsequently, a strategy is developed to identify stages that have critical influences on concerned qualities, defined as critical-to-quality stages. Within these critical-to-quality stages, an algorithm is then further developed to identify the variables that have significant contributions to the quality variations. Finally, based on the identified nature of quality and stage relationships, a set of algorithms is developed for online quality prediction. The applications of the proposed scheme to injection molding show that the proposed analysis and quality prediction are not only effective but are also able to enhance process understanding and identify specific variables and periods for quality improvement.

Covariance‐oriented qualitative and quantitative calibration analysis for multistage batch processes

“…Camacho and Picó (2006a,b) developed a MPPCA algorithm for the detection of the stage segments in a batch process so that each segment can be approximated by a linear model. A variant of k-means clustering algorithm (Lu et al, 2004;Lu and Gao, 2005) has also been developed. Different from the other stage identification algorithms, the clustering algorithm is developed assuming that the alternation between different stages can be revealed by checking the changes of underlying process characteristics.…”

“…The detailed clustering procedure can be referred to previous work (Lu et al, 2004;Lu and Gao, 2005). Moreover, the time-slice covariance matrices calculated during the clustering procedure will be readily employed in the following stage-based statistical analysis to reveal the correlations between process measurements and quality.…”

“…In multistage batch processes, it is common that different stages have different effects on the concerned quality, which are thus classified into the critical and non-critical stages (Lu and Gao, 2005). Critical stages have close relationships with quality so that they can provide accurate and stable regression models and thus give more reliable quality prediction results.…”

“…The concept of critical-to-quality stage is introduced and quality prediction and control can thus be carried out only in those critical ones. Their method (Lu and Gao, 2005) first derives the time-slice sub-PLS models and then calculates their average within the same stage, which will serve as the representative stage-specific regression model. However, it overlooks the time-cumulative effects within the same stage and may not provide stable and reliable enough regression relationships.…”

Mixture of partial least squares experts and application in prediction settings with multiple operating modes

“…In these situations, it would be beneficial for the prediction accuracy and reasonabilily, to consistently train a model for each operating mode of the process [12], or train a model for each set of correlated operating modes [13]; And during online operation, when a new sample is made available, the model which is the most adequate for this new sample is identified and then used to make the prediction. The identification of which model will be used is a key issue in the development [13,14,15], which can be done using expert knowledge [13] or using automatic tools, as finite mixture of Gaussian models (FMGM) [12].…”