(345d) Best Practices on Creating Soft Sensor Models for Batch Processes Utilizing Multi-Way Partial Least Squares (MPLS)

Soft sensors predict product quality at the end of the batch while the batch is still evolving. As a result, accurate soft sensors would allow operators to respond quicker to process deviations that impact product quality and improve process capability and yield.  Data-driven methods, such as multi-way PLS (MPLS), are commonly used in deriving these batch soft sensors.  Normally, a batch data set is comprised of inlet raw material conditions (sample once per batch), in-batch process measurements (sampled in real-time during the batch), and end of batch quality measurements (quality measurements such as conversion, impurity concentration or product purity).  The dimensionality of a typical batch data set is very high and requires pruning prior to modeling and analysis. In addition, variable selection not only improves model performance, it also helps with increasing process interpretability of the resulting models. Nevertheless, due to the dimensionality of batch data set, identifying the most relevant variables (for estimating end of batch quality variables) can be a challenging task. Relevant variables are often observed to change depending upon the way the batch data is structured or organized. The preprocessing steps (critical for variable selection) can be summarized as follows: (1) a proper division of the phases/stages, which are typically defined in the batch recipe and followed by plant operations; (2) restructuring the data in multi-block or single-block arrangements; (3) the trajectory alignment to ensure consistent batch trajectory and identical batch length; and (4) the batch data unfolding, which converts the three-way process data X(I×J×K) into a two-way matrix, where J indicates the number of process variables at K sampling times in I batches.  The combination of these four steps will have a profound impact on the interpretation and identification of the variables relevant for prediction or monitoring.  Specifically, from practical modeling experiences, relying on operating phases defined in the batch recipe do not necessarily result in the most optimal phase division for modeling. This paper will demonstrate that proper phase division and data arrangement are important preprocessing steps in developing a soft sensor model for batch processes, generating a consistent variable selection and acceptable end of batch quality prediction.  An industrial case study will be utilized to validate these best practices for generating a batch soft sensor model.