(348d) Data-Driven Strategy For Real-Time Quality Control And Recovery In Batch Operations

Complex batch processes often suffer a lack of reproducibility from batch to batch due to disturbances such as raw material impurity, variations in start-up initialization and operating conditions. These disturbances and variations have adverse effects on the final product quality but are inherent to the processing and may be difficult for operators to discern. A variety of online process monitoring techniques have been studied in literature for batch processes and have been demonstrated to reliably identify abnormalities, especially when the base case process is well under control.

In this work, we are interested in processes where there is significant variability even in normal operations, so the separating line between normal and abnormal batches in terms of product quality is ambiguous. In such cases where quality variations even among the normal batches is significant; abnormal batches can be rectified and restored if suitable remedial control actions are taken in a timely fashion. The relationship between variations in process variable trajectories as well as the influences of the batch evolution on the final product quality has to be adequately captured and suitably represented. This motivates the development of a model between process variables and final quality variables for with-in batch quality control. A data-driven modeling and real-time quality control strategy is demonstrated in this work. A latent variable model based on partial least squares regression (PLS) is applied to data obtained from historical runs. However, historical data is often not rich in out-of-spec conditions when multiple factors interact. Knowledge of this space is essential for process control targeted at recovery of abnormal conditions. We have therefore developed a suitable design of experiment strategy. A small set of experiments is devised to provide informative process data when several disturbance factors influence the final quality in an interactive or synergistic way. A hierarchical modeling structure that accounts for the fact that data imputation for on-line quality prediction may potentially introduce additional uncertainty in prediction and control.

On-line final quality predictions are made using the proposed model at specific time points and necessary corrections to the manipulated variables calculated based on a model predictive control structure. Plant-model mismatch is considered by using a specially designed objective function. The proposed modeling and within-batch quality control strategy is illustrated in this work using a literature case study.

Keywords: Quality control, PLS, Optimization, Model predictive control, Design of experiment, Batch operations.