(91a) Probabilistic Machine Learning Based Soft-Sensors for Product Quality Prediction in Batch Processes

Statistical machine learning algorithms have been used to analyse industrial data for batch process monitoring and control. The use of statistical learning algorithms has been preferred because of the difficulties in identifying first principles predictive models for these processes. For example, in batch production of formulated products, the goal is typically to keep the batch on specification. However, measurement of key product qualities is often conducted offline, due to the expense of advanced process analytical technology. This generally means one measurement of key product quality is produced from a given batch. This provides difficulty for modelling end-product qualities, as one seeks to find a function mapping from temporal measurements of batch process variables (such as pressures, temperatures and flowrates) to one measurement of end-quality. This data is highly dimensional, multicollinear and nonstationary, which provides challenge to all model construction frameworks.

Recent research has considered the development of soft sensors, based on latent variable models, which automatically quantify the uncertainty associated with their prediction. In this study, we develop a three-step methodology to identify, visualize and systematically reduce data dimensionality for the construction of robust soft-sensors for end-product quality prediction. The approach first screens the entire dataset to identify critical time regions and operational variables, which correlate strongly with the end product quality one wishes to model. We then adopt multiway latent variable modelling to construct a latent space descriptive of the existing batches. Nonlinear estimators are then constructed from the reduced latent space to estimate final product quality, which are able to express model uncertainty automatically. Specifically, in this study, we explore the performance of Gaussian processes (GP), Bayesian neural networks (BNN) and heteroscedastic noise neural networks (HNN).

To highlight the efficiency and practical benefits of the approach, an industrial consumer goods product manufacturing process is presented as an example. The soft sensors were constructed and their performance was assessed via cross validation, using metrics which not only quantify the mean error in prediction, but also the quality of the uncertainty prediction. This assessment was informed with knowledge of the standard error associated with product end-quality measurement. The accuracy, reliability, and interpretability of the soft-sensors is discussed, tested and shown to generalise well on two test datasets. Specifically, the GP and HNN both provide good performing models, with accurate predictions in the mean (10% mean absolute percentage error, MAPE, in cross validation) and uncertainty estimates which reflect the standard error of measurement. Whereas the BNN, has low MAPE, but provides overly confident predictions, relative to the standard error of measurement. This result is also observed in assessment of the soft-sensors on the test set.

Innovations of this approach include ease in data visualisation, ability to identify major operational activities within the factory offline, as well as robustly and interpretably predict product end-quality.