Search

In this article, a digital twin approach is proposed for modeling a pharmaceutical drying process using machine learning techniques, driven by data from different sensors captured in-line. The current difficulty with the drying process is mainly due to the manual operator control for choosing the end-point for terminating the drying step. This results in significant variability, depending on which human operator is supervising, and due to the multi-tasking nature of the job, generally allows a longer processing time to be sure the material is completely dry. The objective is to automate the end-point identification, thus optimizing the processing duration and therefore the overall energy consumption of the process. The point at which the drying is complete is indicated by the temperature difference between the ingoing and outgoing air flow. However, the stochastic nature of the process makes the data modeling a challenge. Firstly, a wide selection of supervised statistical and machine learning algorithms was benchmarked to find the one (CatBoost) which gave the best performance with the data. Next, the set of hyper-parameters was found for CatBoost which gave the optimum performance. This gave a best performance of 0.788 (R2) fitting of the drying end-point prediction with the real values, for a large number of batches (over 700 K records). This is considered a good result taking into account the high residual of data models for these data and the stochastic nature of the process. The approach has been deployed in a real setting digital twin to control the drying process cutoff in the production plant. The results show the viability of the approach for modeling the process and automatically identifying the optimum end-point for the drying process, thus achieving significant energy savings which have been quantified as approximately 3.7 MWh per year for the pharmaceutical company. [ABSTRACT FROM AUTHOR]