A long short-term memory based Quasi-Virtual Analyzer for dynamic real-time soft sensing of a Simulated Moving Bed unit

Abstract

The evaluation of unmeasurable quantities is an issue faced in several fields. One example is in the production of pharmaceuticals, in which, typically, isomer properties related to different effects are found: one enantiomer may kill while the other may cure. Thus, strict quality control is necessary in this field, which requires accurate, reliable and frequent measurements of the system. However, the measurement of the main quality properties associated with the production of some pharmaceuticals has a low frequency. Consequently, in addition to safety-related issues, financial losses are also related to these low frequency of measurements, as the product can easily run towards an out of spec production. In this scenario, the present work proposes a novel Deep Artificial Intelligence structure, which has an intrinsic (Nonlinear Output Error) NOE structure, associated with a (Nonlinear AutoRegressive with Exogenous input) NARX predictor, to be used as an online soft sensor in order to provide information about the main properties of a Simulated Moving Bed chromatographic unit, commonly used in the production of pharmaceuticals, in order to mitigate the low frequency of measurement associated with this unit. The proposed structure is here called Improved Quasi-Virtual Analyzer. The model can adapt itself as the process evolves, having the possibility of online learning through measurements obtained in the laboratory periodically. The proposed structure was tested in a software-in-the-loop scenario and compared with a more traditional alternative. These tests showed a robust capacity of the Improved Quasi-Virtual Analyzer to provide reliable predictions in real-time, as well as to outperform the traditional artificial network structure.