IMA Active is pleased to take part in the Mathworks seminar

IMA Active is pleased to take part in the Mathworks seminar dedicated to the monitoring of production processes and product quality, which will be held online, on November 23, 2021 at 11:00 am CET.

Many companies after using sensors on their production equipment, are now turning to AI to provide greater process efficiency from the product design to the supply chain optimization, from predictive maintenance to the quality control.

The seminar will provide an overview of the MATLAB and Simulink platforms and tools that enable process engineers to employ statistical, signal processing, and machine learning approaches to improve the quality of sensor data to identify anomalies during process monitoring.

Alessandro Ferri, Test and Research Analyst at IMA Active, will give the presentation “Development of a soft sensor using System Identification on a pharmaceutical process machine”.

At IMA Active, we build automatic machines for every manufacturing step of oral solid dose pharmaceutical products such as granulation, mixing, tableting, capsule filling and weight checking, tablet coating, handling of pharmaceutical containers, and washing of machine parts. The systems (and especially the process machines) come with plenty of sensors but some quantities can be challenging to measure due to mechanical constraints, harsh environments, lack of accuracy, or the high price of sensor. Some variables are simply impossible to measure without interrupting the production.

For these reasons, we have investigated the possibility to create a soft sensor to estimate one variable using the correlations between others actually measured. We selected one of our process machines and picked one variable that is critical to the process and currently measured, but at the same time with a presumed high degree of correlation with other quantities. Using the MATLAB System Identification toolbox we built a dynamic model based on historical and newly acquired data. Within MATLAB, we were performed all the preliminary analysis on the data, such as feature selection and best model order estimation. Finally, the most promising model was tested in real time and compared against the real sensor with good results. The project’s success opens the possibility to extend the procedure to other more challenging quantities, including variables where a sensor is not currently available.