(322d) Condition Monitoring of Pharmaceutical Processes Under a Probabilistic Framework | AIChE

(322d) Condition Monitoring of Pharmaceutical Processes Under a Probabilistic Framework


Lagare, R. - Presenter, Purdue University
Ganesh, S., Purdue University
Reklaitis, G., Purdue University
Nagy, Z. K., Purdue University
Condition-based maintenance in pharmaceutical processes requires the continuous monitoring of process parameters and asset performance features to enable predictions of when corrective operational or maintenance action would prove beneficial. Although this data does reflect the current-condition of the process and the equipment/facility, it is never sufficient to provide the complete picture, particularly, of the future state.1 This missing information, coupled with the variability in sensor data, contributes to uncertainty that needs to be explicitly quantified and incorporated into predictive models. Furthermore, identifying parameters and extracting features that best represent the state of the process require the use of training data,2 which may or may not completely represent the state of the process under normal operating conditions or under recurring abnormal conditions.

Even with the abundance of data brought about by the digitalization of processes, uncertainty is still prevalent. Aside from the inability to directly measure the condition of a process, the variability of operating parameters, e.g. raw materials, environment, personnel tendencies, can directly affect a process and hence affect the accuracy of the process model.

The explicit expression of uncertainty, of predictions and beliefs deduced from a model, may be reduced to calculation of the probability of that prediction/belief. With condition monitoring, this means using the data to estimate the probability of the condition of the system, and then updating these estimates to reflect new data in real time. Bayes’ theorem fits well into this paradigm, since it provides a mechanism to correctly update probabilities (as a measure of uncertainty) in light of new data, hopefully reducing uncertainty in predictions.3

In this work, we demonstrate the use of Bayesian methods and the advantages of implementing condition monitoring under a probabilistic framework. We illustrate this approach using cases drawn from the operation of the Continuous Pharmaceutical Tableting Pilot Plant at Purdue University, which is a state-of-the-art high-bay facility that can be configured to produce tablets via direct compaction, dry granulation, or wet granulation. Moreover, the leading software for real-time manufacturing management would be used for these cases, particularly PI System (OSIsoft) and SmartFactory Rx (Applied Materials). The former will be mainly the data historian4 while the latter provides an environment for extracting the data from the historian and processing it for the purposes of operations productivity, analytics and control, knowledge management, and predictive maintenance.5 This environment would be ideal for implementing Bayesian methods for condition monitoring.

1 R. Flage, D.W. Coit, J.T. Luxhøj, and T. Aven, Reliability Engineering & System Safety 102, 16 (2012).

2 C.M. Bishop, Pattern Recognition and Machine Learning (Springer, 2016).

3 O. Martin, Bayesian Analysis with Python: Introduction to Statistical Modeling and Probabilistic Programming Using PyMC3 and ArviZ, 2nd Edition (Packt Publishing Ltd, 2018).

4 OSIsoft, OSIsoft PI System (n.d.).

5 Applied Materials, Applied Materials Automation Software (n.d.).