Intelligent Process Condition Monitoring for industrial tablet coating operation

This paper presents a feasibility study to develop an Intelligent Process Condition Monitoring (IPCM) system for providing a real time ‘health check’ for the tablet coating process in drug product manufacturing. The study fits well under the framework of Intelligent based Manufacturing (IbM) initiated at Pfizer, and intends to move the coating operation from responsive and reactive action to preventive and proactive strategies in ensuring process performance and product quality attributes. The successful implementation will ultimately improve batch yields and reduce coating interruptions.

Complex coating in a controlled-release tablet process is a critical unit operation used to control dissolution and a number of other product quality attributes of some tablets. The variation in coating process and equipment performance, especially sensor/equipment failures, poses a major risk to process outcomes and product quality.

The proposed multivariate condition monitoring system is expected to achieve the following: 1) Real-time monitoring of batch coating evolution and performance; 2) Immediate fault detection and diagnosis, e.g. process deviation or sensor failure; and 3) Enable preventive or corrective actions in conjunction with control systems.

The feasibility study is based on the analysis of historical batch data available in the GE Fanuc Proficy historian. Approximately 30 batches manufactured between 2011 and 2012 were made available for the analysis. The dataset includes a number of batches with desired performance (good batches) as well as those considered atypical. A good batch is defined as one with desirable inert coating process performance, and zero quality issues and equipment faults. Atypical batches are defined as those experiencing equipment faults such as sensor/seal failures, air flow issues, or those with unplanned stoppages (downtime).

Data analysis: condition monitoring of the inert coating process

A multivariate batch monitoring system employs a set of powerful methods for analysing, monitoring, diagnosing process performance and facilitating feedback/feed forward control.

The effectiveness and usefulness of the multivariate batch modelling is demonstrated below by analysing batch process trajectory through observation-level PLS modelling1,2. Several diagnostic tools associated with the model can be used to monitor the batch condition and detect/diagnose faults as the batch evolves, including Score control charts, Hotelling T2 and Distance-to-Model (DModX) control charts. The Hotelling T2 statistic with contributions is derived from score vectors that are significant, whereas the DModX is derived from the insignificant scores or the residuals, and defined as the sum of the squares of the errors. The Hotelling T2 statistic provides a measure of the distance that the current data point lies from the average of all the retained scores, indicating how far normal operating conditions the process is operating. The DModX gives an indication of how similar the current process conditions are to those used in the training data. The DModX and Hotelling T2 together provide a useful framework to determine whether the prediction of a new observation is being performed within the control limits established through the model. More details on these statistics can be found in the literature3,4.

Training – establish control limits

The purpose of training is to build models based on good batches and establish dynamic control limits. Three good batches were included in the model, #19, #20 and #21. More than 1500 time points (Observations) during the inert coating phase were selected, with 15 process parameters included, e.g. Tablet Bed Temperature, Inlet/Exhaust Temperature, Inlet Air Flow, Solution Flow and Drum Load Pressure etc.

The process conditions during these batches are considered to represent desired Normal Operating Conditions (NOC) as these batches all met specifications and operated well within manufacturing instructions acceptance criteria. X-variables include all process para – meters, whereas the Y variable is now batch time/maturity. Data were preprocessed by auto-scaling prior to modelling. A fivecomponent PLS model was established, with R2X of 65 per cent (explained X-variation), R2Y of 90 per cent (explained Y-variation) and Q2Y of 90 per cent (predicted Y-variation accord – ing to cross validation).

Figure 2 shows the dynamic control limits associated with two multivariate control charts, Scores (t1) and Distance-to-Model control chart. Each point in the trajectory plots represents a combination of original variables at a particular time point. The green trace represents average batch over time, and the red traces ±3 sigma control limits. Batches #19, #20 and #21 are coloured in Black, Blue and Yellow.

Prediction – monitor new batches

Once dynamic control limits have been estab – lished based on the good batches, batch control charts can be used to predict, or monitor, new batches against NOC, or detect / diagnose faults and identify problematic variables as batches evolve.

Figure 3 shows predicted process trajectories (represented by DModX) of an atypical batch #25 (marked in blue) as compared against the control limits. It can be seen that the batch has evolved outside control limits at various stages of the batch evolution, indicating process upsets. To diagnose these deviations, the DModX Contribution plot can be used to identify problematic variables contributing to the abnormalities. Exhaust temperature and inlet temperature are seen to show larger contributions, and a further look into individual contributing variables indicates that these two variables appear larger than average after the 200 time point (~14 per cent process completion). This information has a practical use in that operators or a process control system can take preventive or corrective actions.

Figure 4 shows prediction/monitoring of another batch #30, during which there were mechanical issues. A score control chart indicated that the batch evolved mostly along or outside the upper control limit. Score contribution analysis revealed that the biggest contributing variable to the deviations, Drum load pressure, appeared below the average and outside the lower control limit throughout most of the batch duration. With this type of condition monitoring system, these issues could have been detected in real time and a corrective action could have been taken.

Conclusion

Based on historical batch analysis, it is also feasible to develop an Intelligent Process Condition Monitoring (IPCM) system to provide an overall health index of the coating operation. The illustrated IPCM system is an example of intelligence based manu – facturing, which transforms data into process intelligence through a model centric approach. For future implementation, APC software will be connected to the coater via OPC and be used to develop a real-time soft sensor and batch monitoring system. HMI can be developed to display monitoring results in a simple and intuitive way for Operators. Alarms will be triggered, and corrective actions can be taken through open-loop or close-loop control.

Acknowledgement

Steve Hammond, Jacintha Griffin, Orla Markey and Pfizer Newbridge Automation Team are gratefully acknowledged for their support and efforts during this feasibility study.

References

1. Wold, S., et al., Modelling and diagnostics of batch processes and analogous kinetic experiments. Chemometrics and Intelligent Laboratory Systems, 1998. 44(1-2): p. 331-340

2. Chiang, L.H., et al., Industrial experiences with multivariate statistical analysis of batch process data. Chemometrics and Intelligent Laboratory Systems, 2006. 81(2): p. 109-119

3. Nomikos, P. and J.F. MacGregor, Monitoring batch processes using multiway principal component analysis. AIChE Journal, 1994. 40(8): p. 1361-1375

4. Ündey, C., E. Tatara, and A. ÇInar, Intelligent realtime performance monitoring and quality prediction for batch/fed-batch cultivations. Journal of Biotechnology, 2004. 108(1): p. 61-77

About the authors

Dr. Jun Huang is Manager of PAT projects with the Process Analytical Sciences Group (PASG) at Pfizer in Peapack, NJ. His primary role is to implement Intelligence based Manufacturing and PAT within Pfizer Global Supply. Prior to joining Pfizer/Wyeth, he worked at GlaxoSmithKline and PerkinElmer. He is experienced with chemometrics, PAT, multiple analytical techniques, and advanced process control (APC). Jun holds a PhD in chemometrics from the Norwegian University of Science and Technology in Norway.

Dr. Saly Romero-Torres is currently working as PAT project manager at Pfizer Pharmaceuticals. She holds a PhD in Analytical Chemistry which she completed at Purdue University (West Lafayette, IN) in 2006. Her dissertation title is ‘Raman and Chemo metrics for Pharmaceutical Process Analysis’. Before joining Pfizer (legacy Wyeth), she worked at Schering-Plough Pharmaceuticals as a spectroscopist in a physical characterisation team. At Schering Plough, she developed new spectroscopic (IR, NIR and Raman) and chemometrics based analytical methods aimed to characterise and understand chemical and physical attributes that were critical to the quality of pharmaceutical materials.

Liam Ryan is currently working as Operations Lead with Pfizer Ireland Pharmaceuticals, located in the Newbridge Manufacturing Facility. He has worked at Pfizer since 2004 in both technical and operations roles. Liam is a certified Lean Six Sigma Black Belt and he received his MSc in Biomedical Science from the National University of Ireland, Galway in 2000.

Dr Mojgan Moshgbar leads a team focused on Advanced Manufacturing and Innovative Processes in Global Technology Support organisation in Pfizer. She has developed the Intelligence based Manu facturing initiative, now an important element of Pfizer’s technical strategy. In addition, her team has been responsible for development and implementation of advanced integrated PAT platforms for CQV and RTR strategies in both Continuous and Batch API and DP processes. Dr Moshgbar has more than 20 years of experience in Process Modelling, Optimisation, Advanced Controls and Intelligent Manufacturing. Prior to joining Pfizer in 2002, she held a number of senior technical positions in industry and academia. Her education includes a BSc (Hon) in Physics and Astrophysics; Master degrees in Semiconductor Physics; Robotics, Automation, and Artificial Intelligence; and PhD in Advanced Mechatronics, Adaptive Process Control, and real-time Optimisation.