Operators and technicians on the pharma manufacturing floor are constantly responding to process changes as they occur in real-time. Unfortunately, data analysis on the manufacturing floor isn’t real-time, and it’s difficult to understand which of the many production parameters is driving variation in performance. Often a reduction in quality or yield occurs without being anticipated and without a clear reason manufacturing teams can point to. Significant off-line data crunching must be done by manufacturing data scientists to interrogate historical data in the hopes of deciphering the cause. This analysis is only valid for that one point in time, not for any future events that may occur with the same equipment or processes. That’s frustrating for operators as well as process science teams and has caused leading companies in the industry to look for a better way to perform real-time analytics.Performing more sophisticated root cause analysis is more complex than it may first appear. Many pharma processes (especially in biologics manufacturing) aren’t fully characterized, with multiple different process parameters – pressure, temperature, mixing speeds, settling velocities, etc. – simultaneously impacting critical quality attributes (CQAs). These parameters also interact in complex ways with equipment in a multivariate fashion, meaning it’s often unclear how changing a parameter may be impacting output. All of this makes understanding and optimizing pharma manufacturing processes more complex, requiring advanced analytics tools to understand and characterize patterns in the data.
Most often, manufacturing data scientists today understand these patterns using offline and bespoke models in Python, Matlab or R. However, it is difficult to connect these programs to regulated, real-time manufacturing systems. The offline data available is often redacted and summarized, meaning it’s rare for analysts to gain enough insight to optimize the process. Ideally, manufacturing data scientists would be able to build a live analysis based off the current reality—and not the data that was collected months or years ago. Such an approach also enables continuous models that run frequently and can provide fresh insight second-by-second using the most up-to-date data.
Orchestrating data for real-time analysis
SmartFactory Rx® bridges the gap between advanced analytics and real-time data systems. It supports dozens of modeling approaches – including machine learning (ML) techniques – to provide insights and suggestions for process improvement using live models. By linking directly to real-time systems, those live models support direct improvements to generate higher yields, optimize cycle times, and increase overall equipment effectiveness (OEE). The platform adopts a common approach using historical data to develop advanced models and analysis to provide immediate and actionable data about the process.The SmartFactory Rx strategy engine orchestrates all the data and inputs that are coming in from the manufacturing floor and centralizes how that data is processed to be used within the analytics processing stack. There are typically multiple data sources/databases with a combination of time-based process data, equipment/maintenance data, and data about the batch – all of which can impact the CQA being addressed. Centralizing data creates a ‘single source of truth’ for analytics engines, building a unified picture of the process that includes all potential variables to be considered. That in turn allows for the full power of ML or other analytics models to be applied, allowing for more powerful insights.
How challenger models can help achieve real-time continuous process improvement
Models created by SmartFactory Rx take the latest data and continually re-build, re-validate and re-test the current model to see if the inputs that went into that model still have the biggest impact – or if there are new parameters that are more relevant. If there are changes in the process, it also helps to identify whether the change was deliberate, or a matter of process drift. The model helps focus the user on the most critical parameters at the current moment in time: for instance, while five parameters may have been critical to quality when the model was created, there could now be five different, more impactful parameters to focus on.Often, shifts in priority impacts may occur because – as the original parameters are optimized – the impact on the variability of the CQA(s) of interest is addressed and the impacts of other (previously less impactful) parameters can be more easily seen. Being able to see real-time impact of changes made to the process and show operators the new process parameters in real time allows continuous process improvement to take place.
Making it practical for operators and technicians
The deployed model and inputs, as well as the challenger model framework, are monitored through a dashboard designed to simplify the results for users and provide real-time insight on the shop floor. As new models are automatically created, and performance is contrasted with a currently deployed version, the user will receive a notification when the challenger model shows improvement. Challenger model adoption can be automated so that over time, the system automatically evolves to match your changing process. SmartFactory Rx also creates an audit trail so changes to the model are tracked in terms of who made them, and previous versions of the models remain accessible.
Using SmartFactory Rx, when the operator on the floor sees a change in the process, they will be able to determine why it is happening and see recommended actions to take. Rather than realizing the yield is low and conducting a full after-the-fact investigation to find out why, this new generation of ‘learning’ tools allows operators to see real-time issues as well as key contributors that truly represent the critical few things to examine more closely. Operators get simple, timely information while data scientists have real-time data available for powerful in-depth analysis that can continuously evolve to meet the changing needs of the site. Such learning models, when deployed properly, drive significant and continuing improvements to yield, variability and process robustness.