Even when the data infrastructure is sufficient, questions regarding regulatory impact, global acceptance, intended use, risk versus benefit, cost and model maintenance frequency may lead to strong headwinds to adoption of the technology and many hours of interesting debate. This debate becomes exacerbated by three constraints. The first constraint being that it is not possible to determine the final intended use of most process models until their performance has been assessed.
The second constraint is finding resources with the right combination of domain knowledge, modeling and control competencies.
- Berrymans Henry: Living at the Intersection of Need and Art;
- The Colorado Mathematical Olympiad and Further Explorations: From the Mountains of Colorado to the Peaks of Mathematics;
- Passar bra ihop.
This interdisciplinary combination is key when developing models for APC applications. This is particularly true when dealing with complex dynamic systems such as a bioreactor. In fact, it should be the opposite and process modeling should be started early in development to begin to build the necessary data and knowledge plex processes are prone to a higher number of potential failure modes due to having more degrees of freedom.
Thus, in both cases for new and complex processes, process models that can help detect changes from the validated state and pin-point potential variation causalities early in the commercial lifecycle can assist in adding robustness to complex operations and strengthen the overall control strategy.
Subscribe to our e-Newsletters Stay up to date with the latest news, articles, and events. Plus, get special offers from American Pharmaceutical Review — all delivered right to your inbox! Sign up now! This will allow one to develop a model evolution approach that will enable knowledge-based model scoping and intended use maturity, both at the technical and regulatory level. The purpose of this article is to propose an organizational model maturity philosophy and approach. This philosophy approaches model building and intended use in an evolutive manner with the goal of deriving value of APC strategies at various stages of the product lifecycle.
Depending on the level of available process knowledge and characterized knowledge space, various levels of model trust and associated control strategies can be employed. In order to adopt the proposed modeling maturity approach the organization should commit to incorporate process modeling elements and competencies during process development stages. Process modeling and model based process control requires moving away from making decisions that have been traditionally made based on limited and independent data to making decisions based on multivariable relationships Figure 1.
Relevant process variable relationships tend to be in higher dimensional spaces more than three dimensions which is more than what human brains can process. This brings up the need for mathematical and statistical mining strategies to derive values and scorecards that represent multivariable process states and output probabilities.
- Quick navigation?
- Outer Billiards on Kites!
- Tutorial - Simple Nonlinear Model-Based Process Control!
- Techniques of Model-Based Control;
- Table of Contents.
- Particle Accelerator Physics: I Basic Principles and Linear Beam Dynamics II Nonlinear and Higher-Order Beam Dynamics?
- Modern pilgrims: showing the improvements in travel, and the newest methods of reaching the Celestial City.
Modeling with the goal of APC also requires process developers, engineers and operators to start visualizing the manufacturing process as a system which can operate within multiple states leading to a range of outputs, and that within process boundaries, these processes can exhibit infinite continuous possibilities13 Figure 2.
Thus, particularly for complex processes such as a bioreactor, a reductionist approach to seek one-to-one causal relationships is not realistically possible from an intelligent control point of view. When embarking on this journey, the organization needs to be aware that data modeling is not a single task, it is a discipline and that in most cases models will keep on maturing with the process, especially empirical models. Quick statistical analysis and data set description is relatively easy but building models for process control and quality decisions is complex, it requires time, effort and a cross-disciplinary team that speaks roughly the same language.
With regards to language, modeling experts should be mindful that process developers are accustomed to working with a limited number of data points which they evaluate against predetermined ranges and not against each other. Especially, process developers are not used to evaluating new processes using statistical process control tools such as control charts and process capability index due to insufficient representative data.
Thus, the concepts of process variation and control strategy capability may have different meanings to stake holders in the organization depending on their backgrounds and job function. This highlights the imperative need for early language and concept alignment between team members regarding the topics of data distributions, statistical trends and variation relativeness, which are key concepts for empirical model identification.
From a modeling perspective, particularly when working with quantitative machine learning approaches, the main goal of a process model is to characterize plausible variable distributions and complex tendencies i. Thus, from an organizational expectation point of view, understanding that models may not work well to describe all special causes is important to avoid the natural urge of including more complexity in a model to explain improbable states.
Incorporating too much complexity in a quantitative model is cumbersome, expensive and could lead to worse performance when used to describe processes with common cause variability. Fault detection or abnormal event identification is an important element of APC as it allows detecting when a process is performing in an unhealthy manner or outside of process experience.
Fault detection can be achieved by using qualitative, and quantitative approaches. PCM has been defined by Mendonca et. Model Based Process Control MBPC are closed control strategies that use dynamic models to forecast the evolution of the process state around a time period horizon as part of an optimization strategy.
These control strategies use a vector of manipulated variables to optimize the process over a dynamic and receding horizon. In other words, MBPC strategies recalculate new optimal process set-points depending on the process state and some historical knowledge to improve the process outcome. These new set-points are fed into the controller to ensure the output of the process is the best within overall process and time constraints.
Adaptive control strategies fall under the umbrella of performance based control ICH Q12 , in which the focus is not controlling the process to a set of static set-points but controlling the outcome of the process via set-point adaptation.
Figure 4 shows the different elements of APC strategies as well as where they reside within an automation architechture. When evaluating modeling regulatory expectations during submission, the most important factor to be considered is how the model contributes to assuring the quality of the product through the control strategy. As stated by the ICH guidelines, the level of oversight and model documentation should be commensurate with the level of risk associated to the use of the model; this is similar to other types of process or quality controls.
Techniques of Model-Based Control [Book]
Figure 5 depicts the three main categories of model impact as described by ICH standards. Low impact models do not affect process and do not play a role in the product quality assurance strategy. The second category are models with medium impact. Medium impact models can affect the product quality but the effect can be detected through product quality testing, thus there is no risk that the patient will be exposed to potentially impacted product.
The last category described by ICH guidelines is high impact models. High impact models that are the only assurance control of product quality will be scrutinized, validated and maintained with the same rigor as any other analytical method used for quality control. It is important to keep in mind these risk based expectations as pharma starts adopting machine learning approaches and incorporating them into the overall control strategy.
A desired goal of the PAT framework is to develop well understood i. Even though the guideline was written not only from a quality but also from safety, reliability and efficiency perspective much of the PAT efforts have focused on quality measurement lead time reduction. Advanced technologies for quality lead time reduction tend to incorporate high impact models. Some of these high impact models are chemometric models for multivariable analyzer calibration and parametric surrogate models for RTRT and much less on APC fault detection and classification and model-based controls.
Due to the fact that the industry has focused on developing models with the goal of quality measurement lead time reduction, there is a tendency to scrutinize all modeling activities through the lens of high impact models, even when many low impact models can still be included as part of a lower risk APC strategy.
Low to medium impact models within an APC strategy could add tremendous value to supply chain reliability, particularly when used for process state estimation, supervision and scheduling decisions. The output of these models can then evolve into more actionable information once they embed or describe causal or mechanistic knowledge. This maturity model has been developed as a strategic roadmap that should allow us to:. To develop this biopharmaceutical modeling maturity approach, we have drawn inspiration from big data analytics maturity models 22 as well as control strategies used in other highly automated industries.
In our modeling maturity approach, we start with an unsupervised fault detection strategy using high frequency data from the process condition. Our current unsupervised fault detection strategy is based on variable wise batch analysis by means of bilinear modeling. From a modeling difficulty and process expertise standpoint, an unsupervised fault detection strategy requires the least amount of process scientific knowledge and data. These models usually represent a very small subset of the process operational space usually just data from the Proven Acceptable Range, PAR.
The soft sensing layer comes after the unsupervised fault detection layer from bottom to top and it usually employs quantitative models - meaning that the model can estimate an unmeasured variable out of measured ones. To identify data driven soft sensors, data sets with higher variability are needed to characterize an appropriate range of the dependent variable in question.
For example, if we are using process data to estimate the level of an impurity generated in the bioreactor, using data from the PAR alone may not allow us to identify an empirical model or estimating the impurity levels when the process drifts away from typical manufacturing conditions. Thus, this modeling layer requires a more variable data set with an appropriate distribution of the dependent variable even or normal distributions. Appropriate distributions for soft sensor identification are only found in larger historical and perturbed data sets that contain a combination of intentional and un-intentional disturbances and even faults.
As aforementioned, the impact level of a soft sensor will depend on its contribution to assure the quality of the product. Forecaster models can also reside in the soft sensing layer.
Postdoc Dynamic Modeling/Model Based Control of Industrial Chemical Processes
Forecasters are models that can estimate future process outputs depending on earlier process states based on historical or mechanistic understanding. These models are not required to, and in most cases, will not have, the accuracy of a quality measurement of models that are built using an entire process time series. Nevertheless, forecasters could add tremendous value since they could alert plant personnel about the possibility of final quality issues earlier on the process.
Confirmatory measurements triggered by a forecaster can help us to:.
The third layer of our modeling maturity approach is Fault Classification FC. The main goal of a FC strategy is to classify the causation of a fault to enable systematic process recovery if possible or to guide continuous improvement efforts. A fault classification strategy requires a very structured data management approach particularly for complex processes that could exhibit a wide range of possible failure modes. In complex processes, multiple causes can trigger the same symptoms a. Thus, a very thorough system to record faults and linking them to potential root causes and corrective actions is needed to generate a fault classifier.
This fault recording system should collect the information in a machine-readable format and connect it to causation probabilities derived from an expert based assessment or a historical fault database. The fourth layer of our proposed approach to modeling evolution is MBPC for set point optimization.
Model-based advanced process control of coagulation.
Models used to enable this layer require even more data variation and process knowledge a control design space than soft sensors. Jayaraman, V. Sethuraman, P. Raul, and R. Srinivasan, S. Kulkarni, M. Shokrian, G. Shrivastava, and R. Other relatively simple model-based approaches that use first-principles models are GMC generic model control — from the minds of Lee and Sullivan and PFC predictive functional control — from Richalet. GMC using a steady state model could be classified as PI control with output transformation by the model. Advantages are the simplicity of a steady state model and familiarity with tuning and modifying PI.
PFC uses a dynamic model and iterative use of the model to make some future value forecast of the model hit a coincidence point.
Related Model Based Process Control
Copyright 2019 - All Right Reserved