The following discussion is part of an ongoing column called Ask the Automation Pros, authored by Greg McMillan, 2010 ISA Life Achievement Award recipient, industry consultant and author of numerous process control books and articles. Submit your questions to [email protected] with “Ask the Automation Pros” in the subject line. Browse previous questions and answers here. Past Q&A videos are also available on the ISA YouTube channel; you can view the playlist here.
Russ Rhinehart’s question: What projects would you like to see academics explore to support control/automation practice?
It seems to me that academic research in control is led by what academic researchers have published and their practice-irrelevant concepts of what might be possible, but this does not align university creativity with industrial needs and vision. If industry reveals a wish list to academics, it may help steer university research. This would be good: Academic research can provide out-of-the-box solutions that industrial developers might not see. The solutions would be grounded in fundamentals, not heuristics, making them more generically applicable and also confidently accepted. Academic mathematical analysis can provide the underpinning of intuitive solutions that establish credibility. Academics could also provide credible third-party demonstrations and evaluations of new techniques being promoted to get industrial acceptance. Work in investigating possible solutions would better prepare graduates for their careers.
The American Control Conference is planning an Opportunity-Matching event that will have faculty researchers provide abstract presentations of possible solutions to industry’s visions of needs, with feedback and guidance from practitioners. I would like to have your thoughts about practice-directed projects for academics to pursue.
As a starter list of projects, please explore practicable methods to:
- Detect steady-state and transient conditions in process data.
- Use human linguistic models in simulators.
- Automatically identify when controller models or inferential models need to be recalibrated.
- Use one model for all applications — optimization, design, training, process analysis and process control.
Here are a few more ideas that industry would like academics to explore practicable methods to do:
- Validate human linguistic models (or reject folklore) for forecasting process responses.
- Reveal uncertainty in event probability calculations.
- Optimize end-point identification to maximize annual production in batches.
- Develop process controls to shape product property distribution (molecular weight distribution in polymerization, product distribution in cat cracking, particle size distribution in crystallization, passable computer chips on a wafer after etching, etc.).
- Use process data to automatically update process flow diagrams.
Mark Darby’s response
My initial thought for this question was that it would be better to expand the question to more than just academia, although I now see that the motivation for the question is to gather input for a session in the upcoming American Control Conference.
University funding for control and related (process systems engineering) research is increasingly more challenging, and more developments come directly from industrial technology suppliers. Most funding for university research comes from the government; less comes from industry. University funding that does occur often considers applications to newer processes, such as clean hydrogen, nuclear energy and the grid, all worthy topics in my opinion. And not surprisingly, investigators will look at applying newer ideas and techniques, such as AI and machine learning (ML).
However, older techniques need not be abandoned. Sigurd Skogestad, in his paper published in Annual Reviews in Control (2023), “Advanced Control Using Decomposition and Simple Elements,” urged the academic community to “enhance the teaching of advanced regulatory control (ARC) methods and prioritize research efforts in developing theory and improving design methods.” The starting point for a solution need not be limited to the model predictive control (MPC) paradigm: develop a model and an optimization technique that can be solved in real time. Another way of generating development ideas is to ask practitioners questions like:
- What/where are the pain points in your work, where are improvements needed?
- What existing tools and techniques would you like to see improved?
- For academia and industry technology suppliers: What newer techniques (and in what areas) should industry consider applying today? Provide concrete examples.
More generally, I would cite the following development areas:
- Extracting meaningful information and models from historical data
- Hybrid modeling and optimization: combining fundamental knowledge with data-driven techniques
- Abnormal operation detection, including reducing false positives
- Assessment and improvement advisories: process control (e.g., control strategies and tuning) and others
A natural question is this: How can generative artificial intelligence be applied in the above areas?
Greg McMillan’s response
I favor exploring the applications and benefits of using ISA technical reports for test cases with dynamic simulations. The technical reports that are foremost in my thinking of importance for process control are ISA-TR5.9-2023, Proportional-Integral-Derivative (PID) Algorithms and Performance; ISA-TR-75.25.02-2024, Control Valve Response Measurement from Step Inputs; and ISA-106.00.01-2023, Procedure Automation for Continuous Process Operations. These reports and annexes offer extensive, underutilized knowledge.
I would like to see universities develop projects that show how the PID algorithm is best to use for rejecting unmeasured load disturbances and to show that these load disturbances occur as a change in a process input rather than a change in a process output. The 1976 article “Linear Feedback vs. Time Optimal Control. II. The Regulator Problem” by Alan H. Bohl and Thomas J. McAvoy proves what Greg Shinskey and I have tried to teach about load disturbances. I would like to see the value of external reset-feedback for override control, slow secondary loops and slow final control elements, dead time compensation and first tuning the PID for load disturbances and then using setpoint feedforward and setpoint lead-lags or 2 degrees of freedom (2DOF) to achieve the desired setpoint response.
I would also like to see the exploration of the synergy of data analytics, neural networks and first principle dynamic models coupled with the inclusion of measurement and final element dynamics to provide better inferential models based on the knowledge of automation system dynamics and interactions and eliminating cross correlations, gain reversals, bizarre extrapolation and missing relationships.
Studies are needed on how to best control batch and start-up profiles with the realization that decreases in a process control variable that is reactor product concentration is not possible, which makes conventional control unrealizable. Computing the profile slope would address this issue, enabling decreases as well as increases in a controlled variable that is the batch profile slope. It would also enable predicting a batch endpoint by a future value computation. I developed a Mimic Future-Value block that enables rapid updates to a slope calculation with a good noise-to-signal ratio and end point prediction. See the June 2012 Control Talk Blog “Future PV Values are the Future.”
Peter Morgan’s response
The search for a revolutionary control method invariably finds favor as the subject of research projects at places of learning. The work can be stimulating and has the potential of leading to new discoveries that might result in useful applications. In addition, funding for this kind of project is usually more easily obtained, and the research can lead to academic awards.
It occurs to me that an alternative to this kind of project would be to develop projects based on ISA technical report ISA-TR5.9-2023. This report is a compendium of the contemporary knowledge of experts from industry and academia that can directly serve the control community by equipping control engineering students with the knowledge that prepares them for their first assignments in industry and gives them a lasting knowledge of control practice.
Greg also mentions ISA-TR-75.25.02-2024 and ISA-106.00.01-2023. These reports, together with ISA-TR5.9-2023, would complete a comprehensive control practices curriculum option.
If I were to recommend another research project, it would be the development of a plain language, user-configurable diagnostic system to evaluate plant operating vulnerabilities based on equipment availability and operating conditions, and to recommend operator action for risk reduction and upset recovery. Although this is not a new concept, the implementation of such systems has been disappointingly limited, possibly due to the burden of the programming involved, which a new approach might alleviate.
Michael Taube’s response
A challenge/problem I see/hear within the academic community is the “obscene obsession” with “AI.” Even my academic contacts from across both ponds sheepishly admit that the only way they can get research published is to somehow include AI in it. From my perspective, it seems that much/most of the research, as well as products being promoted from it, amount to a “solution in search of a problem.” This is a travesty. Perhaps closely associated with this phenomenon is the perception that “PID is passé.” While I’ve yet to hear anyone actually use that expression, it’s definitely a “vibe” I have perceived for a long time from both industry and academia. Hence, my personal obsession with getting focus back to the foundation of all “advanced” control technologies: the ubiquitous PID.
I’m reminded of the flurry of research that resulted from the publication of the “Tennessee Eastman Challenge Problem” by Downs and Vogel in the early 1990s: numerous papers comparing and contrasting centralized versus decentralized control resulted from it. (In the spirit of full disclosure, I knew nothing about the paper and subsequent research until the late 2010s, well after its heyday. But I personally found it a worthy problem to use as a teaching tool for graduate and undergraduate students, as well as experienced practitioners. Two of my academic colleagues actually pursued using it in an undergraduate course, and the feedback from the students was positive. The paper that was published on this effort is available here. If a “problem” like that was published today — something generic and relevant, that has unique properties/behavior and generates its own data — I suspect that there’d be a great deal of academic activity addressing it. Greg’s suggestion regarding batch control might be a good starting point for developing something similar, as the bio/pharm industry is becoming more automated and, given the historical context for that industry, their understanding of basic process control is pretty limited.
Another aspect of this topic is this: I’m unsure how or even if “research” can address the real problems industry faces. There is an abundance of tools, products and processes by which we can diagnose and address control problems. The real issue isn’t the lack of tools, but the lack of trained and qualified people to implement and use them. There is a “tape worm” that is eating the industry alive from the inside. It started slowly but is rapidly accelerating the process to kill our capacity and capabilities. This parasite is epitomized by the expression “do more with less.” For the better part of 30 years, this business philosophy has decimated staffing levels, training and the technical development of the foundation of every industry: the people at the “Sharp End” who actually perform the work that produces the product/service. This, I believe, is the “elephant in the room” that no one wants or is willing to address. I address this issue consistently in LinkedIn posts and articles, as well as a lengthy paper I presented in 2024 at the Mary Kay O’Connor Process Safety Conference, “Improving Safety Performance: Compliance versus Competence. How will the Process Industries Transcend from the Former and Achieve the Latter?” The paper is available from my LinkedIn profile.
I’ll close with these two quotes from (forgotten?) industry leaders:
- “Human experience shows that people, not organizations or management systems, get things done.” -Admiral Hyman G. Rickover, Father of the Nuclear Navy
- “Without the people to execute, ‘strategy’ is DOA: dead on arrival.” -Larry Bossidy, retired chairman of the board and CEO of Honeywell International
Matthew Howard’s response
I’m not sure I can comment on such a high-level question. I can confirm Michael’s observation about the lean staffing and technical training. I managed the distributed control system (DCS) in a large Kraft pulp mill. Thirty-five years ago, there were four engineers. Since 2010, there has only been one. It is difficult to grow technically with so many system obsolescence problems, capital projects and maintenance problems from breaking in new equipment vying for my time.
I can’t imagine that pulp and paper is unique, though the margins are probably tighter than some other industries.
Héctor Torres’s response
From my experience, dead time, interacting process loops and poorly acting control elements affect process control. University projects should provide practical tools to prepare students to deal with these common situations:
- The effect of dead time on the capacity to control a process and different techniques to minimize its negative impact on control
- Dead time compensation
- Smith Predictor
- MPC
- External-reset feedback
It is common to find interacting control loops. Projects should include exploring cross-coupling between controlled and manipulated variables and its effect on control systems, as well as multivariable control approaches. variable pairing, forward decoupling, inverted decoupling and partial decoupling.
Ideally, these topics should not be isolated exercises but integrated into projects or lab simulations that combine multiple challenges, such as dead time, interaction and sluggish actuators, mirroring what engineers encounter in the field. This approach will produce graduates who understand these bad actors and are confident in selecting the right tool for each problem.
I would encourage universities to partner with process industries to develop case studies and digital twins that reflect real operating conditions. Students equipped with hands-on experience will make faster, more confident decisions when they face these problems in their first industrial role. They will be able to assist with reducing start-up time and improving process reliability from day one.
Pat Dixon’s response
There are some measurements that are particularly challenging in industry. I realize that many vendors are trying to find improvements and solutions, but this could be a good topic for academic research. In the paper industry, consistency (the fraction of solids/wood fiber in pulp) can be difficult under different process conditions and can often go out of calibration. Other pulp properties, such as freeness, kappa number and viscosity, are also difficult. At the reel of a paper machine, there have been several failed attempts to measure strength properties such as tensile, burst, tear and ring crush. In water treatment, biochemical oxygen demand (BOD) is an important property. I am sure many other industries have similar instrumentation challenges for analytical properties. Reliable online measurement can greatly benefit these operations and could be a good candidate for university research.
Julie Smith’s response
Excellent suggestion Pat; I fully support your recommendation. Measurements are the foundation of good process control. There have been many cases where an improved strategy or design has been sidelined by a less-than-reliable measurement. Some properties, such as color, are still measured offline and after the fact in many industrial processes. Yes, vendors are making improvements, to be sure, but they would be much easier to sell if supported by academic research. Such research should also include ways to improve the calibration and maintenance of the measurement instruments, as we see a reduction in the skilled labor ranks.
Bob Heider’s response
Julie Smith’s response captures my thoughts on the subject. I would add emphasis on batch control because most pharma and semi processes are batch.
That said, I still believe every undergraduate chemical engineering student needs a control class, with a focus on instrument fundamentals and basic control as covered in Basic and Advanced Regulatory Control: System Design and Application, third edition by Harold Wade. I see too many installation errors.
George Buckbee’s response
First of all, thanks for starting this conversation. There is a great opportunity for progress if we can gain some alignment between industry and academia.
In my opinion, the biggest opportunity lies in the large-scale integration of plant data and information. The industrial world is awash in data from hundreds of different systems that rarely communicate with each other. They work with completely disparate and seemingly unrelated data, structures, tools and users. Yet, there is a great opportunity to develop something like an intelligent view of the plant if we can find ways to connect the information. The opportunity aligns with these two bullets from your starter list:
- Automatically identify when controller models or inferential models need to be recalibrated.
- Use one model for all applications — optimization, design, training, process analysis and process control.
But I believe the potential can be even larger. Here are some examples, in no particular order:
- Creation of “living” process flow diagrams (PFDs) and P&IDs as digital twins. Combine these with process historian and laboratory data. Validate static models of the process. Update models with actual results. Highlight gaps/errors in the model or in process performance. Identify process shifts over time. Implement automated adaptation of models based on results.
- Automated development of large multivariable (i.e., whole plant) process models. Given access to long-term process and laboratory history, develop a complete static or dynamic model of the process. Of course, if a PFD or P&ID exists, it can guide the process; however, much could also be inferred without those documents.
- Very large-scale studies: Correlate equipment reliability to process history and performance. Review the maintenance, process and equipment history for an entire corporation and then evaluate the effects of equipment cycles, variability, start-up/shutdowns, operator interventions and so on. There could be separate studies for each type of equipment (pumps, fans, compressors, valves, instrumentation, etc.), or for different applications and industries.
- Anomaly detection. Make use of all data: operator logs, process history, alarms, reliability and lab data. Evaluate previous failures and trips and learn from them. Develop predictive analytics for process anomalies. The analysis could be based on whole plant models as identified above, black-box machine learning, or hybrid models (defined model structures with ML parameterization).
Foundational research and technology exist for all of these suggestions. Two aspects are different: application at scale and integration of data from multiple sources. There should also be some level of interest from corporations in funding this sort of research. While some companies may have pursued one-off attempts at these (i.e., turbine reliability studies), the research could also be more generalized to allow for application broadly across industries.
These are just a few thoughts. I’m sure many other similar suggestions could be developed. Please keep me in the loop on your progress.
Michel Ruel’s response
We should remember that process control research is often embedded within other industries such as mining, pulp and paper, and oil and gas. In these sectors, research tends to focus on sensors, instrumentation and application-specific control strategies.
As a standalone discipline, process control offers many new research opportunities driven by emerging technologies, particularly artificial intelligence. Promising areas include abnormality detection, multivariable modeling, root-cause analysis and digital twins — in other words, the layers of control and optimization that sit above traditional PID control.
For example, research could focus on detecting how oscillations and variability propagate through a process area and identifying the control loop responsible for the disturbance. Modern industrial plants generate vast amounts of operational data, making it possible to conduct this type of research using real plant data rather than relying solely on simulations.
Ed Farmer’s response
What I am offering here addresses the more general issue of more effective education for our profession. My ways of looking at things for a teaching application, like mentoring, are a bit different. My original tutoring for that kind of thing came from two top-level engineering managers at Chevron, and from Exxon policies I encountered early in my career. It began with deciding if an inquiry was asking for help with an applied method (called “assisting”) or help selecting, understanding and using an appropriate method (“mentoring”). Depending on the situation, “helping” might begin with “read the manual and let me know if that helps.” Chevron and Exxon were both focused on developing their staff but sensitive to training situations that could result in project completion delays.
My training and experience usually motivated me to a procedure that clearly identified the problem and the more general question, and connected it with a standard practice. If it was beyond “standard practice,” effort moved toward identifying the knowledge needed to address the problem. That often involved a higher level of the engineering department and new “standard practices.”
In most situations, the effort began with the overall situation — the “big picture.” Sometimes the effort needed to include or be directed to “other” or “bigger” systems. Then, a procedure with stuff from an “operations research” study helped ensure that the devised plan considered all the relevant conditions. Sometimes, the devised plan motivated a test program before full implementation. Sometimes the issue had to be escalated to “research and development.” Often, all this effort could impede solving the first situation, but it avoided a rising sea of related problems that need to be resolved. My training and experience have motivated me to carefully understand a situation, consider the “big picture” and then, sometimes, even bigger pictures — all leading to confidence in a “solution.”
I’m not sure this is pertinent to what you’re planning. A lot of problems can be addressed by telling someone to “read Greg’s book.” By the time I began this kind of refinery work, as a consultant, I was never asked to tell people what a book said. I could help them find and follow directions, but their supervisors or “standard practices” were supposed to do that.
This is not meant as “criticism.” It’s meant to suggest a bigger “professional society level” and its pertinence beyond the simple things. It’s always nice to solve a problem with basic effort, but it is not-so-nice to see a situation you were paid to resolve going to R&D or another consultant because of “edges” you didn’t consider. Having a broadly focused group helps ensure a broad vision.
From my earliest “refinery” days, I came to understand the diversity of an “instrument” project. Work ranged from pipe fitters to instrument installers, controller installers, tuning technicians and process-level engineers. The scope involved places in the process, places on the control panel and things upstream and downstream. The work ranged from doing what the instruction manual said, up through engineering and sometimes into fresh learning experiences. A person with overall responsibility might have some really interesting days, perhaps including a visit to Chevron Research.
In mentoring situations, there was usually something to fix and something to learn, which had to be adapted to the various work levels. The engineer in control was responsible for making sure that everyone involved understood their role — one way or another.
All that stuff taught me a lot. From my Army training and experience, I learned to think about “common” tasks and harder ones for which specialists were required. There was also the “team” concept that stressed the importance of broad and deep thinking.
And then there was a focus on the objectives, ranging from safety to overall performance. Somehow, mentor work involved far more than pointing and saying, “Turn that screw.” All that, though, was not supposed to slow the project. “Knowing” often organized and sped up learning and finishing.
In a venture like education and mentoring, it’s important to keep all that stuff in your mind and on the schedule. Sometimes a visit to R&D could teach you things you could use to organize a focused, speed-increasing training program.
Like a lot of things in engineering and construction, it’s hard to teach everything and risky not to, but getting it right the first time saves time and money. A “simple question” mentoring write-up is pretty difficult to do quickly. The “right answer” might need knowledge of related but obscure concepts. Sometimes conditions change the answer — things like “liquid, solid or vapor” or “temperatures?” or…
Covering “everything” might turn out more like a university master’s program. There’s room here for planning, thinking, teaching and doing; along with refining the best ways of doing all those important things.
