Responding to “The truth about controlling processes…”-Part 2

Adding to the conversation of the Truth about controlling processes… post and response by Emerson’s Terry Blevins, is this response from Greg McMillan. Greg is the author of quite a few books on tuning and control which you can see in the DeltaV Bookstore and an author in the monthly Control Talk column in Control magazine.
Greg adds:

The tuning techniques published by David St. Clair are essentially the ultimate oscillation and reaction curve methods developed by Ziegler-Nichols. These models each use two dynamic parameters. An ultimate gain and ultimate period is identified during closed loop tests (controller in automatic) for the ultimate oscillation method. A time delay and ramp rate is identified during open loop tests (controller in manual) for the reaction curve method. Time constants along with the dead time dominant, self-regulating, and integrating responses are also introduced in the publication to help discuss how the ultimate period varies from two to four times the time delay. Laplace transforms are added to explain dynamics and the PID algorithm. The techniques end up with five parameters and three types of processes, which is the same complexity as methods that more explicitly use a model. What seems to be the difference is that the user does not have to identify a process gain. Actually, this process gain is darn easy to find, in that for an integrating process, the integrating process gain is the ramp rate identified in the reaction curve method (same parameter with just an older name). Further, if you multiply this reaction curve ramp rate by the process time constant (largest time constant), you get the process gain for a self-regulating process. Alternately, in an open loop test self-regulating processes, the user could wait for the process to line out and divide the final change in the process variable in percent by the change in controller output in percent. The user just needs to remember to use percent instead of engineering units because the PID algorithm is based on a percent input and output.
The literal use of the reaction curve method assumes the process is lined out. This is rarely the case for tough loops and integrating processes. If a person understands the concept of an integrating process gain, he or she would realize to use the change in ramp rates rather than the lone ramp rate depicted in the reaction curve method.
In fact, all major tuning methods (e.g. Ziegler-Nichols ultimate oscillation and reaction curve, Lambda tuning, and Internal Model Control) for both self-regulating and integrating process reduce to the same form when the user wants maximum performance. Framing tuning techniques in terms of a simple model (process gain. time delay, and time constant) provides this insight. It also allows you to estimate the effect of plant dynamics on performance. If the user understands the sources of time delay and time constants (e.g. transportation delays, thermowell lags, mixing delays, deadband, sticktion, control communication and execution intervals), the ultimate performance of a control system can be improved. My November Control Talk column will explain the unification of tuning methods and the implications.
If the valve size or the calibration span of an instrument is changed, a user knows how to proportionally change the controller gain based on the change in process gain.
Nonlinearities are prevalent in process control but this is more of a reason to have a simple model so that nonlinearity can be identified and quantified. The user then has the option to schedule the controller tuning based on operating regions or use signal characterization. For example, signal characterization of a pH measurement based on the slope of the titration curve (process gain) has proven to be simple and extremely effective. For nonlinear valve characteristics, the combination of a digital positioner and low friction packing combined with signal characterization of the controller output significantly reduces the nonlinearity from deadband, sticktion, and valve trim. In both cases, the improvement in performance is particularly impressive for operation on the flat portions (low gain) portions of the curve.
The introduction of a fast secondary controller can remove most of the nonlinearity seen by the primary controller in a cascade control system. Also, reducing the time delay and improving the tuning enables a controller to stay closer to set point so that the loop sees less of operating point nonlinearities. This is an important technique for pH, reaction, and distillation column control.
Changes in model parameters provide insight as to what has changed in the plant whether it is an increase fouling of a sensor or heat transfer surface or sticktion in a valve. Models offer plant knowledge and allow you to take the blindfolds off. How you use the models are up to you but ignoring them increases war stories and myths and endless meetings where people go around in circle as to what is wrong and could be better. Some processes have been trying to solve the same old problem for decades because there are no models. A model is even more important if you consider process engineers are taught to think steady state, statisticians analyze snapshots of data, and operators want an instantaneous response. The ability to tune a controller from the same model is a plus.
We should not forget the great contributions from the past but we need to move on and seek greater knowledge and performance of our plants.

You can join the conversation here.

Posted Monday, August 28th, 2006 under Education.

2 comments

  1. Thanks for the detailed reply, Greg, once again confirming my “bottom line” in directing our eager beaver towards St. Clair and Nicols and Ziegler. Laplace transforms are used “to explain dynamics and the PID algorithm”. Exactly as I said, “These concepts are useful in understanding how things work.” But they will not lead him to a reliable method of pre-calculating the tuning constants. That is why we are all advising a procedure, not a formula. That is why there are entire books written and there cannot be a simple “rule of thumb” as was requested in the original post. A well known ISA author published a list of six of them in his book but they really amounted to “get Z-N and follow the instructions”. That’s about what I said, having read the book.
    The remarks concerning deadband, sticktion and valve trim are what I meant with the category of “maintenance items” but I did not want to overwhelm the questioner with too much information especially since it was all available in St. Clair.
    Sometimes a bit of oversimplification helps to get the message across. I think the writer of “You can Tune a Controller But You Can’t Tune-a-Fish”” would understand that. 😉
    Wally.

  2. I think we have similar view points and experiences. Like you, I have not had much success in getting detailed estimates of the process time constant, gain, and dead time before a plant is started up. Also, Ziegler Nichols methods worked well for maximum disturbance rejection in the more important type of loops I faced in chemical plants (e.g. pressure, temperature, and composition control of columns, crystallizers, evaporators, neutralizers, and reactors). Finally, I quickly learned the value of simplifying concepts and adding humor to at least get users on the same page and avoid the big mistakes. Hence, my stories, top ten lists, and rules of thumb.
    You undoubtedly understand the significance of what I say next so it is more meant for readers of this dialog.
    I keep going down the path of trying to get instrument, mechanical, and process engineers to understand the relative effects of their design on tuning and performance. For example, a single large well mixed volume provides a large process time constant that slows down the excursion rate of the process variable giving the controller a chance to catch up with it. On the other hand, oversized dip tubes, volumes in series (time constants in series), long piping runs, and cheap shutoff valves with large sticktion and backlash bought as control valves are sources of dead time. This is not easy for people to grasp on their own because the names and nomenclature of the terms are inconsistent and their effects lost in the math. Every time I hear experienced control engineers and professors say that a loop is destined for poor performance because it has a large time constant, I cringe. A large time constant in the process or disturbance path can lead to great performance. A large time constant in the measurement of control valve is bad news because it slows down the ability of the controller to see and correct for a change that is merrily going on its way. Dead time anywhere is undesirable. If the controller is tuned for peformance, the integrated absolute error for disturbances is proportional to the dead time squared divided by the process time constant.
    Two breakthroughs may change the ball game. The first is that most common tuning methods all reduce to the same form for maximum performance and transfer of variability from the process variables to the controller output. The second is that new software can identify the process time constant, gain, and dead time during startup and normal operation. This leads one to hope that we can move forward and understand dynamics enough to do a better job of improving plant design as well as tuning loops for other objectives besides maximizing load rejection, such as promoting coordination and minimizing interaction between loops.

Leave a Reply