The following technical discussion is part of an occasional series showcasing the ISA Mentor Program, authored by Greg McMillan, industry consultant, author of numerous process control books, 2010 ISA Life Achievement Award recipient, and retired Senior Fellow from Solutia, Inc. (now Eastman Chemical). Greg will be posting questions and responses from the ISA Mentor Program, with contributions from program participants.

Vilson Pastor is an automation and control engineer who is working as a project and maintenance Coordinator in the Teadit Group. Vilson has worked in automation projects participating in the build of supervisory systems, control systems, and PLC project commissioning. Now he is in charge of an electrician team and is responsible of new system specifications, machine and substation preventive maintenance, and systems commissioning.

Vilson Pastor’s Question

 How can I quickly identify possible tuning mistakes in common control loops after verifying signals are correct before spending considerable time on testing and identification of the process transfer function with software tools?

Michel Ruel’s Answer

See the ISA Mentor Program Webinar “Loop Tuning and Optimization” for tips on tuning parameters to recognize unrealistic numbers at a quick glance. Basically, if dynamics (mostly dead time) are in seconds, integral time and derivative time should be noted in seconds. Similarly, if in minutes, Ti and Td should be noted in minutes; if oscillations are observed, the period should be in the same order of magnitude as integral time, and so on.

This is true only with series-real and standard-ideal PID forms. Users must take into account the PID tuning units that can vary considerably with supplier and vintage of a DCS and PLC. The series-real and standard-ideal proportional mode setting may be a dimensionless gain or proportional band. The integral mode setting may be a reset time in minutes or seconds or a reset in repeats per minute or repeats per second, and the derivative mode setting may be a rate time in minutes or seconds.

Mark Darby’s Answer

It is quite useful to know the basic characteristics of the process—either from process knowledge and/or from analyzing trend history. For example, is it integrating or lag dominant (which can modeled as an integrator)? Is there significant process dead time? Note: most tuning packages will require you to specify whether the process is integrating.

If you observe a setpoint change, without significant disturbances, and the controller output returns close to its starting value, then the process can be modeled as an integrator. If it is not an integrator, you may be able to estimate the process gain from a setpoint change (process gain is delta setpoint delta divided by delta controller output at steady-state), even if the controller response is somewhat oscillatory.

Historical trends can be very helpful to estimate process model parameters (gain or integrator slope, time constant, and dead time) to use in tuning formulas. You may find that you can avoid both testing and the use of an identification package, or at least have estimates to sanity check the results. A challenge in using historical data is that the scan or update time may be too slow to get accurate results, especially for fast loops like flow and some pressures. I have found 30-second to one-minute data to be adequate for getting at least a rough model for column temperatures and many levels. I have had good success with five-second data—not for identifying a model, but for assessing control and trial and error tuning of flow and pressure controllers. Trying to assess these loops with one-minute data is hopeless due to aliasing effects. Ideally, one should have access to a fast data collection (ideally as fast as one second) that can be set up on the fly for the loops of interest.      

Greg McMillan’s Answer

If you have a data historian update time that is fast enough with no compression, the time difference between a sustained change in the process variable in the right direction and the change in controller output is the dead time. For a data historian update time of one second or less, the average identified dead time should have better than 20% accuracy for processes with a dead time greater than four seconds. The change in controller output is best done by momentarily putting the controller in manual for a couple of seconds and making a step change in output. After the controller is lined out at setpoint, this should be done in the opposite direction. The tests should be repeated for different setpoints and production rates.

For a gain in intelligence literally and figuratively for slow processes, identify the maximum ramp rate during for a time interval of least four dead times after the process starts to respond. If the ramp rate does not decelerate or accelerate, you have a near-integrating or true integrating process. Don’t do this test on a runway process, because of possible acceleration. Most runaway processes should never be put in manual even for a short time.

If, on the returning approach to setpoint, the process variable halters (momentarily hesitates), you have too much gain action. If the process variable momentarily reverses direction when approaching the setpoint, you probably have too much derivative action. Overshoot of the setpoint without any haltering or reversals on the approach is often an indication of too much integral action (setting in seconds is too small or setting in repeats per minute is too large).

If the loop is oscillating, assuming it is not a dead time dominant process, a period about four times the dead time is often due to too high a PID gain. A period about three times the dead time is often due to too much derivative action, and a period five to 10 times the dead time is often due to too much integral action.

A common mistake not readily recognized is having a PID gain that is too small for a near-integrating, true integrating, or runaway process. These processes lack negative feedback (no steady state achieved within 10 dead times when PID is in manual), relying on the proportional action to provide negative feedback. Consequently, there is a window of allowable PID gains where too low of a PID gain causes slow oscillations (e.g., period of about 40 dead times) and too high a PID gain causes fast oscillations (period of about four dead times). Technically, the product of the PID gain and integral time must be greater than four times the inverse of the equivalent integrating process gain. Most people are not accustomed to the PID gain being larger than five, whereas these processes could use a PID gain of 50 or more and people tend to like integral action since it is more gradual. So the problem is often too small of a PID gain. To check this without risking instability, simply increase the integral time by a factor 10 or more and see if the oscillations decay more quickly. Often the integral time is 100 times too small.

Make sure you know the units of tuning settings. You can get in a lot of trouble by not recognizing a proportional mode setting is proportional band in percent versus a dimensionless PID gain, a reset setting in repeats per seconds versus repeats per minute or the inverse in seconds versus minutes, and a rate time in seconds versus minutes.

If the gain has engineering units, you are in deep trouble. This is an indication the PID algorithm is using engineering units, where in industry the PID algorithm is using signals in percent of scales. If there is an integral gain or derivative gain, you have problems, because this is indicative of a parallel-independent PID form that is rarely used in industry, requiring major conversions of these settings. You will also need to convert settings when going between a series-real and a standard-ideal form if the derivative setting is non-zero.

Most notably, for a standard-ideal form, a rate time setting that is greater than ¼ the reset time setting can cause oscillations that get quite bad if the rate time is larger than the reset time. The series-real form inherently prevents this by an interaction of settings in the time domain. The series-real form was used extensively in analog controllers and first editions of DCS controllers, often with a rate setting as large as or even larger than the reset setting due to tuning rules from the last century.

See the ISA Mentor Program Q&A posts “Key Insights to Control System Dynamics,” “How Do You Convert Tuning Settings of an Independent PID,” “What Are the Benefits of Identifying Dead Time and Ramp Rate,” and the three-part ISA Mentor Program webinar “PID and Loop Tuning Options and Solutions for Industrial Applications.”

David De Sousa’s Answer

Greg, Mark, and Michel provided great tips and insight of identifying problems. I would like only to add that you should always make sure that you know the structure of the PID controller architecture being employed by the PLC, PAC, or DCS at your location. The algorithm behind the PID controller can be built in many different ways, and different OEMs will even have multiple variations of each available within different controller. One way is to consult the Process Control Function Library documentation, which will describe in detail the structure, the mathematical model, and the “handshaking” signals employed to interact with other functional blocks. This will also let you know the correct units of the tuning parameters, which is crucial, as described by Greg.

Additional Mentor Program Resources

See the ISA book 101 Tips for a Successful Automation Career that grew out of this Mentor Program to gain concise and practical advice. See the Control Talk column How to effectively get engineering knowledge with the ISA Mentor Program protégée Keneisha Williams on the challenges faced by young engineers today, and the column How to succeed at career and project migration with protégé Bill Thomas on how to make the most out of yourself and your project. Providing discussion and answers besides Greg McMillan and co-founder of the program Hunter Vegas (project engineering manager at Wunderlich-Malec) are resources Mark Darby (principal consultant at CMiD Solutions), Brian Hrankowsky (consultant engineer at a major pharmaceutical company), Michel Ruel (executive director, engineering practice at BBA Inc.), Leah Ruder (director of global project engineering at the Midwest Engineering Center of Emerson Automation Solutions), Nick Sands (ISA Fellow and Manufacturing Technology Fellow at DuPont), Bart Propst (process control leader for the Ascend Performance Materials Chocolate Bayou plant), Angela Valdes (automation manager of the Toronto office for SNC-Lavalin), and Daniel Warren (senior instrumentation/electrical specialist at D.M.W. Instrumentation Consulting Services, Ltd.).