How to tune a PID loop on a running plant
Tuning a loop that is already on production is not the textbook procedure. You do not have a simulator, you do not have permission to drive it unstable, and you usually do not have a clean first-order model. What you have is a trend, a faceplate, and a process that has to keep running.
The order of work is read, then one change, then wait. Most loops that get re-tuned repeatedly were never a tuning problem. Stiction, backlash, a saturated output and a passing valve all produce a trend that looks like a controller fight. Changing constants on those loops makes them slower without making them better.
Start with the manual test
Put the loop in manual and leave the output where it is. Watch the process variable. If the cycling stops, the controller was generating it. If the cycling continues, something else is driving the process — another loop, a cycling supply, or a sticking valve hunting against itself. That split takes half a minute and eliminates about half the possibilities.
If you cannot transfer to manual without the output jumping, fix bumpless transfer before you touch a constant. A mode change that upsets the process is not a tuning problem; it is another problem.
Refuse the mechanical problems
A constant-amplitude cycle that survives manual is almost never tuning. Stiction produces a square-looking output and a PV that barely moves until it jumps. Backlash produces a deadband around the reversal. An output pinned at 0% or 100% has no remaining authority. A valve that will not shut off, with the output on the seat and the PV still above setpoint, is passing. None of those yield to a new gain.
If the trend shows one of those fingerprints, stop. Repair the valve, the supply, or the structure. Come back to the constants afterwards.
One change, then wait until it has finished
If the loop itself is generating a decaying oscillation, the constants are in play. Change one of them. Write down the value you started from. Wait until the response to that change has finished before judging it — seconds on a flow loop, minutes on pressure, much longer on temperature and level. A second change stacked on an unfinished first one produces a picture nobody can read.
Never raise the gain until the loop cycles. That experiment is how Ziegler-Nichols finds the ultimate gain, and it is how you upset a unit and explain a self-inflicted disturbance to the board. Named methods that start from a bump in manual — Lambda, Cohen-Coon, open-loop Ziegler-Nichols — can give you a starting point. They still assume a self-regulating first-order process you have actually measured.
Match the structure to the loop
Flow wants PI and no derivative. Pressure depends on whether the volume is small or large. Most level loops should be averaging, not tight. Temperature is the common case for derivative, on PV, and is usually better as a cascade onto a faster inner loop. Speed loops are often limited by the drive ramps rather than the PID. pH and analyzers are usually structural problems wearing a tuning complaint.
Questions that come up
Can I use a PID tuning calculator at the console?
You can use one to see what a named method produces from numbers you measured. You should not type the result untested. The calculator does not know if the valve is sticking or the process is integrating.
How do I know when to stop?
When a disturbance produces at most one small overshoot and then sits. Chasing the last wiggle is how loops get wound up tighter than the process will tolerate on the next load change.
Related
Tuning guides
Last reviewed 2026-09-10.