The frustrating part of PID tuning is not that there are three gains. It is that the same gains can behave differently on another device, another load, or another sampling period.
A thermal loop heats faster after increasing proportional gain, but starts overshooting. Integral action removes steady-state error, then makes the loop overshoot more. Filtering makes the reading smoother, but the controller reacts late. Replacing a valve, motor, or heater makes yesterday’s stable gains unusable.
It is tempting to treat PID as something almost mystical.
PID itself is not the mystery. The whole loop is:
setpoint
-> error
-> PID
-> limits, ramps, and protection logic
-> actuator
-> plant
-> sensor, filtering, sampling, and communication
-> back to error
PID is only one section of that loop. Delay, noise, saturation, and nonlinearity anywhere before or after it show up as tuning problems.
The Three Terms Solve Different Problems
Proportional, integral, and derivative action are not three independent knobs. They respond to different aspects of error and each one has a cost.
Proportional action looks at current error:
larger error -> larger output
It makes the system react faster to deviation. Too little proportional gain makes the loop slow. Too much amplifies noise, delay, and inertia, and can make the output oscillate around the target.
Integral action looks at accumulated error:
persistent error -> output keeps growing
It removes steady-state error, such as a heater needing a baseline power level or a motor needing extra torque under load. But if integral action is too strong, the internal accumulated error can become large before the physical system catches up, causing overshoot and windup.
Derivative action looks at how fast the error is changing:
faster error change -> earlier correction
It can reduce overshoot, but it depends heavily on measurement quality. Sensor noise, quantization jumps, and sampling jitter can make derivative action noisy.
PID tuning is not simply “increase P if it is slow, increase I if there is offset, add D if it overshoots.” Each term trades response speed, stability, noise sensitivity, and actuator movement against each other.
The Plant Sets the Scale of the Gains
The same PID structure behaves very differently when controlling temperature, motor speed, liquid level, or pressure.
The reason is the plant.
Thermal systems often have large inertia and slow response. After heater power changes, the sensor sees the temperature change later. If proportional gain is too high, the controller keeps heating based on old temperature data, and the stored heat drives the system past the target.
Motor speed responds much faster, but friction, load changes, current limits, and encoder resolution matter. Low proportional gain feels weak. High gain can shake. Integral action can hold speed under load, but can also accumulate too much during current limiting, stall, or low-speed friction.
Liquid level and pressure systems may add valve dead zones, pump minimum flow, pipe delay, and fluid inertia. A changed output command does not always produce a linear physical change.
PID gains do not exist separately from the plant. A slow plant cannot tolerate overly aggressive correction. A nonlinear plant is hard to cover with one fixed set of gains.
The Sensor Decides What Error PID Sees
PID acts on error, and error comes from the measurement.
If the sensor reading is noisy, proportional action turns that noise into output movement, and derivative action amplifies it even more.
If the sensor responds slowly, the controller is looking at the past. It thinks the system is still far from the target and keeps pushing. When the real plant is already near the target, the correction arrives too late.
If the measurement is heavily filtered, the displayed curve may look much better, but feedback reaches the controller later.
Common symptoms include:
- the loop becomes much slower after the reading is made smoother
- increasing proportional gain makes the actuator follow sensor noise
- adding derivative action creates more output spikes
- low measurement resolution makes the target region jump in steps
- moving the sensor changes the behavior of the same gains
Before tuning PID, make sure the error seen by the controller is credible, timely, and connected to the real output.
Sampling Period Changes the Meaning of PID
Discrete PID runs at intervals. That interval is not a detail; it changes integral and derivative behavior.
With a longer sampling period, control actions are less frequent, feedback arrives later, and stability margin shrinks.
With a shorter sampling period, the controller sees changes sooner, but noise, quantization, and scheduling jitter can enter the loop more easily.
If the sampling period is not stable, integral accumulation and derivative estimation are distorted. On devices, common causes include:
- logging overhead
- communication tasks preempting control
- low-power wake intervals changing
- variable sensor conversion time
- bus congestion delaying readings
- poorly chosen control task priority
The same PID gains at 10 ms, 100 ms, and 1 s are not the same controller. Without recording the sampling period, a gain set has little portable meaning.
Actuators Separate Command from Physical Action
PID computes an output command. That is not the same as physical action.
Actuators commonly have:
- maximum output limits
- minimum effective output
- output slew-rate limits
- dead zones, static friction, or backlash
- soft start, protection, current limits, or thermal derating
- mechanical or electrical delay between command and motion
These limits mislead the controller.
For example, integral action may keep pushing the command upward while the actuator is already saturated. The physical output no longer increases, but the controller’s internal state keeps accumulating. When the plant finally approaches the target, the integral term has not unwound yet, and the system overshoots.
Or a valve may not move at small openings. When the error is small, PID output changes have no effect. After the error accumulates, the valve suddenly moves, and the level or pressure goes too far.
These problems are not solved by gain tuning alone. They need output limits, anti-windup, minimum output handling, dead-zone compensation, output ramps, or a different actuator.
Setpoint Changes Can Create Tuning Illusions
Many unstable-looking PID loops are not bad at holding a steady state. They are being hit by setpoint changes that are too abrupt.
If the setpoint jumps from 0 to 100, the error becomes large instantly. Proportional output jumps, integral action starts accumulating, and the actuator may saturate. If the plant has inertia, the system can overshoot first and recover later.
Common engineering fixes include:
- ramping the setpoint instead of stepping it
- ramping the output to limit actuator change
- reducing gain or switching strategy near the target
- disabling integral action while the error is very large
- adding feedforward so PID does not carry the baseline output alone
These are not just tuning tricks. They often matter more than another round of editing Kp, Ki, and Kd.
Measure the Loop Before Tuning It
PID tuning should come after loop validation.
A more reliable order is:
- Test the sensor in open loop: noise, delay, resolution, and placement.
- Test the actuator in open loop: command versus real action, dead zone, saturation, and delay.
- Fix the sampling period: control task timing, sensor reads, and communication timing.
- Disable integral and derivative action temporarily, then use small proportional gain to observe the plant response.
- Increase proportional gain gradually and find the boundary between response speed and oscillation.
- Add a small amount of integral action to remove steady-state error, while watching for saturation.
- Consider derivative action only when measurement quality is good enough.
- Add limits, anti-windup, ramps, dead-zone compensation, and protection logic.
If the open-loop sensor and actuator are not understood, every closed-loop symptom will look like a gain problem.
PID is hard to tune because it exposes the device’s timing, amplitude limits, noise, inertia, and nonlinear behavior.
The real work is to make the loop explainable first, then let the gains handle the part they are actually meant to handle.