control//feedback control//control loop//plant

2026-10-01

The plant is the part of a control loop that the controller acts on and cannot redesign: the oven with its thermal mass, the motor with its load, the tank with its outlet. The name comes from process plants and stuck for anything under control, a car or a drone included. Before tuning anything an engineer needs a model of it, and the cheapest model comes from one experiment: hold everything steady, step the input, and record the output.

Most industrial plants answer such a step test with an S-shaped curve that three numbers describe well enough to tune a controller. The gain KKK is how much the output finally changes per unit of input. The dead time LLL is how long nothing happens. The time constant τ\tauτ is how long the response then takes to cover 63 % of its way. Together they make the first-order-plus-dead-time model (FOPDT), here for a step of size u0u_0u0 at t=0t=0t=0:

y(t)=K u0(1−e−(t−L)/τ)for t≥L,y(t)=0for t<L.y(t)=K\,u_0\left(1-e^{-(t-L)/\tau}\right)\quad\text{for } t\ge L,\qquad y(t)=0\quad\text{for } t<L.y(t)=Ku0(1−e−(t−L)/τ)for t≥L,y(t)=0for t<L.

The model is wrong in detail and right where tuning needs it: how strong, how late and how slow the plant is.

gain K2.11 time constant τ2.9 s dead time L1.5 s A unit step at 2 s into a plant, its output read every 0.1 s. Read off the curve, the response is a gain of 2.11, a dead time of 1.5 s and a time constant of 2.9 s.

Fit the model over the readings with the three sliders until the error falls to the level of the noise, or let the two-point reading do it, then ask for a new plant. The plant behind the dots has two lags and the model only one, so the faster lag comes back as extra dead time: a model is a summary, not a portrait.

The three numbers decide the controller. The gain sets the scale of the controller's gain (a plant with twice the gain needs half the controller gain for the same loop). The ratio L/τL/\tauL/τ says how hard the plant is to control: with a small ratio a tight loop is easy, with a ratio near one or above any aggressive controller oscillates. Classic tuning rules read the gains straight from these numbers (PID tuning).

Not every plant settles. A tank emptied by a pump at fixed flow, or a motor's position, keeps ramping under a constant input: an integrator, with a rate instead of a gain. A few plants first move the wrong way, as the water level in a boiler drum does when steam demand rises, and they punish fast controllers.

What the controller sees is more than the process. Between its output and its input sit the actuator, the process and the transmitter, each with its own lag, and a step test measures the whole chain. That is what tuning needs, and it is why a slow sensor makes a fast process look slow.

A step test has a cost: the process must be steady before it, and the step large enough to stand above the noise and small enough not to upset production. Where that is not allowed, models come from routine operating data or from physics.