Measurement dead time

Why can moving a thermometer one metre further along the pipe turn a rock-steady control loop into an oscillating one?

A temperature loop has run without complaint for years. During a shutdown the thermowell is moved one metre further down the pipe — a better tapping, easier to reach. Controller settings untouched, same sensor, same heater. The loop now hunts, swinging above and below setpoint indefinitely.

Nothing was made less accurate. The thermometer still reads correctly; it simply reads correctly a little later. So how can lateness alone, with no error in the measurement and no change in any gain, destroy a loop that was stable?

The companion explanation here on feedback control ends by pointing at this: delay, not gain, is the real enemy, and gain only decides how badly the delay hurts. That claim is left qualitative there. Here is the arithmetic underneath it.

First, be precise about what was added. Moving the sensor one metre downstream in fluid travelling at half a metre per second means the fluid takes two seconds to carry its temperature there. For those two seconds the thermometer reports the old temperature — not a blurred version of the new one, the old one, exactly — and then the new one. That is dead time, a different animal from a lag: a thermowell's thermal mass makes the reading move immediately and approach the truth gradually, where dead time gives nothing at all, then everything.

Why does that distinction decide stability? Consider a negative feedback loop at some frequency of oscillation. It measures an error and pushes back. If the loop's own dynamics delay that push by half a cycle — 180 degrees of phase — the correction arrives just as the error reverses sign, and the controller now adds to the disturbance rather than opposing it. If the gain around that circuit is at least one there, the swing grows. That is the whole stability criterion.

Now the crucial property. Nearly every element that adds phase lag also attenuates: a thermowell's lag delays the signal and shrinks it, so the further it pushes the phase the less gain remains to sustain an oscillation. That is self-limiting. Dead time is not. A pure delay of theta seconds passes every frequency at full amplitude and shifts each by omega times theta radians — lag in exact proportion to frequency, no attenuation whatever. It hands the loop phase for free.

Put numbers on our metre. Two seconds of delay contributes omega times two radians of lag, or 114.6 times omega in degrees. If this loop crosses over — reaches unity gain around the circuit — at 0.05 radians per second, as a large stirred vessel might, the new delay costs 5.7 degrees of phase margin. Nothing. If it crosses over at 0.5, the same two seconds costs 57.3 degrees, more margin than most loops are tuned to hold. The metre did not change; the loop it was installed in did.

Follow that to its conclusion, because it is a ceiling rather than an inconvenience. At a frequency of pi divided by theta, the dead time alone supplies the entire 180 degrees, before process, valve or controller contributes anything. No gain choice pushes crossover past that, and in practice it must sit well below, since everything else also lags.

So which constraint actually binds? Not the delay, oddly — the siting. The obvious response is to put the sensor close to the heater. But a thermometer at the heater measures the heater, not the process: the fluid there is unmixed and unrepresentative, and a loop controlling that number holds the wrong thing steady. Move downstream for an honest, well-mixed reading and you buy representativeness with delay. That trade has no clean solution, and underneath it sits a wholly non-physical constraint: the tapping points were drawn on a piping isometric long before the control engineer arrived.

The accepted failure follows. Faced with dead time that cannot be removed, the standard answer is to detune — lower the gain, lengthen the integral time — accepting larger, longer excursions after every disturbance in exchange for not oscillating. Sluggish regulation is chosen deliberately. A dead-time compensator predicting what the delayed measurement is about to say buys some bandwidth back, but depends on knowing theta: get the delay wrong and it degrades sharply.

The load-bearing claim is that a pure transport delay contributes phase lag proportional to frequency while attenuating nothing, so it consumes stability margin without any of the self-limiting behaviour a lag has. That is what makes lateness, on its own, destabilising.

The claim is falsifiable and the test is routine. Drive the loop into a controlled oscillation — a relay test does this deliberately — record its period, then insert additional transport delay and repeat. The account predicts the period lengthens in step with the added delay, since the frequency at which total lag reaches 180 degrees has moved down. It also predicts an asymmetry: adding a first-order lag whose time constant equals the delay should cost markedly less stability, because that element attenuates as it lags. The refuting observation would be a loop indifferent to added transport delay, or one where an equal lag and an equal dead time did equal damage.

Two qualifications. Real loops rarely contain pure delay in isolation — a controller's scan interval, an analyser's cycle time and a thermowell's thermal mass are three different things, and only the first two are honestly dead time. And 180 degrees with unity gain is the criterion for a linear loop; a valve that sticks or saturates can sustain a cycle no phase argument predicts, so an oscillating loop is not automatically a dead-time problem.

• How a controller's scan interval and a sampled analyser add dead time that no pipework change can recover.

Every term the collection defines is gathered in the glossary.