Derivative action can sharpen a slow control loop, but on any real, noisy measurement it is almost useless without a filter - because derivative math amplifies noise viciously. Derivative filtering is the practical fix: a filter applied to the measurement or to the derivative term that smooths out the jitter so derivative action responds to genuine trends rather than sensor noise. This guide explains why raw derivative and noise are a terrible combination, how the filter tames it, and how to set the filter without killing the very action you wanted.
Derivative Filtering in one line: Derivative filtering is the application of a low-pass filter to the process value or the derivative term of a PID controller so that the derivative action responds to real rate-of-change in the process rather than to high-frequency measurement noise; because differentiation amplifies noise dramatically, some form of derivative filter is essential for making derivative action usable on noisy flow and pressure signals.
The derivative term acts on the rate of change of the measurement - how fast it is moving, not where it is. That is exactly what makes it dangerous on a noisy signal. Measurement noise is small in amplitude but extremely fast, flickering up and down many times a second, so its rate of change is enormous. A slow, meaningful process trend might have a gentle slope, but the noise riding on top has steep, constantly reversing slopes. Differentiation seizes on those steep slopes and multiplies them, so a tiny bit of noise on the measurement becomes a large, thrashing contribution to the controller output.
The result on an unfiltered loop is a valve that buzzes and hammers, driven not by the process but by the sensor's electrical hash. Flow and pressure signals are the usual offenders - turbulent flow through a meter and pulsating pressure are inherently noisy - which is why derivative action is often left off flow loops entirely. Without filtering, adding derivative to a noisy loop makes control worse, not better: the noise amplification swamps whatever anticipatory benefit the derivative term was supposed to provide, and the actuator pays the price in wear.
The remedy is a low-pass filter that lets slow, real movements through while attenuating the fast noise before or during differentiation. It can be applied to the raw measurement (a PV filter, which also benefits the proportional and integral terms) or built specifically into the derivative path - most industrial PID implementations include a derivative filter as a standard, non-optional part of the derivative block precisely because raw derivative is unworkable. The filter effectively limits how much the controller can amplify high-frequency content, capping the derivative's noise gain so that only genuine trends survive to influence the output.
Two related refinements usually travel with derivative filtering. First, derivative action is commonly taken on the measurement rather than on the error, which eliminates derivative kick - the sharp output spike that would otherwise occur the instant an operator changes the setpoint, since a setpoint step has an effectively infinite rate of change. Second, the filter is characterized by a time constant that sets its cutoff. Together these make derivative action civilized: it anticipates real process moves and rejects both setpoint steps and sensor noise, instead of reacting violently to both.
Setting the filter is a balance. Too little filtering and the noise still bleeds through, thrashing the valve; too much and the filter lags the real signal, delaying the loop's reaction and stealing back the anticipation the derivative was meant to add - the filter itself becomes a source of sluggishness. The filter time constant is usually tied to the derivative time, kept large enough to smother the noise but small enough that genuine process dynamics pass through with little delay. On badly noisy loops the honest conclusion is often to drop derivative action altogether and run PI, rather than fighting a losing battle with the filter.
A cloud SCADA platform is where the noise problem becomes measurable rather than a matter of feel. Merobix trends the raw process value at fine resolution, so an engineer can look at a flow or pressure tag and see directly how much noise it carries, how fast, and whether it is getting worse - a fouling sensor or a developing leak often shows up first as rising noise. That evidence guides the decision of whether derivative action is even viable on a given loop and how aggressively to filter. Merobix does not compute the derivative or hold the filter - that lives in the controller - but by exposing the character of the signal over long horizons it turns filter tuning from guesswork into a decision grounded in what the measurement actually looks like.
Because differentiation responds to rate of change, and measurement noise - though small - changes very fast, giving it a huge rate of change. Unfiltered derivative amplifies that noise into a thrashing output that hammers the valve. A low-pass filter removes the fast noise while letting real process trends through, so derivative action responds to genuine movement rather than sensor jitter.
Usually not. Flow signals are inherently noisy from turbulence and pulsation, and derivative action amplifies that noise badly even with filtering. Most flow loops are run as PI without derivative for this reason. Derivative is more useful on slow, cleaner loops like temperature, where the anticipation it provides outweighs the noise it introduces.
Derivative kick is a sudden output spike that occurs when the setpoint is stepped, because a step has an effectively infinite rate of change that the derivative term seizes on. It is avoided by taking derivative action on the measurement rather than on the error, so setpoint changes no longer feed the derivative term. This is standard practice and pairs with filtering to make derivative usable.
Merobix reads your field devices into a cloud SCADA - the real thing behind these terms, live in days from any browser.