industrial//industrial data layer//deadband

A deadband is a threshold under which a change is not worth reporting: while a value stays within the band around the last recorded sample, no new sample or event is generated; when it crosses the threshold, one is. The small graph with a threshold line in any historian manual is exactly this.

A deadband is a threshold under which a change is not worth reporting: while a value stays within the band around the last recorded sample, no new sample or event is generated; when it crosses the threshold, one is. The small graph with a threshold line in any historian manual is exactly this.

It reduces three things at once: traffic on the network, storage in the historian, and noise in the data. It is the reason a historian can keep years of a signal that changes by a thousandth of a degree. The same idea appears on the wire as report by exception (a device sends only when something changed) and in OPC UA subscriptions as a deadband filter. The trade-off is fidelity: a band set too wide erases the small oscillation a diagnostic later needed. It is easy to take for a smoothing filter such as an EWMA, but smoothing works on a stored series and can always be undone by keeping the raw data, while a deadband decides which samples exist at all and cannot be undone, so it has to be set with the future question in mind.

Information theory says why the saving is so large. A stream that rarely changes carries little entropy: a detector that raises an alarm on 0.1% of its readings carries about 0.011 bits per reading, so almost every sample repeats what the receiver already knew. Sending only the changes, event-based telemetry, transmits close to the information and drops the repetition. A signal that is pure noise has nothing to drop, and a deadband narrower than that noise sends nearly every sample anyway.

What a reader sees between stored points was never measured. A historian or a dashboard asked for the value at an instant with no sample interpolates, usually in a straight line, between the two stored points around it, and a transient that rose and fell inside the band leaves no trace at all. The interpolated curve looks like data; it is the deadband compression artefact, and a model trained on it learns a smoother plant than the real one (industrial labels and their features suffer the same way).

The timestamps become irregular: a quiet tag stores one point an hour, a busy one every second. Any analysis that assumes a regular grid has to resample first, and has to choose how, knowing that each choice invents values.