What Is Deadband Compression?

Back to Blog

Simple Deadband vs Swinging Door

The simplest form is exception deadband: the historian compares each new reading to the last stored value and archives the new one only if it differs by more than the deadband. This is easy and effective for flat signals, but it can be crude on a ramping signal - a steadily rising level triggers a stored point at every deadband step, and a slow drift can be missed if each step stays just under the threshold.

Better historians use a swinging-door algorithm (a technique long used in process historians). Instead of comparing to the last stored point, swinging door checks whether the intermediate points can be reconstructed by a straight line within the deadband. As long as every skipped point lies within a tolerance corridor of a line drawn from the last stored point, they are dropped; when a point falls outside the corridor, the previous point is archived and a new corridor begins. This captures trends and slopes with far fewer stored points than simple deadband, while guaranteeing the reconstructed value never deviates from the real one by more than the deadband.

Choosing the Deadband

The deadband is usually set per tag, in engineering units or as a percentage of range, and it directly trades storage against fidelity. Too wide and you erase real movement - a small but meaningful pressure excursion vanishes into the archive as a straight line. Too narrow and you store instrument noise, bloating the historian and slowing queries with no analytical benefit.

A sound approach is to base the deadband on instrument accuracy and process significance: set it just above the sensor's noise floor and below the smallest change you would actually act on. Many teams also distinguish a collection or exception deadband (what the field device or interface reports) from an archive compression deadband (what the historian keeps), so noise is filtered close to the source and only meaningful changes are both sent and stored.

Deadband Compression in Practice

Deadband compression is lossy by design - it discards data you have declared insignificant - so it must be configured deliberately, especially for tags used in custody transfer, allocation, or regulatory reporting, where you may want tighter deadbands or full-resolution collection. For most monitoring tags, though, sensible compression is what makes years of history affordable and trends fast to load.

Note the close relationship to report by exception: RBE applies a deadband at the field device to save bandwidth, while historian compression applies a deadband at the archive to save storage - the same idea at two points in the pipeline. Merobix stores time-series history for every monitored tag with configurable compression, so operators keep meaningful history at full fidelity without archiving instrument noise.

Checking What Compression Actually Kept

Compression settings should be verified, not trusted, because the failure mode is silent: the archive simply lacks the detail you assumed it had, and you discover it during the incident review that needed it. The direct check is to collect a short window of uncompressed data for a handful of representative tags - a noisy one, a slow-moving one, a fast-cycling one - overlay the archived series on the raw series, and confirm the reconstruction is faithful enough for how each tag is actually used.

The ongoing check is the per-tag historian compression ratio, which makes outliers visible in both directions. A tag storing nearly every sample is either genuinely noisy, carrying a deadband far below its noise floor, or scaled wrong. A tag storing almost nothing is either genuinely flat, carrying a deadband that swallows real movement, or coming from a dead instrument - and the archive alone cannot tell you which, so a near-zero storage rate on a tag that should move is worth a field-side look.

How Compression Interacts with Queries and Aggregates

A compressed archive stores points where the signal changed, which means stored samples are irregularly spaced and cluster around activity. Any calculation that averages stored samples arithmetically inherits a bias toward the busy periods, because that is where the points are. The correct tool over compressed data is a time-weighted average, which weights each value by how long it persisted rather than by how many samples happened to be stored.

Retrieval has the mirror consideration: between stored points, the historian reconstructs the signal, typically by interpolation, and the guarantee behind that reconstruction is exactly the deadband - the true value never differed from the reconstructed line by more than the configured tolerance. That guarantee is what makes compression defensible, and it is also why the deadband must be chosen against how the data will be queried, not just how much disk it saves.

Tags That Deserve Special Treatment

A single site-wide compression policy is a smell; signal types differ too much. A short taxonomy:

Tag type	Sensible handling
Noisy analog (pressures, flows)	Deadband just above the instrument noise floor; swinging door where available
Digital and state tags	Record every transition; a deadband has no meaning on a two-state signal
Totalizers and counters	Very narrow or no compression; a monotonic ramp compresses well anyway and gaps corrupt accounting
Setpoints and operator entries	Record every change; each one is an event, not noise
Custody and regulatory tags	Full resolution or the tightest deadband the contract or rule allows

The pattern behind the table is that compression is a statement about which changes are insignificant, and for event-like signals - transitions, commands, entries - no change is insignificant.

A symbolic example makes the algorithm choice concrete. Take a level tag with span S and a deadband set to a small fraction of S. A flat signal stores almost nothing under either algorithm. A steady ramp is where they part: simple exception deadband archives a point every time the level climbs one deadband step, so the number of stored points grows with the size of the excursion, while swinging door represents the entire ramp with its endpoints because every skipped sample lies within the tolerance corridor of a straight line. Same deadband, same fidelity guarantee, dramatically fewer points - which is why ramp-heavy signals like tank levels are where swinging door pays for itself.

Frequently Asked Questions