When a historian compresses a tag, it makes a choice about fidelity that most people never see stated plainly. Deadband and swinging-door compression are lossy - they discard data that can never be recovered - but in a very controlled way. This guide frames the mechanism the compression pages describe in terms of accuracy: how lossy process compression differs from true lossless storage, and why the compression deviation is really a guaranteed maximum reconstruction error.
Lossless vs Lossy Compression in one line: Lossless compression stores data so it can be reconstructed exactly, bit for bit, while lossy compression discards detail that cannot be recovered. Historian deadband and swinging-door compression are lossy, but they are bounded-error lossy: the compression deviation you set becomes a guaranteed maximum for how far any reconstructed value can differ from the original reading.
Compression in computing usually means lossless - a ZIP file, a database page compressor, an image format that reproduces every pixel. You shrink the storage, and when you decompress you get back exactly what you put in, with no loss of information. The savings come purely from encoding redundancy more efficiently. If a historian archived every raw sample and then packed the numbers with a lossless codec, that would be lossless compression: smaller files, identical data.
Process historian compression means something different. Deadband and swinging-door do not merely re-encode the samples; they decide that some samples are not worth storing and drop them permanently. When you query a time that was not stored, the historian reconstructs a value by interpolating between the points it kept. That reconstruction is an estimate, not the original reading, which makes the scheme lossy - the discarded samples are gone and can never be reproduced exactly.
The distinction matters because people often assume a historian's numbers are exact when they are actually reconstructed. A trend drawn from a compressed archive is a faithful approximation of the real signal, not a replay of every measurement the collector took. For most process data that is completely fine, but knowing it is lossy is essential the moment fidelity is questioned - for instance when a stored value is used as evidence in a measurement dispute.
What saves process compression from being sloppy is that its loss is bounded, not arbitrary. When you set a compression deviation - the deadband or swinging-door tolerance - you are not just telling the historian how aggressively to compress. You are setting a promise: the algorithm will only discard a sample if the value reconstructed for that moment stays within the deviation of the real reading. In effect the deviation is a guaranteed maximum reconstruction error for the tag.
This is what separates intelligent process compression from blindly throwing samples away. The swinging-door algorithm keeps a point precisely when dropping it would let the reconstructed line drift outside the deviation, so the error bound holds across ramps and curves, not just flat sections. If you set the deviation to a value you consider engineering-significant for that measurement, you have a mathematical assurance that no stored value misrepresents reality by more than that amount.
That guarantee is only as good as the deviation you choose, and choosing it is an accuracy decision. Set it too wide to save storage and you loosen the error bound until real events blur; set it tighter than the instrument's own uncertainty and you store noise while gaining no real fidelity, because the extra precision is below what the sensor can actually resolve. The sensible target is a deviation a little above the instrument's noise floor - small enough to preserve every meaningful change, large enough not to archive jitter.
For oil and gas monitoring, the bounded-error framing is what lets operators trust a compressed archive for real decisions. A cloud SCADA platform storing years of data from hundreds of remote sites cannot keep every raw sample forever, so it compresses - but if the deviation is set per tag to an engineering-significant tolerance, the trends and reports built from that archive carry a known accuracy guarantee rather than an unknown one. That is the difference between compression as a storage trick and compression as a controlled measurement decision.
It also clarifies when lossy compression is inappropriate. Some values must be exact and cannot tolerate any reconstruction error - a discrete alarm state, an event sequence, a legally significant custody counter. These are stored losslessly or on true change, without a deviation band, because for them there is no acceptable error to bound. A mature platform applies lossy bounded-error compression to continuous analog process data and lossless storage to discrete and event data, matching the method to what each tag actually is.
Merobix compresses continuous process tags with a bounded-error deviation set per tag, so every retained value honors a known maximum error, while discrete states and events are kept exactly. For an operator, that means a Merobix trend of a remote wellhead pressure is a faithful reconstruction within a tolerance you chose, and an alarm or a custody count is stored precisely - the right fidelity model for each kind of data.
It is lossy. Deadband and swinging-door compression permanently discard samples the historian judges insignificant, and reconstruct in-between values by interpolating the points it kept, so a queried value is an estimate rather than the original reading. It differs from an arbitrary lossy scheme because the loss is bounded: the compression deviation guarantees a maximum error, so no reconstructed value strays further than that from reality.
It sets the maximum reconstruction error for the tag. The algorithm only drops a sample if the value reconstructed for that moment stays within the deviation of the real reading, and it keeps a point precisely when dropping it would exceed that bound. So if you set the deviation to an engineering-significant tolerance, you have a mathematical assurance that no stored value misrepresents the true measurement by more than that amount.
When any reconstruction error is unacceptable. Discrete alarm states, sequence-of-events records, and legally significant custody counters must be exact, so they are stored losslessly or on true change rather than with a deviation band. Bounded-error lossy compression is appropriate for continuous analog process data, where a small, controlled error is fine, but not for values that must be reproduced precisely.
Merobix reads your field devices into a cloud SCADA - the real thing behind these terms, live in days from any browser.