industrial//industrial data layer//historian

A **process historian** is a time-series database specialized in industrial process data, with compression, holding millions or billions of samples of tag, timestamp, value and quality. The historic products are OSIsoft PI (now AVEVA PI) and AspenTech IP.21.

A process historian is a time-series database specialized in industrial process data, with compression, holding millions or billions of samples of tag, timestamp, value and quality. The historic products are OSIsoft PI (now AVEVA PI) and AspenTech IP.21.

Its classic trait is compression: it does not necessarily store every sample when the value barely changes. A temperature reading 100.000, 100.001, 100.002, 100.000, 100.001 is five nearly identical points; a deadband keeps the shape of the signal with a fraction of the storage. The best known algorithm, swinging door, stores a new point only when the value leaves a band around the straight line it has been drawing since the last stored point, so a slow ramp costs two points. The results are irregular timestamps, values interpolated when read, and fine detail lost for good.

It is the archive behind every trend on a SCADA screen and the raw material of any later analysis. It coexists with, rather than being replaced by, the modern time-series database: historian, data lake or cloud, TSDB and SQL each keep a role in a real plant. The historian was built for a plant's tag namespace, deadband compression and decades of retention, the TSDB for labels, SQL joins and observability-style queries, and the interesting question is which one the MES reads from.

Compression is configured per tag, and some tags must not be compressed at all: a regulated batch record, an energy meter used for billing or a safety-relevant trip signal may have to be kept sample by sample for audit, which is a decision for quality and compliance, never a storage default.

What it stores is the tag; what it usually lacks is the business context that makes a tag comparable across lines or plants (data contextualization).

A historian keeps what happened, more or less, when it saw fit.

Compression ate the transients a detector was meant to find, interpolation invented values in the gaps, the timestamp is often the time of storage rather than the event time, and halfway through the series a sensor was replaced along with its bias. Before training a model on it, ask for the raw data of a short period and compare.

The trap is quiet because the data look clean. A model trained on compressed history learns a smoother plant than the real one, and then meets raw signals in the field, which is one reason predictive projects ask the edge gateway for features or raw captures instead (industrial labels, time synchronization).