The new raw material of the plant

People talk about artificial intelligence, digital twins, predictive maintenance. But they all rest on one thing: data. A plant produces colossal amounts of it — every sensor emits a value several times a second. The problem is not having it, it is having good data, well organised and understandable. Before any AI project, you must understand industrial data.

Data born in the field

Industrial data comes from the process: a temperature transmitter, a meter, a valve state, a controller fault. It rises through the controllers (PLC, DCS, SCADA), often via the standard OPC UA protocol that unifies access to equipment. It is real-time, raw data, expressed in the sensor’s unit — and not always reliable: a sensor drifts, clogs, fails.

Time series and the historian

Most industrial data is a time series: a sequence of “timestamp + value” pairs. It is archived in a specialised database, the historian, optimised to store millions of points and replay them fast. It is the plant’s memory: it lets you review a drift, analyse an incident, compare two production campaigns. The volumes are huge: a single sensor at 1 Hz generates 86,400 points per day; a plant with 10,000 tags produces ~1 billion per day. Hence the compression (often 10–20×) and adaptive sampling typical of historians.

Contextualising: giving the measurement meaning

A value alone — “73.4” — means nothing. You must know what it refers to (which equipment), in what unit, in what context (which product, which batch, which operating mode). That is contextualisation: linking raw data to the model of the plant. Without it, analysis is impossible. It is often the longest and most underestimated step of a data project, and it is the role of an MES to supply that production context.

OT and IT: two worlds of data

Field data (OT) and business data (IT) have neither the same pace nor the same priorities. OT favours real time and availability; IT favours consistency and analysis. Making the two talk — without opening a security breach, see the Purdue model — is one of the great challenges of the connected plant and of the IIoT.

From data to decision

The ultimate goal is not to pile up points, but to inform a decision: a performance indicator, a drift alert, an optimisation. The chain is always the same: measure, transport, archive, contextualise, analyse, decide. Artificial intelligence only comes at the end — and is never better than the quality of the data feeding it.