industrial//industrial data layer

The industrial data layer is the part of a plant's architecture that carries data from its machines to many consumers, and it turns a plant designed to control machines into an industrial data infrastructure. The plant went from one chain with one consumer, built to control, to one source with many consumers, built to exploit data as well. In the automation pyramid data climbs level by level; the data layer cuts across it, taking every signal at the edge and handing it sideways to whoever needs it, which is why modern architectures are drawn as a hub rather than a ladder.

The industrial data layer is the part of a plant's architecture that carries data from its machines to many consumers, and it turns a plant designed to control machines into an industrial data infrastructure. The plant went from one chain with one consumer, built to control, to one source with many consumers, built to exploit data as well. In the automation pyramid data climbs level by level; the data layer cuts across it, taking every signal at the edge and handing it sideways to whoever needs it, which is why modern architectures are drawn as a hub rather than a ladder.

Before1Sensor2PLC3SCADAone consumer

Now1Sensor2PLC3Edge gateway or IPC4MQTT or an OPC UA server5SCADA, the historian, MES, analytics, cloud and AIin parallel

The second lane is the typical IoT telemetry pipeline of a plant. A drone has its own version: telemetry travels by radio at a few hertz (MAVLink on open autopilots) and the full log stays on board.

The unit of data is the tag with value, timestamp and quality; the other kind of datum is the event. Storage is the historian with deadband compression, alongside a time-series database.

The data an analysis needs lives in systems never designed to talk to each other. Process variables sit in the historian, events and stops in the logs and alarms of the PLC and SCADA, the product and batch in the MES, the maintenance work orders (where the failure labels are) in the CMMS, and fleet telemetry arrives late, out of order or never. Building one training table means aligning all of it in time, and that alignment, more than any model, is where a predictive project spends its months.

The hard part is meaning rather than moving bytes: data contextualization is what turns a register value into motor M17, batch B562, customer X, and fault semantics is why control protocols cannot be swapped for a REST API. OPC solved interoperability first; OPC UA added the information model.

Several neighbouring notes widen the view: IIoT connects assets to get more from them, telemetry carries their variables out, fleet data compares many plants at once, and edge AI runs the analysis next to the machine. All of it lives on the boundary described in OT.