SCADA and PLC
The plant floor already has the data. It just never leaves the plant.
Historians hold years of process data that no business system can reach, while planning runs on numbers typed into a spreadsheet at shift end. The gap between OT and IT is where most manufacturing analytics projects quietly die.
The short answer
Connect plant systems through OPC UA where equipment supports it, or through the historian's own interface where it does not — and always in one direction, out of the plant. Reading process data into enterprise software is a well-understood problem; writing back into control systems is a different discipline with safety implications and should not be in scope.
Interfaces
How it actually exposes data.
The interfaces, the pitfalls, and how we handle each.
OPC UA
The modern standard, and the right default where equipment supports it. Structured, secured and vendor-neutral, which matters on a floor with machines from six manufacturers bought across twenty years.
OPC DA via a gateway
Older equipment speaks the classic Windows-only protocol. A gateway bridges it rather than touching the machine — the correct move, since nothing on a production line should be modified to satisfy a reporting requirement.
Historian interfaces
PI, Wonderware and similar hold the time-series record and expose it through their own APIs. Usually the best source: already collecting, already trusted by the plant team, and reading from it disturbs nothing.
Weighbridge, MES and quality systems
Often the most commercially valuable and the least standardised — serial output, a local database, or a file drop. Each is bespoke and each carries a number the business genuinely needs.
What goes wrong
The failures that are not in the documentation.
Every one of these has cost somebody a release. They are the reason the design below looks the way it does.
Treating OT like IT
A control network is not a corporate network. Scanning it, patching into it, or adding unplanned load is unacceptable to a plant engineer for good reasons. Integration happens through a defined boundary with the plant team's agreement, or it does not happen.
Time-series volume underestimated
A few hundred tags at one-second resolution is billions of points a year. Designing this as ordinary relational data produces a system that is fine in pilot and unusable in month four.
Tag names that mean nothing
Historian tags are cryptic and their meaning lives with the engineer who configured them. Building the tag dictionary is a real project task with real plant-team time attached, and estimates that omit it are wrong.
Assuming clean, continuous data
Sensors drift, fail and get bypassed during maintenance. Analytics that assume good data produce confident nonsense, and the plant will notice before the board does.
Method
What we build in front of it.
- 01
Agree the boundary with the plant team first
What is read, from where, at what rate, and through which network path. Signed off by whoever is accountable for the line. Nothing else starts before this.
- 02
Build the tag dictionary
Tag to physical meaning to business metric. Unglamorous, and the thing that decides whether anything downstream can be trusted.
- 03
Land it in storage built for time series
With downsampling and retention decided up front, because retrofitting them means reprocessing everything already collected.
- 04
Join to business context
Process data becomes useful when it meets the work order, the batch, the shift and the material. That join is where OEE, yield and energy-per-tonne come from — and it is the deliverable, not the raw tags.
Questions
What people ask before starting.
Next step
Need to connect to SCADA and PLC?
Bring the constraint everyone has been working around. Forty-five minutes with an engineer who has done this before is usually enough to tell you whether it is a week or a quarter.
- No sales deck
- An engineer on the call, not an account manager
- NDA before you share anything