Where Data Is Born: The Birth Certificate Your Chain Inherits

Ask where plant data comes from and most diagrams answer “sensors”.

A sensor produces a signal. It does not produce data. Data is born in whatever device first samples that signal, attaches a timestamp and a name, and holds the result in memory: a PLC scanning its inputs, an RTU logging a remote well, a power meter accumulating registers, a vibration module computing spectra, a gateway polling them all.

That birthplace decides properties the data carries for life — resolution, rate, timestamp meaning, quality flags, buffering behaviour. And the hard part, the part that makes this the least reversible section of an architecture: no downstream layer can add what the birthplace never recorded.

Architecture reviews that start at the historian have already skipped the decisions that matter most.

Five birthplaces, five sets of habits

The acquisition layer is every device that turns signals into named, timestamped values. In a real plant there are five families of them, and each arrives with its own strengths and its own caution.

The controller is the dominant birthplace in most plants, and the convenience is real: it already scans the I/O that matters for control, holds it in addressable memory, and exposes it by protocol. Two cautions come attached.

The first is capacity. The controller’s job is control; its data-serving capability is a budget, and that budget belongs to the automation owner.

The second is subtler and costs projects more. The controller knows only what control needed. A census may find four hundred analogue points wired, but how many are scanned at rates useful for the new decision, with deadbands set for trending rather than for control loop stability, is a per-point question the project has to actually ask. Points are frequently present, connected, and useless for the purpose being proposed — not because anything is broken, but because they were configured correctly for a different job.

The certificate the chain inherits

For each birthplace, six properties are fixed at the moment of birth and inherited by everything downstream.

• → Rate: how often the device looks at the signal. Everything between looks is gone, and fast phenomena sampled slowly do not vanish — they alias into slow, plausible-looking artefacts.
• → Threshold: what change was large enough to be recorded. A drift smaller than the deadband is invisible to the entire rest of the chain.
• → Timestamp origin: whether the stamp records the observation or its receipt, and the gap between the two ranges from milliseconds to hours after a buffered outage.
• → Clock: which time source produced the stamp, and how far that clock might be from true.
• → Quality: whether a status or quality code exists at all, and travels with the value or is dropped at the first hop.
• → Buffering: what the device does when the link fails — hold, drop oldest, drop newest, refuse — and whether the drained data still carries its original time.

None of these can be repaired downstream. A historian cannot recover a pulse the acquisition layer sampled instead of counted. A lake cannot restore a quality flag the gateway discarded. An analytics platform cannot recover the excursion that fell inside a deadband set in 2011.

This is why the map of birthplaces is the artefact worth producing first. It is also, reliably, more capable and more constrained than anyone assumed — capable because IEDs and smart transmitters hold data nobody knew was there, constrained because the rates and thresholds serving control are rarely the ones a new decision needs.

Reading without breaking

Every birthplace raises the same governance question: the new data path reads from a device whose first duty is operations. Three disciplines keep the read honest.

1. 01
   Capacity, measured before and after

   Every poll consumes controller CPU, network segment budget or serial bandwidth that control traffic also needs. The project measures the before-state, presents its load against a ceiling the site's automation owner sets, and measures the after-state at commissioning.

2. 02
   Change control, without exception

   Enabling a server, opening a port, adding a routing entry or upgrading firmware to obtain a feature are modifications to an operating control system. They pass through the site's management-of-change and security processes — never through a data project's enthusiasm or its deadline.

3. 03
   Ownership, by name

   Each read path gets a named owner on the OT side. When the controller CPU alarms at two in the morning, 'the data project' is not a phone number, and the person who answers needs to know the path exists before they meet it.

The architectural instinct that follows is short: read once, close to the birthplace, and fan out downstream.

Five systems each polling the same PLC quintuples the load and creates five versions of the truth — five sets of scaling assumptions, five timestamp conventions, five outage behaviours, and five separate arguments the next time two reports disagree. One governed collection path feeding a shared layer does neither.

Getting there in a brownfield plant is a journey most sites will not finish. But every new connection can at least refuse to make the spaghetti worse, and that refusal is free.

An exercise that costs an afternoon

Pick five points that already feed a screen someone looks at. Not a system — five points.

For each, walk backwards to the device that first named and stamped the value, and fill the six certificate fields from configuration rather than from memory. Screenshots of the actual settings, not recollections of what they should be.

Then ask two questions of the result.

Where did the timestamp come from? If it was applied by a poller or a gateway rather than at source, write that down — not as a defect, but as a property consumers need to know. A stamp applied at collection answers when we knew, and the trend that assumes it answers when it happened is quietly wrong by the polling interval, or by the length of the last outage.

Was any value re-stamped in transit? Test it deliberately: compare a source stamp at the device with the stamp on the same value after conversion. Gateways that replace source time with “now” are common, and silent, and the test takes an hour.

What usually turns up is a mixed estate — some device families stamping at source, some stamped by pollers, one or two gateways quietly re-stamping, and no monitoring on any of it. The remediation is typically a change-controlled configuration pass and an alert rule. The alternative is discovering it during an investigation, at the worst possible hour, when the data is all you have.

Which of your five points can name the device that gave it its timestamp?

The five acquisition birthplaces, the birth-certificate framing, the access disciplines of capacity, change control and ownership, and the read-once instinct are from Industrial IoT and Data Architecture: From Sensor to Historian to Dashboard (Part 2: Edge and Acquisition, section 2.1).