Seven production lines publish live data here. This page is the contract they honour — the OPC-UA address space, the MQTT topic namespace, the command channel — and what it takes to point the same pipeline at a real machine instead of a simulated one.
Data leaves the machine and never comes back the same way. Commands travel on their own channel, and the dashboard never opens a connection to a machine.
A station with a state, a cycle, measurements and faults. Here it is simulated tick by tick; the interface is what matters.
One OPC-UA server per line for SCADA-style clients, and an MQTT publisher for the event stream. Same source, two shapes.
Broker on :1883. A topic filter decides which lines are forwarded — a new line needs its prefix added there.
Flows normalise each payload and persist it. This is where the historian lives, ~2.1 M rows a day.
Reads the database and the live topics. Read-only by design: it renders state, it does not own it.
Why two protocols and not one. OPC-UA answers the question a SCADA client asks — “what is the value of this tag right now, and may I write it?”. MQTT answers the question a data platform asks — “tell me when something happens, and let ten consumers hear it at once”. A station that only spoke one of them would force the other side into polling.
Each line runs its own endpoint, so a client can subscribe to a single line without seeing the rest of the plant.
Address space
Writes are real. Writing CycleTime or SpeedMultiplier on a
station changes the running model immediately and the change shows up on the next MQTT publish — the same way a
set-point write reaches a real machine. Every other node is read-only: a dashboard has no business writing counters.
Three levels: the station, the line, and the plant. Payloads are JSON, published on QoS 0 with no retain — the historian is the memory, not the broker.
| Topic | When | Carries |
|---|---|---|
| {line}/{station}/status | on state change | state, cycle time, parts produced, fault code |
| {line}/{station}/measurements | per part | one value per characteristic, with nominal and limits |
| {line}/{station}/transaction | per event | operationStart / operationComplete / scrap / rework, with the serial |
| {line}/line/buffer | on change | marriage-buffer level, min, max, status |
| {line}/line/summary | every 5 s | output, OEE, FPY, bottleneck |
| mes/{line}/state | every 5 s | line running or stopped, per-station snapshot |
| mes/alerts | on event | breakdown, faultCleared, quality, maintenance, starving, blocked |
| mes/wip | every 5 s | work in progress per station: on board, in queue, queue capacity |
| mes/materials/status | every 5 s | warehouse racks, line stores, AGV missions, shipping |
| mes/operators/state | every 5 s | crews, shift, breaks, per-operator role and state |
Event-driven, with a heartbeat
State changes, measurements, transactions and alerts are published the moment they happen — nothing
is sampled. On top of that a summary tick runs every PUBLISH_INTERVAL_MS (5 s by default) so a consumer
that connects late gets a complete picture without asking. The client reconnects on its own every 5 s if the broker
disappears.
Commands arrive on mes/commands/{verb} as JSON. Anything that changes the plant answers on
mes/commands/response; anything that asks a question answers on the topic that owns the data.
One boot-time caveat. The bridge subscribes to the command topics about 30 seconds after the process starts, once the lines are seeded. A command published before that is lost — there is no queue in front of it, deliberately: a machine that replays an hour of stale set-points on restart is worse than one that ignores them.
Two things every event-driven integration gets wrong the first time, both fixed here at boot.
{line}/{station}/status is only published on a state change. A station that has not
moved since the last process died keeps that process's value on the broker — one station showed FAULT for
25 minutes while sitting idle. Boot now publishes one snapshot per station, once.
Downtime is reconstructed by pairing a breakdown alert with the next
faultCleared. A process that dies mid-fault never sends the clear, so the stoppage is reported as
ongoing forever and the downtime chart grows without limit. A restart genuinely ends every fault, so boot emits
one clear per station; unmatched clears are ignored by the pairing logic.
Nothing above the broker knows that a simulator is producing this data. Replacing it is an integration job, not a rewrite.
Modern controllers expose a server natively. Point a collector at it, map its nodes onto the station model above, and publish the same topics. This is the shortest path and the one to prefer.
S7, Modbus, EtherNet/IP — a gateway (Kepware and equivalents) lifts the tags to OPC-UA. The mapping table becomes the contract, and it is where the project lives or dies.
An edge box publishing JSON straight onto the namespace skips OPC-UA entirely. Cheapest to build, and the topic layout above is the whole specification.
What has to survive the translation
| Element | Why it is not negotiable |
|---|---|
| station id | The id is the join key across telemetry, quality, downtime and the traceability ledger. It must be the same string everywhere, forever. |
| state vocabulary | Downtime, OEE and andon are derived from an ISA-88 state machine. A machine that reports its own words needs a mapping table, not a new state. |
| units + limits | A measurement without its nominal and its two limits is a number, not a characteristic — SPC and first-pass yield both fall over. |
| the serial | Genealogy is the point. An e-axle carries the serial of its rotor, its reducer and its inverter; drop it at one station and the thread breaks for the whole unit. |
| timestamps | Event time, in UTC, from the source. The broker's arrival time is not the event time, and the difference shows up as impossible cycle times. |
The same transparency rule as everywhere else on this site: what is missing is written down, not hidden.
| Not here | What production needs |
|---|---|
| endpoint security | The OPC-UA endpoints accept anonymous connections and the broker is open on the host. Production means certificates, user tokens, a signed and encrypted policy, and TLS on 8883. |
| delivery guarantees | Everything publishes at QoS 0 with no retain. A metering or genealogy feed needs QoS 1 with a persistent session, or an outbox on the edge. |
| store and forward | If the broker is down, those messages are gone. A real edge buffers to disk and replays with the original timestamps. |
| write-back to the plant | The dashboard is read-only by design. Commands here drive a simulation; driving a machine is a safety conversation, not a feature. |
| schema versioning | Payloads are plain JSON with no version field. Anything long-lived wants a schema and a migration path — Sparkplug B if you want it standardised. |
The architecture is the deliverable. Every limit above is a configuration or a component away — none of them changes the shape of the contract on this page. That is the whole reason the contract comes first.
The lines on this site are publishing right now, on exactly the topics documented above.