Sixty million ids a day, every one observed

The plant, and what flows through it

Nine moulding lines — three each of forks, spoons and knives — each ending in a vision station that judges every piece. Thirty-two stackers take a fork, a spoon and a knife from the lines' output and close a stack that its own vision station judges; two stackers feed every wrapper because a stack takes twice as long as a wrap; sixteen wrappers seal a stack onto a plate; four palletizers take 240 wraps to a pallet. The customer's arithmetic is 60,041,666 ids a day: 30,000,000 pieces, 10,000,000 stacks, 10,000,000 plates, 10,000,000 wraps and 41,666 pallets.

How an id enters: one group notify, one event, one record

Nothing is minted from a count. The station is the source of truth: its vision system decides whether the piece is good, stamps the serial and the four readings with one source time, and publishes them as one OPC UA group. The agent subscribes to every inspection tag at full rate, assembles the tags that share a source time into one event, and writes it in bulk — the unit, its inspection row and, for a stack or a wrap, the containment of its members — with no HTTP in the path. A piece that fails never reaches a stack. OPC UA only notifies a value that changed, so a tag that repeats from one event to the next is filled from what the station last sent; a group whose serial never arrived is recorded as partial and counted, never dropped.

What the vision stations judged, and where the pieces went

Every piece carries four readings and a verdict; every stack and every wrap the same. The verdict is a pass word whose bits name the attribute that failed, so why a piece was scrapped is a query. The containment records say which marker's pieces went into which stacker's stacks, which stacker fed which wrapper, and what reached which pallet — and because a failing piece is never offered to a stacker, the number of stacks containing a failed piece has one right answer.

The dimensional checks, and the certificate a pallet leaves with

Every fifteen minutes an inspector measures five characteristics of each utensil through the same API a browser uses. A pallet's certificate of analysis states, for each characteristic, the checks inside the pallet's production window, the Cpk over them (sigma from the mean moving range) and whether the process was stable — and says no capability outright when the window holds fewer than the twelve points the method needs. Under that it lists every wrap, the stack and the plate in each, and every piece in each stack, rendered from the containment record and issued as an immutable document: 1,440 units under one pallet serial, gathered in 37.3 ms.

What the line did

The bookings say what each station made, the state intervals say what it was doing, and the scorecard says whether the MES saw the breakdowns the script put in — with a fault it recorded after its window withheld rather than scored missed.

For IT: the systems, the loads, and where they bend

Five processes and a database server stood in for a plant: an OPC UA server replaying the line's per-second truth and publishing the inspection groups; the OPC agent taking the groups, writing units and inspections in bulk, booking from counter deltas and recording states; PostgreSQL holding one database for the run; the API serving the floor, the screens and the certificates; and the operations loop being the people. Storage is a module: the same models and migrations run on one SQLite file for a small plant, and on PostgreSQL when a plant like this one switches it on.

Findings

What the simulator cannot say

• Its stations' vision readings are drawn around a nominal; a real station's distribution has the shape the process gives it. The method — limits, a pass word, Cpk over a window — does not depend on the shape.
• The OPC UA server is a Python replay. It publishes what the plant would at 1x; at higher speeds the emitted count against the expected count says how much of the load it could produce, and only what was published can be scored.

Provenance

The notebook that produced every figure, the runner that produced the evidence, and the plant description are in the repository under labs/cutlery/. Re-running the notebook re-derives the page.