The problem

Every reading a plant needs is already being taken. What is missing is the sentence that says what each one is, and it lives in people, binders and old drawings.

Every reading a plant needs is already being taken. What is missing is the sentence that says what each one is.

And that sentence is not in the data. It is in a person who is on shift, a binder in a cabinet, and a drawing from the year the line was built.

297 tags per plant
5 systems that disagree
5 clocks that disagree
one person who knows

An operator station

31TC0149.SP

An ISA loop tag. The letters say a temperature, the number says nothing at all, and the description field on this one is empty. Which of the twenty-two cells is it on?

A control centre

AI:1064

An RTU point index. Not a name, an address. Everything about what it measures and what it is attached to lives in a spreadsheet somebody keeps.

A laboratory testbed

P1_B2004

From a published testbed. It is a pressure setpoint on a heat exchanger, and the only place that sentence exists is page 13 of a manual.

1

A working plant measures everything. Tens of thousands of points, every one of them logged, most of them for decades.

2

What each point is was written down once, by a contractor, in a naming scheme that made sense to them, and then the plant changed and the contractor left.

3

So every project that wants to use that data starts by asking a person what the tags mean. Every dashboard. Every model. Every agent. The same question, from the beginning, each time.

4

That person is the bottleneck, and they are on shift. It is why industrial AI stalls, and it is not the models.

5

The tools that move this data leave the meaning to a person. The best of them will match tags to an asset register you already have. Most plants do not have one.

This is not rhetoric, and it cuts both ways. Pointed at a physical laboratory testbed with no connector written for it, Auge placed 0 of 86 tags. It refused every one rather than guessing, and so would anything else, including a person who has not read the manual. The meaning is not in the data. Getting it out of the binder and the drawing and the head of the man on shift, and doing it without inventing anything, is the whole product.

One sentence in here has no competitor. We checked, and it cost us two of the four claims we had written.

1 We tell you which of the tags nobody can name are moving right now.

Every tool in this category can show you a list of unmatched tags; one of them has a checkbox for it. Nobody crosses that list with liveness. An unmatched tag sitting still is a backlog item. An unmatched tag that changed value in the last hour is a fault nobody can be told about, by us or by anyone else. In our own aluminium works 7 of 30 unidentified tags are moving, and two of them are a rectifier's cooling circuit failing, flow falling 70 m³/h, arriving as 31TI0365.PV and 31FT0368 with nothing on site saying what they are attached to.

2An answer from your engineer places more than the tag it is about, and stays on your side of the fence.

Answering the refusal list is the first afternoon. On our aluminium works, 104 answers placed 143 tags: 32 were placed without anybody asking, because naming one machine tells the compiler which loop numbers belong to it, and all 32 were right. 0 correct placements were broken by an answer, on either sector. On the grid substation it adds nothing, 1.01 per answer, because a bare RTU index carries no structure, and we say so. The answers are a text file the plant owns. A tool that has not compiled the plant has nothing to turn an answer into.

3 We read structure out of the wiring, not out of the names.

Recovering a hierarchy from tag names is a crowded field and we should not pretend otherwise: it has been sold as regular expressions for years and now ships with a model behind it. This is a different signal. A potline is a series circuit, so every cell carries the same current measured by its own shunt, an identity that holds at every instant to 0.13%. Two work centres of nine cells each fall out of it with nobody having drawn them. It also caught us merging 7.4 real machines per site into each other.

4 Nothing leaves the plant, and that is tested rather than promised.

One command seals the process against every address that is not the machine it runs on, proves the seal bites by reaching for a hostname, a bare address and an HTTP request and requiring all three to be refused, then runs a full working day on two plants inside it. 555 tags compiled, two shifts handed over, nothing sent anywhere.

5 We publish the experiments that failed, which is the part that cannot be copied quickly.

Eight of twelve came back under the number they had to beat, and all eight are published with the reasoning we got wrong. A competitor can copy any algorithm here in a week. What takes a year is knowing which four of the twelve were worth keeping.

A sector we have never seen costs a file, not a release.

We measured the ceiling honestly: pointed at a plant whose sector nobody wrote a connector for, the compiler places 0 to 8.6% of tags on a machine, against 86 to 91% with the right one. The quantity is 100% either way. So we tried to remove the need for a connector by working the naming convention out from the tags themselves, and that experiment is falsified: a convention like an ISA loop tag only implies the machine rather than writing it down, and nothing recovers what was never written.

So a connector is now a text file rather than our code. Both of ours were rewritten as files and checked: 2,787 tags held out, every one placed identically to the software it replaces. The two sectors share a core that was never allowed to learn anything about either: the scoring refuses to run if it has. A plant that will not send us its tag list can write its own and never talk to us.

The format is published, with the one trap that cost us 174 tags named in it. This is a smaller answer than inferring the convention would have been, and it is the one the measurements support.

Two claims that were on this page yesterday are gone. Confidence scores with evidence we had called a differentiator; it is published prior art, including work from a large cloud vendor that emits a score and a natural-language justification per point. Hierarchy from tag names we had called ours; it is sold by at least five products. We still do both, they are still absent from most industrial tools, and they are no longer the reason to buy this.

Back to the front page