A vendor stands at the front of the room with a dashboard open. A gauge reads ninety-one percent confidence that a bearing on line four will fail within the week. The plant head asks the obvious question: how do you know? The vendor says the model has seen this pattern before.
The maintenance lead in the back of the room says nothing, because he already knows why that reading looks the way it does. He changed that sensor himself three weeks ago, and it runs a little high until it beds in. He wrote it in the log. Nobody asked him before the number went on the slide.
Ninety-one percent is not a wrong number. It is a number that will not say what it is made of. Some of it is a fact about the machine — a tolerance, a known drift — true regardless of who is in the room that day. Some of it is a pattern the model picked up from months of history that happen to line up, no more certain than a hunch. Blended into one score, the plant head cannot tell which kind of sure he is looking at.
Why we build it this way
Most systems sold into a plant are trained on everything at once — sensor readings, maintenance logs, technician notes, past failures — and asked to find whatever correlates with the outcome. That treats a physical limit and a coincidence as the same kind of evidence, because to a model that only sees numbers in a table, they are. It has no way of knowing that a tolerance is fixed by the part, and a pattern in the notes is fixed by nothing more than having happened before.
We start from a different place. There is what the machine is — a matter of measurement, written down before any model is trained, because a shaft's tolerance and a line's takt time do not change with whatever data happens to be on hand. And there is what the plant has learned — a matter of memory, held by the people who have run the line the longest, most of it never written down because nobody had a way to record it that did not mean filling in a form.
A model trained on both at once will always tell you it is sure. It will never tell you which kind of sure.
So we build the two separately, on purpose. One is constrained by what is physically true about the machine, set before it ever sees live data. The other is built from judgement a person has actually confirmed — a diagnosis agreed with, a step corrected, an option rejected and why — recorded as it happens, never guessed at afterwards. The two share one thing only: the same words for the same fault, so a reading from the machine and a memory from a person can be set side by side, not mixed together.
How it works, roughly
Kept apart, each one reads only what it is qualified to read:
- The machine side reads what can be measured directly — vibration, load, cycle time, tolerance — against limits taken from the physics of the part, not learned from a training set
- The memory side reads what a person accountable for the line has already confirmed, corrected or rejected about this exact kind of fault before
When the two agree, the answer is quiet: a measured fact and a remembered pattern pointing the same way. When they do not — the physics say the bearing is within tolerance, but the pattern looks like the last three that failed — that disagreement is what gets shown, not a blended number that hides which side is talking. A person reads both and decides, the same person accountable for the line, and the decision itself is recorded, so the next disagreement has one more answer to draw on.
A year in, the machine's limits have not moved, because they were never a guess to begin with. The plant's memory has grown instead of leaving with whoever retires. The two stay separate on the screen, and a person still reads both before anything changes on the floor.
Read the thesis →