The queue that matters
In March 2012 a 12.5 MVA generator at a sugar mill was running at 13 mm/s axial on the coupled-side bearing — along the shaft, that is — two and a half times the radial. Everyone looked at the bearing, once that is where a radial spectrum points. The mill was in crop and could not stop. I had a day with the machine running, a run-up, a coast-down with the power off, and two accelerometers, what is the wrong instrument for a shaft orbit. The curves rose and fell smoothly with load and temperature. No step anywhere. Whatever it was did not need a temperature to happen.
That already removed a lot. A loose fan, a rub, anything that waits for heat and then jumps, shows as a step in the curve. What was left was what the spectra said: bearing clearance, in the radial; angular misalignment, in the axial. The base was rigid, every foot inside the norm, no relative motion. So far, one reading, and one reading is not a diagnosis.
The second signal came from a different machine. The gearbox bearing and the generator bearing were vibrating axially 180° out of phase — each pushing the other way. The third came from looking: a mark on the shaft showed the rotor running outside its magnetic centre. The stator pulls the rotor back to centre, and that force is axial. The cause was electromagnetic; the symptom was mechanical; one system between them. For the first hour I would have bet on the bearing too.
The report says what the instrument did not see: the orbit was measured with accelerometers, which explains the residue in the plot. And it says what to do: inspect the bearings at the off-season stop, put the coupling back on centre, and do the restart together with the turbine people. The diagnosis took a day. The fix waited four months for the crop to end.
On 9 September Hydrocarbon Engineering ran a special report from UptimeAI on a refinery where a reboiler fouled and the feed pump loaded up to hold throughput. An agent built the causal chain across column, reboiler and pump. The product page gives the number: investigation cycles "from weeks to hours." I sell a system that does this. The chain reads right to me. What I want is to open the number.
"Weeks to hours" has a verb, and the verb is investigate. The clock starts at the alarm and stops at the diagnosis. At the mill that clock read one day. The clock that cost money started after it: the outage window, the part, the turbine crew's calendar. Hours against a day is a good ratio for a vendor. Either of them against a season is the ratio the owner of the machine lives with. If investigation drops from weeks to hours and the stop is still booked for the off-season, what changed was who waits, not how long.
I am not saying the number is false. I am saying what it measures.
A month after that visit, a consultant I had met at the mill wrote to the director of a different client, a 25 MW machine under warranty. He asked that I be put on the team, and attached the report. He did not recommend the company. He recommended the document. The conversation left no trace; the report did. I am not sure what to do with that, but I keep it.
What this costs us is the second number. A diagnosis in two hours is a vendor's number, and I am a vendor. The number that goes with it is what moved because of it: the stop brought forward, the part ordered, the hours that came off someone's week. And hours only count once they became a person not hired, or a person moved to other work. Otherwise they were absorbed, and an absorbed hour is not a saving, however good it looks in a proposal. That second number is the customer's. It is smaller than ours and it comes later than ours, and on a machine that cannot stop until the crop ends it can be four months late.
What I'd bet on
The queue that matters starts when the diagnosis ends. Weeks-to-hours is the vendor's clock; the customer's clock runs from the diagnosis to the stop, and the money is on the second one. We will report the first only with the second, and we will count a saved hour only when it became a hire not made or a person doing something else. This costs a good number on every proposal. It costs less than explaining, at the off-season stop, why a machine diagnosed in hours vibrated for a whole crop.
Also this month
- 4 Sep — a semiconductor equipment weekly puts TSMC's tool procurement list at nearly twice its December 2025 baseline; revenue recognition still waits on acceptance criteria.
- 10 Aug — the CSB's final report on the Clairton coke works finds the valve-washing practice had run "for at least three years, on an ad hoc basis." There was a procedure. Nobody had written it.
Diego Mercadal started as a commissioning engineer on high-voltage motors and generators, joined an offshore drilling contractor as a rig hand, worked several positions in the drilling crew, and ended up running its AI/ML function. He is now co-founder and CEO of Wonder DataLabs.