Use casesPredictive maintenance
Know which asset fails next, with a watch list your crews can actually work.
Every plant and every fleet has the data: loading, temperature, vibration, the maintenance log. What they do not have is a ranked list of what to inspect this week, sized to the people available. We build failure models that score every asset daily, explain each score in plain words and are graded on failures they never saw.
What gets in the way
A machine fails on a Friday night. The part has a twelve-week lead time. The sensor data that would have shown it warming up for a month was sitting in a historian nobody opened.
A failure model that flags everything is a list nobody reads. A model that flags nothing until the week before is a model that does not exist. The useful one is sized to the inspections your crew can do.
And a prediction with no reason is an argument with the maintenance lead. If the model cannot say why, the work order does not get written.
What we build
A score per asset per day, a reason per score and a watch list sized to the work you can do.
- Every monitored asset gets a daily score for failing within a window you choose, read from the signals you already collect: loading, temperature, vibration, cycles, age and the maintenance history.
- Each unit on the watch list carries a plain reason: running hot for 30 days, 20% above rated load, three alarms this month. Dispatching an inspection writes the work order back.
- The list is sized to crews, not to the model. Fifty inspections a week means a fifty-unit list, ranked, with the catch rate at that size reported honestly.
- It runs in your warehouse or on your historian’s export, on a schedule, and the run stops itself when the sensor feed goes quiet.
- A storm or an accident is not wear-out. The model says what it predicts and what it does not, so nobody expects it to see a tree fall.
How we know it works
Graded the way it will be used: trained on one stretch of time, tested on the failures that came after.
- On Grid Intelligence, the transformer failure model scored an AUC of 0.81 on a test fleet. A 50-unit watch list caught 58% of failures, and the riskiest tenth of units held 6.5 times their share. A plant gets its figures in the same form.
- The limits are published beside the results. The model predicts wear-out, not storm damage, and every screen on the test fleet says the fleet is synthetic.
- For a client with thin failure history, we inject known faults into real sensor data so the catch rate is measured before anyone trusts a list.
- The catch rate at your list size is on the dashboard, not in a sales deck, and it is rechecked as real failures arrive.
Built from this work
No plant runs this under a client’s name yet. The failure model and the watch list are live in the platform below, measured on a test fleet built from public records.
58% of failures caught by a 50-unit watch list on the test fleet
8.8% load forecast error on the test fleet, against 11.8% for yesterday’s load
4 of 4 summer peaks caught in an out-of-sample replay on the test fleet
Read the story
Where it fits
Manufacturing
AI for manufacturers, where the count has to be right.
Questions
We only have a few failures a year. Is that enough?
Sometimes not for a pure model, and we say so. Thin history leans on the anomaly layer and on injected faults to measure the catch rate. The model’s reported confidence reflects what it was given.
Which data do we need?
Whatever your historian and your maintenance system already hold: sensor series, alarms, work orders, failure dates, asset age and rating. The first week is an inventory of exactly that.
How does it reach the crew?
As a ranked list in the tool they already use, with the reason on each row. Marking an inspection done writes back, so next week’s list knows.
What if the model is wrong?
It will be, some of the time, and the catch rate at your list size says how often. That number stays on the dashboard, where the crew can see it.
Bring us the project that matters most.
Tell us what you’re building and what’s in the way. You’ll hear back within 24 hours.