273 hours of downtime anticipated, from a log the plant was already collecting.
Aerospace & defence supplier · production line, 4 machines · regulated environment. No new sensor, no intervention on the machines, and no data preparation asked of the client.
A backtest, not a projection. The model was calibrated on an earlier period and then scored against six months it had never been shown. How we validate →
“Does a serious stop start within the next hour?”
One question, asked of one line, against data the plant had been writing to disk for years.
- Input
- The run/stop event log alone. No sensors, no process data, nothing added to the machines.
- Signal
- Recent micro-stop activity — how many, how long, and whether they came in bursts or evenly spread.
- Output
- One score per machine every 10 minutes, against a threshold that adapts to that machine.
- Causal
- Only the past is ever used, so the model that was backtested is the model that deploys.
Two machines paid off. Two did not.
The gain is concentrated. V4 and V2 carry it; on V1 and V3 the log held almost nothing the model could use. Rather than average that away, here is the split.
The median warning time matters as much as the hours. On V4 it is 5.9 hours — a shift’s notice. On V2, 37 minutes: enough to finish a part and stop cleanly, not enough to plan. On V1 and V3, around five minutes, which is not a warning at all. The per-machine economics →
What survives the false alerts.
Alerts that precede nothing cost real time to check. That cost is deducted below rather than left out, which is why we publish two numbers instead of one.
Nothing was installed. That is the point.
The plant had already spent years instrumenting. The event log existed, it was complete, and until this ran it had never been used to decide anything.
No new sensor
Not one. The deployment added no hardware and required no intervention on the line.
No client man-days
No data preparation was asked of the client’s team, and no engineering time was booked against it.
Regulated environment
Micro-stop counts only — no named operator data, so the client’s GDPR scope was unchanged.
Name one line — we read the log you already have.
Within a week we tell you whether it carries a precursor, before you commit to anything.
