The typical situation
The machining line stopped without warning roughly twice a week. Each stoppage cost between two and six hours and, because the failing part was rarely in stock, sometimes a day. The plant had tried shorter preventive-maintenance intervals; that reduced failures a little and increased planned downtime a lot.
The maintenance head kept a thorough register: seven years of breakdowns, causes and parts used, in Excel. The PLCs logged spindle load and temperature every 30 seconds to a server nobody opened.
What the data looked like
Seven years of breakdown registers in Excel, PLC logs at 30-second intervals, shift production records and the spare-part issue register. No linking key existed between them; the first three weeks were spent building one from timestamps and machine codes.
What we built
- A daily failure-risk score for each of the 12 machines, with the three signals driving it.
- Alerts to the maintenance head's phone when a machine crossed the threshold, with the likely component.
- A spare-part reorder list derived from predicted failures.
- Everything on a small server inside the plant network, maintained by the plant's IT engineer after handover.
Typical outcome
| Measure | Before | After | Period |
|---|---|---|---|
| Unplanned downtime hours | 112 / month | 77 / month | 3-month average before vs 3 months after |
| Emergency spare purchases | 9 / month | 3 / month | same period |
| Planned maintenance hours | 48 / month | 41 / month | same period |
What usually trips these projects up
We would start the sensor data pipeline in week one rather than week three; the linking work took longer than expected and delayed the first model. The temperature signal turned out to matter far less than the register's text notes, which we nearly ignored.