Predictive maintenance across three production lines

The challenge

Unplanned downtime cost roughly €240,000 a year, and maintenance was carried out at fixed intervals, regardless of the actual condition of the equipment.

Sensor data existed but was discarded after 72 hours, in a system that allowed no historical analysis.

The solution

We built the full pipeline, from signal acquisition to the alert reaching the maintenance team.

• Collection from 190 machines, across different industrial protocols
• An analytical store keeping three years of time series
• An anomaly detection model, evaluated monthly on real data
• Integration with the existing maintenance ticketing system

The model went into production only after three months of parallel running, to validate the false alarm rate.

The outcome

Unplanned downtime fell by 38% in the first full year of operation. The false alarm rate settled at 7%, below the 10% threshold agreed at the outset.

The maintenance team moved from fixed intervals to interventions planned around the real condition of the equipment.