An automotive moulder suffered regular hydraulic failures on older presses, each costing a shift or more. Sensors and anomaly detection on 12 presses predicted 17 failures in the first year, cut unplanned downtime from 6.2% to 2.4% and saved about $80,000 in maintenance.
Key numbers
An automotive moulder ran 12 older hydraulic presses. They were reliable most of the time, but every few weeks one would fail: a pump, a valve or a seal, usually without warning. Each failure cost a shift or more of production, often with emergency repairs and overtime to recover lost output for automotive customers who expected on-time delivery.
The situation
Unplanned downtime averaged 6.2% of scheduled time across the presses. Preventive maintenance was carried out at fixed intervals, but many failures happened between services, and some parts were replaced long before they needed to be. Maintenance cost was around $410,000 a year.
What changed
The company introduced condition monitoring and anomaly detection, following the approach in predictive maintenance for moulding machines:
- Sensors for oil temperature, hydraulic pressure and pump motor current were added to all 12 presses.
- Models were trained on each press's normal behaviour across products and operating conditions, so they could flag unusual patterns rather than relying on simple fixed thresholds.
- Alerts were linked directly to maintenance work orders, so every anomaly was reviewed and acted on by a technician.
- Daily operator checks, introduced as part of TPM, continued alongside the monitoring.
Results
In the first year, 17 developing failures were detected in advance, allowing repairs to be planned during scheduled downtime. Unplanned downtime fell from 6.2% to 2.4% of scheduled time, a reduction of 61%. Maintenance cost fell by about $80,000, to around $330,000 a year, as emergency repairs and unnecessary replacements both decreased.
Why it worked
The warning signs had always been there, in rising oil temperatures and changing pump current, but nobody could watch twelve presses continuously. The models did the watching. Maintenance expertise did the rest: deciding what each alert meant and planning the right intervention.
Lessons for your operation
- Start with the equipment where failures hurt most.
- Use data already available from controllers, then add simple sensors.
- Train models on normal behaviour for each machine, to reduce false alarms.
- Link alerts to work orders, so every warning leads to action.
- Combine predictive maintenance with basic operator care, not instead of it.
