Predictive maintenance is one of the most mature industrial applications of AI. In automotive, where a production line stoppage costs an average of 50,000 euros per hour, the stakes are immediate.
Stellantis: from sensors to decision in 90 seconds
At its Sochaux plant, Stellantis deployed a real-time monitoring system across 2,400 critical pieces of equipment. Sensors continuously collect temperature, vibrations, power consumption and pressure. A model analyses these streams and generates a predictive maintenance alert when an anomaly is detected.
The average window between alert and actual failure is 72 hours, enough time to plan an intervention without stopping production. Since deployment, unplanned failures on covered equipment have fallen by 38%.
BMW: integration into production planning
BMW chose to integrate maintenance alerts directly into its production planning system at the Leipzig plant. When maintenance is recommended, the system automatically identifies a slot in the schedule where stopping the equipment will have the least impact on output.
The maintenance teams describe this integration as the real change. It is not the prediction that creates value; it is the ability to act on the prediction without disrupting production.
Renault: the limits of generalisation
Renault ran a pilot across three plants with uneven results. The system performs well on standardised equipment used in the same way across sites. It is less effective on older machinery (over 15 years old) whose sensor data is incomplete or non-standardised.
The group has since decided to make any deployment conditional on a prior audit of data quality for each piece of equipment.
Key takeaway
Predictive maintenance works. Its conditions for success are known: reliable sensors, sufficient historical data, and above all, integration into existing operational processes. Deployed in a silo, it generates alerts nobody follows.