Case study · Industry
Predictive maintenance that tells you every morning what will fail
For an international automotive manufacturer, we combined years of failure, repair and sensor data from five systems. Every morning, maintenance teams now receive an overview of the equipment most at risk of failure, together with recommended action, before the line stops.
Challenge and solution
Challenge
Every unplanned stop of an assembly line costs production time and money. Maintenance had plenty of data, but it was scattered across different systems: failure and repair records, preventive maintenance, control system alarms and sensor measurements.
What was missing was a way to identify in time which equipment was likely to fail, and to plan interventions accordingly.
Solution
We developed a custom automated predictive maintenance system in Python. We connected historical data going back to 2020 and set up daily processing.
- Integration and cleansing of data from 5 source systems, including the SAP enterprise system and the SCADA control system
- Machine learning models predicting failures and downtime duration
- A daily e-mail report and dashboards with alerts for technicians and managers
What every morning looks like
The whole process runs automatically. By the time the morning shift starts, technicians know what to focus on.
Data is ready
New records from the previous day arrive from the source systems.
Processing and prediction
The system cleans the data, recalculates the models and assesses the risks.
Report by e-mail
Technicians receive a summary of the top risks and recommended actions.
9 reports, three approaches
Each report answers a specific maintenance question. Together they provide a view of the past, the present and the future.
Prediction and trends
What is likely to fail?
- Equipment failure probabilityFailure risk 7 days ahead, including estimated downtime duration and priority.
- Specific fault predictionWhich fault is likely to recur on a given device and location.
- Abnormal temperature developmentEarly detection of fluctuations and rising temperatures from machine sensors.
Statistics and frequency
Where do failures recur?
- Failure frequency and seasonalityCritical locations, months with the most failures and recommended preventive steps.
- Failure and downtime cyclicalityEquipment reliability through MTBF and MTTR and assets with repeated downtime.
Alarm overview
What needs attention first?
- Alarm frequencyThe most frequent alarms by location, priority and type, with charts and heatmaps.
- Alarm durationThe longest-lasting alarms, separating warnings from real failures.
The four alarm reports cover two different systems, which brings the total to nine.
A real-world example
This is how a technician works with the reports for a manipulator on the assembly line:
Prediction: a report flags a high probability that the manipulator will fail next week.
History: a chart of failures, repairs and preventive maintenance shows how they were related in the past.
Specific fault: another report predicts which type of fault is likely to occur.
Recommendation: the system suggests checking the device and carrying out preventive maintenance before the line stops.
Project approach
From analysing the current state to deployment in the client’s environment.
What the project delivered
Technologies used
Looking for a similar solution?
We will discuss your data and goals and propose the way forward. The first step is a no-obligation consultation.