Forecast
Predictive Maintenance ML System
ML pipeline for industrial equipment failure prediction. Deployed on edge nodes across manufacturing sites — 91% prediction accuracy, 48-hour advance warning, €2.1M annual savings.
91%
Prediction Accuracy
48h
Advance Warning
€2.1M
Annual Savings
Context
A heavy manufacturing group was spending €6M annually on reactive maintenance across 14 facilities. Equipment failures caused 12–40 hour shutdowns per incident, cascading into supply chain delays.
The Challenge
Sensor data was noisy, imbalanced (failures are rare by design), and varied by machine type and age. Models trained in the lab degraded quickly in production due to concept drift. Network constraints made cloud-first inference impractical.
The Solution
Built a distributed ML pipeline: data collection on edge Raspberry Pi gateways, feature engineering in a streaming Kafka pipeline, model training in the cloud (MLflow managed), and model deployment as ONNX artifacts on edge nodes. An automated retraining trigger monitored prediction confidence and scheduled retraining when drift was detected.
Methodology
- Time-series feature engineering: rolling statistics, FFT decomposition, signal power
- Ensemble model combining XGBoost and LSTM for multi-horizon predictions
- ONNX export for portable edge deployment without Python runtime dependency
- Automated drift detection and scheduled retraining via MLflow + Airflow
Impact
Deployed across 14 facilities within 6 months. Prediction accuracy stabilized at 91% on held-out validation sets. Annualized maintenance cost savings of €2.1M in year one.