How we combined Azure IoT Hub sensor telemetry with a TensorFlow machine learning pipeline to predict equipment failures 72 hours in advance for a heavy equipment manufacturer — cutting unplanned downtime by 72% and delivering $3.5M in annual savings.
A global manufacturer of heavy industrial equipment operated assembly and test facilities across five continents. Unplanned machine downtime on critical CNC machining centres and hydraulic test rigs was costing the business an average of $4.6M per year in lost production, emergency parts procurement, and overtime labour. Maintenance schedules were purely calendar-based — machines were serviced every 90 days regardless of condition, meaning both over-maintenance of healthy equipment and under-maintenance of deteriorating assets.
The engineering team knew their machines were generating vast amounts of vibration, temperature, pressure, and cycle count data, but it was being discarded in real time. There was no infrastructure to capture, store, and analyse it, and no data science capability in-house.
Catalogued 500+ sensors across 3 flagship facilities. Designed an Azure IoT Hub ingestion architecture with MQTT edge gateways on the shop floor. Established data contracts and telemetry schemas for each equipment class.
Built an Apache Kafka streaming pipeline within Azure Event Hubs to ingest 28,000 sensor readings per second. Data was routed to Azure Stream Analytics for real-time anomaly flagging and to cold storage in Azure Data Lake for model training.
Partnered with senior maintenance engineers to catalogue known failure signatures for 12 critical failure modes across 6 machine types. Derived 80+ engineered features from raw sensor data including rolling FFT spectral analysis and thermal gradient patterns.
Trained a multi-class LSTM recurrent neural network on 3 years of historical sensor data and matched failure records. The model outputs a 72-hour failure probability score per machine, per failure mode. Achieved 89% precision on held-out test data.
Integrated model predictions with SAP PM (Plant Maintenance) via REST API — automatically raising work orders when failure probability crossed a configurable threshold. Built a real-time Power BI dashboard for maintenance supervisors.
The predictive maintenance platform transformed maintenance from a time-based cost to a data-driven investment. Within six months, the engineering team had more confidence in machine health data than in any manual inspection programme they had previously operated.
"We used to find out about failures when the machine stopped. Now we know three days in advance, we schedule the work, order the parts, and the machine never goes down. The ROI on this project paid for itself in 5 months."
— VP Global Manufacturing Engineering, Heavy Equipment Group
Let’s discuss how we can solve your unique challenges and deliver measurable impact.
Start a Conversation View All Case Studies