Predictive Maintenance: 35% Less Unplanned Downtime, $1.2M Saved
A mid-size auto parts manufacturer was losing $100K+ per unplanned equipment failure. We built a sensor data pipeline and predictive models that detect failures 48 hours before they happen.
The Challenge
What was getting in the way
- 01
Three CNC machines and two press lines had no monitoring beyond basic alarms. When something broke, the line stopped for 8 to 14 hours while parts were ordered and installed
- 02
Unplanned downtime averaged 6 incidents per quarter, each costing $80K to $120K in lost production, overtime labor, and rush-ordered spare parts
- 03
The maintenance team ran on fixed schedules. They replaced parts every 90 days whether they needed it or not, wasting $200K a year in premature part swaps
The Solution
How we solved it
We retrofitted 14 machines with vibration, temperature, and current sensors feeding into an IoT gateway. Sensor data streams into AWS IoT Core, gets processed by Kinesis, and lands in a time-series database (TimescaleDB). We trained gradient-boosted models on 8 months of historical failure data combined with live sensor readings. The models score every machine every 5 minutes and flag anomalies 24 to 48 hours before likely failure. Maintenance gets a Slack alert with the machine ID, predicted failure type, and recommended action. We also built a Grafana dashboard showing machine health scores, upcoming maintenance windows, and historical trends. The team went from reactive firefighting to planned maintenance with a 48-hour heads-up.
Technologies
What We Built
A look inside the project
CNC-01
CNC Mill
HealthyPress-A
Hydraulic Press
HealthyLathe-03
CNC Lathe
WarningBearing wear detected - Replace within 48hrs
Weld-B2
Robot Welder
HealthyLive Sensor Readings
Recent Alerts
Predicted failure prevented
Part replaced on schedule
Anomaly detected and resolved
Illustration based on actual project deliverable
The Process
Step-by-step delivery
Sensor Retrofit
Install vibration, temperature, and current sensors on 14 machines
Data Pipeline
Stream sensor data through IoT Core, Kinesis, and TimescaleDB
Model Training
Train on 8 months of failure history + live sensor feeds
Alert System
Score machines every 5 min, alert on Slack 48 hours before failure
Dashboard & Insights
Grafana dashboards for machine health, trends, and planning
The Results
The numbers
Less Unplanned Downtime
Annual Cost Savings
Early Failure Detection