AI in manufacturing
What Data Do You Really Need for Predictive Maintenance?
Sensors, history, labels and context: the minimum data set before any model is worth building.
Trikaan Research & Engineering4 min read
KEY TAKEAWAYS
- Models learn from failures; without labelled history they guess.
- Choose signals tied to known failure modes.
- Operating context — load, product, speed — explains many anomalies.
- Start with thresholds and trends before complex models.
Predictive maintenance projects often begin by installing sensors everywhere. Months later, there is plenty of data and no reliable predictions. The missing pieces are usually failure history and context.
Here is the minimum data set worth having.
Signals tied to failure modes
| Failure mode | Useful signals |
|---|---|
| Bearing wear | Vibration, temperature |
| Motor issues | Current, temperature |
| Pump problems | Pressure, flow, vibration |
| Belt or chain wear | Speed variation, vibration |
| Clogging | Pressure drop, current |
Sampling and quality
Some signals need high-frequency sampling (vibration), others only periodic readings (temperature). Check sensor placement, calibration and gaps in data before modelling.
Failure history and labels
Each failure needs a record: date, machine, component, failure mode and what was found. These labels tell a model what to look for. Clean breakdown records are as valuable as sensors.
Operating context
- Product running and speed.
- Load and throughput.
- Ambient conditions where relevant.
- Maintenance actions taken.
Start simple
Thresholds based on OEM limits and trend alerts on rising signals catch many problems early. Add statistical and machine-learning models once you have enough labelled events.
Frequently asked questions
How much data is needed?
Enough to include several examples of each failure mode you want to predict, plus normal operation across conditions.
Are OEM limits useful?
Yes — they provide initial thresholds before models are trained.
Can we use existing PLC data?
Often yes — currents, temperatures and pressures are frequently available.
When do ML models help?
When patterns are subtle and you have labelled history across conditions.
- #predictive maintenance data
- #condition monitoring data
- #vibration analysis data
- #failure prediction
- #maintenance machine learning
Trikaan Research & Engineering
Trikaan
Written by the researchers and engineers who build Anvaya, ASTR and Seed Track, with input from Trikaan's delivery and engineering teams.
