Predictive Maintenance Architecture: From Data to Work Order

Predictive maintenance has two failure modes: too many false alarms (the team stops listening) or too few alarms (the team stops believing). Engineering the precision recall trade off is the whole game.

Sensor & Signal Stack

• Vibration: tri axial accelerometers at bearing housings, sampled at 5, 25 kHz.
• Current: clamp on or panel mounted CTs.
• Temperature: surface RTDs at bearings, motor windings.
• Acoustic emission for early stage cracks.
• Process signals from PLC, the most under used data source.

Feature Engineering

Raw signals are not features. Engineered features include:

• Vibration: RMS, peak, kurtosis, frequency domain spectra (FFT).
• Current: harmonics, total harmonic distortion (THD).
• Temperature: rate of rise, delta T.
• Cross signal: load normalised vibration.

Model Patterns

• Anomaly detection (unsupervised) when failure history is sparse, most Indian plants start here.
• Remaining Useful Life (RUL) regression when failure history is rich.
• Multi modal models combining vibration + process data for highest precision.

Closing the Loop with CMMS

An alarm without a work order is just noise. The pattern that works:

1. Model flags an anomaly above threshold.
2. System auto creates a work order in the CMMS with the asset, the signal, and the recommended action.
3. Maintenance team accepts / rejects / schedules.
4. Outcome (was it a real failure?) feeds back to the model.