Full Article
Instrumenting equipment is the easy part. Vibration, thermal and current sensors are cheap, gateways are commodity hardware, and dashboards are a solved problem. Yet most predictive maintenance programmes plateau at monitoring — plenty of data, few predictions anyone acts on.
The Missing Ingredient
Predicting failure requires examples of failure. A model trained on six months of healthy operation has learned what normal looks like, which supports anomaly detection but not remaining-useful-life estimation. Teams consistently underestimate how long it takes to accumulate labelled failure events, and how much value sits in historical maintenance records that were never digitised.
A Workable Sequence
- Start with anomaly detection. It works from healthy data alone and delivers value while failure data accumulates.
- Digitise maintenance history. Past work orders are labelled failure data — usually the largest untapped asset in the plant.
- Instrument for the failure modes that matter. Sensor placement should follow a failure mode analysis, not convenience.
- Close the loop. Every maintenance action taken on a prediction must feed back as a label, or the model never improves.
Edge Versus Cloud
High-frequency vibration data is expensive to transmit and mostly uninteresting. Extracting features at the edge — spectral peaks, RMS, kurtosis — and forwarding only those cuts bandwidth by orders of magnitude while preserving diagnostic content. Raw waveforms are worth keeping only around detected events.
Measuring Success Correctly
Model accuracy is the wrong headline metric. What matters is unplanned downtime avoided, maintenance labour redirected from inspection to repair, and parts consumed per operating hour. A model with modest precision that catches the expensive failures beats a sharper one that catches trivial ones.
Tags
