Predictive Maintenance: What Can AI Actually Predict – and What Can’t It?
Having spent years working on predictive maintenance with smart sensors, I can tell you where the real value is – and where the marketing promises quietly exceed what physics and data allow.
The idea is compelling: instead of maintaining machines on a fixed schedule (often too early, wasting parts and downtime) or after failure (too late, causing unplanned stops), you monitor their actual condition and intervene exactly when needed.
What AI does well here rests on a physical truth: most mechanical degradation announces itself. Bearings developing faults change their vibration signature weeks before failure; motors under stress show thermal and current anomalies; pneumatic systems leak progressively. Sensors capture these signals; models learn to distinguish normal operating fingerprints from early degradation patterns – often long before a human would notice anything.
Three levels of ambition, in increasing difficulty:
- Anomaly detection – “this machine is behaving unusually.” The most robust and accessible level: the model learns normality and flags deviations. It doesn’t tell you what’s wrong, but it tells you where to look, early.
- Diagnosis – “this pattern looks like bearing wear on axis 3.” Requires labeled failure history or physics-informed features. Powerful, but demands data most plants don’t systematically collect.
- Remaining Useful Life (RUL) – “this component will fail in 12 days.” The holy grail of the brochures, and the hardest to deliver honestly. Precise RUL requires many recorded run-to-failure trajectories of the same component under comparable conditions – data that well-maintained plants, by definition, rarely have.
What AI cannot do: predict failures that leave no measurable precursor (sudden electronic failures, external damage), predict with data you don’t collect, or compensate for sensors placed where the signal isn’t. And it cannot skip the cold-start problem: you need months of baseline data before models become trustworthy.
The playbook: start with anomaly detection on critical assets, build the habit of recording maintenance outcomes (they are your future labels), and grow toward diagnosis. Distrust anyone promising day-one RUL predictions on your equipment. Physics doesn’t do marketing.
