The honest sequencing
Predictive maintenance is sold as a model that tells you a bearing will fail in nine days. That requires a long history of failures with sensor data attached, which most operations simply do not have — and cannot generate quickly, because failures are rare by design.
The sequence that works: monitor condition, alert on thresholds, record every failure properly, and build prediction once you have enough history to validate it.
Stage one: condition monitoring
Vibration, temperature, current draw, pressure and run hours, captured continuously and compared against normal ranges. No machine learning required, and it catches a substantial share of the failures that would otherwise be a surprise.
A temperature sensor with a sensible alert threshold has prevented more unplanned downtime in smaller operations than any predictive model we have deployed. Start there and be honest that it is the highest-return step.
Stage two: record failures properly
The reason prediction is impossible in most operations is that failure records are unusable: free text, inconsistent categories, no timestamps of when degradation started, no link to the sensor data.
- Failure mode from a fixed list, not free text
- Timestamp of failure and, if known, when symptoms began
- What was replaced or repaired, and the cost
- Linked to the asset and to its sensor history
Twelve to twenty-four months of this makes prediction feasible. Without it, no vendor can deliver what they promise, whatever the demo shows.
Stage three: prediction, if the numbers justify it
With sufficient history, models can identify degradation patterns ahead of failure. Whether this is worth building depends on the cost of a failure: unplanned downtime on a bottleneck machine justifies a great deal; a spare pump that can be swapped in an hour does not.
Calculate the cost of failure per asset before deciding where to invest. In most operations, two or three assets account for most of the downtime cost and everything else can stay on condition monitoring.
What the payback actually looks like
| Stage | Typical cost | What it prevents |
|---|---|---|
| Condition monitoring on key assets | £5,000–£20,000 | A meaningful share of surprise failures |
| Structured failure recording | £3,000–£10,000 | Nothing directly — it enables everything later |
| Predictive models | £25,000–£80,000 | Failures with warning signs in the data |
The middle row is the one businesses skip and the one that determines whether the third is possible.
The organisational part
An alert nobody acts on is worse than no alert, because it trains people to ignore the system. Before deploying anything, agree who receives alerts, what they do, and what authority they have to stop a machine.
Also agree the false alarm tolerance. Too many and the system is ignored within a month; too few and you miss real failures. This needs tuning with real data over the first quarter.