Predictive maintenance is sold as an artificial intelligence problem and is mostly a discipline problem. The marketing suggests you need a data platform, a model and somebody with a statistics degree. For the large majority of industrial plants, the first eighty percent of the available benefit comes from measuring a handful of parameters consistently and having a named person look at the trends every month.
The four maintenance strategies and where prediction fits
Reactive maintenance fixes things when they break. It is the cheapest strategy to administer and the most expensive to live with, because failures occur at the worst moment and take adjacent components with them.
Preventive maintenance replaces or services components on a schedule. It is predictable and it is wasteful, because a fixed interval must be conservative enough for the worst case, so most components are replaced with useful life remaining.
Condition based maintenance measures the actual state of equipment and intervenes when a measurement crosses a threshold. This captures most of the value of prediction with none of the modelling complexity.
Predictive maintenance extends that by estimating remaining useful life, which allows the intervention to be planned rather than merely triggered. The distinction matters less in practice than the literature suggests, because the operational benefit, which is knowing in advance and planning the work, arrives with condition based monitoring.
What to measure, in order of value
Start with running hours and machine state, because a condition threshold means little without knowing duty. Then temperature, which is cheap and where the trend is almost always the signal. Then motor current or power, which usually requires no new hardware at all since drives and protection relays already measure it, as described in our note on Industry 4.0 sensors, what they actually measure.
After that, vibration on rotating equipment large enough to justify it, and oil analysis on anything with a substantial lubricant volume. Oil analysis is particularly underrated in small plants, because a laboratory report on a sample costs very little and identifies wear metals, contamination and additive depletion with a specificity that no simple sensor matches.
The approach that works without a data team
Establish a baseline
Measure every machine when it is known to be healthy, under a recorded load condition. Without a baseline, no later reading means anything, because a vibration level that is alarming on one machine is normal on another. This single step is what most programmes skip and what most later confusion traces back to.
Measure the same way every time
Same point on the housing, same orientation, same load condition, same instrument. Inconsistent measurement produces variation that looks like a developing fault and wastes the credibility of the whole programme on false alarms.
Trend, do not threshold
A reading compared with a manufacturer’s limit tells you whether you have a problem now. A reading compared with the same machine six months ago tells you whether you are developing one. Most of the value is in the second comparison, and it requires nothing but a spreadsheet and consistency.
Name somebody and give them time
This is the step that determines success. A monthly review of trends by a person whose job explicitly includes it. Programmes fail not because the analysis is hard but because reviewing data is nobody’s actual responsibility and it quietly stops happening after the enthusiasm of the first quarter.
Where the money comes from
The saving is rarely in the repair itself. It is in four other places. Planned downtime instead of unplanned, which may mean a weekend instead of a production shift. Secondary damage avoided, because a bearing caught early does not destroy a shaft and a seal. Parts ordered at normal lead time and normal price rather than expedited. And useful life fully used, because components are replaced when their condition warrants rather than when the calendar says.
That last point is the quiet one. Many plants adopting condition monitoring find their first saving is extending intervals on equipment they were over maintaining, not catching failures they were missing.
Where it does not pay
Be honest about this, because enthusiasm leads programmes into places where they cost more than they return. Equipment that is cheap to replace and quick to swap does not justify monitoring. Fully redundant equipment where failure causes no production loss does not either. Machines that run rarely accumulate too little data for trending to mean anything. And components with genuinely random failure modes, where there is no gradual degradation to detect, cannot be predicted by any method.
Apply it where failure is expensive or disruptive, where degradation is gradual and detectable, and where the equipment runs enough to generate a trend. That is usually a small subset of the asset list, and concentrating effort there beats instrumenting everything.
Scaling up, if it earns the right
Once a manual programme is running and trusted, automation of data collection becomes worthwhile, because the limiting factor becomes the labour of taking readings. Continuous sensors, automatic collection and software that flags deviations all remove that burden.
Machine learning becomes genuinely useful at the point where there are enough machines and enough history that patterns across the fleet exceed what a person can hold in mind. That is a real threshold and most small plants are nowhere near it. Reaching for the tool before the threshold produces a project, not a benefit, which is the same trap described in our note on the real payback period on a factory automation project. General background is summarised in the overview of predictive maintenance.
Frequently asked questions
Do we need machine learning to do predictive maintenance?
No. Consistent measurement and monthly trend review capture most of the available benefit. Machine learning helps once there are enough machines and enough history to exceed what a person can track.
What is the most important first step?
Establishing a baseline on healthy equipment under a recorded load. Without it, later readings cannot be interpreted.
Why do these programmes usually fail?
Because reviewing the data is nobody’s explicit responsibility. The analysis is not the hard part, the sustained attention is.
Where does the saving actually come from?
Planned rather than unplanned downtime, avoided secondary damage, parts bought at normal lead time, and full use of component life rather than calendar based replacement.
Which equipment should not be monitored?
Anything cheap and quick to replace, anything fully redundant, anything that runs too rarely to trend, and anything whose failure mode is genuinely sudden rather than gradual.