Predictive maintenance is often told as an algorithm problem: you imagine a model that, watching the data, predicts a failure weeks in advance. It is a fascinating but misleading picture, at least for the vast majority of SMEs. In reality, the leap in value comes long before machine learning: it comes when machine data is collected continuously and turned into timely, actionable alerts.
The three levels of maintenance
- Reactive: you intervene after the breakdown. It is the most expensive and unpredictable level, because downtime arrives when you least expect it, often at the worst possible moment.
- Preventive: you intervene on a schedule, at fixed intervals. A step forward, but it means replacing components that are still good while offering no protection against failures that happen between one service and the next.
- Predictive: you intervene when the data suggests it. Vibrations, temperatures, electrical absorption, cycle times, duty cycles drifting outside expected values become the signal that something is changing.
Thresholds come before models
A rule system built on well-chosen thresholds already intercepts a surprising share of anomalies: the micro-stop that starts repeating, the temperature creeping up day after day, the cycle time stretching by a few percentage points. These are weak signals that an operator, busy with production, does not notice — but that a system watching tirelessly picks up immediately.
Only on top of this recorded data history does it make sense, later on, to graft forecasting models that learn the "normal" patterns and flag the drifts. The order matters: first you collect and react on thresholds, then — once the history is broad enough — you refine with models. Starting from the algorithm without data is the recipe for a project that never takes off.
The critical point is delivery
There is an underrated aspect: an alert has value only if it reaches the right person, on the right channel, at the right moment. A threshold breach that ends up in a log nobody reads is useless. The alert must reach the maintenance technician or the shift leader with the reason code, the history and the context attached, so whoever receives it knows immediately what to look at and with what priority.
This is where predictive maintenance stops being a conference slogan and becomes, very concretely, less machine downtime: fewer emergency interventions, spare parts ordered in time, maintenance planned in the right windows instead of suffered.
A realistic path
For an SME the sensible path is incremental: start from the most critical machine, collect the relevant measurements (the ones the maintenance technician already looks at), define a few thresholds with the people who know the plant, route the alerts to the right people. After a few months of history, evaluate where models add value. VMES45 was born to make this path practical without a dedicated IT department. Get in touch to talk about it.
