Three ways to maintain a machine

Before talking about prediction, you need the three main maintenance strategies:

• corrective: you repair after the breakdown. Simple, but the stop is imposed, often at the worst moment.
• systematic preventive: you replace at a fixed interval (calendar, running hours), wear or no wear. You avoid the breakdown, but you sometimes change parts that are still good.
• predictive (or condition-based): you monitor the real condition and act only when degradation requires it.

Predictive aims for the optimum point: neither too early (waste), nor too late (breakdown).

The principle of predictive

The idea is simple: a machine announces its failure before it stops. A wearing bearing vibrates differently; a straining motor heats up; a poorly lubricated bearing emits ultrasound. By measuring these signals continuously and tracking their evolution, you detect the drift and estimate the remaining time before failure. You then schedule the intervention at the right moment, parts and crew ready.

The signals that announce failure

The variables monitored depend on the machine: vibration (the classic for rotating machines), temperature, electric current, oil analysis, acoustic emissions. Raw data rises from the field (see “industrial data”); you compare it to a healthy signature, and any lasting deviation becomes an alert. This is where machine learning adds value: it learns a machine’s “normal” and spots signal combinations a simple threshold would miss.

Availability, the quantified target

All of this aims at one indicator: availability, which links mean time between failures (MTBF) and mean time to repair (MTTR):

$A = \frac{\mathrm{MTBF}}{\mathrm{MTBF} + \mathrm{MTTR}}$

A machine with an MTBF of 2,000 h and an MTTR of 8 h shows $A = 2000/2008 \approx 99.6\%$. Predictive acts on both levers: it extends MTBF (you act before breakage) and cuts MTTR (you prepare the intervention, parts and crew ready).

What it takes to set up

Predictive is not just an algorithm. It assumes: reliable, well-placed sensors, a clean data history, knowledge of failure modes — often from a FMECA analysis — and a tool to trigger and trace interventions, the CMMS. Without that foundation, the model predicts into the void. It is the spirit of the ISO 55000 asset management standard: decide based on risk and value, not on habit.

When it pays off — and when not

Instrumenting and analysing is expensive. So reserve predictive for critical equipment: that whose failure stops production, threatens safety, or takes long to repair. For redundant, cheap or consequence-free equipment, corrective or preventive is enough. The right reflex: reason by criticality, supported by indicators like MTBF (mean time between failures) and MTTR (mean time to repair), to predict only where it pays.