Predictive maintenance with vibration and temperature: where to start

Predictive maintenance means measuring the condition of a machine while it runs and acting when the data shows it is degrading, before it fails. You do not need a huge platform or artificial intelligence to begin. You need to choose the right machines, the right variables, and decide who acts when something changes.

What is the difference between corrective, preventive and predictive maintenance?

All three coexist in almost every plant. The point is knowing which one suits each asset.

Type	When you act	Advantage	Drawback
Corrective	After the breakdown	No upfront cost	Unplanned stoppages, secondary damage
Preventive	By calendar or running hours	Easy to plan	Parts replaced while still healthy, or too late
Predictive	By measured condition	Intervention when it is actually needed	Needs sensors, data and discipline

Corrective maintenance is still sensible for cheap, redundant or easily replaced equipment. Predictive maintenance makes sense where a stoppage is expensive or a failure causes knock-on damage.

Why start with vibration, temperature and motor current?

Because together they cover most failure modes of rotating equipment, using affordable and well-understood sensors.

• Vibration. Imbalance, misalignment, looseness and bearing defects produce recognisable vibration patterns. Bearing faults, for example, show up at characteristic frequencies that depend on the bearing geometry and shaft speed.
• Temperature. A poorly lubricated bearing, an overloaded winding or a strained coupling heats up. Temperature is usually a later signal than vibration, but very easy to measure.
• Motor current. Changes in current draw point to changes in mechanical load, supply problems or, with finer analysis, certain electrical and mechanical defects.

Combining the three gives context: rising vibration with steady current means something different from rising vibration with the load climbing.

Which machines should I monitor first?

Use a simple criticality rule: cost of stoppage × likelihood of failure.

1. List your rotating assets: motors, pumps, fans, compressors and gearboxes.
2. For each one, score the cost of a stoppage (lost production, repair, related damage) from 1 to 5.
3. Score the likelihood of failure (breakdown history, age, operating conditions) from 1 to 5.
4. Multiply the two scores and sort from highest to lowest.

For example, a pump with no standby that feeds an entire line might score 5 × 4 = 20, while a duplicated fan scores 2 × 3 = 6. Start at the top of the list.

Which sensors do I need and how should they sample?

Accelerometers: wireless or wired

Wireless sensors install quickly with no cabling and usually send overall values or spectra at intervals (every few minutes or hours) to save battery. They are good for trends.

Wired sensors allow continuous acquisition with more bandwidth. They are preferable on highly critical or high-speed machines, or where a fault can develop quickly.

Temperature

RTDs (such as Pt100) are accurate and stable across the usual range for bearings and housings. Thermocouples cover higher temperatures and are robust, with somewhat lower accuracy. On accessible surfaces, a well-fixed contact sensor is usually enough.

Current

Current transformers (CTs), including split-core types, let you measure without breaking the circuit. Any installation inside power panels must be carried out and certified by a licensed electrician.

Sampling basics

• Temperature changes slowly: one reading a minute is usually more than enough.
• For overall vibration values (RMS velocity), a reading every few minutes is enough to see trends.
• For spectral analysis, the sampling rate must be at least twice the highest frequency you want to see, and in practice considerably more. Bearing defects may need several kHz.
• Always log speed or running status: comparing a stopped machine with one under load tells you nothing.

Why do you need baselines and trends before AI?

A single reading says little. What matters is knowing how that machine behaves in normal conditions and spotting when it drifts away from it.

A baseline is a record of vibration, temperature and current over several weeks of healthy operation, covering different shifts and loads. With it you can see slow drift, sudden jumps or changes in pattern.

Machine-learning models need exactly that: months of clean data and, ideally, documented failures. Without that foundation, a model only adds complexity and false alarms.

How do I set alarm thresholds?

Two references complement each other:

1. Vibration severity standards. The ISO 10816 and ISO 20816 series define severity zones (from newly commissioned to a level that may cause damage) according to machine type, size and mounting. They are a reasonable starting point when you have no history yet.
2. Change against the machine’s own trend. Once you have a baseline, the most useful alarms are relative: for example, alerting when overall vibration rises steadily above its usual value, even if it is still within an “acceptable” zone of the standard.

Use two levels: warning (inspect and plan) and alarm (intervene soon). Require the condition to persist for a set time so a single spike does not trigger it.

What does a realistic staged plan look like?

Stage	Indicative duration	What happens
1. Selection	1–2 weeks	Criticality list, choose 3–5 machines, define variables
2. Installation	1–3 weeks	Sensors, communications, local data logging
3. Baseline	8–12 weeks	Logging without aggressive alarms, weekly data review
4. Alarms and routine	Ongoing	Thresholds, owners, procedure for each alert
5. Expansion	Based on results	More machines, spectral analysis, CMMS integration

Use open protocols (Modbus, OPC UA, MQTT) so your data is not tied to a single vendor. If you are unsure which fits, see Modbus, OPC UA or MQTT: which protocol to choose. If the machine is old, machine retrofit: modernise or replace is also worth reading.

What are the most common mistakes?

• Lots of sensors, no plan. Fitting sensors across the whole plant without knowing which failure you are looking for on each machine produces data nobody looks at.
• Nobody owns the alerts. If an alert goes to a generic inbox, it ends up ignored. Each machine needs an owner and a defined action.
• Over-sensitive alarms. Ten false alarms a week destroy trust in the system.
• Not recording interventions. If you do not note what was found when the machine was opened up, you cannot validate or improve the thresholds.
• Relying on the cloud for everything. If logging and basic alarms do not work without internet, a connection outage leaves you blind.

Checklist before you start

• Criticality list done and 3–5 machines chosen
• Expected failure mode identified for each machine
• Variables and sampling rates defined
• Local data logging that works without internet
• Owner and procedure for each alert level
• A way to record what was found at each intervention

How we approach it at Kinetifab

We start by measuring: we choose the critical machines with your team, fit sensors using open protocols and local logging, and build the baseline before proposing alarms or expansion. The data is yours. You can see how we work in industrial automation or tell us about your case.