What sensors do I need for predictive maintenance? A practical guide

A bearing failure on a logistics conveyor in Ghent costs around €8,000 per hour of downtime. Caught three months in advance with a vibration sensor: €120 of preventive maintenance. The arithmetic is not controversial — the open question is which sensor goes on which machine, where exactly it should be mounted, and how often someone needs to look at the data. That is where most predictive maintenance programs succeed or quietly fail, and where the time of any maintenance director evaluating the move from preventive to predictive is best spent.

The four sensor families that matter

Roughly 90% of detectable failures in an industrial plant fall into four monitoring categories. Knowing what each family detects, and where it shines, prevents buying the wrong technology.

Vibration sensors (accelerometers, ISO 13373). The backbone of any serious program. A piezoelectric accelerometer mounted on a bearing housing detects the onset of inner-race spalling 2 to 6 months before catastrophic failure. The spectral signatures are specific: imbalance at running speed (1×RPM), misalignment at 2×RPM, bearing damage at much higher frequencies (BPFI, BPFO depending on geometry). ISO 13373 covers acquisition and analysis. Natural use cases: electric motors, centrifugal pumps, fans, gearboxes, spindles. Installed cost per point: €400 to €1,200, depending on whether it is a portable route point or wired into a continuous system.

Thermography and embedded thermocouples (ISO 18434). An IR camera spots hot points on switchgear, loose connections, overloaded motors and starved bearings. The practical rule: a +10°C delta versus a symmetric component or historical baseline is already an alert; +30°C demands immediate intervention. Classical thermographic inspection is periodic (monthly or quarterly by a level I/II certified technician), but fixed sensors exist for critical points. Use cases: electrical cabinets, transformers, motors, heat exchangers, furnaces.

Airborne and structural ultrasonics. Detects emissions above 20 kHz that the human ear cannot hear. An ultrasonic detector finds compressed-air or gas leaks above 40 kHz, partial discharges in electrical insulation, arcing in HV equipment, and bearing lubrication issues even before the accelerometer does. In a typical plant with a 7 bar compressed-air network, undetected leaks routinely account for 20% to 35% of compressor electrical consumption.

Oil and lubricant analysis. Combines periodic laboratory samples (spectrometry, ferrography, ISO 4406 particle counting) and, on critical machines, in-line viscosity, moisture and particle-count sensors. Detects abnormal wear from 50–200 ppm of metal particles depending on the component, water above 200 ppm, and additive degradation. Key applications: gearboxes, hydraulic systems, diesel engines, mineral-oil transformers.

Choosing the right sensor for each machine

There is no universal sensor. The choice starts from the dominant failure mode of the machine. A centrifugal pump typically fails through bearings, cavitation or misalignment — vibration handles 80% of the diagnosis. An electrical cabinet fails through loose contacts and overloads — thermography. A screw compressor combines vibration (gears and bearings) with oil analysis (air end). A hydraulic system depends almost entirely on fluid condition — in-line oil sensors.

A simple decision matrix to start from:

• Rotating equipment (motors, pumps, fans, centrifugal compressors): vibration as the primary sensor, monthly thermography as secondary.
• Switchgear and electrical cabinets: quarterly thermography. On critical installations, fixed IR windows allow inspection without de-energising.
• Compressed air and gases: ultrasonics. Full annual audit plus quarterly walks of critical lines.
• Gearboxes and reducers: vibration plus oil analysis. The combination catches both bearing faults and tooth wear.
• Hydraulics: in-line particle and moisture sensors, complemented with lab samples every 500–1,000 hours.

For genuinely critical machines (the pump that stops a whole production line) continuous combined monitoring is justified: permanent accelerometers, bearing thermocouples, and an in-line oil sensor. For the rest, a monthly route with a portable instrument is usually enough.

Where to install them and how often to read

Sensor placement determines the quality of the diagnosis. The rule for vibration: as close as possible to the bearing load zone, in the direction of the dominant force (radial for cylindrical rolling-element bearings, axial for thrust bearings), mounted on a machined surface with controlled torque. An accelerometer stuck via magnetic base on flaking paint introduces parasitic resonances and masks the real signatures. The thermography equivalent: always shoot from the same angle and at the same load, so the historical comparison stays valid.

Sampling rate depends on criticality. Continuous systems (wired into PLC or IoT gateway) record per-second and apply automatic alarms — appropriate for the top 5–10 critical machines. Portable systems on a monthly route cover the rest of the fleet effectively. Daily routes only make sense on machines already in known degradation, kept running until a planned shutdown.

The wireless versus wired debate has effectively been settled in practice: industrial wireless sensors (LoRa, sub-GHz, NB-IoT) now offer 3–5 year battery life and enough precision for trend monitoring. Wiring is reserved for high-resolution spectral analysis or environments with severe RF interference. Integration with the CMMS (computerised maintenance management system) closes the loop: a sensor alarm automatically generates a work order with history, identified spare and assigned technician.

Real ROI: a practical calculation

A typical case we see often in mid-sized Belgian and French plants. A food plant in Charleroi running 24/5: one sanitary pump critical for CIP cycles and five auxiliary motors on mixers and dosing pumps.

Initial investment: six wireless accelerometers at roughly €800 per installed point, an IoT gateway, a trend-analysis software licence and initial training. Total: about €8,000–€9,000 in year one. System upkeep and quarterly remote-analysis service: around €2,500/year.

Return is calculated against unplanned downtime. An unplanned CIP-pump stop means an incomplete cleaning cycle, food-safety risk and a delayed start of the next shift: typical loss of 6 hours of production at about €5,000/hour = €30,000. A catastrophic motor failure on a mixer is 4 hours at €5,000 plus a €3,500 replacement = €23,500.

Avoiding a single major stop in year one covers the investment with margin to spare. In practice, serious programs avoid 1 to 3 major stops per year and reduce the volume of small reactive interventions, because repairs are now planned with the right parts staged. Typical ROI: 3–6 months for the system, 4–8 months including training. Belgian and EU regulations on food safety and traceability (HACCP, CE marking on installations) reinforce the case: a stop caused by mechanical-failure cross-contamination carries regulatory cost on top of production cost.

The five most common implementation mistakes

After deploying programs across Belgian, French and Spanish sites for over a decade, the same five mistakes keep showing up.

Wrong sensor placement. Accelerometer mounted away from the bearing load zone, on thin sheet-metal housing, or near unrelated vibration sources (the next pump along, resonant pipework). The result is noisy spectra and masked real signatures.

Setting alarm thresholds without a baseline. A default "5 mm/s RMS" threshold will trip on healthy machines and stay quiet on degraded ones. Each unit needs 4–12 weeks of data in normal operation before reasonable alarms are set. ISO 17359 covers exactly this flow: failure-mode identification, parameter selection, baseline establishment, criteria tuning.

Acting on raw values instead of trends. A spot reading says little; what matters is the slope. A vibration rising from 2 to 3 mm/s in four weeks is more concerning than a steady 4 mm/s for months.

Forgetting calibration drift. Cheap sensors, harsh conditions (vibration, heat, dust) and two years later the absolute values no longer compare to the original baseline. Schedule recalibration or replacement every 2–3 years.

Nobody is looking at the dashboards. The error that invalidates all the others. Without a named technician reviewing alerts weekly and closing the loop with a work order, data accumulates and failures arrive on schedule as before.

When predictive isn't worth it

Honesty helps. Predictive monitoring does not apply universally. A 1.5 kW redundant sump pump in a pit that takes five minutes to access does not need an accelerometer — direct replacement is cheaper. An auxiliary fan with three parallel units and automatic failover does not either. Equipment with low replacement cost (below about €3,000) and immediate availability rarely justifies the sensor.

The practical rule: if the expected annual cost of unplanned stops on a given machine is below €5,000, planned corrective maintenance is usually more efficient. Predictive shines on machines where downtime is expensive or where reactive repair is physically slow (difficult access, long-lead-time spare).

Where to start

The honest recommendation for a plant starting its program: install accelerometers on the five most critical rotating machines, schedule a quarterly thermography inspection of switchgear, integrate everything with the CMMS and dedicate one hour per week to reviewing trends. With that foundation, the rest of the program is built on real data rather than intuition.

If you want a free audit of which sensors to install in your plant and an ROI calculation tuned to your operational reality, get in touch and we will review your critical equipment.

Frequently asked questions

How long before a sensor produces actionable insights? Between 4 and 12 weeks of baseline in normal operation. Before that, thresholds are indicative and alerts should be treated cautiously.

Can I retrofit sensors on old equipment? In the vast majority of cases, yes. Accelerometers mount on any housing with a machined surface or M6/M8 bolt. 20–30 year old machines are perfect candidates because they tend to be the most critical and the most poorly maintained.

What is the difference between condition monitoring and predictive maintenance? Condition monitoring is measuring current state (is this bearing hot?). Predictive maintenance projects the trend to anticipate the failure and plan the intervention. Without the trend, you have data but no prediction.