Predictive Maintenance With Vibration Monitoring: Where to Start With Ten Machines, Not a Thousand
November 11, 2026 · 6 min read · Etzal Group Team
A food factory buys two hundred wireless sensors and a platform licence for predictive maintenance. A year later, sensors are installed on forty machines, the dashboard shows thousands of red alerts, maintenance has stopped looking at it, and a critical blower fails without anyone noticing its trend. The technology worked. The project started with a purchase and not with a question. Vibration monitoring pays when it starts small, on the machines that matter, with alarms that people trust.
What vibration tells you
A healthy rotating machine vibrates in a stable pattern. As a fault develops, the pattern changes, usually long before the machine stops.
Vibration is good at revealing:
- Imbalance: a fan with deposits on its blades, a worn impeller.
- Misalignment between motor and driven equipment.
- Looseness: foundation bolts, mountings.
- Bearing wear, often the earliest and most valuable signal.
- Gear problems in gearboxes.
It is weak or blind for:
- Sudden electrical failures.
- Process problems unrelated to rotation, such as a blocked filter, unless they change the load.
- Slow or intermittent machines, which need different methods.
- Faults that develop in minutes.
Pair vibration with temperature at the same point. The two together are far more convincing than either alone.
Choosing the first ten
List candidate machines and score each.
- Consequence of failure: does the line stop? For how long? What does an hour cost?
- History: how many unplanned stops in the past two years?
- Repair cost and lead time for parts.
- Detectability: is it rotating equipment running steadily?
- Access: can someone safely mount a sensor and reach it later?
Pick the ten with the highest product of consequence and history. Typically: main process pumps, blowers and fans, compressors, critical gearboxes, large motors.
Do not instrument spare-backed, cheap equipment first. A small pump with an installed standby is not where the money is.
Sensors and data
Two levels of measurement
- Overall vibration level: a single number, such as velocity in millimetres per second, plus temperature, sent every few minutes or hours. Inexpensive, long battery life, enough to see a trend.
- Spectrum or waveform: detailed data that lets a specialist identify which fault is developing. More data, more power, more cost.
A sensible start: overall level and temperature on all ten, with the ability to capture a spectrum on demand or periodically for the most critical.
Practical choices
- Wireless battery sensors are quick to install. Check the stated battery life at your reporting interval and the range through your plant's steel and concrete.
- Wired sensors suit the most critical machines and hazardous or hot locations.
- Mounting matters more than the brand. Fix the sensor rigidly, close to the bearing, on a clean flat surface. A sensor on a cover plate measures the cover plate.
- Hazardous areas need suitably certified equipment.
- Record the installation: machine, position, orientation, photo, date.
Baseline before alarms
Each machine has its own normal. Two identical pumps can run at different vibration levels and both be healthy.
- Collect two to four weeks of data in normal operation.
- Note operating conditions. Vibration changes with speed and load. A machine on a variable speed drive needs comparisons at similar speed.
- Set the baseline per machine, per measuring point.
- Use published severity guidance for machine classes, such as the ISO 20816 series, as a sanity check for absolute levels.
Alarms based on a generic number for all machines are the source of the thousand red alerts.
Alarm levels people trust
Three levels are enough.
- Watch: clearly above baseline, sustained. No action beyond looking more often.
- Plan: rising trend, well above baseline. Inspect and schedule work at the next convenient stop.
- Act: high or rising fast. Inspect now and decide whether to stop.
Rules that prevent false alarms:
- Require persistence. A level must hold for a period, not spike once.
- Ignore readings while the machine is stopped or starting.
- Alert on rate of change, not only on level.
- Route each alarm to a named person, with the machine, the point and the trend in the message.
An alarm that fires weekly and means nothing teaches the team to ignore the one that matters.
From alarm to work order
Data without action is decoration.
- Alarm received by the maintenance planner.
- Check: look, listen, measure by hand, review the trend.
- Diagnose: if needed, capture a spectrum and have it analysed.
- Plan: create a work order, order parts, choose a stop.
- Repair.
- Confirm: vibration returns to baseline.
- Record: what was found, with a photo of the failed part.
Step 7 builds the evidence and, over time, the failure history that any later analytics will need.
Proving the value in ninety days
Agree on the measures before starting.
- Catches: faults found early and repaired in a planned stop.
- Avoided cost per catch: hours of downtime avoided multiplied by the cost per hour, plus secondary damage avoided, estimated conservatively and agreed with production.
- False alarms: how many, and the trend.
- Missed failures on monitored machines.
One documented catch on a critical machine often pays for the pilot. If ninety days pass with no events at all, extend the pilot. Do not expand on faith.
Scaling up
Expand when the first ten have:
- Alarms the team trusts.
- A working path from alarm to work order.
- At least a few documented catches.
Then add machines in order of the same ranking. Consider analytics and machine learning when there are enough machines that people cannot review trends by hand, and enough recorded failures to learn from.
Where the data should live
- Connected to the maintenance system, so alarms create work requests.
- With process data, such as speed and load, for context.
- Owned by you. Confirm in the contract that you can export raw data and that it stays available if you change supplier.
Common mistakes
- Buying sensors for the whole plant first.
- Instrumenting easy machines and not critical ones.
- Poor mounting.
- Generic thresholds for every machine.
- No baseline.
- Alarms to a shared inbox nobody owns.
- No link to work orders.
- Expecting machine learning to work without failure history.
- Not recording what was found at repair.
- Judging success by sensors installed, not failures avoided.
We design and deliver condition monitoring, from sensor selection to alarms and integration with maintenance systems, as part of IoT and Control Systems. If you want to start with the ten machines that matter, contact us.
Summary
Start vibration monitoring on the ten rotating machines whose failure costs the most and has happened before. Measure overall vibration and temperature, mount sensors properly, build a baseline per machine, and use three alarm levels with persistence and a named owner. Connect alarms to work orders, record what each repair finds, and judge the pilot by documented catches and false alarm rate before scaling or adding machine learning.
Frequently asked questions
Which machines should get vibration sensors first?
Rotating equipment whose failure stops production or is expensive to repair, and which has failed before: motors, pumps, fans, blowers, gearboxes and compressors. Rank by the cost of an unplanned stop multiplied by how often it has happened. Start with the top ten, not with every machine in the plant.
Does predictive maintenance need artificial intelligence?
Not to begin. Most early value comes from trends and thresholds: overall vibration rising against the machine's own baseline, and temperature rising with it. Machine learning becomes useful later, when there are many machines and enough failure history to learn from. A project that waits for a model before delivering alarms usually delivers nothing.
How far ahead can vibration monitoring warn of a failure?
It depends on the fault and the machine. Bearing wear, imbalance, misalignment and looseness typically develop over weeks or months and are well suited to early detection. Sudden failures, such as an electrical fault or a foreign object, may give little or no warning. Vibration monitoring reduces surprises. It does not remove them.
Related service
WhatsApp AI Chatbot
We build, train, and run a WhatsApp AI agent that answers customers 24/7, qualifies leads, and follows up.
