Condition monitoring: a practical guide to predictive maintenance
Director & MLOps Lead
Predictive maintenance specialist, industrial data analysis, vibration-based condition monitoring, applied AI for manufacturing and automotive operations
Condition monitoring is the measurement of machine health (vibration, temperature, oil, sound and electrical current) to spot developing faults before they cause a failure. It is the data source for predictive maintenance: instead of servicing on a fixed schedule or after a breakdown, you plan work when the measurements show it is needed.
Key takeaways
- Start with a criticality analysis. Monitor the assets whose failure stops production, risks safety or takes months to replace.
- Match technique to failure mode: vibration for rotating machines, thermography for electrical gear, oil analysis for gearboxes and hydraulics, ultrasound for leaks and early bearing wear.
- The US Department of Energy's O&M guide cites 8–12% savings for predictive over preventive maintenance, and 30–40% where a site relies mainly on reactive repair.
- Value comes from acting on alerts. Connect monitoring to your CMMS so an alert becomes a work order with a due date.
- Keep data in formats and platforms you control. Managed analytics services can be withdrawn, as Amazon Lookout for Equipment was in October 2026.
What is condition monitoring?
It is the regular or continuous measurement of parameters that change as a machine degrades. A bearing that starts to wear vibrates at characteristic frequencies long before it seizes. A loose electrical connection runs hot before it arcs. Oil from a gearbox carries metal particles as gears wear.
Measurements are compared with a baseline and with alarm limits. When a trend crosses a limit, the maintenance team investigates and plans a repair, ideally in a planned stop and with parts already on the shelf.
How does it relate to predictive maintenance?
Monitoring provides the data; predictive maintenance is the strategy that uses it. The terms are often used together, but they are not the same thing.
| Strategy | When work is done | Strength | Weakness |
|---|---|---|---|
| Reactive (run to failure) | After breakdown | No upfront cost; fine for cheap, non-critical items | Unplanned downtime, secondary damage, overtime and rush parts |
| Preventive (time-based) | On a calendar or run-hour schedule | Simple to plan | Parts replaced too early or too late; failures between intervals |
| Condition-based | When a measurement crosses a limit | Work done when needed | Needs sensors, routes and trained analysts |
| Predictive | When trends and models forecast a failure window | Longest warning, best planning | Needs history, data quality and analytics skills |
Most plants use all four. Run-to-failure is right for a light fitting; predictive is right for a compressor that stops the line.
Which monitoring techniques should you use?
Pick techniques by the failure modes you want to catch. Most programmes combine two or three.
Vibration analysis
Accelerometers on bearing housings measure vibration. Spectrum analysis then separates imbalance, misalignment, looseness, gear mesh problems and bearing defects by their frequencies. It is the core technique for motors, pumps, fans, compressors and gearboxes, and it often gives weeks to months of warning on bearing wear.
Infrared thermography
Thermal cameras show hot spots in switchgear, motor control centres, transformers, steam systems and refractory linings without contact. A scan under load finds loose or overloaded connections quickly.
Oil and lubricant analysis
Laboratory or online sensors measure wear metals, particle counts, water, viscosity and additive depletion. It shows internal wear in gearboxes, engines, hydraulics and turbines without opening them, and it tells you when oil really needs changing.
Airborne and structure-borne ultrasound
Ultrasonic detectors pick up high-frequency sound from compressed air and steam leaks, faulty steam traps, electrical arcing and corona, and early-stage bearing friction. It is cheap to start and often pays back first through energy savings on air leaks.
Motor current and electrical signature analysis
Measuring current and voltage at the motor control centre can reveal broken rotor bars, eccentricity, supply problems and some driven-load faults. It is useful when the motor itself is hard to reach.
Process data
Pressures, flows, temperatures and power draw already in your control system are free indicators. A pump drawing more power for the same flow is telling you something. Combining process data with the techniques above improves diagnosis.
What does a monitoring programme cost and return?
The best-known public figures come from the US Department of Energy's Operations & Maintenance Best Practices Guide (Release 3.0). It estimates that a properly functioning predictive programme saves 8–12% compared with preventive maintenance alone. Sites that rely heavily on reactive maintenance could see savings above 30–40%.
The same guide quotes industry surveys of average results: maintenance costs down 25–30%, breakdowns down 70–75% and downtime down 35–45%. Treat these as directional. They date from a 2010 guide and from surveys, not controlled studies. The guide is also frank about cost: much of the equipment costs over USD 50,000, and training staff to use it well needs considerable funding.
Wireless sensors and cloud software have since lowered the entry cost for continuous vibration and temperature monitoring. Analyst time, not hardware, is now often the largest cost. Build your own business case from your downtime cost per hour, failure history and spare-part lead times.
How do you implement a monitoring programme?
- Criticality analysis: rank assets by safety, production, environmental and cost impact of failure, plus repair lead time. Monitor the top tier first.
- Failure modes: for each critical asset, list how it fails and which measurement detects each failure early. This links the technique to a real risk.
- Route-based or continuous: use monthly or quarterly routes with portable instruments for most assets, and permanent sensors for critical, hard-to-reach or fast-failing machines.
- Baselines and alarm limits: record healthy readings after installation or overhaul and set alarm levels using the relevant standard (for example ISO 20816 for vibration severity) and the manufacturer's guidance.
- Pilot: run on 10–20 assets for a few months, tune limits, and check that alerts lead to sensible actions.
- Integrate with the CMMS: each confirmed alert should create a work order with a priority and a recommended action.
- Train and certify: vibration and thermography analysts should be qualified, commonly to the ISO 18436 series.
- Review: track avoided failures, false alarms and repeat faults, and feed findings into root cause analysis.
ISO 17359 gives general guidelines for setting up a programme along these lines, and ISO 13374 covers how monitoring data is processed and communicated.
What does the system architecture look like?
A modern continuous-monitoring system has five layers:
- Sensors: accelerometers, temperature probes, current clamps, ultrasonic and oil quality sensors.
- Edge: data acquisition units or gateways that sample, compute features such as RMS and spectra, and buffer data if the network drops.
- Network: wired fieldbus or Ethernet, industrial wireless, private 5G or cellular. Keep it separate from office IT and follow your OT security policy.
- Platform: a time-series store and analytics, on premises or in the cloud. AWS IoT SiteWise, for example, is a managed service to collect, store, organise and monitor data from industrial equipment at scale.
- Applications: dashboards, alerting, analyst workbenches and the CMMS integration.
Plan for vendor change. AWS announced it would discontinue support for Amazon Lookout for Equipment on 7 October 2026, after which its console and resources are no longer accessible. Store raw and feature data in open formats in storage you control, and keep models and alarm logic documented, so you can change analytics tools without losing history.
Where do AI and machine learning help?
Machine learning is good at learning normal behaviour across many signals and flagging deviations, especially on fleets of similar machines. It helps an analyst triage hundreds of assets.
It does not replace physics-based diagnosis. A model can say a pump is behaving unusually; an analyst still has to say it is cavitation and not a bearing. Models also need clean, labelled history, which most plants have to build first. Start with rule-based alarms and trending, then add anomaly detection where data volume justifies it.
What are the common pitfalls?
- Monitoring without action: alerts that never become work orders waste the investment.
- Too many alarms: poorly set limits cause alert fatigue. Tune them during the pilot.
- Wrong technique for the failure mode: temperature alone finds bearing faults late.
- No skills plan: data needs qualified people or a service partner to interpret it.
- Weak OT security: new sensors and gateways add network entry points. For operators covered by NIS2, they belong in your risk assessment.
How Opsio helps
Opsio builds the data side of monitoring programmes: IoT predictive maintenance platforms, cloud data pipelines and dashboards, CMMS integration and anomaly detection, plus IoT asset management and 24/7 operations of the platform. We work alongside your reliability engineers and certified analysts rather than replacing them. Our teams are in Karlstad, Sweden and Bangalore, India.
Frequently asked questions
What is the difference between condition monitoring and predictive maintenance?
Monitoring is measuring machine health. Predictive maintenance is the maintenance strategy that uses those measurements and their trends to schedule work before a failure. You cannot do the second without the first.
What are the main types of monitoring techniques?
Vibration analysis, infrared thermography, oil analysis, ultrasound, motor current analysis and process data trending. Many plants also use NDT methods such as ultrasonic thickness measurement for corrosion.
Which machines should be monitored first?
Those whose failure stops production, creates a safety or environmental risk, or has a long repair or spare-part lead time. A criticality analysis ranks them.
Is continuous monitoring better than periodic routes?
Not always. Continuous sensors suit critical, remote or fast-failing machines. Monthly or quarterly routes with portable instruments are cheaper and enough for most balance-of-plant equipment.
How quickly does a programme pay back?
It depends on your downtime cost and failure history. The first avoided failure on a critical asset often covers a pilot. Calculate it from your own data rather than relying on industry averages.
Related reading: emerging NDT technologies, which covers permanent sensors from the inspection side.
Related reading: How Predictive Maintenance Services Transform Modern Manufacturing.
Related reading
Related Articles
About the Author
Director & MLOps Lead
Vaishnavi leads machine learning operations initiatives at Opsio, enabling ML and predictive capabilities for industrial and automotive operations. Her expertise spans predictive maintenance, industrial data analysis, vibration-based condition monitoring, and applied AI — with a focus on practical, experiment-driven solutions designed for real operational environments.
Editorial standards: This article was written by a certified practitioner, named in the byline above. We revise it when the technology or regulation it covers changes; the dates above show the latest update. Opsio maintains editorial independence — we recommend solutions on technical merit, not commercial relationships.
