Predictive maintenance (PdM) means using measured condition data, such as vibration, temperature and motor current, to forecast when a machine is likely to fail so you can repair it at the right time rather than too late or too early. It pays off on a short list of critical assets with known failure modes, and only when the data actually reaches a person who can act on it. According to Siemens' True Cost of Downtime 2024 report, almost half of the large industrial firms it surveyed now have dedicated PdM teams, twice the proportion seen in 2019, and nine out of ten do some form of condition monitoring.
What is the difference between reactive, preventive, condition-based and predictive maintenance?
The four strategies differ in what triggers the work: a failure, the calendar, a measured limit, or a forecast. Most plants run a mix, and the sensible goal is to put each asset in the cheapest strategy that its failure consequences allow.
| Strategy | What triggers the work | Good for | Main weakness |
|---|---|---|---|
| Reactive (run to failure) | The machine breaks | Cheap, non-critical, easily replaced items | Unplanned stops, secondary damage, overtime |
| Preventive (time or run-hours) | A fixed interval | Lubrication, filters, wear items with predictable life | Replaces parts that still have life; misses random failures |
| Condition-based | A measured value crosses a limit (for example a vibration alarm) | Assets where one or two variables show degradation | Warns late if the limit is crude; no forecast of time left |
| Predictive | A trend or model forecasts remaining useful life | Critical rotating equipment with slow, detectable degradation | Needs good data, baselines and a process to act on alerts |
The US Department of Energy's Federal Energy Management Program published an Operations and Maintenance Best Practices Guide in 2002 that puts rough numbers on the progression. It estimated that preventive maintenance can save 12% to 18% over a purely reactive programme, and that a properly functioning predictive programme can save a further 8% to 12% over preventive alone. Treat these as dated, indicative estimates from a US facilities context, not as a promise for your plant. The same guide is candid about the downside: predictive equipment was expensive to start with, and staff training needed real funding.
What does unplanned downtime cost?
Siemens' report estimates that the world's 500 biggest companies lose almost $1.4 trillion a year to unplanned downtime, equivalent to 11% of their revenues. It puts the cost of a lost hour at $36,000 in fast-moving consumer goods and $2.3 million in automotive, and says SME manufacturers can lose up to $150,000 an hour at the top end. Large plants in its sample averaged 25 downtime incidents and 27 lost hours a month, down from 42 incidents and 39 hours in 2019.
Read those figures with care. They come from a survey run by a vendor that sells predictive-maintenance software, they are in US dollars, and they describe very large plants. Your own number is better: lost margin per hour, plus idle labour, scrap, expedited parts and any contractual penalties, multiplied by realistic hours lost per year on each asset. That calculation, not an industry headline, decides where PdM is worth the effort.
Which sensors and data do you need?
Most useful PdM data comes from three families of measurement, each catching different faults.

| Measurement | What it tends to reveal | Typical source | Easy to bring into a PLC? |
|---|---|---|---|
| Vibration (acceleration or velocity) | Bearing wear, imbalance, misalignment, looseness | Accelerometers, often with a dedicated analyser or gateway | Overall levels yes; raw waveforms usually no |
| Temperature and thermal imaging | Overheating bearings or windings, loose electrical connections | RTDs and thermocouples on the machine; handheld thermal camera on rounds | Yes for fixed probes |
| Motor current and power | Load changes, blockages, pump problems, some rotor and mechanical faults | Variable frequency drive diagnostics, current transformers | Yes, often over PROFINET or Modbus |
| Run-hours, starts, cycle counts | Wear that follows usage rather than time | The PLC itself | Yes, already there |
Vibration is the workhorse for rotating machines. The ISO 20816-1 standard sets out general conditions and procedures for measuring and evaluating machine vibration, which is why many sites express alarm limits as vibration velocity bands by machine class. A thermal camera on a monthly route is one of the cheapest ways to start, and it needs no integration at all. Other signals, such as ultrasound, oil condition and pressure or flow, matter for specific equipment.
One rule saves many projects: always record the operating state alongside the measurement. The same vibration level means different things at idle, at half load and at full speed, so a model that does not know the machine's state will raise false alarms.
Where do PLCs and SCADA fit in predictive maintenance?
The PLC and SCADA layer is where PdM data is cleaned, given context and turned into action, even though the heavy analysis often runs elsewhere. A PLC already holds run status, start counts, run-hours, drive current, temperatures, valve travel times and alarms, all time-aligned with what the machine was doing.
A typical architecture looks like this:
- Sensors and drives feed values to the PLC, or to a dedicated condition-monitoring gateway for high-rate vibration data.
- The PLC scales and timestamps signals, applies simple health logic (start counts, run-hour limits, high-high alarms) and publishes the data.
- SCADA and a historian trend the values, add context such as product, recipe and speed, and show alarms to operators.
- An edge device or analytics platform pulls the data, typically over OPC UA, runs anomaly detection and returns a health score. The OPC Foundation describes OPC UA as a platform-independent, service-oriented architecture that protects communication with encryption and certificates, which is why it is the usual bridge between the control network and analytics.
- The maintenance system receives a work order when a health score crosses a defined threshold.
Keep machine-learning workloads out of the PLC scan. Control and safety logic should stay deterministic and separate from experimental analytics. For deeper background, see our guides to OPC UA connectivity, SCADA systems and IO-Link smart sensors.
How does the machine learning actually work?
At a practical level, PdM machine learning is feature extraction followed by either anomaly detection or forecasting, and most plants should start with the simpler anomaly side. The steps are:
- Features: raw signals are reduced to numbers that describe health, such as RMS vibration, spectral band energy, temperature rise above ambient, or current compared with load.
- Baseline: a few weeks or months of normal operation across the real range of loads and products.
- Anomaly detection: an unsupervised model learns what normal looks like and flags departures. Scikit-learn's documentation describes this as novelty detection: the training data is not polluted by outliers, and you want to know whether a new observation is an outlier. Isolation Forest and one-class support vector machines are common starting points.
- Classification and remaining-useful-life models: these need labelled examples of faults or run-to-failure histories, and most plants have very few. Without them, a forecast is guesswork.
You can practise on public data before touching a plant. Case Western Reserve University's Bearing Data Center offers vibration recordings from a 2 hp test motor with faults seeded on the inner raceway, rolling element and outer raceway at diameters from 0.007 to 0.040 inches, at motor loads from 0 to 3 horsepower. Python, covered in our guide to Python for automation engineers, is enough to build a first detector.
The honest limits: models drift when products or speeds change, false alarms destroy trust faster than missed faults, and a clever model with no owner and no work-order route delivers nothing.
How should a UK plant adopt predictive maintenance?
Start small, prove value on a handful of assets and only then scale. A realistic sequence:
- Rank assets by consequence. Combine downtime cost per hour with how often each asset fails. Pick five to ten.
- Use the data you already have. Check PLC tags, drive diagnostics and the maintenance system's failure history before buying anything.
- Pilot on three to five assets. Begin with route-based thermal and vibration readings, then fit permanent sensors only where they earn their keep.
- Define the action. Who receives an alert, what they check, and what work order follows. Decide how you will count avoided failures and false alarms.
- Integrate and secure. Bring signals into SCADA and the historian over segmented networks, and keep analytics access read-only.
- Scale or stop. If the pilot did not avoid a failure or improve planning, fix the process before adding more sensors.
For smaller manufacturers, Made Smarter, described on its website as a national movement backed by businesses and the UK government, says on its website that it helps manufacturing SMEs digitalise their processes through regional delivery partners, and what is available varies by region. Check your local partner before budgeting.
How EDWartens can help
PdM sits on top of solid control, instrumentation and troubleshooting skills. The EDWartens Professional Module is a five-day live online class in Siemens TIA Portal covering PLC programming, HMI and WinCC SCADA, followed by self-paced Basic Engineering modules that include instrumentation and sensors and maintenance and reliability. It does not teach machine-learning-based predictive maintenance. To build the groundwork for free, try the process control simulator and the PLC electrical troubleshooting course, then see the Professional Module or sign up.




