BlogIoT and edge AI

Predictive maintenance with IoT and AI: a practical guide

Predictive maintenance uses sensor data from a machine to spot a developing fault early enough to plan the repair. In practice it starts with condition monitoring and anomaly detection on a few signals, because most plants have too few recorded failures to train a failure-prediction model on day one. Prove it on one class of asset, send the alerts into the maintenance system, then extend.

From reactive to predictive

Maintenance strategies differ in what triggers the work.

Strategy Work is triggered by What you need Where it fits
Reactive (run to failure) The breakdown Spares and a fast response Cheap equipment, or anything with a standby
Scheduled (preventive) The calendar or running hours A plan and the discipline to follow it Wear that tracks usage closely: filters, belts, lubrication
Condition-based A measured value crossing a limit Sensors or inspection rounds, and thresholds Assets that show symptoms before they fail
Predictive A model’s view of a developing fault or of remaining life Continuous data, history, analytics and a feedback loop Critical assets with enough units or history to learn from

These are not rungs on a ladder that every asset must climb. Run to failure is the correct strategy for a great deal of equipment, and a plant that puts sensors on everything has wasted money on the extractor fan in the canteen. The ranking question is simple: what does an unplanned stop of this asset cost, and does the asset give a measurable warning? ISO 17359, the general guideline for condition monitoring, starts in the same place, with an audit of which equipment matters and how it fails before any measurement is chosen.

Most of what is sold as predictive maintenance is condition-based maintenance with better statistics. That is fine. It is where the value starts.

Which signals matter?

Each signal sees different faults at different lead times. Maintenance engineers talk about the P-F interval: the time between the point where a fault becomes detectable and the point where the machine stops doing its job. A signal is useful only if that interval is longer than the time you need to plan a repair.

  • Vibration is the main signal for rotating machines. Imbalance shows at running speed, misalignment at one and two times running speed, and rolling bearing defects as repeating high-frequency impacts that envelope analysis can pick out long before the bearing gets noisy or hot.
  • Temperature is cheap and late. A bearing that is hot to the touch is near the end. Use it as a backstop and for electrical connections, not as the early warning.
  • Motor current needs nothing mounted on the machine, only a clamp in the control cabinet. Current signature analysis can reveal broken rotor bars, and changes in load point to trouble in the driven equipment.
  • Acoustic and ultrasound sensors catch lubrication problems, leaks and early bearing wear, and work where a housing cannot be reached.
  • Process data already in the control system, such as pressure, flow, speed and valve position, supplies the context. Without it you cannot tell a fault from a change in duty.

We start with vibration and the process data that already exists, and add current where motors are hard to reach. Add nothing you cannot name a use for.

Sensors and gateways

A few choices decide whether the data will be worth analysing.

Sensor type. Piezoelectric accelerometers are the established choice, with wide bandwidth and low noise at high frequencies. MEMS accelerometers cost less, use little power and are good enough for most pumps, fans and motors. Either way, bandwidth matters more than brand: seeing bearing faults needs several kilohertz of usable bandwidth and a sampling rate more than twice that. A sensor that reports one overall number per hour will find only the problems a technician would have heard anyway.

Mounting. Fix the sensor rigidly to the bearing housing, as close to the load path as you can get. A sensor on a guard or a thin cover measures the cover.

Wired or wireless. Battery-powered wireless nodes install in minutes. They wake up periodically, capture a few seconds, and send features or a spectrum. That suits faults that develop over weeks. Machines that are critical or that fail fast justify wired sensors and continuous capture.

The gateway. It collects the readings, computes spectra and features locally, buffers when the network drops and forwards the results. This is edge AI in its plainest form. MQTT is the usual transport, and Sparkplug adds birth and death messages on top so the receiving system knows whether a device’s data is live or stale. Process values usually come from the PLC over OPC UA, or Modbus on older equipment.

The asset register. Every sensor must be tied to an asset ID, a measurement point and an orientation. It is dull work, and getting it wrong is the most common cause of confusion six months in.

The data problem: too few failures

Supervised failure prediction needs examples of failure. A well-run plant has few. A given pump model may have failed three times in five years, each time differently, and nobody wrote down the exact date. Public datasets exist, such as the bearing and turbofan sets in NASA’s prognostics data repository and the Case Western Reserve bearing data, where faults were machined into the bearings of a 2 hp motor. They are good for learning the methods. They will not transfer to your gearbox.

So start with anomaly detection: learn what normal looks like for each machine in each operating state, and flag departures.

  1. Baseline. Record several weeks that cover the normal range of speed and load.
  2. Features. From each capture compute overall velocity, crest factor, kurtosis, energy in bands around the running-speed harmonics, and peaks in the envelope spectrum. Engineers can read these, which matters when an alert has to be believed.
  3. Operating state. A pump at 60 per cent speed is a different machine from the same pump at full speed. Group the data by state, or model each feature as a function of speed and load. Skipping this step is the largest single source of false alarms we see.
  4. Model. Rolling statistical limits per feature and per state go a long way. Isolation Forest is a sensible next step when features need to be judged together: it isolates unusual points with random splits and needs no failure labels. Autoencoders on spectra can wait until the simpler methods have clearly run out.
  5. Absolute limits. Keep the vibration severity zones from ISO 20816-3, the standard for industrial machines above 15 kW, as a safety net. A machine that was already in poor health during the baseline will otherwise look normal for ever.

Labelled faults accumulate from confirmed alerts. Once there are a few dozen for one fault type, a classifier that names the fault becomes realistic, and that is ordinary predictive ML. Remaining-useful-life estimates come last, and only for failure modes with enough run-to-failure history. Be suspicious of anyone promising them in the first month.

A worked example: one cooling-water pump

Take a centrifugal pump driven by a four-pole motor on a 50 Hz supply. It turns at about 1,480 rpm, or 24.7 Hz. The impeller has six vanes, so the vane-pass frequency is near 148 Hz.

We would fit accelerometers on the motor’s drive-end bearing housing and the pump’s inboard bearing housing, and a current clamp on one phase. From the PLC we take discharge pressure, flow and the running signal. The node captures a few seconds of vibration at 20 kHz or more every 15 minutes while the pump runs.

Six weeks of baseline give limits for each feature at each duty point. Then, in week eleven, the envelope spectrum from the pump bearing shows a small peak near 89 Hz, about 3.6 times running speed. It is not a whole multiple of shaft speed, so it is not a harmonic of imbalance or misalignment, and it matches the outer-race defect frequency calculated from the bearing’s geometry. Overall vibration is still well inside its limits and temperature has not moved. Over the next ten days the peak grows, harmonics of it appear and the anomaly score stays above the alert threshold on three consecutive captures.

The alert that reaches the planner names the asset and the measurement point, states the suspected fault, shows the trend and suggests an inspection at the next planned stop. The bearing is changed during a scheduled shutdown. The fitter finds spalling on the outer race and records it on the work order. That record is the programme’s first labelled example.

The same pump can tell a different story. Rising vane-pass vibration and broadband noise together with falling discharge pressure point to cavitation or impeller wear, not to a bearing. This is why the process data is worth the trouble of collecting.

Alert fatigue

More projects die here than in modelling. Twenty alerts a day are ignored within a month, and the one that mattered is among them.

  • Raise one alert per asset and fault, not one per sensor or per feature.
  • Require persistence across several consecutive captures.
  • Gate by operating state: no alerts during start-up, shutdown or cleaning cycles.
  • Suppress alerts while the asset is under maintenance. The maintenance system already knows when that is.
  • Use two or three severity levels with different routes: advisories go to a weekly review list, urgent alerts go to a phone.
  • Attach the evidence and a suggested action to every alert.
  • Ask the planner to mark each alert as confirmed, false or unclear, and track the share of alerts that led to a finding.

Agree an alert budget with the maintenance team, meaning the number they can review properly in a week, and tune to it. A system that stays inside the budget gets read.

Connecting to the maintenance system

An alert in a separate dashboard is one more place to look, and nobody has time for that. Send it into the CMMS or asset management system the planners already use. Begin by creating a notification or an inspection request, and let a person turn it into a work order. Automatic work orders can come once the alerts have earned trust.

The return path matters as much. Pull back work order status, failure codes, parts used and the fitter’s findings, because without them the model never learns whether it was right. Check spares lead times too. Three weeks of warning is worth little if the bearing takes six weeks to arrive, and that should influence which failure modes you target first.

Expect this data plumbing between sensors, asset register and maintenance system to take more effort than the model.

How to start: one asset class

  1. Pick one class of asset with at least a dozen similar units, a known cost of failure and failure modes that give a measurable warning. Pumps, fans and conveyor drives are common first choices.
  2. Write down the failure modes you expect to catch, and the ones you will not.
  3. Define success before installing anything: faults detected and confirmed on inspection, warning time against planning time, false alerts per week, failures missed.
  4. Install, record a baseline for four to six weeks, then run with alerts going to a named person.
  5. Keep going for at least three months, longer for slow failure modes. Review every alert and every failure, including the ones the system did not see.
  6. Put your own numbers on the result: downtime hours, call-outs and secondary damage avoided, at your plant’s costs. Generic savings percentages come from other people’s plants.
  7. Decide whether to extend to the next asset class, change the approach or stop.

The first decision is the asset class, not the algorithm, and the discipline is the same as for any AI proof of concept that has to survive real data. Sensor selection, gateways, models and the maintenance integration are what our IoT and edge AI team builds.

Questions

Asked often.

What is the difference between condition-based and predictive maintenance?

Condition-based maintenance acts when a measured value, such as vibration or temperature, crosses a limit. Predictive maintenance goes a step further and uses models on the same data to recognise a developing fault earlier or to estimate how long the machine can keep running. In practice most predictive programmes begin as condition-based monitoring with better statistics.

How much failure data do you need for predictive maintenance?

None to begin with. Anomaly detection learns what normal operation looks like from a few weeks of healthy data and flags departures from it. Naming the fault type needs a few dozen confirmed examples per fault, and estimating remaining useful life needs run-to-failure histories that most plants do not have for most machines.

Which sensors are used for predictive maintenance on motors and pumps?

Vibration sensors on the bearing housings do most of the work, because imbalance, misalignment and bearing wear all show up in vibration before anything else. Temperature is a cheap backstop, a current clamp in the control cabinet sees electrical and load problems, and pressure and flow from the control system give the context needed to interpret the rest.

How long does a predictive maintenance pilot take?

Plan for three to six months on one class of asset. The first four to six weeks go on installing sensors and recording a baseline across normal operating conditions. The rest is needed for real faults to develop, for alerts to be checked against inspections and for the team to tune thresholds until the false alarm rate is acceptable.

Not on the list? Ask us directly

Keep reading

More notes from the build.

All blogs

Let’s build your stack.

Send a brief and we’ll reply with next steps.