Predictive Maintenance With AI and Machine Learning: How It Works

AI predictive maintenance uses sensor data and machine learning to detect a developing fault and estimate when a machine will fail, so the repair is planned before a breakdown. It works when the fault gives early warning in a measurable signal, the data is collected under known conditions, and someone owns the alert.

In this lesson by RealPars, predictive maintenance is explained from first principles. It sits in the use-case map module of our free AI for Industrial Automation course, where the task is to place a proposal among the use cases and name the label each one needs. This article covers the concepts. If you want to build a model today, the hands-on companion is Predict Machine Failure From Sensor Data in Google Colab.

Four ways to maintain a machine

Predictive maintenance is one strategy among four, and it is not always the right one.

• Reactive: run it until it breaks. Sensible for a cheap fan with a spare on the shelf.
• Preventive: service it on a calendar or after a number of run hours. Works when wear tracks usage, as with filters or belts.
• Condition-based: measure something and act when it crosses a limit, such as vibration velocity above an alarm level.
• Predictive: use history and models to forecast the failure, including how long is left.

Most plants use all four. The question is which assets justify the extra sensors, data work and attention that the last one needs.

The P-F curve

The P-F curve is the idea everything else rests on. P is the point where a failure first becomes detectable. F is functional failure, when the machine can no longer do its job. The time between them is the P-F interval, and it is the warning you have to plan the repair.

Different measurements see the same fault at different points on the curve. On a rolling-element bearing, vibration and ultrasound usually show damage well before the bearing runs hot, and heat and audible noise often come late. So the choice of sensor decides how much warning you get. A temperature probe on a bearing housing may give you hours; an accelerometer may give you weeks.

The data a predictive maintenance model uses

• Vibration from accelerometers: the main signal for bearings, imbalance, misalignment and looseness on rotating machines. Severity is commonly judged against the ISO 20816 series.
• Motor current: current signature analysis can reveal broken rotor bars and some load faults without touching the machine.
• Temperature: winding and bearing temperatures, and thermography of electrical panels.
• Oil analysis: wear metals and contamination in gearboxes and hydraulics.
• Process context from the PLC or SCADA: speed, load, product type and run state. Without this, a model mistakes a change of recipe for a fault. Getting that context lined up with the sensor data is most of the work; see data engineering for industrial AI.
• Maintenance records: work orders from the maintenance system, which become your labels.

Four families of AI predictive maintenance models

Anomaly detection learns what healthy looks like and flags departures from it. It is the usual first step because it needs no failure examples, only clean data from normal running across the loads the machine sees. In machine learning terms this is novelty detection, and scikit-learn's guide to novelty and outlier detection describes the standard methods, such as one-class SVM and local outlier factor.

Fault classification names the fault: bearing, imbalance, misalignment. It needs labelled examples of each, which most plants do not have in quantity at the start.

Remaining useful life estimates how long until failure, by regression on degradation trends or by survival analysis. It needs run-to-failure histories. NASA's C-MAPSS turbofan data is the standard practice set for this.

Forecasting a condition indicator forward in time, and alerting when the forecast crosses a limit, sits between condition-based and predictive maintenance and is often the easiest to explain to a maintenance team.

Where machine learning helps, and where it does not

A fixed vibration alarm on one pump, set from a standard, catches many faults with no machine learning at all. Start there. Machine learning earns its place when:

• one machine runs at many speeds and loads, so a fixed limit is either too tight or too loose;
• several signals together tell the story, and no single threshold does;
• you have many similar assets and want to compare each one with its peers.

A model of how the machine should behave, compared with how it does, is also the idea behind using a digital twin in manufacturing as a fault detector.

How a project runs

Choose one critical asset with an owner who will act on the alert. List its failure modes, ideally from an FMEA, with what each costs. For each mode, pick the sensor that sees it earliest on the P-F curve. Collect healthy data across normal operation before you model anything. Turn raw signals into condition indicators such as RMS velocity, kurtosis or the energy in a band around bearing defect frequencies; these are easier to trend and explain than raw waveforms.

Then start with anomaly detection, add classification when you have labels, and decide where the model runs: beside the machine, as described in edge AI in industrial automation, or centrally. Send alerts only when the condition persists over several readings, so one noisy sample does not wake anyone up. Finally, count what matters: faults caught in time, false alarms, and the cost avoided.

Mistakes that sink pilots

• Judging on accuracy. Failures are rare, so a model that always says "healthy" scores high accuracy and catches nothing. Use recall and precision on the failure class.
• Leakage. A column that records what happened, or a random split that puts the same machine in training and test, inflates the score. Split by machine or by time.
• Ignoring operating context. A load change looks like a fault if the model does not know about it.
• Alerts with no owner. A correct prediction nobody acts on saves nothing.

Standards worth knowing

ISO 17359 gives general guidelines for condition monitoring and diagnostics of machines. ISO 13374 covers how monitoring data is processed and communicated. ISO 13381-1 covers prognostics. The ISO 20816 series sets out how machine vibration is measured and evaluated. You do not need to memorise them, but a business case that cites them is taken more seriously.

Learn predictive maintenance free

Predictive Maintenance with Machine Learning runs from the P-F curve and failure modes through spectra, condition indicators, anomaly detection, fault classification, remaining useful life on NASA's engines, and alerts that get acted on. It needs Machine Learning with Python first. For the maintenance side, Vibration Analysis Training teaches reading spectra, and AI for Industrial Automation builds the business case. All are on the free AI courses for engineers list, and a free account saves your progress.

The courses are free in full. The optional EDWartens Certificate of Completion for an intermediate course is a small one-off fee, with a code anyone can check at edwartens.com/verification. It is not a vendor certification or an accredited qualification.

Take the free course

Free

AI and machine learning · Intermediate · Free

Predictive Maintenance with Machine Learning

Free

AI and machine learning · Intermediate · Free

AI for Industrial Automation

Free

Condition monitoring · Intermediate · Free

Vibration Analysis Training

Free

AI and machine learning · Beginner · Free

Machine Learning with Python