Key Points
- AI predictive maintenance learns how each machine normally behaves and flags changes in its vibration, current, temperature, and acoustic signals long before a fixed alarm threshold would trip.
- The core work happens in feature extraction: RMS, kurtosis, and FFT spectra turn raw sensor data into patterns a model can match to specific failure modes such as bearing wear, winding degradation, and cavitation.
- ROI comes from converting emergency failures into planned repairs; you can estimate it with four inputs your plant already tracks: failure count, cost per failure, detection rate, and program cost.
A fixed vibration alarm on a pump bearing usually trips when the bearing is already close to the end of its life. By then, the planner has days, not weeks, to order parts and schedule a stop.
AI predictive maintenance, one of the most established uses of industrial AI, closes that gap. Instead of waiting for one number to cross one line, a model compares the machine's current signals against its own learned baseline and against known fault signatures, so a developing defect shows up while the reading is still "in range."
How AI Differs from Traditional Predictive Maintenance
Traditional predictive maintenance relies on fixed alarm limits and periodic expert review; AI predictive maintenance learns each asset's normal behavior and detects deviations and fault patterns continuously. The practical difference is earlier warning and fewer alarms that turn out to mean nothing.
Traditional predictive maintenance still catches many faults. Its limits are frequency and scale: a technician with a handheld collector might visit a fan once a month, and a single overall vibration value cannot separate a bearing defect from a loose base.
| Factor | Threshold alerting | AI pattern recognition |
|---|---|---|
| Decision logic | One value compared to one fixed limit | Many features compared to a learned baseline and to fault signatures |
| Baseline | Generic limit, often from a standard or OEM sheet | Specific to each asset, its speed, and its load |
| Typical warning stage | Late, when overall levels are already high | Earlier, when patterns change but overall levels may still look normal |
| Load and speed changes | Can trigger false alarms or mask real faults | Context from speed and load is part of the model |
| Diagnosis | "Vibration high" with no cause | Likely failure mode, such as outer race defect or misalignment |
| Data frequency | Often monthly route readings | Continuous or near-continuous sensor data |
| Analyst workload | Review of each alarm, many of them nuisance | Review focused on prioritized, classified alerts |
Why a fixed threshold misses a developing fault
Consider an illustrative case. A centrifugal pump bearing runs at an overall velocity of 2.0 mm/s for months. The alarm is set at 7.1 mm/s.
An early outer race defect starts producing small impacts. Overall velocity rises to 2.6 mm/s, which is still well below the alarm. But the kurtosis of the acceleration signal jumps, and a peak appears in the envelope spectrum at the bearing's outer race defect frequency.
A threshold system sees nothing. An AI model trained on that pump's baseline sees a 30% rise in RMS, an impulsive waveform, and a spectral peak that matches a known bearing signature. It flags a likely outer race defect with weeks of lead time instead of days.
These numbers are illustrative; real values depend on the machine, mounting, and sensor.
What Data AI Uses
AI predictive maintenance uses four main signal types: vibration, motor current, temperature, and acoustic or ultrasonic data. Each one reveals different failure modes, and combining them catches faults that a single signal would miss or misclassify.
Industrial IoT sensors mounted on motors, pumps, compressors, fans, and conveyors collect these signals and send them for analysis.
Vibration
Vibration is the most widely used signal for rotating equipment, and industrial vibration analysis has the deepest library of known fault signatures. Accelerometers measure how the machine moves, and the frequency content of that movement points to specific mechanical causes.
Imbalance shows at running speed (1x). Misalignment often shows at 2x and in the axial direction. Bearing defects produce impacts at calculable frequencies. Gear problems show at gear mesh frequency and its sidebands.
Motor current
Current reveals the electrical side of a motor and also reflects the mechanical load it drives. Broken rotor bars create sidebands around line frequency. Stator winding problems create imbalance between phases.
Because the pump or fan load passes back through the motor, current can also show process issues such as cavitation or a clogged filter.
Temperature
Temperature changes more slowly than vibration, but a bearing running hot alongside rising vibration is a stronger case than either reading alone.
Winding temperature is a direct indicator of insulation stress. Many motor insulation classes lose expected life quickly as they run above their rated temperature.
Acoustic and ultrasonic signals
High-frequency acoustic and ultrasonic signals pick up friction and impacts before they show in standard vibration ranges. Poor lubrication is a common example: metal-to-metal contact produces ultrasonic noise early, and adding grease brings it back down.
Why multimodal matters
Multimodal monitoring combines multiple signal types into one view of asset health, so a fault that is ambiguous in one signal becomes clear when another signal confirms it. Tractian's multimodal platform pairs the Smart Trac Ultra sensor (vibration, ultrasound, temperature, and magnetic field) with electrical signals to identify bearing faults, winding failures, cavitation, imbalance, and misalignment, and to tell them apart.
A practical example: elevated 2x line frequency vibration on a motor can come from an electrical problem or from a mechanical problem such as soft foot. Current data helps separate the two, which changes whether you send an electrician or a millwright.
How AI Processes Sensor Data
AI processes sensor data in five stages: acquisition, feature extraction, model training, inference, and alert routing. Most of the diagnostic value comes from feature extraction, where raw waveforms become measurable indicators of machine condition.
Step 1: Acquisition
The sensor samples a signal at a set rate. Vibration for bearing analysis needs high sampling rates, often in the kilohertz range, to capture impacts and resonances. Temperature can be sampled far less often.
Each sample also needs context: shaft speed, load, and whether the machine was running.
Step 2: Feature extraction
Feature extraction turns a waveform into numbers that describe it. Three features do much of the work:
- RMS (root mean square): the overall energy of the signal. It tracks general deterioration and is the basis for severity charts such as those in ISO 20816.
- Kurtosis: a measure of how "spiky" the signal is. A smooth, random signal has a kurtosis near 3. Repeated impacts from an early bearing defect push it higher, often before RMS moves much.
- FFT (Fast Fourier Transform): converts the time waveform into a frequency spectrum. Peaks at specific frequencies point to specific causes, which is what makes diagnosis possible.
Other common features include crest factor, peak-to-peak amplitude, and envelope spectra. Our deeper guide to AI fault detection covers each feature with worked bearing frequency calculations.
Step 3: Model training
Models learn in two ways. Unsupervised models learn what "normal" looks like for one asset and flag deviations. Supervised models learn from labeled examples of known faults and classify what they see.
Step 4: Inference
Inference is the model running on new data. Each reading is scored against baseline and known fault patterns.
This can run on the sensor or gateway (edge), in the cloud, or split between them. The tradeoffs are covered below, and the AI condition monitoring hub covers how these systems plug into CMMS, ERP, and SCADA.
Step 5: Alert routing
Good systems attach the likely failure mode, the evidence (trend, spectrum, waveform), and a recommended inspection.
Routing then sends the alert to the reliability engineer or planner, often as a work order in the CMMS, ranked by asset criticality.
Failure Types AI Detects
AI predictive maintenance is designed to detect the failure modes that leave measurable physical signatures: bearing wear, winding degradation, cavitation, imbalance, and misalignment are five of the most common on rotating equipment. Each produces a distinct pattern in vibration, current, temperature, or acoustic data.
Understanding the physics behind each pattern explains why the AI can tell them apart, and why these modes account for so much equipment failure on motors, pumps, and fans.
Bearing wear
Rolling element bearings fail in predictable stages. A small spall on the outer race, inner race, or a rolling element creates an impact each time a rolling element passes over it.
Those impacts repeat at defect frequencies set by bearing geometry and shaft speed: BPFO for the outer race, BPFI for the inner race, BSF for the rolling elements, and FTF for the cage. Early on, the impacts are too small to raise overall vibration, but they excite high-frequency resonances that envelope analysis can isolate.
Inner race defects often show sidebands spaced at running speed, because the defect rotates in and out of the load zone. As damage spreads, kurtosis can fall again while RMS climbs, which is one reason a single feature is not enough.
Winding degradation
Stator winding insulation breaks down under heat, moisture, contamination, and voltage stress. A turn-to-turn short creates a local hot spot and an imbalance in the magnetic field.
In current data, this shows as imbalance between phases and a rise in negative sequence current. Temperature rises in the affected winding, and vibration at twice line frequency (120 Hz on a 60 Hz supply) can increase.
Early detection leaves time to plan a rewind before the short spreads to a phase-to-ground failure.
Cavitation
Cavitation happens when pressure at the pump inlet drops below the liquid's vapor pressure. Vapor bubbles form, then collapse violently as they move into higher-pressure zones near the impeller.
Each collapse is a tiny impact. Together they create broadband, random, high-frequency vibration and noise, and they erode the impeller surface over time.
AI picks this up as a raised high-frequency noise floor in vibration and acoustic data, unsteady motor current as the hydraulic load fluctuates, and a link to process conditions such as low suction level or a partly closed inlet valve.
Imbalance
Imbalance means the mass center of a rotor does not sit on its axis of rotation. The resulting force rotates with the shaft.
It shows as a dominant peak at 1x running speed in the radial direction, with a stable phase. Because the force grows with the square of speed, imbalance on variable-speed fans becomes far more severe at high speed.
Misalignment
Misalignment means two coupled shafts do not share a common centerline. Angular misalignment tends to produce high 1x vibration in the axial direction. Parallel (offset) misalignment tends to produce high 2x vibration in the radial direction.
Misalignment also loads bearings and seals, so left alone it shows up later as bearing wear.
The AI separates misalignment from imbalance by the ratio of 1x to 2x, the axial component, and the phase relationship.
Edge AI vs Cloud AI for Predictive Maintenance
Edge AI runs models on or near the sensor and sends only results; cloud AI sends data to remote servers for heavier analysis. Most industrial deployments use a hybrid: edge processing for fast screening and bandwidth reduction, cloud processing for model training and fleet-wide comparisons.
| Factor | Edge AI | Cloud AI |
|---|---|---|
| Latency | Low; analysis happens on site | Depends on network connection |
| Connectivity | Keeps working through outages | Needs a reliable link |
| Bandwidth | Sends features or alerts, not raw waveforms | Raw or high-resolution data must be transmitted |
| Compute power | Limited by device hardware | Large models and long histories |
| Fleet learning | Limited to local data | Compares similar assets across sites |
| Model updates | Pushed to devices | Updated centrally |
Why bandwidth matters more than speed for most faults
Bearing wear and winding degradation develop over days to months, so a few seconds of latency changes nothing.
Bandwidth is the bigger constraint. A high-resolution vibration capture from one sensor can run to hundreds of kilobytes. Multiply that by hundreds of sensors and frequent captures, and sending raw data from a plant with limited connectivity becomes impractical. Computing features at the edge solves this.
Where low latency does matter
A sudden vibration spike on a high-speed machine, or a rapid temperature rise that signals seizure, needs a fast response, and edge processing can flag these locally.
Tractian's sensors capture data on the asset, a gateway relays it over 4G/LTE, and AI analysis runs in Tractian's cloud, with no plant Wi-Fi required. The edge AI guide for industrial plants covers architectures, cellular connectivity, and how hybrid systems split work between the device and the cloud.
ROI: What to Expect
The ROI of AI predictive maintenance comes from turning unplanned failures into planned repairs. You can estimate first-year value from the number of failures you have, what each one costs, how many the system catches early, and what a planned repair costs instead.
Step-by-step ROI framework
- Count unplanned failures. From the CMMS or downtime log, count unplanned failures per year on the assets you plan to monitor. Focus on failure modes with physical signatures (bearings, motors, pumps, fans).
- Calculate cost per failure. Add lost production (downtime hours multiplied by margin per hour), emergency labor and overtime, expedited parts, and secondary damage such as a seized bearing that scores a shaft.
- Estimate early detection rate. Estimate the share of those failures that leave a detectable signature with enough lead time to plan. Use a conservative figure and validate it during a pilot.
- Calculate the planned repair cost. A planned repair usually costs a fraction of an emergency one: standard parts, scheduled labor, and a stop that fits a production window.
- Subtract program cost. Include sensors, software subscription, installation, integration, and team time to review alerts.
First-year value = (failures per year × early detection rate × (cost per unplanned failure − cost per planned repair)) − program cost
Illustrative worked example
The figures below are hypothetical and exist only to show the arithmetic. They are not Tractian customer data; for benchmark figures, see the Verdantix study below.
| Input | Illustrative value |
|---|---|
| Unplanned failures per year on monitored assets | 12 |
| Average cost per unplanned failure | $40,000 |
| Share detected early enough to plan | 60% |
| Average cost of the same repair when planned | $8,000 |
| Annual program cost | $90,000 |
Failures converted to planned repairs: 12 × 0.6 = 7.2.
Savings per converted failure: $40,000 − $8,000 = $32,000.
Gross savings: 7.2 × $32,000 = $230,400.
First-year value: $230,400 − $90,000 = $140,400.
Run the same calculation with your own figures, and test a pessimistic case. If the program still pays off at half your detection estimate, the business case holds up.
For a deeper look at licensing models and hidden costs, see the guide to AI condition monitoring cost.
Benchmarks and customer results
A Verdantix study commissioned by Tractian modeled a 401% ROI over three years, an 8-month break-even point, and $3.3 million NPV for a modeled manufacturer with 12 sites. Verdantix estimates continuous monitoring reduces unplanned downtime by up to 10% and planned downtime by up to 5%.
Case studies add plant-level results: Danone avoided up to $40,000 in maintenance costs when a homogenizer fault was caught early, and ICL recovered 400+ tons of production per year after removing one of two annual shutdowns.
Which assets to start with
Start where a failure hurts most and where AI has a clear physical signal to read. This decision table helps rank candidates.
| Asset profile | Priority | Reason |
|---|---|---|
| Critical, no spare, history of bearing or motor failures | High | Large downtime cost and strong vibration and current signatures |
| Critical, installed spare, frequent failures | Medium | Lower downtime cost, but repeated repair cost adds up |
| Non-critical, rarely fails | Low | Little cost to avoid |
| Hard to reach or in hazardous areas | High | Remote monitoring replaces risky manual routes |
| Low-speed (under roughly 100 rpm) or reciprocating machines | Check with vendor | Signatures differ and need specific sensor and model support |
Frequently Asked Questions
What is AI for maintenance?
AI for maintenance is the use of machine learning to analyze equipment data and decide when maintenance is needed. In practice, models read sensor signals such as vibration, current, and temperature, learn each asset's normal behavior, and flag developing faults. The output is a prioritized list of assets that need inspection and the likely cause.
How is AI used in maintenance operations?
AI in maintenance operations detects developing failures from sensor data, classifies the likely failure mode, and routes alerts to the right team, often as CMMS work orders. It also helps planners rank work by asset criticality and fault severity. Some teams use it to extend or shorten preventive maintenance intervals based on actual condition.
Which AI do you recommend for maintenance?
For rotating equipment, look for AI that combines several signal types (vibration, current, temperature, and acoustic), classifies specific failure modes rather than only flagging anomalies, and explains each alert with the evidence behind it. It should integrate with your CMMS and run reliably on your plant's connectivity. Our guide to the best AI predictive maintenance software compares platforms on these criteria.
What is the difference between machine learning maintenance and predictive maintenance?
Predictive maintenance is the strategy: perform maintenance based on the measured condition of equipment. Machine learning is one way to carry it out. Traditional predictive maintenance relies on analysts and fixed limits, while machine learning maintenance automates pattern recognition across many assets and signals.
How long does AI need to learn a machine's normal behavior?
Anomaly detection models typically need a baseline period of steady operation, often measured in days to a few weeks, depending on how much the machine's speed and load vary. Fault classifiers trained on labeled data from other machines can recognize known signatures sooner. Machines with many operating states need longer baselines to avoid false alarms.
Can AI predictive maintenance detect failures in motors and pumps?
Yes. Motors and centrifugal pumps are among the best-suited assets, because their main failure modes (bearing wear, winding degradation, cavitation, imbalance, and misalignment) produce clear vibration, current, and temperature signatures. Detection depends on sensor placement and the failure modes the model is trained to recognize.
Does AI replace vibration analysts?
No. AI handles the volume of screening and first-pass diagnosis, which lets analysts spend their time on confirmation, root cause analysis, and difficult cases. Plants without in-house analysts get a first diagnosis they could not produce before, but complex or high-stakes calls still benefit from expert review.
How accurate is AI predictive maintenance?
Accuracy depends on the failure mode, the sensor data available, and how well the model fits the asset. Faults with strong physical signatures, such as bearing defects and imbalance, are detected more reliably than rare or slow process-related faults. Ask vendors for accuracy evidence from comparable assets, and measure false positives and confirmed catches during a pilot.
Do I need a CMMS to use AI predictive maintenance?
You do not need one to detect faults, but a CMMS makes the alerts actionable. Integration lets an AI alert create a work order with the diagnosis attached, and failure history from the CMMS helps confirm what the model found. Without it, alerts depend on someone manually carrying them into the work order process.
See how Tractian detects failures before they happen. Explore Condition Monitoring