Key Points
- AI fault detection identifies developing machine faults by analyzing features of vibration, current, temperature, and acoustic signals, and matching them to the physical signatures of specific failure modes.
- Fixed thresholds react to overall signal levels, which often stay normal until a fault is advanced; features such as kurtosis, envelope spectra, and current sidebands change much earlier.
- Supervised classifiers name known faults, unsupervised anomaly detection catches unfamiliar ones, and the analysis can run close to the machine (edge) or in the cloud, with tradeoffs in bandwidth and response time.
AI fault detection is an application of industrial AI: the use of machine learning to find developing faults in equipment from its sensor signals, and to name the likely cause. On rotating machines, that means reading vibration, current, temperature, and acoustic data for the patterns that bearing defects, winding faults, cavitation, imbalance, misalignment, and looseness leave behind.
Fixed thresholds miss developing faults because they watch the wrong thing. An overall vibration value averages the whole signal, so a few sharp impacts from a tiny bearing spall barely move it. By the time the overall level crosses the alarm, the spall has grown and the remaining life is short.
Signal Types Used
AI fault detection uses four signal types: vibration for mechanical faults, motor current for electrical and load-related faults, temperature for friction and insulation stress, and acoustic or ultrasonic data for early friction and impacts. Each catches faults the others see poorly.
Industrial IoT sensors mounted directly on the machine capture these signals continuously.
| Signal | What it reveals | Faults it is strongest on | Limits |
|---|---|---|---|
| Vibration | Forces and impacts in rotating parts | Bearing defects, imbalance, misalignment, looseness, gear wear | Weak on purely electrical motor faults |
| Current | Motor electrical condition and driven load | Rotor bar defects, winding imbalance, cavitation, overload | Needs access to motor supply cables |
| Temperature | Heat from friction, electrical loss, poor cooling | Lubrication failure, overloaded bearings, insulation stress | Changes slowly; often confirms rather than leads |
| Acoustic and ultrasonic | High-frequency friction and impacts | Early lubrication problems, early bearing damage, cavitation, leaks | Sensitive to background noise and mounting |
Combining signals reduces misdiagnosis. Elevated vibration at twice line frequency can come from an electrical fault or from a mechanical issue; current data separates the two. Multimodal monitoring, which combines several signal types into one view of asset health, gives the model this cross-check. Tractian takes this approach: its Smart Trac Ultra sensor reads vibration, ultrasound, temperature, and magnetic field, and the platform correlates them with electrical signals so a fault flagged in one signal can be confirmed or ruled out by another.
Feature Extraction: What AI Actually Analyzes
AI analyzes features: numbers computed from raw signals that describe machine condition. The core vibration features are RMS, kurtosis, peak-to-peak amplitude, FFT spectral components, and envelope spectra; each responds to a different aspect of how faults change a signal.
Industrial vibration analysis has used these features for decades; AI applies them continuously across many machines.
RMS
RMS (root mean square) measures the overall energy of a signal. Velocity RMS in mm/s or in/s is the basis for severity zones in ISO 20816.
RMS tracks general deterioration well but responds late to localized defects, because a few short impacts add little energy to the total.
Kurtosis
Kurtosis measures how impulsive a signal is. A healthy machine with smooth, random vibration has a kurtosis close to 3. Repeated sharp impacts, such as those from an early bearing spall, push it well above 3.
Kurtosis often rises before RMS does. As damage spreads and impacts blur into general noise, kurtosis can fall back toward 3 while RMS keeps climbing. A model that reads both features can tell early damage from late damage.
Peak-to-peak and crest factor
Peak-to-peak is the distance between the highest and lowest points of the waveform. Crest factor is the peak divided by RMS.
Both rise when impacts appear on top of normal vibration. They are useful for detecting looseness and bearing impacts, and peak-to-peak displacement is a common measure on slow machines.
FFT spectral components
The FFT (Fast Fourier Transform) converts a time waveform into a frequency spectrum. Each peak shows how much vibration occurs at that frequency, and specific frequencies point to specific causes.
Running speed (1x) and its harmonics show imbalance, misalignment, and looseness. Gear mesh frequency shows gear condition. Twice line frequency points to electrical issues in motors.
Envelope analysis
Envelope analysis isolates repeating impacts that are buried in other vibration. It filters the signal around a high-frequency band where impacts excite structural resonances, extracts the envelope (the pattern of amplitude changes), and runs an FFT on that envelope.
The result is a spectrum where bearing defect frequencies stand out clearly, often while they are still invisible in the standard spectrum.
Worked example: bearing defect frequencies
Bearing defect frequencies are calculated from geometry and shaft speed. This illustrative example uses geometry commonly published for a 6205 deep groove ball bearing; use your bearing manufacturer's data for real calculations.
Inputs (illustrative):
- Number of rolling elements (n): 9
- Ball diameter (d): 7.94 mm
- Pitch diameter (D): 39.04 mm
- Contact angle (φ): 0°
- Shaft speed: 1,780 rpm, so running frequency (fr) = 1,780 / 60 = 29.67 Hz
- Ratio d/D = 7.94 / 39.04 = 0.203
Formulas and results:
| Defect frequency | Formula | Multiple of fr | Result at 1,780 rpm |
|---|---|---|---|
| BPFO (outer race) | (n/2) × fr × (1 − (d/D) cos φ) | 3.585 | 106.4 Hz |
| BPFI (inner race) | (n/2) × fr × (1 + (d/D) cos φ) | 5.415 | 160.7 Hz |
| BSF (rolling element) | (D/2d) × fr × (1 − ((d/D) cos φ)²) | 2.357 | 69.9 Hz |
| FTF (cage) | (fr/2) × (1 − (d/D) cos φ) | 0.398 | 11.8 Hz |
If the envelope spectrum of this motor's bearing shows a peak near 106 Hz with harmonics at about 213 Hz and 319 Hz, the model has strong evidence of an outer race defect. A peak near 161 Hz with sidebands spaced about 30 Hz apart points instead to an inner race defect.
These frequencies are not whole multiples of running speed, so they are hard to confuse with imbalance or misalignment.
Supervised vs Unsupervised Fault Detection
Supervised fault detection trains classifiers on labeled examples of known faults and names the fault it sees; unsupervised fault detection learns normal behavior and flags departures from it. Supervised models give a diagnosis, unsupervised models catch the unexpected, and production systems usually use both.
| Supervised (labeled fault classifiers) | Unsupervised (anomaly detection) | |
|---|---|---|
| Training data | Signals labeled with confirmed faults | Normal operating data from the asset |
| Output | Named fault and confidence | Deviation score |
| Strength | Tells the technician what to inspect | Works without fault history; catches rare faults |
| Weakness | Needs many labeled examples; misses untrained faults | Does not say what is wrong; can flag normal operating changes |
| Best fit | Common faults with clear physics, such as bearings and imbalance | New assets, unusual machines, rare failure modes |
When each applies
Use this decision table to see which approach carries most of the load in a given situation.
| Situation | Lead approach | Why |
|---|---|---|
| Standard motors, pumps, and fans with common failure modes | Supervised | Large libraries of labeled faults exist for these machines |
| Newly installed or custom machinery | Unsupervised | No fault history yet; baseline is learned on site |
| Failure mode seen once or twice in the plant's history | Unsupervised, then supervised | Anomaly flags it; confirmed cases become labels |
| Variable speed and load operation | Both, with operating context | Speed and load inputs keep anomalies from firing on normal changes |
| High-criticality assets | Both | Anomaly detection for coverage, classification for fast action |
The feedback loop matters. When a technician confirms a fault on a work order, that confirmed case becomes a label, and the supervised model improves over time.
Fault Signatures AI Detects
AI detects faults by matching features to their physical signatures. The six most common on rotating equipment are bearing defects, winding faults, cavitation, imbalance, misalignment, and looseness, and each one changes signals in a distinct, explainable way.
Knowing these signatures makes AI alerts easier to verify. For background, see equipment failure.
Bearing defects: inner race, outer race, rolling element
A defect on any bearing surface creates an impact each time a rolling element contacts it.
- Outer race: The outer race is usually stationary, and the defect sits in a fixed position in the load zone. Impacts repeat at BPFO with fairly constant amplitude, so the envelope spectrum shows BPFO and its harmonics, usually without running speed sidebands.
- Inner race: The inner race turns with the shaft, so the defect moves in and out of the load zone once per revolution. Impact strength rises and falls with each turn, which creates sidebands at 1x spacing around BPFI.
- Rolling element: A defective ball or roller strikes both races, so energy often appears at twice BSF. Because the element travels with the cage, sidebands at FTF are common.
Kurtosis and ultrasonic levels typically rise first, envelope peaks become clear next, and overall RMS and temperature rise late.
Winding faults
Stator winding insulation degrades with heat, moisture, contamination, and voltage spikes. A short between turns creates a hot spot and unbalances the motor's magnetic field.
Current shows imbalance between the three phases and a rise in negative sequence current. Temperature rises locally, and vibration at twice line frequency (120 Hz on a 60 Hz supply) can increase.
Rotor faults have their own current signature. A broken or cracked rotor bar creates sidebands around line frequency at f × (1 ± 2s), where s is slip. For a 4-pole, 60 Hz motor running at 1,770 rpm, slip is (1,800 − 1,770) / 1,800 = 0.0167, so sidebands appear near 58 Hz and 62 Hz. The deeper those sidebands sit below the line frequency peak, the healthier the rotor.
Cavitation
Cavitation occurs when pressure at the pump inlet falls below the liquid's vapor pressure, usually because available net positive suction head (NPSHa) drops below what the pump requires (NPSHr). Vapor bubbles form, travel into higher pressure zones, and collapse violently against the impeller.
The collapses produce broadband, random, high-frequency vibration and noise rather than a clean peak. AI recognizes cavitation from a raised high-frequency noise floor, fluctuating motor current, erratic readings at vane pass frequency, and a link to process conditions such as low suction tank level.
Imbalance
Imbalance means the rotor's center of mass is offset from its axis of rotation, creating a rotating force.
The signature is a dominant radial peak at 1x running speed with steady phase and little axial vibration. The force grows with the square of speed, so doubling a fan's speed roughly quadruples the imbalance force.
Misalignment
Misalignment means coupled shafts do not share a centerline.
Angular misalignment produces strong 1x vibration in the axial direction. Parallel, or offset, misalignment produces strong 2x vibration in the radial direction. Phase readings across the coupling often differ by about 180 degrees. A high ratio of 2x to 1x and a large axial component help the model separate misalignment from imbalance.
Looseness
Looseness changes how the machine responds to normal forces.
Structural looseness, such as loose hold-down bolts or a cracked base, often shows at 1x with a large difference between horizontal and vertical readings. Rotating looseness, such as a bearing loose in its housing or on its shaft, produces a series of 1x harmonics, sometimes with half-order peaks at 0.5x, 1.5x, and 2.5x, and a time waveform that looks clipped or erratic.
Signature summary
| Fault | Key features | Distinguishing pattern |
|---|---|---|
| Outer race defect | Kurtosis, envelope spectrum | BPFO and harmonics, no 1x sidebands |
| Inner race defect | Kurtosis, envelope spectrum | BPFI with 1x sidebands |
| Rolling element defect | Envelope spectrum | 2x BSF with FTF sidebands |
| Stator winding fault | Phase current, temperature | Phase imbalance, 2x line frequency vibration |
| Rotor bar defect | Current spectrum | Sidebands at line frequency × (1 ± 2s) |
| Cavitation | High-frequency vibration, acoustic, current | Broadband noise floor rise, fluctuating current |
| Imbalance | FFT | Dominant 1x radial, steady phase |
| Misalignment | FFT, axial readings | High 1x axial or high 2x radial |
| Looseness | FFT, waveform | Many 1x harmonics, half-order peaks |
Edge Inference for Real-Time Detection
Edge inference means running fault detection models on the sensor or a nearby gateway instead of in a remote data center. It matters because it cuts the data that must travel over the network, keeps detection running during connectivity outages, and allows fast local alerts for events that develop in seconds.
Why latency matters
For most faults, latency is not the main concern: a bearing defect developing over weeks does not need a millisecond response.
Some events do develop fast: a sudden jump in vibration on a high-speed machine, a lost balance weight, or a bearing temperature climbing toward seizure. Local inference can flag these immediately, without waiting for data to reach the cloud and an alert to come back.
How edge AI runs on sensor hardware
Sensor hardware has limited processing power, memory, and battery, so edge models are built small. The common pattern is:
- The sensor samples the raw signal at high rate for a short capture.
- On-device firmware computes features such as RMS, kurtosis, peak values, and selected spectral bands.
- A compact model or rule set scores those features against the asset's baseline.
- The sensor sends features and scores on a regular schedule, and sends a fuller capture when something changes.
- Heavier analysis, such as detailed classification and model retraining on fleet data, runs in the cloud.
This split keeps bandwidth and battery use low. Some systems skip on-device models entirely: in Tractian's setup, the sensor captures data on the asset, a gateway relays it over 4G/LTE, and the AI analysis runs in Tractian's cloud. Our guide to edge AI in industrial plants covers architecture options and connectivity in more depth.
Where AI fault detection fits
Fault detection is one part of a predictive maintenance program: it tells you what is wrong and how urgent it is, and planning turns that into a repair.
For how detection fits into the full program, including ROI, see our guide to AI for predictive maintenance. For deployment and integration choices, see the AI condition monitoring hub.
Frequently Asked Questions
What is AI fault detection?
AI fault detection is the use of machine learning to identify developing faults in equipment from sensor data and to name the likely cause. On rotating machines, models analyze features of vibration, current, temperature, and acoustic signals and match them to known fault signatures. The goal is to flag problems early enough to plan the repair.
Can industrial AI predict equipment failure using vibration data?
Yes. Many mechanical faults change vibration in predictable ways long before failure: bearing defects create impacts at calculable frequencies, imbalance raises 1x vibration, and misalignment raises axial or 2x vibration. AI tracks features such as kurtosis, envelope spectra, and RMS over time, which can give weeks of warning for faults that develop gradually.
What industrial AI software can detect machine failures from vibration data?
Software that detects failures from vibration data combines continuous sensor data with models trained on fault signatures, so it can name a likely fault rather than only flagging a change. Look for envelope analysis for bearings, spectral analysis for imbalance, misalignment, and looseness, and support for current and temperature data to confirm diagnoses. Tractian's condition monitoring platform combines vibration with temperature, current, and acoustic signals for this purpose.
How does AI detect bearing faults from vibration?
AI detects bearing faults by looking for repeating impacts at the bearing's defect frequencies (BPFO, BPFI, BSF, and FTF), which are calculated from bearing geometry and shaft speed. Envelope analysis isolates those impacts from other vibration, and rising kurtosis confirms the signal is impulsive. The pattern of harmonics and sidebands shows whether the defect is on the outer race, inner race, or a rolling element.
What is the difference between machine learning predictive maintenance and threshold alarms?
Threshold alarms compare one value to one fixed limit and alert when it is crossed, which usually happens late. Machine learning predictive maintenance compares many features to each asset's learned baseline and to known fault patterns, so it can flag changes while overall values still look normal. It also attaches a likely cause, which a threshold alarm cannot do.
Does AI fault detection need historical failure data?
Not necessarily. Unsupervised anomaly detection only needs a period of normal operation to learn a baseline. Supervised classifiers need labeled fault examples, but these can come from similar machines in other plants rather than your own history. Over time, confirmed faults from your own work orders improve the models further.
Can AI fault detection work on variable speed equipment?
Yes, if the model uses speed and load as inputs. Fault frequencies such as BPFO scale with shaft speed, so the system needs to know the speed at each reading to place peaks correctly. Without operating context, normal speed changes can look like faults and real faults can be masked.
See how Tractian detects failures before they happen. Explore Condition Monitoring