Every passing train puts tonnes of dynamic force through each wheel-rail interface. Flats, cracks, spalling, and polygonal wear are not merely cosmetic — a single undetected wheel defect can damage hundreds of metres of track, destabilise a vehicle at speed, or trigger a derailment. For decades, detection relied on scheduled depot inspections and manual wayside checks: periodic, slow, and blind to the hours of operation between visits. Machine learning has changed this equation entirely — placing defect detection on the track itself, scoring every wheel on every pass, and surfacing faults before they reach the threshold that puts safety at risk.
Predictive Maintenance · Rolling Stock · Wayside AI · Infrastructure Health
Wheel and Axle Defects Are Failing Silently. ML Can Catch Them Every Single Pass.
iFactory's infrastructure AI platform connects continuous asset health scoring with rolling stock condition data — so wheel deterioration is caught in motion, not discovered during a depot visit after damage is done.
The Defect Types That Put Rolling Stock — and Track — at Risk
Wheel and axle defects do not appear suddenly. They develop progressively — from microscopic surface fatigue to visible damage that generates dangerous impact forces. The six defect categories below represent the most operationally significant faults that ML models are trained to identify.
Wheel Flats
Caused by wheel lock-up during braking. The flat section strikes rail with each revolution, generating a sharp impact force detectable acoustically from up to 15 metres away. Progressive and worsening with every km run.
Spalling
Rolling contact fatigue causes subsurface cracks that propagate and break through the tread surface, detaching material. Spalling patches alter the wheel's vibration signature — a reliable ML detection signal.
Shelling
Surface layer separation across a wider tread area than spalling — often appearing in clusters. Vision-based ML models identify shelling from high-resolution wayside imagery with accuracy that exceeds manual inspection.
Polygonal Wear
Out-of-roundness where the wheel profile develops geometric facets instead of remaining circular. Generates characteristic vibration harmonics that ML frequency analysis models can identify and quantify with high precision.
Axle Cracks
Fatigue cracks in the axle body — typically near stress concentration points at wheel seat transitions. Deep learning CNN models now detect crack depth severity in addition to presence, enabling risk-stratified maintenance responses.
Tread Discolouration
Thermal damage from braking or electrical arcing. While not structurally critical in early stages, discolouration indicates conditions that accelerate other defect types. Computer vision classifiers identify it from wayside images as a precursor alert.
100%
Wheel flat detection accuracy achieved by unsupervised ML wayside systems in validated field tests
90%+
Wheel flat severity classification accuracy with a single wayside accelerometer and ML
91%
Wheel condition classification accuracy using CNN models trained on vibration-derived recurrence plots
98%
Defective wheel detection rate achieved with zero false alarms in large-scale Union Pacific Railroad field deployment
Two Sensing Approaches — And Why the Best Systems Use Both
ML wheel defect detection is not a single technology. It is built from two sensing modalities — acoustic and vision — each capturing different categories of fault, and each requiring different ML architectures to process. Understanding both is essential to understanding what a complete wayside detection system actually covers.
Approach 01
Acoustic & Vibration ML
Detects: flats, polygonal wear, spalling, bearing faults
Trackside accelerometers and acoustic emission sensors capture the vibration signature of every wheel-rail contact event as a train passes at speed. ML models — trained on labelled datasets of healthy and defective wheel signals — classify each wheelset's condition in real time. Unsupervised learning approaches detect anomalies without requiring labelled training data: detecting wheel flats with 100% accuracy by identifying outlier vibration events, then using Hidden Markov Models to localise which specific wheelset on a consist is the source of the signal.
Signal types processed
Time-domain vibration · Spectral kurtosis · Acoustic emission bursts
ML models used
SVM · Random Forest · k-means · Hidden Markov Model · CNN
Key advantage
Works at full operational speed · No line blockage required
Approach 02
Computer Vision ML
Detects: shelling, cracks, discolouration, surface damage
High-resolution cameras mounted in wayside enclosures capture images of every wheel tread as trains pass — at speeds where the shutter time must be measured in microseconds to avoid motion blur. Deep learning models — primarily YOLO-family object detectors and transformer architectures — analyse these images to classify and localise surface defects. Recent work, including the FaultSeg dataset and YOLOv9-based evaluation, demonstrates that computer vision now handles four defect categories simultaneously: flats, shelling, cracks/scratches, and discolouration, each mapped to a specific region of the wheel image.
Image processing steps
Histogram equalisation · GAN augmentation · HSV conversion · Background removal
ML models used
YOLOv9 · RTD Transformer · EfficientNet-B7 · CNN with TTDCNN
Key advantage
Detects surface morphology invisible to vibration sensors · Generates visual evidence record
From Signal to Maintenance Action: The ML Detection Pipeline
A complete ML wheel defect detection system does more than raise an alarm. It turns raw sensor data into a prioritised, actionable maintenance output — telling operators not just that a defect exists, but which vehicle, which axle, how severe, and what to do about it.
1
Signal Capture
Acoustic, vibration, or image data acquired from wayside sensors as train passes
2
Feature Extraction
Vibration level, temperature, spectral kurtosis, or image region features extracted from raw input
3
ML Classification
Model outputs defect type, severity class, and confidence score for each wheelset on the consist
4
Localisation
HMM or position-tracking logic identifies which axle on which vehicle carries the detected defect
5
Asset Health Score
Defect finding updates the rolling stock asset's condition record and triggers an infrastructure health alert
6
Maintenance Action
Prioritised work order issued: monitor, plan depot visit, or flag for immediate withdrawal from service
Asset Health Platform · Rolling Stock · Wayside Integration
Your Wheels Are Telling You Something. Is Your Infrastructure Platform Listening?
iFactory integrates wayside defect detection signals with its infrastructure health scoring engine — so wheel condition data drives your maintenance planning, not just your alarm inbox.
Which ML Model for Which Defect Type?
Not all defects are best approached with the same algorithm. The choice of model depends on whether the detection problem is supervised or unsupervised, signal-based or image-based, binary classification or multi-class severity estimation.
Model
Best for
Typical accuracy
Key strength
Random Forest
Vibration feature classification, flat and spalling detection
High — most consistent across feature sets in comparative studies
Handles mixed feature types well; robust to noise in sensor data
Support Vector Machine
Binary healthy/defective classification from acoustic signals
High in structured feature spaces with clear class separation
Effective in small-to-medium labelled datasets; well-validated in rail
CNN (1D + 2D)
Multi-class defect classification from vibration and image inputs combined
91%+ on wheel condition classification; state of the art for image inputs
Learns spatial and temporal patterns end-to-end without manual feature engineering
YOLO / Transformer
Real-time visual defect detection: shelling, cracks, discolouration
Leading performance across 4 defect categories in field imaging
Simultaneous detection and localisation in a single forward pass
Unsupervised (k-means + HMM)
Wheel flat detection and localisation without labelled training data
100% flat detection accuracy with zero false alarms in field validation
Deployable on networks without historical defect annotation datasets
Why Wheel Defects Are an Infrastructure Problem as Much as a Rolling Stock Problem
A defective wheel does not just put the vehicle at risk. It damages the infrastructure it runs over. A flat wheel generates impact forces many times the static wheel load at each revolution — forces that accelerate rail fatigue, crush ballast, and overload fasteners. The connection between wheel condition and track condition is direct, measurable, and increasingly quantifiable through ML.
Without ML Detection
Defective wheel runs for hours or days before detection at scheduled depot inspection — damaging kilometres of track in the interim
Track maintenance teams receive corrugation and joint damage reports with no information about the rolling stock source that caused them
Infrastructure renewal cycles are driven by rail wear data, not by knowledge of which vehicle-wheel combinations are accelerating that wear
With ML Detection
Every passing train has its wheels scored in real time — defects flagged within seconds of detection, before the vehicle completes another circuit
Track health monitoring systems receive wheel defect alerts linked to specific vehicle IDs — enabling correlation between rolling stock condition and rail degradation rates
Infrastructure renewal planning is informed by actual wheel-track interaction data — prioritising sections with highest exposure to high-impact wheel passes
The literature from 2019 to 2025 shows a clear trend toward machine learning and deep learning approaches, often reaching nearly 100% accuracy in validated settings. The priority for deployment teams is now moving from demonstrating accuracy to achieving robust real-world generalisation across different train types, speeds, and environmental conditions.
— Vibroacoustic Methods for Wheel-Flat Detection Review, Applied Sciences, 2025
Deploying ML Wheel Detection: What It Takes in Practice
Academic results are necessary but not sufficient. Deploying ML wheel detection on a live network involves hardware installation, data pipeline design, model training, and integration with maintenance management systems. Here is what each element requires.
Sensor Hardware
A single rail-mounted or sleeper-mounted accelerometer is sufficient for vibration-based flat detection. Vision systems require a wayside enclosure, triggered illumination, and high-shutter-speed cameras calibrated to the track speed range.
Training Data
Supervised models require labelled defect examples. Unsupervised approaches can operate without prior labelling — particularly useful on networks where historical defect annotation records do not exist. An 80-20 train/test split is standard for model validation.
Edge vs. Cloud Processing
Real-time wayside systems typically process signals at the edge — classifying each passing wheelset before the next train arrives. Aggregated results are pushed to cloud platforms for fleet-level trending, reporting, and integration with asset management systems.
Maintenance Integration
Detection outputs must connect to maintenance management systems — not just generate alerts. iFactory's infrastructure platform provides the asset health scoring layer that translates individual defect findings into prioritised work orders with severity-based response timelines.
In-motion
Detection at full operational speed
ML wayside systems score every wheelset as the train passes at line speed — no speed restriction, no dedicated inspection slot required.
1 sensor
Sufficient for whole-consist flat detection
A single wayside accelerometer, combined with HMM localisation, identifies both the presence and the position of wheel flats across a full train consist.
4 classes
Simultaneous defect categories from vision models
Latest YOLO and transformer-based vision systems classify flats, shelling, cracks/scratches, and discolouration in a single inference pass per image frame.
Conclusion
Machine learning has transformed wheel and axle defect detection from a periodic, manual activity into a continuous, automated process that runs with every train movement. The evidence is now substantial: unsupervised acoustic systems achieving 100% flat detection, multi-sensor CNN models reaching 91% classification accuracy, and large-scale deployments demonstrating 98% defective wheel detection rates. The technology works — and it works in the field, not just in laboratory conditions.
What determines whether these detection capabilities translate into operational and financial outcomes is how defect findings connect to the wider maintenance management process. iFactory provides the infrastructure health intelligence layer that closes this gap — scoring every asset's condition continuously, integrating rolling stock defect signals, and turning raw detection outputs into prioritised maintenance actions. Book a Demo to see how iFactory connects wayside defect detection to your maintenance workflow, or sign up free to see your first infrastructure health scores today.
Frequently Asked Questions
Every defective wheel that completes another circuit is damaging your track. Does your maintenance system know about it before the next scheduled inspection?
iFactory connects wayside defect detection signals with continuous infrastructure health scoring — so wheel condition findings drive your maintenance planning, not just your alert log.
