PaperlayerView exact revision on arXiv

Source-linked AI summary

Damage detection in railway wheels from impact-induced ultrasonic ring-down using self-referenced decay features

Aashish Shaju, Steve Southward, Mehdi Ahmadian

arXiv:2610.30541v1eess.SP

TL;DR

Stationary wheel screening needs response features that can be used without a prior measurement of each wheel. This study tests air-coupled ultrasonic ring-down features and threshold transfer across two sites, correctly classifying all 11 wheelsets in the other dataset when features are fixed and 21 of 22 under nested wheel-level validation.

  • Problem

    Wheel screening often lacks an undamaged model or prior measurement of the same wheel, motivating response features usable across differing wheel dimensions and service histories.

    Source: §1 · p. 2

  • Method

    The detector combines effective damping normalized by dominant response frequency with strike-to-strike variation in decay-fit quality, then applies calibrated thresholds.

    Source: §1 · p. 2, §3.3 · p. 6

  • Results

    With the two features fixed, thresholds calibrated on either dataset correctly classified all eleven wheelsets in the other; nested validation correctly classified 21 of 22 wheelsets.

    Source: §7 · p. 17

  • Takeaways & Limitations

    The two features provide complementary information from current ring-down recordings and support threshold transfer between the tested datasets for stationary wheel-damage screening.

    Source: §7 · p. 17

    Only five healthy wheelsets were included, all new, and Norfolk Southern support condition coincided with class; worn serviceable wheels and both support conditions require evaluation.

    Source: §6.4 · p. 17

Abstract

from arXiv · show

This study evaluates a non-contact screening method for stationary railroad wheelsets using air-coupled ultrasonic measurements excited by hammer impacts. Two decay features are considered. First, an effective damping ratio is obtained by normalizing the measured decay rate by the dominant response frequency. Second, strike-to-strike variations are measured to assess the quality of the energy-decay fit for the overall wheel. The method was evaluated using 22 wheelsets at two railroad facilities. The analysis indicates that a random forest trained on one dataset and tested on another performed poorly. This result motivated a simpler approach based on thresholds applied to the two decay features, which fixed the feature definitions and calibrated the thresholds on either dataset. This approach correctly classified every wheelset in the other dataset. A more stringent wheel-level validation repeated second-feature selection and threshold calibration while each test wheelset was withheld. This procedure correctly classifies 21 of 22 wheelsets. All defective wheelsets are identified, while one healthy wheelset yields a false positive. The results further show that the two-feature detector transferred between the tested datasets for stationary assessment of wheel damage. Further evaluation is required using worn but serviceable wheels, controlled support conditions, and additional independent datasets.

1. Introduction

Stationary impact testing offers controlled, repeatable excitation, while response-based wheel screening needs features that do not require a prior measurement of each wheel. This study examines two-feature detection and cross-dataset transfer, including validation that withholds test wheelsets during feature selection.

1. Introduction

  • Wheel condition measurements depend on excitation during passage, whereas stationary impact testing lets operators choose the excitation location and repeat measurements.Source: §1 · p. 1
  • Acoustic wheel measurements can reflect contact conditions and material-related sources as well as damage.Prior work examined wheel-flat acoustics and acoustic-emission responses in different test settings.Source: §1 · pp. 1–2
  • A current-response feature is useful because wheel screening usually lacks an undamaged model or an earlier measurement of the same wheel.The motivation applies across wheels with different dimensions and service histories.Source: §1 · p. 2
  • The study evaluates transfer between two independently acquired datasets and a two-feature detector alongside a larger-feature model.Earlier studies evaluated performance within the development population.Source: §1 · p. 2
  • The study defines an effective damping ratio and a repeated-impact variability feature, then assesses threshold transfer and nested wheel-level feature selection.The nested assessment withholds each test wheelset during second-feature selection.Source: §1 · p. 2

2. Measurement System and Test Campaigns

The test campaign used a purpose-built acquisition system for air-coupled ultrasonic measurements and evaluated 22 wheelsets across two facilities. The setup supported repeatable sampling, signal-quality checks, and wheel-level comparison across measurement positions.

2.1 Requirements on the Acquisition System

  • Fixed-interval sampling and complete records were required because timing errors or missing samples would affect decay-rate estimates.Source: §2.1 · p. 3
  • Each strike needed a peak at least 25 dB above noise, and the decay fit needed to span a 20 dB decrease in the energy decay curve.The measured system noise was approximately 20 µV RMS, with a dynamic range of 103.2 dB.Source: §2.1 · p. 3
  • The acquisition system streamed identified strike records to a host for on-site signal-quality checks and repeat measurements when needed.Records included wheelset, position, and strike identifiers for wheel-level and cross-dataset evaluation.Source: §2.1 · p. 3

2.2 Signal Chain

  • The acquisition chain used an air-coupled microphone, driver stage, 24-bit converter at 500 kSPS, programmable logic, and Ethernet transfer to the host.The driver stage supplied the only analog anti-alias filtering.Source: §2.2 · p. 3, Figure 1 · p. 3

2.3 Wheelsets, Excitation Protocol and Datasets

  • The campaign measured 22 wheelsets across MxV and Norfolk Southern, with 11 wheelsets at each facility and 396 strikes overall.Each wheelset was struck at six positions, with three strikes per position, using non-contact response measurement.Source: §2.3 · p. 4
  • At Norfolk Southern, healthy wheelsets were tested on rail and defective wheelsets on the ground, confounding support condition with class.All MxV wheelsets were measured on rail.Source: §2.3 · p. 4

3. Signal Processing and Feature Definitions

The section corrects the measured front-end response and derives two decay features from compensated hammer-impact recordings. It then combines position-level feature values into a wheelset decision.

3.1 Front-End Compensation and Band Selection

  • The measured front-end response is converted to a regularized inverse correction, and compensated signals are analyzed from 20 to 80 kHz in four sub-bands.The damping branch uses a selected sub-band; the variability branch uses the full 20–80 kHz band.Source: §3.1 · pp. 4–5

3.2 Energy Decay Curve and Noise-Floor Truncation

  • Backward integration of squared compensated pressure produces a decreasing energy curve, while Lundeby truncation limits noise-floor energy that would flatten the fitted decay.The crossover is estimated with a 5 dB margin above the noise floor and recalculated for each frequency branch.Source: §3.2 · pp. 5–6

3.3 Frequency Normalization and the Effective Damping Ratio

  • The effective damping ratio normalizes the fitted 20–35 kHz decay rate by the dominant frequency from the same strike, with signal-strength and fit-quality checks governing retention.A retained damping estimate requires signal peaks at least 25 dB above broadband noise and 22 dB above in-band noise, plus a fit R² of at least 0.95.Source: §3.3 · p. 6
  • Damping ratios from the 20–35 kHz band are used; damping ratios from the other three sub-bands did not separate condition classes.Source: §3.3 · p. 7

3.4 Strike-to-Strike Energy Decay-Fit Variability

  • The variability feature is the interquartile range of retained broadband EDR_R2 values across repeated strikes, capturing fit-quality differences that damping may miss.A signal-qualified strike with a fit R² below 0.95 is excluded from damping but retained for variability.Source: §3.4 · p. 7, Figure 5 · p. 7

3.5 Aggregation and Decision Rule

  • At each position, a flag is triggered when either feature exceeds its threshold; a wheelset is defective when at least three of six positions are flagged.Thresholds are calibrated on training data; with fewer than six positions, a strict majority is required, and a position with both features missing counts as flagged.Source: §3.5 · p. 8, Figure 6 · p. 8

4. Evaluation Design

The evaluation design calibrates feature thresholds on training positions, selects the second feature using wheel-level performance, and separates three validation schemes by which data are withheld. It reports class-balanced metrics to avoid masking poor healthy-wheel performance.

4.1 Threshold Calibration

  • Trial thresholds are placed between consecutive training values, and the threshold with the highest position-level balanced accuracy is selected, using the lowest threshold for ties.Missing values do not define trial thresholds or trigger flags during single-feature calibration.Source: §4.1 · p. 9
  • The final two-feature detector combines position flags to classify a wheelset, so one flagged position alone does not make it defective.Thresholds are calibrated separately for each feature using training positions.Source: §4.1 · p. 9

4.2 Validation Schemes

  • The three validation schemes vary which wheelsets or datasets are withheld, separating threshold transfer from whether second-feature selection transfers.Cross-dataset transfer fixes the features selected using both datasets; nested validation withholds a wheelset from both feature selection and threshold calibration.Source: §4.2 · pp. 9–10

4.3 Selection of the Second Feature

  • The effective damping ratio stays fixed, while candidate second features are selected by wheel-level balanced accuracy from inner leave-one-wheel-out classifications.Thresholds are calibrated using position-level performance, but feature selection uses wheel-level performance.Source: §4.3 · p. 10

4.4 Metrics

  • Balanced accuracy averages healthy- and defective-class recall, giving the classes equal weight despite five healthy and seventeen defective wheelsets.A detector labeling every wheel defective would have plain accuracy of 0.77 but balanced accuracy of 0.50.Source: §4.4 · pp. 10–11

5. Results

The two-feature threshold detector transferred across the tested datasets, and nested hidden-wheel validation correctly classified 21 of 22 wheelsets. In contrast, the random forest transferred poorly, and single-dataset feature selection did not reliably identify a transferable second feature.

5.1 Pooled Evaluation of the 109-Feature Model

  • The pooled 109-feature random forest achieved healthy-class recall of 0.73, defective-class recall of 0.79, and balanced accuracy of 0.76.Both datasets were represented in every training fold.Source: §5.1 · p. 11, Figure 8 · p. 11

5.2 Cross-Dataset Evaluation

  • The cross-dataset random forest achieved healthy-class recall of 0.06, defective-class recall of 0.90, and balanced accuracy of 0.48.Removing 24 frequency-related predictors did not change these results to two decimal places.Source: §5.2 · pp. 11–12, Figure 8 · p. 11

5.3 The Two-Axis Feature Audit

  • The variability feature was the only candidate meeting the audit criteria, with dataset sensitivity of 0.06 and damping complementarity of 0.68.The complementarity score corresponded to an AUC of approximately 0.84.Source: §5.3 · p. 12

5.4 Cross-Dataset Transfer of the Two-Feature Detector

  • 11 of 11 test wheelsets were correctly classified in both cross-dataset transfer directions using thresholds calibrated on the other dataset.The damping thresholds independently calibrated on the two datasets differed by approximately one percent.Source: §5.4 · p. 13, Figure 9 · p. 13
  • The variability feature identified W7, W8, and W9, which damping missed, while elevated damping identified W19, which variability missed.Every defective wheelset was identified by at least one feature.Source: §5.4 · p. 13

5.5 Nested Hidden-Wheel Validation

  • 21 of 22 wheelsets were correctly classified in nested hidden-wheel validation, with all defective wheelsets identified and W5 the only error.Fixing the variability feature in advance produced the same classifications; the balanced accuracy was 0.90.Source: §5.5 · p. 14, Figure 10 · p. 14

5.6 Feature Selection from a Single Dataset

  • Single-dataset feature selection chose different second features in the two directions, and each detector correctly classified only eight of eleven test wheelsets.Fixing the variability feature identified using both datasets yielded eleven correct classifications in both directions.Source: §5.6 · p. 14, Figure 11 · p. 15

6. Discussion

The 109-feature model transferred poorly between sites, while thresholds transferred after the two decay features were fixed; feature selection and threshold stability remain sensitive to the available wheelsets. The screening results are limited by a small healthy sample, confounded support conditions, and the detector’s binary output.

6.2 Cross-Dataset Performance and Feature Selection

  • Thresholds transferred between sites after the two features were fixed, but the variability feature was selected using data from both sites, leaving single-site feature selection unconfirmed.The 109-feature model also performed poorly in cross-dataset testing; the authors caution that these results do not establish independent feature selection from one site.Source: §6.2 · p. 16

6.3 Threshold Stability and Screening Decisions

  • The variability threshold differed by approximately one-third between site calibrations and 63% across leave-one-wheel-out folds, making it more calibration-sensitive than the damping threshold.The damping threshold differed by approximately one percent between site calibrations and spanned 2.8% across fixed-feature folds; the authors recommend assessing the thresholds separately as data grows.Source: §6.3 · p. 16
  • Poor exponential-fit quality is the measured quantity for the variability feature, so excluding every poor-fit strike would remove variation that the detector is intended to capture.The elevated variability in three identified MxV wheelsets was not explained by weak recordings: their median signal-to-noise ratios were 53.0–55.8 dB, and the correlation with decay-fit quality was +0.05.Source: §6.3 · p. 16
  • The OR rule and flagging positions with neither feature available favor further inspection, but can increase healthy-wheel referrals.The authors recommend reporting healthy-class recall, missing-measurement frequency, and referral rate in future evaluations.Source: §6.3 · p. 17

6.4 Limitations

  • Only five of the 22 wheelsets were healthy, and all five were new, so the healthy class does not represent worn wheels still acceptable for service.One error changes healthy-class recall by 0.20 and balanced accuracy by 0.10; additional healthy wheels are needed to estimate referral rates and calibrate thresholds more reliably.Source: §6.4 · p. 17
  • Support condition was confounded with class at Norfolk Southern, and the detector provides only a binary screening decision, not defect type, location, or severity.Healthy wheels were tested on rail and defective wheels on the ground; controlled testing of both classes under both support conditions and prospective evaluation on an independent dataset are required.Source: §6.4 · p. 17

7. Conclusions

The study evaluates hammer-excited, air-coupled ultrasonic screening of stationary railway wheelsets using two complementary decay features. Fixed-feature thresholds transferred across the tested datasets, while nested validation classified 21 of 22 wheelsets; broader evaluation remains necessary.

7. Conclusions

  • Thresholds using the two fixed decay features correctly classified all eleven wheelsets in the other dataset, and nested validation correctly classified 21 of 22 wheelsets.The 109-feature Random Forest performed better with wheelsets from both sites in training but declined when tested on a different site.Source: §7 · p. 17
  • Feature selection restricted to a single dataset was less reliable: the resulting detectors correctly classified only eight of eleven wheelsets in the other dataset.The results support threshold transfer after features are fixed, but do not confirm that the same features can be selected independently from a single-site dataset.Source: §7 · p. 17
  • Evaluation was limited to 22 stationary wheelsets from two sites, with only five healthy wheels, all new; further work should include worn serviceable wheels and controlled support conditions.An independent dataset is also needed, and moving-wheel use requires additional work on excitation, background noise, and repeated measurements.Source: §7 · p. 17
Loading 2610.30541v1…