John S. PopovicsThe University of Illinois at Urbana-Champaign

William W. Hay Railroad Engineering SeminarFebruary 5, 2021

Assessment of Rail Track Condition Using Novel Nondestructive

Evaluation Technologies

My research work focuses on sensing, testing and imaging for infrastructure materials and structures

We must consider complex material characteristics, large size and limited access to engineer the next generation of cost effective and accurate health monitoring technologies

for effective forensics and rehabilitation ensuring sustainable infrastructure 2

In situ characterization of concrete rail tie condition using air-coupled

ultrasonic measurements

Mr. Sai Evani and Dr. Agustin Spalvier

Monitoring (concrete) rail tie condition is important

Conceptual vision of contactlessscanning approach for effectiverail tie inspection

Defective rail ties fail to provide required support to the rails, resulting in development of additional bending stresses in the rail and alteration of track geometry, both of which increase the risk of derailment.

Condition assessment of rail ties is typically performed by an experienced technician though visual inspection.

This process is time consuming and inefficient in cases where the defect does not manifest visibly on the surface, or when deterioration occurs in inaccessible locations, such as the rail seat area.

Our approach for monitoring tie condition: air-coupled ultrasonic surface wave inspection

Surface waves scatter when they encounter inclusions in the rail tie e.g. cracks, delaminations, and other damage. Fully air-coupled sensors transmit and receive mechanical

waves propagating along the free surface of a rail tie without physical coupling.

Forward Backward

Air-coupled testing configuration: transducer sender and MEMS array receivers

Contactless sender

coherent

Incoherent (scatter)

Contactless MEMS array

1 23 4

5 6 7

Conceptual visualization of generated surface wave fields

Ultrasonic scan data were collected from each end of concrete rail tiesA scanning frame placed on top of the rail controlled the position of the sensors, where the transmitted position was fixed at the tie center and the receiver array scanned across the

shoulder portion of the tie to form a scan line

Center of the tie

Position 1Position 35

35 points @ 5mm spacing

MEMS Transmitter

20cm

MEMS 1

A

Aluminum frame

106cm

2cm

Surface waveScanregion

A

Path A-AMEMS 2

C CB B

Data were collected from a set of concrete rail ties with varying visual (ground truth) condition: healthy, transverse cracking, longitudinal cracking, RSD, corrosion

Data from 81 healthy and 93 damaged scan lines were collected

Signal data were analyzed in both time and frequency domains to extract information about propagating and backscattered wave fields

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Time (ms)

-5

-4

-3

-2

-1

0

1

2

3

4

Mag

nitu

de

0 50 100 150 200 250 300

Time ( s)

0

50

100

150

200

250

300

350

Spac

e (m

m)

-400 -300 -200 -100 0 100 200 300 400

Wavenumber (rad/m)

20

30

40

50

60

70

80

Freq

uenc

y (k

Hz)

0

0.2

0.4

0.6

0.8

1

Stacked signals over space

2-dimensional Fourier transform (over time and space). Color bar shows

normalized amplitude

Time signal Time-space domain

Frequency-wavenumber domain

Coherent pulse

Coherent pulsecontent

Speed and energy characteristics can be extracted from t-x domain data

t-x scan from a healthy rail tie t-x scan from a damaged rail tie

A robust wavefront fitting algorithm identifies the first four wavefronts, fit a straight line and consider the average as wave speed. Wave energy is calculated by integrating the amplitude of the time signal. Both wave speed and wave energy are expected to decrease in the presence of damage

Speed and energy characteristics can also be extracted from f-k domain data

f-k scan from a healthy rail tie f-k scan from a damaged rail tie

The peak amplitude within the f-k domain serves as a central point for this: the magnitude of the peak amplitude provides energy while speed is provided by 2πfpeak / kpeak

Speed (f-k)=2691m/s Speed (f-k)=2143m/s

How to analyze the various data? Machine learning!

• We have four signal parameters collected across 174 scan lines (81 healthy and 93 damaged)

• The whole data set was distributed randomly into a training set (50-healthy and 50-damaged) and a test set (31-healthy and 43-damaged)

• The training set is used to estimate a threshold boundary between healthy and damaged conditions (a “decision space”) using the support vector machine (SVM) algorithm. The test set is used to evaluate accuracy of the decision space.

• The SVM algorithm computes a hyperplane by maximizing the distance between proposed decision boundary and training data points near the decision boundary (called support vectors) such that most of the data points in the training set are correctly classified (i.e., the predicted label is same as the ground truth label)

Example two-dimensional decision spaces computer from the training data

We can also create a four-dimensional decision space using all the extracted parameters

Performance of four-dimensional decision space

Prediction accuracy with test set is about 80%

Performance on training set

Damaged Healthy Accuracy (%)

Damaged 40(True positive)10

(False negative) 80

Healthy 12(False positive)38

(True negative) 76

Performance on test set

Damaged Healthy Accuracy (%)

Damaged 34(True positive)9

(False negative) 79

Healthy 5(False positive)26

(True negative) 84

Predicted condition

True condition

Predicted condition

True condition

Summary

o Air coupled sensors provide high quality signal data while using a platform that provides practical in situ implementation.

o Four signal parameters, extracted from t-x and f-k domains, demonstrate sensitivity to the presence of damage in concrete rail ties; each of these parameters showed global reduction in value in the presence of damage.

o Machine learning algorithms help us sort out a very complicated data set and draw conclusions from it. The four-dimensional decision space provided predictions with highest accuracy (80%) among all the decision spaces considered in this study.

In situ estimation of rail stress using nondestructive vibration

measurements

Dr. Xuan Zhu, Mr. Chi-Luen Huang, Mr. Marcus Dersch, and Mr. Sangmin Lee

Monitoring longitudinal rail stress in continuously welded rail is important

Rail buckling event caused by restrained thermal expansion

Derailment in Northbrook, ILcaused by rail buckling (2012)

Rail track thermal buckling is a long-standing safety issue for continuously welded rail structures (CWRs), usually caused by excessive longitudinal stress developed by extreme temperatures.

Effective rail neutral temperature (RNT) management is essential to track buckling prevention. There is a demand from the industry for in-situ rail stress/RNT measurement to facilitate maintenance decision making.

Microphone

Impactor

A B

C

D

E

Our approach for monitoring longitudinal rail stress: contactless vibration measurements

Acoustic measurements of rail vibration provide a rich data set that contains information about the temperature and stress state of the rail, but also contains disruptions from other

influences. We believe that a machine learning approach can extract useful information about rail stress from the complex acoustic data set.

Contactless tests Computational models Supervised learning

Contactless test configuration using microphones

Condensermicrophone

Impactor

A BC

D

E

BA

C

DE

Microphone: Effective frequency response with ±2dB ranging from 4 to 70000 Hz

o 5 locations across a cross-section for impacting and receiving

o 25 combinations of impacting and receiving

o “E-A” configuration illustrated here

A BC

DE

2.29 kHz 5.78 kHz 6.46 kHz 9.72 kHz 14.02 kHz

19.78 kHz 22.72 kHz 31.75 kHz 32.95 kHz

35.4 kHz 38.57 kHz 42.24 kHz 46.39 kHz

Fixed supporting condition

FEM vibration simulation showing vibration mode shapes and frequencies

Field tests on in-service rail

o BNSF main line near Streator, IL

o Instrumented with temperature and strain sensors before rail cut

o Multiple data collection visits during the period from early August 2019 to June 2020

o Two testing locations: East and West (each at 2m from the cut)

o Different support conditions: Tie spacing, rail anchor, spike tightness

West location

East locationCut

location

Detail about sensors and stress and temperature data acquisition

Additional strain gauges for train detecting

Welding strain gauges in full-bridge configuration

Example rail condition data from the east test location throughout the day of Aug. 5, 2019

*Rail neutral temperature (RNT) is the railtemperature at which zero axial stress is present inthe rail.

Temperature

Strain

RNT*Times of vibration data

collection

Field tests on in-service rail

Contactless microphone provides high quality, reliable vibration data

The instrumented rail system provides continuously streaming, high quality stress, temperature, and RNT data

Multiple vibration resonance frequency values across a broad range of frequencies are affected by temperature-induced stress in the rail -- this provides a rich data set for us to use

~37 kHz ~76.5 kHz

@ Highest temperature and strain

@ Highest temperature and strain

76 kFrequency (Hz) Frequency (Hz) 36.5 k 37.5 k3.25 k 4.25 k Frequency (Hz) 77 k

~3.8 kHz

Sensitive More sensitiveInsensitive

We need help extracting meaningful signal features from the entire spectrum: we use non-negative matrix factorization (NMF)

36781 Hz8873 Hz 76627 Hz

NMF looks for patterns and consistency in rich data sets, identifies them and groups them

1st component2nd component

3rd component

Stress and temperature prediction using AI using selected features as input values

Hidden layers

Stress

1st sensitive: f45 – f9

2nd sensitive: f28 – f15

4th sensitive: f102 – f36

5th sensitive: f95 – f50

3nd sensitive: f56 – f27

…… …… Temperature or RNT

o The neural network will be trained using field data from the east location

o The neural network performance will be evaluated using data from the west location

• Input: differential frequencies with top sensitivity to strain and temperature

• Inputs labeled by strain and temperature

• Supervised learning with feed-forward neutral network (FNN)

• Output: stress and temperature and RNT

Summary

o Air coupled acoustic sensing allows us to collect lots of meaningful data, with potential to be implemented on a moving test platform.

o Vibrational resonance data show correlation with temperature induced stress.

o Machine learning algorithms help us sort out a very complicated data set and draw conclusions from it.

Acknowledgments

o The work reported here represents important contributions from a number of collaborators: Dr. Xuan Zhu, Mr. Chi-Luen Huang, Mr. Marcus Dersch, Mr. Sai Evani, Dr. Agustin Spalvier and Mr. Sangmin Lee, access to the rail test site provided by BNSF railways, and overall assistance from RailTEC.

oThe work was made possible through support from the US National Academies (Rail safety IDEA) and the American Association of Railroads (Tech Scan).

Slide Number 1Slide Number 2Slide Number 3Slide Number 4Slide Number 5Slide Number 6Slide Number 7Slide Number 8Slide Number 9Slide Number 10Slide Number 11Slide Number 12Slide Number 13Slide Number 14Slide Number 15Slide Number 16Slide Number 17Slide Number 18Slide Number 19Slide Number 20Slide Number 21Slide Number 22Slide Number 23Slide Number 24Slide Number 25Slide Number 26Slide Number 27Slide Number 28Slide Number 29