F06 Austin SCC Detectiodsn

OAK RIDGE NATIONAL LABORATORYU. S. DEPARTMENT OF ENERGY

Guided Waves for Stress Corrosion Crack Detection in Pipelines – Feature Selection

and Classification

Austin Albright, Venugopal K. Varma,

Raymond Tucker, and Philip Bingham


Outline Introduction and System Overview Current Methodology & Challenges Features Selection/Manipulation Our Technique & Features Results from Synthetic & Real-World

SCC Samples Lessons Learned & Future Work


Stress Corrosion Cracks (SCCs) are a growing concern for the Nation’s aging infrastructure.

Major contributors to creation of SCCs are: Repeated Stressing and Relaxation of the

System e.g. Thermal and operating pressure

variations, and other mechanical influences.

Environment Soil pH, moisture level, coating break down


Our Focus is to detect SCCs in large diameter Natural Gas Pipelines, specifically 26-inch and 30-inch diameter pipelines.

Corrosion + Cyclical Loading = SCC SCC generally found in colonies SCC are very hard to see with the naked eye The majority of SCCs run along the axial

direction of pipes We are looking for SCCs on the outside of the

pipes… from inside the pipe


Liquid Fluorescent Magnetic Particle Inspection allows SCCs to be visualized on pipes that have been removed from service & cleaned Suspension of fine metal particles in an aerosol spray can The suspect area is sprayed, then a strong magnetic field

is applied This draws the metal particles into any “depression” in

the surface such as nicks, scratches, and SCCs – The metal particles glow under a blacklight – The suspension is white paint and the particles are black

White Paint & Black Particle Version Blacklight Version

One or the other not both


Magnetic Flux Leakage (MFL) is the “standard” pipeline inspection technique used today

Used on active, buried natural gas pipelines MFL creates a magnetic field axially along the

pipe Unfortunately, the axial orientation of SCCs

combined with their small size result in little to no flux leakage.


As an alternative, we can inspect buried pipes using Electromagnetic Acoustic Transducers (EMATs) Allows ultrasonic Inspection without the need

of a liquid coupling agent EMATs can be designed to fit almost any

diameter pipe ORNL EMATs

have been designed specifically to create an ultrasonic guided wave to detect SCC (axial defects)

produce a shear wave traveling circumferentially in the pipe wall

utilize “pitch-catch” mode (one is the transmitter, the other is the receiver)


The ORNL Shear EMAT

f = J x B

f – is body force per unit volume

J – is current density [Amp/m2]

B – is magnetic flux density [Tesla]

Lorentz Force:

N

S

Magnet

Current Coils

Pipe WallShear

N

S

N

S N

S

N

S N

S

Shear Wave

Permanent Magnets

Pipe Wall

Current Coils

Aluminum Frame

Magnets

EMAT Coil

EMAT for 30” diameter pipes


ORNL’s Test Platform

Resolver

EMATs

Computer

Signal Conditioning unit

System Carriage


ORNL Platform Inside a 30 inch Pipe


The received EMAT Signatures are functions of axial position and time.

Axial Position [inches] can also be in [signatures]

Amplitude Color Indicates Amplitude

Dis

cret

e D

ata

Poi

nts

in T

ime

The features are extracted from the data in this boxed range

Features are numerical quantities that describe an event, object, or trait.

“Good” features should improve the discernability between classes

There are two categories describing feature sets:Supervised and Unsupervised

• “Supervised learning” uses known data sets (i.e., a signature in the set is known to be a “SCC signatures” or a “non-SCC signatures”) and determines what features discriminate these two classes.• “Unsupervised learning” assumes nothing about the class of the signatures making up the data • Supervised Feature Sets are difficult to create for real-world applications.


Wavelet filtering is a useful method for extracting information from transient signals, such as our EMAT signatures. Wavelet Analysis

Time-Frequency Decomposition Transients can be resolved Basis Function can be created or selected to “target” a

signal

Example Wavelet Decomposition Tree Details of our Wavelet Decomposition• 4-level Decomposition (yields 5 “pieces”)• Using “semi-custom” Basis Function


We still must “distill” features from the wavelet transform

• Energy – Percentage of Energy in each Wavelet Level.

• Entropy – Percentage of Entropy in each Wavelet Level.

• Difference Measure – the mean of each wavelet level is calculated from the “no-defect” set and subtracted from the matching level of the current signature and then summed.

• Point-by-Point Mahalanobis Distance (MD) – treat each discrete point of a wavelet level as an actual feature.

• (Point-by-Point MD)2 – square each wavelet level’s point-by-point MD value , for each signature.

If all the features were “good,” then using all the features would produce the best classification.

Feature Problems A feature can be noise Redundant Misrepresentative

Related Problems Overfitting Dimensionality Poor Classification

Common Techniques to Reduce these Problems are:

• Find Better Features

• Linear Discriminant Analysis (LDA)

• Principal Component Analysis (PCA)

Principal Component Analysis (PCA) also known as Karhunen-Lóeve transform

• Reduces the feature space dimensionality

• PCA is an Unsupervised Technique

To project: xey Τ

set

Set (matrix) of eigenvectors being used

Projected data Original data

eeS

where,

is the covariance matrix of the data set

is an eigenvector(s) of

is the eigenvalue(s) corresponding to the eigenvectors

S

e

S

Linear Discriminant Analysis (LDA)also known as Fisher Discriminant Analysis

To project: xwy Τ

set

Set (matrix) of eigenvectors being used

Original dataProjected data

Seeks to minimize the in-class variance while at the same time reducing the between-class variance

Projects the data

in to a matrix LDA is a Supervised

Technique

nc 1

wSwS WB

where,

is the between-class scatter matrix

is the within-class scatter matrix

is the eigenvectors of

is the eigenvalue of , each corresponds a single eigenvector in

BS

w

WS

BS

BS

w


LDA versus PCAEfficient Discrimination or Efficient Space


PCA+LDA, as a sequential operation applied to the original feature set

PCA is used to remove stochastic noise and lower the dimensionality

LDA is used to provide improve discernability between the classes

FeatureMatrix(n x m)

PCA LDA Classifier

Example: n – features m – samples 2 classes

k x m

k < n

1 x m

ClassA

ClassB


Which technique should be used? It depends on what your problem is, but the basic criteria are…

If you have a noise problem and/or have high dimensionality…

at least use PCA. If you have a class separability issue…

use LDA. If you have both problems…

use PCA+LDA. If you just plain aren’t sure what is going on…

try each of these techniques.


The Mahalanobis Distance (MD) calculated from a set of non-SCC signatures is used to classify the unknown signatures.

Statistically the non-SCC (“good”) signatures are well represented, which is why we use MD from the “good” signatures for our classifier.

MD returns a scalar value indicating a signature’s distance from the “good” cluster’s centroid.

We refer to the MD value as the “Flaw Distance,” since the larger the distance the more “flaw-like” the signature under test.


The feature transformation techniques presented have improved our defect detection.

Synthetic SCC Test Bed Unused (new) 10-foot long, 30” diameter pipe 4 scan lines were machined in to it 2 lines of parabolic cuts

circumferential widths of 8, 12, 16, and 20 mils

2 lines of rectangular cuts circumferential widths of 8, 12, 16, and 20 mils

Cuts ranged in depth from 10-75% of the wall thickness (0.375”).


Comparison of Results from a Parabolic Cuts Scan Line

MD using all of the original features

MD using 2 hand-picked features

MD after using PCA (13 out of 25 eigenvectors)

MD after using LDA

MD after using PCA+LDA


Blind Test of a Decommissioned Natural Gas Pipeline Section (known to contain SCCs)

Conducted at the Battelle Pipeline Simulation Facility (PSF) in Columbus, OH.

Given a pipe and specific areas from which we were to report our findings (from those areas only.)

Allowed to make multiple scans of the pipe to collect data.

Given 2 weeks to submit our findings.

Then the PSF staff distributed the “answer keys.”


January 2006, Blind Test Results


Every defect response in the MD plot corresponds to a defect and/or combination of defects

SCCsCorrosion

Corrosion & SCC

NOT SCCsSuperficial

Manufacturing “Handling” Marks

B

A

SCC


Conclusion/Summary We can detect SCCs in any orientation e.g. axial We also can detect pitting and corrosion

patches Our detection threshold is tied to the volume of

the defect, not merely depth of penetration PCA+LDA significantly improved the

discernability of both synthetic and real SCC Mostly by suppressing the responses generated by

metallurgic variations and small changes in gap between the EMAT and the pipe wall.

Most Important is, when the projection matrices calculated from the synthetic SCC data are used on REAL SCC data, the same improvements are achieved.


Future Work Automate identification of real defect

Develop a method for differentiating between corrosion patches and SCCs, and possible SCCs in a corrosion patch as well.

Investigate the benefits of “normalizing” the energy across an entire scan data set, in hopes of removing the false positives due to changes in the air gap (coupling) between the EMATs and the pipe wall.


Questions ?

THE END

ORNL Sensor System

Hardware Electromagnetic

Acoustic Transducers (EMATs) – the sensors

Resolver – for position

Signal Conditioning Devices – amps & matching networks

Computer – dual Xeon®

Data Acquisition Tone Burst Generator

Software Online

O/S Windows 2000®

LabVIEW™ (National Instruments) All Data Acquisition is

handled via LabVIEW

Offline Matlab®

(Mathworks)

Data Processing,Flaw Identification, Visualization, etc

We still must “distill” features from the wavelet transform.

15 and 14 13, 12, 11, jfor ,2))10()10(

(1

kjSM

kjiS

n

kijF

Difference Measure – the mean of each wavelet level is calculated from the “good” set and subtracted from the matching level of the current signature and then summed

5 and 4 3, 2, 1, jfor ,2

1

5

1

2

1

ijkpS

n

kp

ijkS

n

kij

F

Energy – Percentage of Energy in each Wavelet Level

10 and 9 8, 7, 6, jfor ,))5(

ln()5(1

kjiS

kjiS

n

kijF

Entropy – Percentage of Entropy in each Wavelet Level

We still must “distill” features from the wavelet transform, Continued.

20 and 19 18, 17, 16, jfor ,

))15()15(

(

)1)15(1)15(

(

1 )])15()15(

( )1)15(1)15(

[(

kjGM

kjiS

jGM

jiS

kjGM

kjiS

jGM

jiSM i

Point-by-Point MD – treat each discrete point of a wavelet level as an actual feature.

(Point-by-Point MD)2 – square each wavelet level’s point-by-point MD value , for each signature.

• GM – mean of level j of the “good” set• Γ-1 – the “good” set’s inverse covariance matrixNOTE: j is “formed” so that all the features can be calculated in one pass. GM is accessible as a whole level or point-by-point

• j – the wavelet level• i – signature number• k – number of discrete points in each wavelet level e.g. detail-4 k = 32• p – number of wavelet levels• S – signature data vector

Notation


PCA can… reduce stochastic noise lower dimensionality by representing the

“information” in the most efficient space

LDA can… improve classification by better separating the

classes lower dimensionality, returns feature matrix as

a (Num. Classes – 1) x (Num. Samples) matrix

PCA+LDA can… remove stochastic noise lower the dimensionality Improve discernability between classes

Documents

F06 Austin SCC Detectiodsn