AUTOMATIC CLASSIFICATION OF DEFECTS IN TIME-OFFLIGHT-DIFFRACTION DATA

O. Zahran and W. Al-Nuaimy

Department of Electrical Engineering & Electronics,
University of Liverpool, Brownlow Hill, Liverpool
Abstract: Ultrasonic Time-Of-Flight Diffraction (TOFD) is a recent
innovation that has proved highly effective for the inspection of steel plates
and tubular pipelines and has started to take its way to replace the other
ultrasonic testing techniques. TOFD technique has a lot of advantages
which make it the preferable technique in material testing. This technique
gives accurate sizing, positioning and characterising of weld and other
defects with a high probability of detection. TOFD still suffers from the
difficulties associated with the interpretation of the images. Feature analysis
and shape detection have been applied to TOFD images. Combining the
obtained features with an artificial intelligent system allows the full
automation of the interpretation process. The results of an initial
investigation are presented and discussed.
1. Introduction: Ultrasonic techniques are still the most popular nondestructive testing method. TOFD has successfully been applied to the
inspection of steel plates and tubular pipelines. TOFD has a lot of
advantages which make it the preferable technique in material testing. It
gives accurate sizing, positioning and characterising of weld and other
defects with a high probability of detection.
Currently most of the TOFD data interpretation is done manually, requiring
operator skill, experience and most significantly time. In light of the
industrial pressure, the recent trend is to partially or fully automate the
inspection and data interpretation process. This could potentially improve
these procedures by adding and element of robustness and consistency by
utilising computational tools that are better suited to discriminating between
subtle variations in visual and spectral properties of the data. Furthermore,
this could potentially save money, effort and time (1).
Each defect class has its main characteristics. The main clues of
characterising each defect class can be used for the classification of the
detected defects. The classification of these defects may be possible using
advanced signal and image processing to build an artificial intelligent
system able to use these main clues and perform the automatic classification
of detected defects.
2. TOFD: TOFD first appeared in 1977 and started to take its way to
replace the other ultrasonic testing techniques. This technique has a lot of
advantages which make it the preferable technique in material testing (2, 3,
4). There are many successful examples for applying TOFD technique,
which show that TOFD is a powerful testing tool which gives accurate
sizing, and characterising of weld defects.

TOFD is based on measurement of the time of flight of the diffracted

echoes of ultrasonic waves on the tips of discontinuities (defects) which are
directly related to the true position and size of the defect instead of
geometrical reflection from the interface of the discontinuities in traditional
methods (5). This technique uses two probes in a transmitter-receiver
arrangement as shown in Fig. (1). When ultrasound is introduced into the
material, each defect edge works as a point source of diffracted waves. The
received signals can be visualized in an A-scan presentation or stacked
together to give a 2-dimensional image called a B or D-scan representation
as shown in Fig. (2).

Figure 1: TOFD technique principle

The most important advantages of TOFD technique are that, TOFD defect
detection does not depend on the defect orientation, in contrast to the other
techniques, defect height can be exactly determined, depth sizing is very
accurate with a high probability of detection up to 95%, and very low cost
(6, 7, 8,9).

Figure 2: TOFD (a) A-scan presentation (b) D-scan presentation

3. Phase relation: When a wave is reflected at the interface between two
media from higher acoustic impedance to lower acoustic impedance, there
will be a 180 phase difference. Therefore when the backwall signal is
reflected at the interface between steel and air, there will be a phase
difference 180 between lateral wave and backwall echo.
When the material under test containing a defect, there will be a 180 phase
difference between the signal from the top of the defect and the lateral wave
as if it had undergone a reflection which means that the phase is like the
backwall echo. The signal from the bottom of the defect is still in phase
with the lateral wave which means there will be a 180 phase difference
between the two defect echoes as shown in Fig. (3).
Theoretically if there is a 180 phase difference between two adjacent
diffracted signals, this means they must have a continuous crack between
them. Rarely in a very few cases the top and bottom diffraction signals may
not have this phase change, but in general they will.

Figure 3: Phase relationships

It can be concluded that the determination of phase relation between defects
echoes and comparing them with the lateral wave and backwall echoes are
very important especially for characterising of some defect classes and also
to achieve the most accurate defect sizing. Therefore a phase determination
algorithm based on the maximum correlation between the two signals has
been investigated in order to be used to decide whether the two signals are
in phase or out of phase.
4. Defect characterisations: The common defects in welds can be
classified into four main categories, planar flaws, volumetric flaws, threadlike flaws and point flaws (10). Each category has special characterisations
and patterns but there are some similarities between these categories which
makes the discrimination between these categories not an easy task. It is
very important to study these characterisations and patterns carefully which
may be helpful in providing an automatic interpretation system.
Planar flaws include cracks and lack of fusion. Planar flaws may be open to
upper surface, breaking the lower surface or internal.
The planar flaws open to the upper surface show up as an echo from the
bottom edge of the flaw with a higher frequency content usually
accompanied by a loss or a weakening of the lateral wave signal and the
phase is still as the lateral wave. Also, apparent migration of lateral wave
echo to deeper depths usually appeared as shown in Fig. (4). The defect
signature is this case is similar to a reflected moustache shape. The spectral
analysis and shape detection may be used to classify this class of defects.

Figure (4): Upper surface breaking

pattern
While the planar flaws breaking the lower surface show up as an echo from
the top edge usually accompanied by an increasing delay in and/or
weakening of the backwall signal. The effect on the backwall depends on

how much the deep of the crack as shown in Fig. (5). The phase of the echo
is still the same as the backwall echo. The defect signature is this case is
similar to a moustache shape. The shape detection may be used to classify
this class of defects.

Figure (5) Lower surface breaking defect

Internal planar flaws show as two echoes with a distinct 180 phase
difference between the echoes from the top and bottom tips of the flaw. The
phase of the upper tip echo is the same as the backwall echo while the lower
one is the same as lateral wave. Both echoes have a similar amplitude and
defect signature. The defect signature in this case is also similar to a
moustache shape as shown in Fig. (6). The shape detection may be used to
classify this class of defects. Lack of fusion is very similar to the internal
cracks and both have two echo signals with 180 phase change. More
analysis techniques may be used to discriminate between these two defects
which are included in the same class.

Figure (6): Internal crack pattern

Volumetric flaws include lack of penetration and large slag lines. The
echoes from reflectors of this type also show the features and phases
outlined for internal planar flaws but the echo from the upper surface is
greater than the diffracted around the lower surface. The target signature of
the large slag lines look like the straight line as shown in Fig.(7). Therefore
the shape detection may be used to classify this class of defects.

Figure (7): Large slag line

Thread-like flaws include flaws with significant length but little through
wall extent such as lamellar flaws and near horizontal area lack of fusion.
The reflector appears as an apparent upper edge echo in phase with
backwall echo without lower edge echo. Long narrow slag shown in Fig. (8)
can be considered as an example of this categories. The target signature of
this category look like the straight line therefore the shape detection may be
used to classify this class of defects.

Figure (8): Threadlike flaw

Point flaws category includes pores and small pieces of slag. These flaws
are most common in welds and their echoes have similar pulse
characteristics to the volumetric or thread like flaws but have no resolvable
length. Point flaws give multiple echoes but with no other co-linear echoes
at greater or lesser depth in the specimen which are similar to patterns of
acoustic noise. This class of defects may be classified using structure
features. The defects of this category produce signals which look like arcs
on D-scan as shown in Fig. (9).

Figure (9): Point flaws

The automatic discrimination between these categories can be achieved by
applying a combination of analysis techniques in order to look for the
distinguishable characteristics of TOFD signals for each defect category
(11).
6. Interpretation: To fully automate the interpretation process of TOFD
data, the interpretation process can be divided into four main stages as
shown in Fig. (10). These stages are pre-processing stage, defect detection
or pre-classification segmentation stage, defect classification stage and
post-processing stage. Each stage contains a number of processes which
will be applied in order to prepare the input of the next stage until obtaining
the final output which is the interpretation result.
Raw
image

Preprocessing

Image
segmentation

Image
classification

Figure (10): Automatic interpretation system

6.1 Pre-processing stage

Postprocessing

Interpretation
result

In pre-processing stage there are two processes to be done over the raw
image in order to prepare it for the segmentation stage. These two processes
are cutting the mode converted areas from the raw image and global
background removal as shown in Fig. (11).
The mode converted echoes may be useful in the case of shallow flaws
which may be hidden by the lateral wave. In that case mode converted
echoes can be used as an indication of the presence of the shallow defect
but it is better to remove the mode converted areas because the detection
process can not relay on mode converted echoes because it sometimes give
false indication without the presence of the defect. In addition to the
significant reduction in processing time by removing the mode converted
echoes.
There are two important features in all TOFD scans, the backwall echo and
the lateral wave signal. These two features have to be removed with the
background in order to reduce the amount of data processed in the next
steps. If the lateral wave and backwall are regular and are not affected by
the presence of defect, they can be removed by subtracting the mean of the
image scans from each scan in the image. If there are some irregularities in
lateral wave or backwall which is common, the process of global
background removal is considered not enough for removing all the lateral
wave and backwall echoes and therefore intelligent background removal
have to be applied in the next stage.
Raw
image

Cut mode-converted
areas

Image ready for

segmentation

Global (standard) background

removal

Figure (11): Pre-processing stage of the automatic interpretation system

6.2 Segmentation stage
Segmentation is a very important stage in TOFD interpretation because only
a small fraction of the collected data actually represents defects, whereas
the majority of the data is considered redundant. The image segmentation
permits more exhaustive image processing to be applied to the image in the
next stages.
In segmentation stage there are three processes to be done over the output
image from the pre-processing stage. These processes are statistical
detection, intelligent background removal and then statistical detection
again as shown in Fig. (12).
Image after preprocessing

Statistical
detection

Intelligent (informed)
background removal

Statistical
detection

Image ready for

classification

Figure (12): Segmentation stage of the automatic interpretation system

The first process in segmentation stage is statistical detection by
considering the variation between statistical features in the defect areas and
background areas in order to detect only the defect areas. This process
result in a segmented image with highlighted defects areas but also there
will be small highlighted areas which related to the places of irregularities
in lateral wave and backwall echoes. To remove these small detected areas
which are not defects, an intelligent background removal has been applied.
This intelligent background removal depending on determining the A-scans
which not detected in the previous detection and then subtracting the mean
of each adjacent group from these A-scans from the whole A-scans in the
original image. After that the statistical detection is done again. The output
at that time is the segmented image with highlighted only the defects areas.
6.3 Classification stage
The most important stage in interpretation process is the classification. It is
a very complex stage; therefore little research has been done in this field.
To achieve this stage a combination of analysis and discrimination
techniques may be used as shown in fig.(13).
Texture
features
analysis

Shape
detection
analysis

Point flaws
category

Porosity
Small pieces of slag

Phase
determination
algorithm

Straight line

Thread-like flaw category

Segmented
image

Fuzzy logic
Moustache
Hyperbolic

Planar
flaws
category

Upper surface breaking crack

Lower surface breaking crack
Internal crack
Lack of fusion

Figure (13): Classification stage of the automatic interpretation system

Statistical texture analysis computes local features at each point in the
segmented image, and calculates a set of statistics from the distributions of
the intensity levels. The local feature is defined by the combinations of
intensities at specified positions relative to each point in the image. Because
of point flaws category has special texture than other categories, texture
features analysis may be used to identify this category, especially the
second order statistics. Point flaw category includes pores and small pieces

of slag. Fuzzy logic may be used to discriminate between these two types of
defects.
As explained previously, some flaw categories have a special defect
signature shape. Therefore these defects may be classified based on the
shape they produce in D-scans. There are three expected shapes as
explained previously, straight line, moustache and hyperbolic.
The straight line signature represents thread-like flaw category if there is an
apparent upper edge echo in phase with backwall echo without lower edge
echo, or represent a large slag line (volumetric flaw category) if there are
two echoes from the reflector which show the phases outlined for internal
planar flaws but the echo from the upper surface is greater than the
diffracted around the lower surface.
The moustache shape signature represents the planar flaws category. If
there are only one moustache shape signature, this means there is a lower
surface breaking flaw as indicated in Fig. (5). If there are only one reversed
moustache shape signature, this means there is an upper surface breaking
flaw as indicated in Fig. (4). If there are two moustache shape signatures,
this means there is an internal crack or lack of fusion. The phase
determination at that time is important to confirm the classification decision
by approving the phase relation between defect echoes, lateral wave and
backwall.
The hyperbolic shape signature may represent the lack of fusion or internal
crack (planar flaws category) if there two echoes with different phases.
Therefore the phase determination at that time is important to confirm the
classification decision by approving the phase relation between defect
echoes, lateral wave and backwall. If there is only one echo with no
resolvable length, the hyperbolic shape may represent a point flaw.
6.4 Post-processing stage
After detection and classification stages, the final stage is coming. In postprocessing stage there are three processes to be done in order to obtain the
final output of the system which is interpretation result. These three
processes are sizing, calibration and then applying the acceptance codes as
shown in Fig. (14).
Detected and
classified defects

Sizing

Calibration

Acceptance
codes

Interpretation
results

Figure (14): Post-Classification stage of the automatic interpretation system

The first process in post-processing stage is sizing of the detected defects.
The sizing is this stage is related to the number of pixels not to the actual

size. Therefore this process should be followed by calibration process to

ensure that the measured size expresses the actual size of the defects. The
last process in this stage is applying the acceptance codes to decide is this
defect accepted or not.
7. Results
An example of the preliminary results is shown below in Fig. (15). It shows
the output of each stage which are explained above. The final output image
shows that the desired defect areas are successfully detected and classified
with acceptable level of accuracy in a fully automatic and un-supervised
manner. The results have so far been promising.

(a)
(b)

(c)
(d)
Figure (14): Output images of each stage (a) Original image (b) Image after
pre-processing stage (c) segmented image before intelligent background
removal (d) Final detected and classified image
8. Conclusion
A comprehensive automatic interpretation of TOFD data can be achieved
with satisfactory levels of accuracy and reliability which could potentially
save money, time and effort. This can be done by combining the main
characteristics of each defect category with advanced signal and image
processing techniques to build an artificial intelligent system. The results of
preliminary results are presented. The obtained results were quite good and
promising in terms of accuracy. It is expected that such techniques will
greatly reduce the degree of reliance on the trained operator during initial
site investigations.

Acknowledgments
The authors would like to express their gratitude and appreciation to Karl
Quirk of Phoenix Inspection Systems Ltd and to Tim Armitt of Lavender
International NDT for providing the TOFD data and their guidance on the
interpretation of the TOFD data.
References
1. O Zahran, S Shihab and W Al-Nuaimy, Recent developments in
ultrasonic techniques for rail-track inspection, BINDT 2002, ISBN no.
(0-9031-3230-3), pp 55-60, September 2002, United Kingdom.
2. A Erhard and U Ewert, The TOFD method between radiography and
ultrasonic
in
weld
testing,
(www.ndt.net/article/v04n09/erhard/erhard.htm), Vol 4, No 9,
September 1999.
3. R Krutzen, Evaluation of currently applied ultrasonic sizing techniques
for stress corrosion cracks in steam generator tubes, 17th EPRI Steam
Generator NDE Workshop, Breckenridge, Colorado, USA,
(www.nuson.nl/news/do05pres.html), August 1998.
4.

N Trimborn, The time-of-fight-diffraction technique, NDTnet,

(www.ndt.net/article/tofd /trimborn/ trimborn.htm), Vol 2, No 9,
September 1997.

5. M Silk, The rapid analysis of TOFF data incorporating the provision of

standards, (www.ndt.net/article/tofd/Silk/Silk.htm), Vol 2, No 9,
September 1997.
6. F Betti, A Guidi, B Raffarta, G Nardoni, P Nardoni, and L Nottingham,
TOFD - the emerging ultrasonic computerized technique for heavy
wall
pressure
vessel
welds
examination,
(www.ndt.net/article/v04n09/nardoni/ nardoni.htm), Vol 4, No 9,
September 1999.
7.

F Betti, G Zappavigna, C Pedrinzani, G Nardoni, and P Nardoni,

Accuracy capability of TOFD technique in ultrasonic examination of
welds, (www.ndt.net/article/wcndt00/papers/ idn634/ idn634.htm),
2000.

8. A Hecht, Time of fight diffraction technique (TOFD) - An ultrasonic

testing
method
for
all
application,
(www.ndt.net/article/tofd/hecht/hecht.htm), Vol 2, No 9, September
1997.
9. S Webber, Five years of testing using the semi-automated ultrasonic
time
of
fight
diffraction
system,
(www.ndt.net/article/apcndt01/papers/898/898.htm), 2001.

10. British standard Institution, The British TOFD standard BS 7706,

British Standards Institute, 1993.
11. O Zahran, S Shihab and W Al-Nuaimy, Discussion of the ability of
defect detection classification in weld inspection using ultrasonic timeof-flight diffraction technique, PREP 2004 conference, April 2004,
United Kingdom.