J. V. JOHNSONSELV A, J. RAJA SEKAR: DEFECT CHARACTERIZATION OF METALLIC MATERIALS USING ...
719–728
DEFECT CHARACTERIZATION OF METALLIC MATERIALS
USING SELF-PARAMETERIZED DENSITY-BASED CLUSTERING
AND COMPUTATIONAL INTELLIGENCE TECHNIQUES
KARAKTERIZACIJA NAPAK V KOVINSKIH MATERIALIH Z
UPORABO SAMOPARAMETRIZIRANE GOSTOTE GROZDOV IN
TEHNIK RA^UNALNI[KE INTELIGENCE
J. V. Johnsonselva
*
, J. Raja Sekar
Department of Computer Science and Engineering, Mepco Schlenk Engineering College, Sivakasi, India
Prejem rokopisa – received: 2023-09-30; sprejem za objavo – accepted for publication: 2024-09-30
doi:10.17222/mit.2023.982
Rapid growth of manufacturing industries is propelled by transformative technologies such as machine intelligence, autonomous
computing and non-destructive testing (NDT). During the manufacturing of wrought products, there is no guarantee that the fi-
nal product is 100-% flawless. Thus, all final products are subjected to quality checking to identify and eliminate defective prod-
ucts. In industries, most of internal defects are identified using NDT techniques, which fail to precisely characterize the defects.
In this paper a novel algorithm, called Self-Parameterized Density-Based Clustering (SPDBC), is proposed for defect character-
ization. The proposed clustering method uses spatial parameters to identify the size and position of defects by filtering out the
noise and other data that correspond to the non-defect area. Using these filtered data, computational intelligence techniques are
employed to predict the defect type. SPDBC achieved Jaccard indices of 97.02 % and 98.78 % for identifying the defect size
and position, respectively. Gradient boosting regression trees (GBRT) achieved a maximum accuracy of 97.44 % in predicting
the defect type. As a result, the proposed approach can assist NDT experts in various sectors to differentiate between problem
severities faster and replace defective parts before any major breakdown occurs.
Keywords: material defect characterization, ultrasonic non-destructive testing, density-based clustering, artificial intelligence
Hitro rast industrij razli~nih izdelkov poganjajo nove prodirajo~e tehnologije, kot so strojna inteligenca, avtonomno
ra~unalni{tvo in neporu{no testiranje materialov (NDT; angl.: Non-Destructive Testing). Med proizvodnjo surovih izdelkov je
te`ko zagotoviti, da bodo le-ti 100 %-no brez napak. Zato se kon~ne proizvode oz. izdelke obvezno kontrolira z namenom, da se
odstrani slabe, preden se jih dobavi naro~nikom ali potencialnim kupcem. V industriji se ve~ina notranjih napak ugotavlja z
NDT tehnikami, toda le-te pogosto niso dovolj natan~ne, da bi z njimi lahko identificirali vse napake v dolo~enem materialu. V
tem ~lanku avtorji opisujejo razvoj novega algoritma imenovanega samo-parametrizirana gostota na osnovi skupljanja (SPDBC;
angl.: Self Parameterized Density Based Clustering) in ga predlagajo kot metodo za karakterizacijo napak v materialih.
Predlagana metoda uporablja prostorske parametre za identifikacijo velikosti in polo`aja napak in nato izvr{i filtriranje »hrupa«
in ostalih podatkov, ki predstavljajo podro~ja brez napak. S tehnikami ra~unalni{ke inteligence nato uporabijo filtrirane podatke
za napoved vrste napake. Z metodo SPDBC so avtorji dosegli Jaccardov indeks 97,02 % pri identifikaciji velikosti in 98,78 % za
identifikacijo polo`aja napak. Gradientno oja~ana regresijska drevesa (GBRT; angl.: Gradient Boosting Regression Trees) so
dala maksimalno natan~nost 97,44 % za napoved vrste napak. Posledi~no avtorji predlagajo, da bi lahko ta pristop pomagal
ekspertom s podro~ja NDT tehnik pri hitrej{em razlikovanju nevarnosti prisotnih napak in hitro zamenjavo z novimi rezervnimi
deli, {e preden pride do resnej{ih po{kodb in/ali zlomov kon~nega izdelka.
Klju~ne besede: karakterizacija napak v materialih, ultrazvo~ne neporu{ne tehnike testiranja materialov, gostota na osnovi
zdru`evanja v grozde, umetna inteligenca
1 INTRODUCTION
In general, defects are defined as any kind of un-
wanted irregularities present inside the material struc-
ture. In an industrial environment, material defects pose
a serious threat of causing equipment malfunctions.
Some defects of industrial materials originate from the
earlier stages of the manufacturing process such as cast-
ing and moulding. During quality control (QC),
1
most of
defective materials are identified and removed but some
may pass through because of their small size. There is a
high possibility that defects which slipped through QC
may become larger during further processing of wrought
products. Thus, an industrial material may contain both
internal and external defects. While an external defect is
easy to observe with the naked eye, internal defects are
hard to identify. Industrial equipment failures are com-
mon due to the growth of internal defects, which occur
due to uneven temperature and stress in the material
structure. Since internal defects are uncertain, it is chal-
lenging to accurately detect their presence without
breaking the material structure. To identify internal de-
fects precisely, non-destructive testing (NDT) is adopted
in industries.
2,3
Some industries employ magnetic-based
NDT to identify defects but it cannot be used for
non-magnetic alloys and multi-layered composite mate-
rials. To overcome the limitations of magnetic-based
NDT, ultrasonic-based NDT with the pulse-echo tech-
Materiali in tehnologije / Materials and technology 58 (2024) 6, 719–728 719
UDK 666.9.019:004.896 ISSN 1580-2949
Original scientific article/Izvirni znanstveni ~lanek Mater. Tehnol.
*Corresponding author's e-mail:
johnsoninboxx@gmail.com (J. V. Johnsonselva)

nique is widely used in industrial environments. Ultra-
sonic equipment operates very well in a harsh industrial
environment.
4
Ultrasonic-based NDT uses through-trans-
mission and pulse-echo techniques to analyze materials.
The through-transmission technique is used on compos-
ite materials,
5
concrete,
6
aluminum alloys,
7
carbon fiber,
8
etc. The pulse-echo technique is used on metals and its
alloys. The proposed work is concentrated only with
metals, and so the pulse-echo technique is employed.
In the industrial domain, there are numerous chal-
lenges in performing a NDT analysis.
9
As per the litera-
ture, the ultrasonic NDT techniques are widely used for
observing the integrity of metals. The researchers have
developed a method to observe discontinuity in adhesive
bonds using the ultrasonic pulse-echo technique.
10
This
technique succeeds in observing the discontinuity even
in multi-layered metal composites. A research proposal
included an automatic approach for the inspection of
composite materials using the ultrasonic pulse-echo tech-
nique.
11
However, the success rate of an NDT analysis
depends upon the method and ability to collect the data.
12
The most challenging and complex problem with ul-
trasonic NDT is that it needs a level-II expert to diagnose
defects precisely. Defect characteristics that are collected
with the ultrasonic equipment need to be analyzed manu-
ally by a human expert. Some methodologies for defect
characterization using computational intelligence tech-
niques are presented in the literature.
13,14
To deal with in-
complete data, the system uses the Improved Mean Im-
putation Clustering Algorithm
15
and Kernel-Based Fuzzy
C-Means Algorithm
16
with a considerably positive out-
come. Research experts have developed a defect classifi-
cation method using Probabilistic Neural Networks
(PNNs).
17,18
The Convolution Neural Network (CNN)
based ultrasonic testing is used to detect and classify
railhead surfaces and subsurface defects.
19
A method has
been developed using a neural network-based solution
and radiographic images to categorize defects, but it is
limited only to the identification of cracks.
20
The existing
methods presented in the literature are limited only to the
classification between defect and non-defect signals. The
computation of defect characteristics such as defect size,
position and type are not addressed. There are currently
no studies in the literature covering automatic computa-
tion of defect characteristics such as defect size, position
and type using computational intelligence techniques.
Hence, an effective and automatic approach is required
to aid the NDT experts in defect characterization.
In this work, a novel Self-Parameterized Density
Based Clustering (SPDBC) algorithm is proposed to
identify the defect size and position in wrought products.
Several classifiers are employed in this work to catego-
rize the defect type from the clustered data. This will as-
sist the NDT experts to identify the severity of defects
more effectively.
21
It is extremely beneficial for the in-
dustries that produce heavy-duty vehicles to locate de-
fects in wrought products before they are subjected to the
manufacturing process.
2 METHODOLOGY
The proposed design methodology contains loosely
coupled hardware and software modules, which can de-
tect, transmit and process the material defect informa-
tion.
The existing ultrasonic hardware design was incorpo-
rated in this system which is commonly used in indus-
trial applications. Still, this hardware model needs a
level-II expert for precise evaluation of defects. To en-
able the user to operate the ultrasonic machinery without
any expertise, a software model with high accuracy and
precision in finding defects is required. Thus, a software
model is proposed, which includes two stages, covering
defect size, position calculation and defect type predic-
tion as depicted in Figure 1. To implement the first
stage, a novel algorithm called Self-Parameterized Den-
sity Based Clustering (SPDBC) is proposed to compute
the size and position of a defect. The second stage deals
with the defect type prediction using computational in-
telligence techniques.
J. V. JOHNSONSELV A, J. RAJA SEKAR: DEFECT CHARACTERIZATION OF METALLIC MATERIALS USING ...
720 Materiali in tehnologije / Materials and technology 58 (2024) 6, 719–728
Figure 1: Workflow of the proposed methodology

2.1 Hardware model set-up
The hardware model is responsible for gathering raw
data using ultrasonic NDT and storing it in the primary
storage. The ultrasonic NDT in this research work is per-
formed with the Epoch6LT ultrasonic kit, using the
pulse-echo technique to detect the abnormalities in the
internal material structure. The ultrasonic pulse-echo
technique can be utilized using any ultrasonic probe with
an oscillating frequency of 2–20 MHz. In ultrasonic
NDT, the penetration ability of the ultrasonic wave is un-
affected by changes in the material’s magnetic proper-
ties, making it suitable for both magnetic and non-mag-
netic materials. Due to a unique behavior of ultrasonic
waves in different materials, internal defects in a wide
range of materials can be studied. Using the NDT data
observed with the ultrasonic pulse-echo technique, struc-
tural details of defects can be clearly observed. Although
the ultrasonic pulse-echo technique is suitable for all
types of industrial materials, it needs a trained expert to
diagnose defect characteristics in detail. Industrial re-
search should be focused on honing ultrasonic NDT us-
ing knowledge mining algorithms. Additionally, defects
need to be categorized to assess the extent of damage ac-
curately. This classification helps prevent an unnecessary
replacement of an entire part when the defect is minor.
Hence, an effective method is proposed combining com-
putational intelligence techniques and ultrasonic NDT to
analyze and characterize internal defects found in indus-
trial materials in order to support NDT experts in the in-
dustry.
To gather NDT data using the ultrasonic pulse-echo
technique, the ultrasonic probe is placed on the mate-
rial’s surface with a layer of lubricant between them and
a high-power ultrasonic beam of 4 MHz is passed inside
the material as seen in Figure 2a. Due to the changes in
the aquatic impedance between the material and atmo-
spheric air, ultrasonic waves bounce back to the probe at
the end of the material, forming a pulse-echo signal. The
same reflection takes place in the presence of impurities,
air molecules and voids inside the material structure.
These pulse-echo signals may contain data that corre-
spond to defects, non-defects, and outliers. Outliers often
appear in data due to random noise, data loss, data mis-
understanding, etc. In NDT, these outliers have to be
handled effectively to increase the result accuracy as it
degrades the performance of a prediction. A standard
pulse-echo signal contains initial echo (IE), defect echo
(DE) and back-wall echo (BE) as illustrated in Fig-
ure 2b. If there is no defect present inside the material,
and nothing interferes with the beam path, there is no de-
fect echo in the signal. The presence of a defect can be
described with three different cases as follows below.
Case 1: Bigger back wall and smaller defect echo
Defects like cracks and discontinuities are thin defec-
tive structures present inside materials. Under ultra-sonic
NDT, such defects only block a very small portion of the
ultrasonic beam path. As a result, the defect echo of the
pulse-echo signal is much smaller than the back-wall
echo.
Case 2: Same size of the defect and back-wall echo
It is rare that both back wall and defect echo have the
same size. This pattern can be seen in the verification of
drill holes in an industrial environment and in the pres-
ence of an internal structure similar to a blister, porosity
and shrinkage.
22
This kind of defects is classified as
highly critical and need to be taken care of as soon as
possible.
Case 3: Absence of the back-wall echo
This case is only possible if a defect is bigger than
the beam path. Materials with bigger defects pose a seri-
ous threat to the material structure and the operating en-
vironment. The scenario of having no back wall during
ultrasonic NDT is considered as most critical case in an
industrial environment.
The data observed using the Epoch6LT ultrasonic kit
is stored in its primary storage unit to be retrieved as
š*.xml’ and š*.csv’ files for further analysis. The hard-
ware model outputs several parameters, and among
those, nine parameters are chosen for further processing.
They include position ID, initial echo amplitude, defect
echo amplitude, defect echo position, back-wall echo
amplitude, back-wall echo position, angle, velocity, and
frequency, as described in Table 1. These nine parame-
J. V. JOHNSONSELV A, J. RAJA SEKAR: DEFECT CHARACTERIZATION OF METALLIC MATERIALS USING ...
Materiali in tehnologije / Materials and technology 58 (2024) 6, 719–728 721
Figure 2: a) Ultrasonic pulse-echo principle, b) pulse-echo structure

ters of the pulse-echo signal significantly represent the
characteristics of defects observed using ultrasonic NDT.
Table 1: Description of the ultrasonic NDT dataset
Attribute Type Value
Position ID (PID) Integer [1–8]
Initial echo amplitude (IEA) Integer [0–110] A
Initial echo position (IEP) Integer [0–2] mm
Defect echo amplitude (DEA) Integer [10–100] A
Defect echo position (DEP) Float [0–200] mm
Back-wall echo amplitude
(BEA)
Integer [10–100] A
Back-wall echo position (BEP) Float [1–200] mm
Angle Integer [0°, 45°, 60°]
Velocity Integer [1200–6300] m/s
Frequency Float [5–12] Hz
Table 2: Description of the materials utilized
Material
ID
Sample
type
Defect
type
Length,
mm
Width,
mm
Height,
mm
S_001 Industrial Ali 123 64 15
S_002 Industrial Bli 74 82 15
S_003 Industrial La 74 195 15
All the defective materials utilized in this work were
collected from broken industrial equipment and for a
better understanding of the samples, their details are de-
scribed in Table 2.
2.2 Software design
The software model retrieves the parameters stored in
the primary storage and processes them using machine
learning techniques. Initially, the SPDBC algorithm uses
the defect echo position and back-wall echo position to
cluster all defects individually and to exclude non-defect
areas. From the clustered defect data, the size and posi-
tion of defects are calculated using the SPDBC algo-
rithm. Meanwhile, the algorithm filters out the parame-
ters that correspond to the non-defect areas. Following
that, computational intelligence techniques are used to
accurately predict the defect type.
2.2.1 Proposed SPDBC algorithm for defect size and
position identification
The density-based clustering method works based on
the spatial metrics between data. This method is well
known for its ability to cluster spatial data and filter out
noise/outliers. It identifies a cluster as a region of high
point density, separated from regions of low point den-
sity. The idea here is to segregate groups of data with
similar traits and assign them to the same cluster. By us-
ing this method for a defect analysis, multiple defects
within a material can be identified precisely, with exact
positions and sizes. The accuracy of density-based clus-
tering can also be increased by adjusting the spatial pa-
rameters. However, the limitation of density-based clus-
tering lies in the fact that it uses more computational
time as it iterates through all possible parametric values
and requires human assistance.
To overcome this limitation and achieve accurate re-
sults effectively, the SPDBC method is proposed. It is a
clustering method derived from density-based clustering
like Density Based Dynamically Self-Parameterized
Clustering for Material Inspection (DBDSPCMI).
23
Un-
like other density-based clustering methods, SPDBC
uses a classification and clustering approach. This ap-
proach uses a minimum threshold (T
min
) value of 10 and
a maximum threshold (T
max
) value of 100 to identify the
data as belonging to defects and non-defects. The values
of T
min
and T
max
are determined to be the possible ampli-
tudes of the data gathered from defective portions of ma-
terials as mentioned earlier in Table 1. The data from the
defective portion of a material is considered valid (V),
while the data from the non-defective portion is consid-
ered invalid (I), using Equation (1). This approach helps
us to improve the speed and accuracy of SPDBC by
identifying defect and non-defect data during clustering.
fx
VD E AT T
I
()
,
,
min max
=
>< ⎧ ⎨ ⎩ and
Otherwise
100
(1)
During SPDBC, effective spatial parameters such as
reachability (R) and density factor (D) can be computed
automatically. R is the distance metric used to identify a
possible connected neighbor (q) around the data point (p)
where D is the minimum density of q required by p. Data
points p and q are said to be connected neighbors only if
they are within the q distance and have a minimum of D
neighbors within R, as illustrated in Figure 3. These spa-
tial parameters are computed using Equations (2) and
(3).
R =
M a t e r i a lL e ngt hi nmm
Total No. of Samples Observed
⎛ ⎝ ⎜ ⎜ ⎞ ⎠ ⎟ ⎟ (2)
D
R
=
1
(3)
The defect size can also be roughly determined using
Equation (4). If a defect is identified to be big, then each
cluster is considered as a defect boundary. In case of a
J. V. JOHNSONSELV A, J. RAJA SEKAR: DEFECT CHARACTERIZATION OF METALLIC MATERIALS USING ...
722 Materiali in tehnologije / Materials and technology 58 (2024) 6, 719–728
Figure 3: Illustration of SPDBC

big defect, defect boundaries are reconstructed as a sin-
gle defect to observe its size and position.
fx
DEA BEA
()
,
=
≤ ⎧ ⎨ ⎩ Smaller,
Bigger Otherwise
(4)
We define x as the data, where x(i,j) are the data
points. For each data point x(i,j) the number of connected
neighbors within R should be D
n
, This can be expressed
as
fx
DDD
()
,
,
min max
=
< ⎧ ⎨ ⎩ 1
0
n
<
Otherwise
(5)
In Equation (5), the number 1 stands for a defect
whereas number 0 stands for a non-defect/outlier. Also,
 D ( D = D
max
– D
min
) depends on the number of obser-
vations per millimeter and varies accordingly.
2.2.2 Computational intelligence techniques for a
defect-type prediction
It is observed that the amplitude of the defect echo
and back-wall echo varies depending on the defect struc-
ture even though the initial echo amplitude is kept as a
constant. Using the parametric changes in the pulse-echo
structure that varies based on the defect’s properties, the
type of a defect that resides inside the material can be
recognized precisely. In machine learning and data sci-
ence, the computational intelligence techniques are pri-
marily used for data classification and prediction. To
identify the optimal existing model for predicting de-
fects, a set of algorithms such as K-Nearest Neighbor
(KNN),
24
Support Vector Machine (SVM),
25
Decision
Tree (DT),
26
Naive Bayes Classifier (NBC),
27
Random
Forest Classification (RFC),
28
Adaptive Boost
(AdaBoost),
29
Stochastic Gradient Descent (SGD),
30
Ar-
tificial Neural Network (ANN)
31
and Gradient Boosting
Regression Trees (GBRT),
32
are used. The goal is to de-
termine, which algorithm can forecast defects more ac-
curately during ultrasonic NDT by comparing their ap-
plicability based on their features.
2.2.3 Performance evaluation metrics
The performance of the proposed SPDBC algorithm
is statistically evaluated in terms of the Jaccard index
(JI) and the classifiers are comparatively evaluated using
the accuracy metric. The Jaccard index, often known as
the Jaccard similarity coefficient, is a metric used to
measure the similarity between two sets of data as given
in Equation (6). Accuracy is the ratio of correctly classi-
fied samples to the total number of samples as given in
Equation (7).
JI
TP
TP FP FN
=
++
(6)
Accuracy =
+
+++
TP TN
TP FP TN FN
(7)
Abbreviations TP , FP , TN, and FN stand for true posi-
tive, false positive, true negative, and false negative, re-
spectively. These values are obtained from the confusion
matrix.
3 RESULTS AND DISCUSSIONS
Here, the structural integrity of defect samples used
and the rate, at which the ultrasonic pulse-echo observa-
tions are made are discussed in detail. After that, defect
characterization using the proposed SPDBC algorithm
and computational techniques are observed.
3.1 Details of defect samples employed in the work
The sample materials used in the analysis are made
of low-carbon iron and include defect such as
alligatoring (Ali), blister (Bli) and lamination (La). For a
better understanding of defect samples used in the pro-
posed work, some of their cross-sections are shown in
Figure 4. In an ideal environment, these defect structures
are not necessarily considered as a threat, but under in-
dustrial circumstances, stress and temperature changes
cause them to weaken the material structure. Thus, mate-
J. V. JOHNSONSELV A, J. RAJA SEKAR: DEFECT CHARACTERIZATION OF METALLIC MATERIALS USING ...
Materiali in tehnologije / Materials and technology 58 (2024) 6, 719–728 723
Figure 4: Cross-sections of (a–b) S_001, (c–d) S_002, (e–f) S_003

rials with internal defects need to be eliminated from
wrought products during manufacturing to prevent fail-
ure.
Ali is a defective structure, caused by accidental split-
ting of the material during hot rolling. Although defect
Ali in Figure 4a and 4b seems to appear at the edge of
the material structure, ultrasonic inspection reveals the
presence of microcracks around it. Thus, it is not safe to
just trim off the defective edges as they already caused
microcracks on the material. These microcracks cause
structural failure under stress even after the welding of
the surface. During casting, gases like nitrogen are ab-
sorbed by the material during heating and released dur-
ing cooling. If they have no space to move out of the ma-
terial, the released gases are trapped inside, forming
porosity.
When a material with porosity is processed as a
wrought product, trapped gases expand along with the
material and form a Bli defect inside the final product, as
shown in Figure 4c and 4d. During material processing,
rapid heating and cooling cause the material to expand
and contract. This kind of expansion and contraction of
the molecular structure causes shrinkage or cavity inside
the material. When it is processed as a wrought product
using milling, the cavity will be elongated in the material
structure, forming La as shown in Figure 4e and 4f.
3.2 Experimental analysis of data gathering
To visualize a defect’s structure with ultrasonic NDT,
the number of observations per millimeter (n/mm) must
be sufficient to define the defect structure accurately.
Also, the distance for each observation must be as con-
sistent as possible. To select the effective n/mm, a series
of inspections with different n values such as n {0.1,
0.2, 0.5, 1.0 and 2.0} is carried out on S_000, as seen in
Figure 5. As the sample is fabricated in a laboratory, its
dimensions are well known, as illustrated in Figure 5a.
The visual similarity between the observed defect
structure and the actual known structure can be well dif-
ferentiated by plotting the DEP against the material
length. This also shows how effective the ultrasonic
NDT can be if properly used on materials. From the
analysis, it is observed that only 0.1’s/mm, as in Figure
5b, can recognize a defect. Increases in the pulse-echo
signal such as 0.2’s/mm and 0.5’s/mm make it difficult to
identify the geometrical structure of a defect as these ob-
servations do not cover all the defect structure clearly.
Figure 5c and 5d implies that the increase in n’s/mm in-
creases the visualization of defect accuracy. However, it
also confirms that the increase in n’s/mm increases the
similarity between the actual and observed defect. As we
increase n’s/mm, a higher accuracy in visualizing defects
is achieved using 1.0’s/mm and 2.0’s/mm, as shown in
Figure 5e and 5f. To identify the effective n’s/mm, the
similarity measure for all series of observations is com-
puted using the Jaccard coefficient or Jaccard index,
Minkowski distance, Manhattan distance and Cosine dis-
tance, as shown in Figure 5g. By multiplying these simi-
larity measures by one hundred, the Jaccard index of the
size and position identification can be computed. All
similarity measures imply that the convergence has been
achieved at 1.0’s/mm. Comparably, the outcome after
convergence gives better results than before convergence.
To achieve accurate results of defect identification, any-
thing higher than 1.0’s/mm is acceptable. By inspecting
all available samples at n’s/mm, a labeled dataset with a
total of 74800 observations is created for further analysis
of defect prediction.
J. V. JOHNSONSELV A, J. RAJA SEKAR: DEFECT CHARACTERIZATION OF METALLIC MATERIALS USING ...
724 Materiali in tehnologije / Materials and technology 58 (2024) 6, 719–728
Figure 5: a) Illustration of a defect in S_001; (b, c, d, e, f) series of different numbers of observations/mm; (g) similarity evaluation curves for
different observations/mm

3.3 Defects size & position computation using the pro-
posed SPDBC algorithm
To identify a defect inside the S_001 material sam-
ple, 350 pulse-echo observations are made at a rate of
2.8 s/mm on the material surface and the data is filtered
using Equation (1). In the clustering process, only the
data identified as V are clustered using parametric values
R = 0.35 and D = 2.8 obtained from Equations (2) and
(3). A plot of DEP against PID is created. On relation-
ship graph between the defect and back-wall amplitude
in Figure 6a shows that the defect echo is much smaller
than the back-wall echo in all instances, and the pattern
represents the presence of smaller defects such as
microcracks or discontinuity as per Equation (4). Thus,
each cluster in the spatial representation is considered as
an individual defect. From the spatial representation of
defects observed in Figure 6d, the size and position of
defects can be analyzed precisely. The result shows that
material S_001 has 9 defective structures (A1-A9).
The S_002 sample is subjected to the same
pulse-echo technique using a 4 MHz frequency and 180
observations made at a rate of 2.4 s/mm. The observed
data is filtered using Equation (1) and processed using
SPDBC. In the clustering process, the data identified as
V are clustered using parametric values R = 0.41 and
D = 2.4 obtained with Equations (2) and (3). The ampli-
tude relationship graph in Figure 6b shows that DEA
and BEA intersect and that there is no back wall at some
sites representing a bigger defect as per Equation (4).
Thus, the clusters present in the spatial representation in
Figure 6e are considered as the boundaries of the defect.
J. V. JOHNSONSELV A, J. RAJA SEKAR: DEFECT CHARACTERIZATION OF METALLIC MATERIALS USING ...
Materiali in tehnologije / Materials and technology 58 (2024) 6, 719–728 725
Figure 6: (a-c) Amplitude relationship graphs, (d-f) SPDBC results
Table 3: Defects observed on S_001, S_002 and S_003
Material ID RD
Amplitude (A)
Defect ID
Position, mm Size, mm
DEA BEA X-axis Y-axis X-axis Y-axis
S_001 0.35 2.8 4–24 26–83
A1 0–2.6 7.25–8.5 2.6 2
A2 6–17 8.25–9.25 11 1.4
A3 15.5–36.6 0.5–1.75 21.1 1
A4 35–56.5 2.45–4.5 21.5 2.2
A5 55–59.5 12.45–13.5 4.5 1.9
A6 66.25–73.5 11.6–12.6 7.25 1.5
A7 70.5–89 14.45–15.55 18.5 1
A8 85.5–91 11.25–12.25 5.5 1.7
A9 118–123 5.75–6.7 5 1.5
S_002 0.41 2.4 4–67 5–110 B1 18.5–48 3–14.1 29.5 11.1
S_003 0.37 2.7 4–67 11–110 L1 7.5–52 11–14.5 44.5 3.5

Defect boundaries are reconstructed to visualize de-
fect geometry from the point of contact of the ultrasonic
probe. The result shows that the S_002 material includes
only one defective structure (B1). On the S_003 sample,
198 observations are made using the pulse-echo tech-
nique at a rate of 2.7 s/mm and the data is filtered with
Equation (1).
During clustering, the data identified as V are clus-
tered using parametric values R = 0.37 and D = 2.7 ob-
tained with Equations (2) and (3). From the amplitude
relationship graph in Figure 6(c), it is observed that DEA
and BEA intersect each other, but there is no sign of the
absence of the back wall. As per Equation (4), it is in-
deed a big defect but not big enough to block the beam
path. The result shows that the S_003 material includes
only one defective structure (L1), as illustrated in Fig-
ure 6f. For a better understanding, a detailed description
of defects observed during SPDBC is included in Ta-
ble 3. The accuracy of defects identified using SPDBC is
98.78 % (Jaccard index: 0.9878) in terms of the defect
position and 97.02 % (Jaccard index: 0.9702) in terms of
the defect size, obtained with the help of reference de-
fects fabricated in our laboratory.
3.3.1 Comparison of the proposed algorithm with
existing techniques
The ultrasonic data gathered in this work was studied
using also different existing clustering algorithms such
as K-means,
33
hierarchical algorithm,
34
affinity propaga-
tion clustering (APC),
35
DBSCAN
36
and DBDSPCMI.
23
Individual industrial materials might include several de-
fects, like in the case of S_001. Clustering methods like
K-means and hierarchical algorithm need to be initial-
ized with a number of clusters. As a result, they fail to
classify the data about materials with multiple defects as
it would lead to misclassification. On the other hand,
APC gives better results in finding defects by clustering
data based on affinity otherwise termed as similarity. The
only issue with using APC is that it cannot remove the
noise from the data. As ultrasonic data includes noise as
well, density-based clustering methods, such as
DBSCAN and DBDSPCMI, are used.
Table 4: Comparative analysis using the Jaccard index (in %)
Algorithm
Defect size
(Jaccard index)
Defect position
(Jaccard index)
Single
defect
Multi-
ple de-
fects
Aver-
age
Single
defect
Multi-
ple de-
fects
Aver-
age
K-means
33
95.91 13.09 54.50 94.53 11.51 53.02
Hierarchical
34
96.02 17.51 56.80 94.07 11.74 52.91
APC
35
93.06 92.51 92.79 93.43 92.79 93.11
DBSCAN
36
95.87 95.00 95.44 95.12 94.65 94.89
DBDSPCMI
23
97.23 96.70 96.97 95.68 94.80 95.24
SPDBC 99.23 98.32 98.78 97.68 96.36 97.02
From these results, it is observed that these algo-
rithms successfully classify data from the materials with
multiple defects, using spatial metrics and removing the
noise as well. Using density-based clustering methods,
the defect position and size can be observed precisely.
The Jaccard index, used as an accuracy measure, shows
that the existing methods are inferior to the proposed
method in identifying a defect structure accurately, as
described in Table 4. It is observed that the highest accu-
racy, in terms of the Jaccard index, among the existing
methods is 96.97 % for defect size calculation and
95.24 % for defect position computation. However, the
proposed SPDBC algorithm achieves Jaccard indices of
98.78 % and 97.02 % for defect size and position com-
putations.
3.4 Defects type prediction using computational intelli-
gence techniques
After filtering out the parameters of non-defect areas
using the proposed SPDBC algorithm, the remaining pa-
rameters only partly correspond to defects Ali, Bli and
La. The nine parameters that correspond to them are only
used as input data for the classifiers. An analysis is per-
formed using two-fold, three-fold, five-fold, ten-fold and
twenty-fold cross-validation ratios and a common train-
ing and testing ratio of 7:3. The input data contains
74,800 row vectors and nine column vectors for defect
characterization. A row vector indicates the number of
observations and a column vector indicates the nine pa-
rameters. The input data include five different classes
such as Ali, Bli, La, empty and outlier. The empty class
refers to the data gathered from the non-defect portion of
the material, and the outlier is random data, considered
as an error of data collection. Testing results for the clas-
sifier are derived from the confusion matrix and perfor-
mance metrics such as accuracy, area under curve
(AUC), F1 score and precision, calculated as shown in
Tables 5 to 9.
Table 5: Testing results (two-fold, 7:3 ratio)
Model Accuracy AUC F1 Precision
KNN 90.22 % 0.961 0.891 0.888
SVM 86.22 % 0.966 0.850 0.845
DT 95.11 % 0.959 0.938 0.926
NBC 81.77 % 0.940 0.835 0.861
RFC 92.44 % 0.962 0.919 0.904
AdaBoost 93.33 % 0.951 0.930 0.929
SGD 80.88 % 0.839 0.771 0.805
ANN 80.44 % 0.954 0.749 0.799
GBRT 94.66 % 0.972 0.940 0.934
Table 5 displays testing results for the two-fold
cross-validation ratio, where DT predicts a defect with a
high accuracy of 95.11 %. In comparison, the results ob-
tained with GBRT are higher than those obtained with
DT in terms of AUC, F1 score and precision value.
J. V. JOHNSONSELV A, J. RAJA SEKAR: DEFECT CHARACTERIZATION OF METALLIC MATERIALS USING ...
726 Materiali in tehnologije / Materials and technology 58 (2024) 6, 719–728

Table 6: Testing results (three-fold, 7:3 ratio)
Model Accuracy AUC F1 Precision
KNN 90.22 % 0.963 0.891 0.886
SVM 85.78 % 0.964 0.947 0.935
DT 96.00 % 0.958 0.947 0.935
NBC 88.00 % 0.963 0.879 0.878
RFC 96.00 % 0.976 0.954 0.943
AdaBoost 93.78 % 0.953 0.941 0.949
SGD 85.33 % 0.892 0.863 0.864
ANN 86.22 % 0.955 0.849 0.845
GBRT 96.44 % 0.997 0.961 0.958
Table 6 displays testing results for the three-fold
cross-validation ratio. Here, the performance results ob-
tained with GBRT are higher than those of other algo-
rithms in terms of accuracy, AUC, F1 and precision
value. It is evident that an increase in the folding value
from two to three increases the accuracy of the entire al-
gorithm considerably, except for KNN, which shows no
significant changes.
Table 7: Testing results (five-fold, 7:3 ratio)
Model Accuracy AUC F1 Precision
KNN 92.00 % 0.972 0.908 0.898
SVM 85.84 % 0.960 0.850 0.845
DT 96.42 % 0.964 0.951 0.939
NBC 88.89 % 0.961 0.884 0.879
RFC 96.00 % 0.976 0.956 0.943
AdaBoost 96.00 % 0.967 0.951 0.943
SGD 86.22 % 0.888 0.861 0.861
ANN 86.67 % 0.955 0.854 0.851
GBRT 96.44 % 0.992 0.961 0.958
Table 7 displays testing results for the five-fold
cross-validation ratio. Here, we see no great changes in
the GBRT parameters, but AdaBoost seems to perform
better than with three-fold cross-validation. Still, GBRT
achieves the highest accuracy in all aspects, while KNN,
SVM, NBC, AdaBoost, SGD and ANN also show in-
creased accuracy.
Table 8: Testing results (ten-fold, 7:3 ratio)
Model Accuracy AUC F1 Precision
KNN 91.77 % 0.977 0.904 0.895
SVM 86.77 % 0.969 0.850 0.845
DT 93.82 % 0.966 0.951 0.939
NBC 83.82 % 0.963 0.870 0.870
RFC 93.53 % 0.992 0.966 0.973
AdaBoost 94.12 % 0.969 0.957 0.954
SGD 86.47 % 0.891 0.861 0.861
ANN 87.94 % 0.957 0.850 0.844
GBRT 95.29 % 0.989 0.974 0.974
Tables 8 and 9 display testing results for the ten-fold
and twenty-fold cross-validation ratios. For the ten-fold
cross-validation, most algorithms first start to reach con-
vergence, after which accuracy starts to decrease. At this
stage, SGD and ANN still show an increase in accuracy,
but their performance is too low compared to GBRT. The
final testing result for the twenty-fold cross-validation
shows that GBRT achieves the highest accuracy and pre-
cision along with the highest AUC and F1 score.
Table 9: Testing results (twenty-fold, 7:3 ratio)
Model Accuracy AUC F1 Precision
KNN 92.00 % 0.981 0.908 0.899
SVM 87.11 % 0.973 0.859 0.853
DT 95.56 % 0.959 0.945 0.934
NBC 88.89 % 0.961 0.889 0.892
RFC 97.33 % 0.988 0.960 0.948
AdaBoost 96.89 % 0.975 0.962 0.961
SGD 86.67 % 0.886 0.857 0.857
ANN 88.89 % 0.967 0.876 0.870
GBRT 97.78 % 0.993 0.974 0.974
Thus, the proposed work uses an ultrasonic-based
NDT technique, which is extremely suitable for all types
of metals as well as their alloy compounds, achieving
high accuracy of defect characterization. The prime sig-
nificance of the proposed SPDBC algorithm is that it ex-
cludes non-defect parameters from the whole set of pa-
rameters covering both defect and non-defect areas.
Since the proposed SPDBC algorithm automatically se-
lects defect parameters, the time needed for the entire de-
fect analysis decreases. The main advantage of the pro-
posed model is that it helps an analyst detect a material
defect faster and more effectively, without acquiring the
knowledge of NDT.
4 CONCLUSIONS
In this work, a hybrid approach combining hardware
and software models was developed to detect the internal
defective structures of metals and metal alloys, using ul-
trasonic NDT. The proposed work aims to determine the
size and position of a defect using the proposed SPDBC
algorithm, followed by computational intelligence tech-
niques to predict the defect type.
The proposed SPDBC algorithm accurately identified
a defect with Jaccard indices of 98.78 % for defect posi-
tion computation and 97.02 % for defect size computa-
tion.
By removing non-defect data prior to clustering, the
SPDBC algorithm clusters defects 10.89 % faster than
the existing DBSCAN and DBDSPCMI algorithms.
The system achieved the highest accuracy of 96.44 %
in predicting the defect type using the GBRT computa-
tional intelligence technique.
Therefore, the proposed approach will assist NDT
professionals in industries in identifying and differentiat-
ing the severity of faults faster so that they can replace
defected parts before a significant breakdown may occur.
This model is extremely useful for locating defects in
wrought products before they are used in the industries
that produce heavy-duty vehicles.
J. V. JOHNSONSELV A, J. RAJA SEKAR: DEFECT CHARACTERIZATION OF METALLIC MATERIALS USING ...
Materiali in tehnologije / Materials and technology 58 (2024) 6, 719–728 727

5 REFERENCES
1
A. Sifa, A. S. Baskoro, S. Sugeng, B. Badruzzaman, T. Endramawan,
Identification of the Thickness of Nugget on Worksheet Spot
Welding Using Non Destructive Test (NDT) – Effect of Pressure,
IOP Conf. Ser.: Mater. Sci. Eng., 306 (2018), 012009, doi:10.1088/
1757-899X/306/1/012009
2
S. Sambath, P. Nagaraj, N. Selvakumar, S. Arunachalam, T. Page,
Automatic detection of defects in ultrasonic testing using artificial
neural network, Int. J. Microstruct. Mater. Prop., 5 (2010)6 ,
561–574, doi:10.1504/IJMMP.2010.038155
3
D. Selvathi, I. H. Nithilla, N. Akshaya, Image Processing Techniques
for Defect Detection in Metals using Thermal Images, Proc. of the
3
rd
Inter. Conf. ICOEI, 7 (2019) 2, 939–944, doi:10.1109/ICOEI.
2019.8862616
4
A. El Kouche, H. S. Hassanein, Ultrasonic non-destructive testing
(NDT) using wireless sensor networks, Procedia Comput. Sci., 10
(2012) 2, 136–143, doi:10.1016/j.procs.2012.06.021
5
M. Jolly, Review of Non-Destructive Testing (NDT) Techniques and
their Applicability to Thick Walled Composites, Procedia CIRP, 38
(2015) 1, 129–136, doi:10.1016/j.procir.2015.07.043
6
Y . Jia, L. Tang, P. Ming, Y . Xie, Ultrasound-excited thermography
for detecting microcracks in concrete materials, NDT E Int., 101
(2019) 2, 62–71, doi:10.1016/j.ndteint.2018.10.006
7
A Savin, Influence of the defect characteristic in aluminum alloy
7075 in non-destructive electromagnetic testing, IOP Conf. Ser.: Ma-
ter. Sci. Eng., 126 (2022) 2, 012029, doi:10.1088/1757-899X/1262/
1/012029
8
E. Jasiûnienë, L. Ma`eika, V . Samaitis, V . Cicënas, D. Mattsson, Ul-
trasonic non-destructive testing of complex titanium/carbon fibre
composite joints, Ultrasonics, 95 (2019) 1, 13–21, doi:10.1016/
j.ultras.2019.02.009.
9
H. Towsyfyan, A. Biguri, R. Boardman, T. Blumensath, Successes
and challenges in non-destructive testing of aircraft composite struc-
tures, Chinese J. Aeronaut., 33 (2020) 3, 771–791, doi:10.1016/
j.cja.2019.09.017
10
R. Leiderman, A. M. B. Braga, Scattering of guided waves by defec-
tive adhesive bonds in multilayer anisotropic plates, Wave Motion,
74 (2017) 3, 93–104, doi:10.1016/j.wavemoti.2017.05.007
11
T. D’Orazio, M. Leo, A. Distante, C. Guaragnella, V . Pianese, G.
Cavaccini, Automatic ultrasonic inspection for internal defect detec-
tion in composite materials, NDT E Int., 41 (2008) 2, 145–154,
doi:10.1016/j.ndteint.2007.08.001
12
M. E. Buchanan, Methods of data collection, AORN J., 33 (1981)1 ,
137–149, doi:10.1016/S0001-2092(07)69400-9
13
J. Peters, Computational Intelligence: Principles, Techniques and
Applications, Comput. J., 50 (2007) 6, 758–765, doi:10.1093/comjnl/
bxm073
14
L. Sepulvene, Performance Evaluation of Machine Learning Tech-
niques for Fault Diagnosis in Vehicle Fleet Tracking Modules,
Comput. J., 65 (2021) 8, 2073–2086, doi:10.1093/comjnl/bxab047
15
H. Shi, P. Wang, X. Yang, H. Yu, An Improved Mean Imputation
Clustering Algorithm for Incomplete Data, Neural Process. Lett., 54
(2020) 5, 203–212, doi:10.1007/s11063-020-10298-5
16
D. Q. Zhang, S. C. Chen, Clustering Incomplete Data Using Ker-
nel-Based Fuzzy C-means Algorithm, Neural Process. Lett., 18
(2003) 3, 155–162, doi:10.1023/B:NEPL.0000011135.19145.1b
17
T. M. Meksen, B. Boudraa, M. Boudraa, Defects clustering using
kohonen networks during ultrasonic inspection, IAENG Int. J.
Comput. Sci., 36 (2009)3 ,1 – 4
18
S. Arivazhagan, J. Jasline Tracia, N. Selvakumar, Mater. Res. Ex-
press, 6 (2019) 9, 096539, doi:10.1088/2053-1591/ab2d83
19
I. Ghafoor, P. W. Tse, N. Munir, A. J. C. Trappey, Non-contact detec-
tion of railhead defects and their classification by using
convolutional neural network, Optik (Stuttg.), 253 (2022) 2, 168607,
doi:10.1016/J.IJLEO.2022.168607
20
V . A. Golodov, A. A. Maltseva, Approach to weld segmentation and
defect classification in radiographic images of pipe welds, NDT E
Int., 127 (2022) 3, 102597, doi:10.1016/J.NDTEINT.2021.102597.
21
Y . Ahmad Al-Maharma, P. Sandeep Patil, B. Markert, Effects of po-
rosity on the mechanical properties of additively manufactured com-
ponents: a critical review, Mater. Res. Express, 7 (2020) 12, 122001,
doi:10.1088/2053-1591/abcc5d
22
S. Sambath, P. Nagaraj, N. Selvakumar, Automatic Defect Classifica-
tion in Ultrasonic NDT Using Artificial Intelligence, J. Nondestruct.
Eval., 30 (2011) 1, 20–28, doi:10.1007/s10921-010-0086-0
23
P. Radha, N. Selvakumar, J. Raja Sekar, J. V. Johnsonselva, Den-
sity-Based Dynamically Self-Parameterized Clustering for Material
Inspection, Comput. J., 66 (2021) 2, 416–428, doi:10.1093/comjnl/
bxab169
24
N. S. Altman, An introduction to kernel and nearest-neighbor non-
parametric regression, Am. Stat., 46 (1992) 3, 175–185, doi:10.1080/
00031305.1992.10475879
25
H. F. Chen, In silico log p prediction for a large data set with support
vector machines, radial basis neural networks and multiple linear re-
gression, Chem. Biol. Drug Des., 74 (2009) 2, 142–147,
doi:10.1111/j.1747-0285.2009.00840.x
26
J. R. Quinlan, Induction of decision trees, Mach. Learn., 1 (1986)1 ,
81–106, doi:10.1007/bf00116251
27
Z. Muda, W. Mohamed, M. D. Nasir, I. U. Nur, K-Means Clustering
and Naive Bayes Classification for Intrusion Detection, Journal of IT
in Asia, 4 (2016) 2, 13–25, doi:10.33736/jita.45.2014.
28
T. K. Ho, Random decision forests, Proc. Int. Conf. Doc. Anal. Rec-
ognition, ICDAR, 5 (1995) 1, 278–282, doi:10.1109/ICDAR.
1995.598994
29
Y . Freund, R. E. Schapire, A Decision-Theoretic Generalization of
On-Line Learning and an Application to Boosting, J. Comput. Syst.
Sci., 55 (1997) 1, 119–139, doi:10.1006/jcss.1997.1504.
30
Y . Tian, Y . Zhang, H. Zhang, Recent Advances in Stochastic Gradi-
ent Descent in Deep Learning, Mathematics., 11 (2023) 3, 682,
doi:10.3390/math11030682
31
E. Grossi, M. Buscema, Introduction to artificial neural networks,
Eur. J. Gastroenterol. Hepatol., 19 (2007) 12, 1046–1054,
doi:10.1097/MEG.0b013e3282f198a0
32
D. D Nguyen, T. H. Nguyen, GBRT-based model for predicting the
axial load capacity of the CFS-SOHS columns, Asian J. Civ. Eng., 24
(2003) 4, 3679–3688, doi:10.1007/s42107-023-00743-w
33
S. Na, L. Xumin, G. Yong, Research on k-means Clustering Algo-
rithm: An Improved k-means Clustering Algorithm, ISIIT&SI’10, 5
(2010) 3, 63–67, doi:10.1109/IITSI.2010.74
34
X. Ran, Y . Xi, Y . Lu, Comprehensive survey on hierarchical cluster-
ing algorithms and the recent developments, Artif. Intell. Rev., 56
(2003) 7, 8219–8264, doi:10.1007/s10462-022-10366-3
35
B. J. Frey, D. Dueck, Clustering by passing messages between data
points, Science, 315 (2007) 5814, 972–976, doi:10.1126/sci-
ence.1136800
36
J. R. Sekar, N. Selvakumar, P. Radha, J. V . Johnsonselva, Supervised
and unsupervised learning for characterizing the industrial material
defects, International Journal of Business Intelligence and Data
Mining, 21 (2022) 3, 233–246, doi:10.1504/IJBIDM.2022.10039148
J. V. JOHNSONSELV A, J. RAJA SEKAR: DEFECT CHARACTERIZATION OF METALLIC MATERIALS USING ...
728 Materiali in tehnologije / Materials and technology 58 (2024) 6, 719–728