Available online at www.sciencedirect.

com

ScienceDirect
Procedia Computer Science 50 (2015) 316 322

2nd International Symposium on Big Data and Cloud Computing (ISBCC15)

Use of Machine Learning Algorithms for Weld Quality

Monitoring using Acoustic Signature
A.Sumesh* a, K.Rameshkumar b, K.Mohandas c, R. Shyam Babu d
a,b,c,d

Department of Mechanical Engineering, Amrita Vishwa Vidyapeetham, Coimbatore 641 112

Abstract
Welding is one of the major joining processes employed in fabrication industry, especially one that manufactures boiler, pressure
vessels, marine structure etc. Control of weld quality is very important for such industries. In this work an attempt is made to
correlate arc sound with the weld quality. The welding is done with various combinations of current, voltage, and travel speed to
produce good welds as well as weld with defects. The defects considered in this study are lack of fusion and burn through. Raw
data points captured from the arc sound were converted into amplitude signals. The welded specimens were inspected and classified
into 3 classes such as good weld and weld with lack of fusion and burn through. Statistical features of raw data were extracted
using data mining software. Using classification algorithms the defects are classified. Two algorithms namely, J48 and random
forest were used and classification efficiencies of the algorithms were reported.
2015
2015 The
The Authors.
Authors.Published
Publishedby
byElsevier
ElsevierB.V.
B.V.This is an open access article under the CC BY-NC-ND license

Peer-review under responsibility of scientific committee of 2nd International Symposium on Big Data and Cloud Computing
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
(ISBCC15).under responsibility of scientific committee of 2nd International Symposium on Big Data and Cloud Computing
Peer-review
(ISBCC15)
Keywords:SMAW; Weld Quality, Acoustic Signature; Weld Defects; Burn-Through; Lack of Fusion;

1. Introduction
Physics of arc welding is of highly complex nature, which makes it difficult to develop a mathematical model to
correlate the quality factors to the process variables or emissivity characters such as spectroscopic, arc sound etc. It is
necessary to study the root cause of weld defects and how the weld parameters, like current, voltage, speed, arc sound
etc. influence the weld defects. Weld defects can be predicted by employing machine learning tools, such as Decision
Trees, Artificial Neural Network (ANN), Fuzzy Logic, Support Vector Machines etc.

* Corresponding author. Tel.: 91-422-2685000; fax: +91-422-2656274.

E-mail address:[email protected]

1877-0509 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of scientific committee of 2nd International Symposium on Big Data and Cloud Computing (ISBCC15)
doi:10.1016/j.procs.2015.04.042

A. Sumesh et al. / Procedia Computer Science 50 (2015) 316 322

317

Literature review indicates that there exist many computational intelligent techniques for automated weld quality
control. A decision support system1 was used to assess the quality of resistance seam welds of steel strips by statistical
analysis. They assessed weld quality by correlating both mechanical and electrical variables involved in the welding
process monitored with previously recorded historical data of similar welds. The radiation emitted by the plasma2, 3, 4,
5
present in the electric arc have been captured using spectroscopy sensors and analyzed to predict the weld quality.
Acoustic sensing methods6, 7 was used to monitor the welding process. Investigations using the arc sound as signature8
for quality monitoring of GMAW welding process was also reported in the literature. They conducted qualitative and
quantitative study of arc sound signals9 and also uses of arc temperature10 as signature for weld quality monitoring.
The scope of electrical impedance11, 12 as parameter for weld quality monitoring was also reported in the literature.
In this work an attempt is made to correlate arc sound with the weld quality. An experimental setup was established
and experiments were conducted using Shielded Metal Arc Welding (SMAW) of Carbon Steel plates. The welding
is done with various combinations of current, voltage, and travel speed to produce good welds as well as weld with
defects.
2. Experiment Setup and Methodology
Welding trials were carried out to characterize the nature of various defects such as burn through and lack of fusion.
An experimental set up has been established to carry out SMA Welding of CS plates and to capture arc sound during
welding. Schematic of the experimental set up is shown in the figure.1 Details and specification of the SMAW test
specimen is given in table.1

Table 1: Test Specifications.

Base Material

C S plates of size 150100 6mm thick

Electrode

AWS A 5.1, E6013, Dia 3.15

Position

1G

Joint type

Single V Butt weld

Root face thickness

1.5 mm

Root Gap

1.5 mm

Groove Angle

60 o

No of Samples

20

Figure 1: Experimental Setup

Initial trials were carried out to establish the range of welding parameters for obtaining defect free welds and welds
with burn through and lack of fusion. Experiments were designed and conducted using the three level Central
Composite Design (CCD). CCD for conducting the experiments are based on the data obtained from the initial trials.
CCD adopted for conducting the experiments is shown in the Table 2. The input weld parameters chosen in this study
are current, voltage and travel speed. Based on CCD, twenty experiments were planned. Carbon Steel specimens were
prepared as per the American Welding Society (AWS) structural welding code. Input weld parameters considered in
this study are Current, Voltage and Travel Speed. After the welding, weld samples are cleaned and quality of the weld
was examined using Non Destructive Testing. Carbon Steel weld samples of good weld, weld with burn through and
lack of fusion were identified. A good quality microphone is used to record the arc sound during the welding process.
Table 2: CCD Experiments.

318

A. Sumesh et al. / Procedia Computer Science 50 (2015) 316 322

Amps

Volts

Speed
(mm/s)

Heat Input
(kJ/mm)

Result Weld Quality

80

28

2.13

1.05

Lack of Fusion

105

28

2.63

1.12

Good Weld

80

18

2.13

0.68

Lack of Fusion

105

30

3.03

1.04

Good Weld

105

16

1.85

0.91

Lack of Fusion

80

28

2.04

1.10

Good Weld

105

23

2.13

1.14

Good Weld

105

23

2.00

1.21

Good Weld

68

23

1.35

1.16

Lack of Fusion

10

105

23

2.50

0.97

Good Weld

11

82

24

2.00

0.98

Lack of Fusion

12

138

36

1.61

3.08

Burn Through

13

130

28

1.39

2.62

Burn Through

14

130

28

1.79

2.04

Burn Through

15

130

28

1.45

2.51

Burn Through

16

105

23

2.17

1.11

Good Weld

17

154

26

2.27

1.76

Burn Through

18

80

18

1.90

0.76

Lack of Fusion

19

105

23

2.50

0.97

Good Weld

20

105

23

2.27

1.06

Good Weld

Sample

Figure 2 (a) : Welded Samples

The welded joints are shown in the Figure 2 (a). Methodology adopted in this study to classify the weld defects
using the arc sound is shown in the Figure 3. Using the microphone, raw sound signals are captured during the welding
process. From the sound signals, dominating statistical features are extracted using the decision tree algorithm. The
best features provided by the decision were fed to classifier to classify the input data as good weld or weld with lack
of fusion or burn through. After welding, the specimens were allowed to cool and cleaned using the wire brush. Weld
specimens were inspected and classified as good weld, weld with lack of fusion and weld with Burn through. The
experimental results were shown in Table 2.

A. Sumesh et al. / Procedia Computer Science 50 (2015) 316 322

319

Figure 2 (b): Methodology

3. Arc Signature Analysis
Analysis of arc sound signals captured was carried out in two stages.
1.
2.

Statistical Feature Extraction

Classification using J48 - decision tree algorithm and Random Forest Algorithm

3.1 Statistical Feature Extraction

The sound signals captured were converted in to wav format. Matlab software is used for extracting raw data
points of the sound signal. The raw data points generated using Matlab were exported in to Excel spread sheet. Each
signal was divided in small segments containing 10,000 data points. These data are of little use for performing process
analysis and quality evaluation. Hence statistical features for each segment were computed using data analysis function
Excel Spread sheet. The feature extraction process was automated using excel macro.
3.2 Classification using Decision Tree
A decision tree is a powerful tool used for data mining applications. It is a predictive machine-learning model that
decides the target value (dependent variable) of a new sample based on various attribute values of the available data.
The internal nodes of a decision tree denote the different attributes. The branches between the nodes tell us the possible
values that these attributes can have in the observed samples, while the terminal nodes tell us the final value
(classification) of the dependent variable.
3.2.1 Classification using J48 Algorithm
In order to classify a new item, it first needs to create a decision tree based on the attribute values of the available
training data. So, whenever it encounters a training set, it identifies the attribute that discriminates the various instances

320

A. Sumesh et al. / Procedia Computer Science 50 (2015) 316 322

most clearly. This feature that gives more information gain is selected and classified based on it. Now, among the
possible values of this feature, if there is any value for which there is no ambiguity, that is, for which the data instances
falling within its category have the same value for the target variable, then we terminate that branch and assign to it
the target value that we have obtained. For the other cases, we then look for another attribute that gives us the highest
information gain. Hence we continue in this manner until we either get a clear decision of what combination of
attributes gives us a particular target value, or we run out of attributes. In the event that we run out of attributes, or if
we cannot get an unambiguous result from the available information, we assign this branch a target value that the
majority of the items under this branch possess. Classification efficiency of J48 algorithm is found to be 70.78% . The
result in the form of confusion matrix is shown in Table 2(a). In the confusion matrix a means Lack of fusion
condition, b is the good weld condition and c denotes burn through condition.
3.2.2 Classification using Random Forest Algorithm
Random forest is an ensemble of decision trees whose predictions are combined to make the overall prediction for
the forest. For many data sets, it produces a highly accurate classifier and runs efficiently on large databases. The tree
visualization option is not available in this algorithm when compared to J48 algorithm and works similar based on the
information gain of the variables.. This method is used for balancing error in class population in unbalanced data sets.
Classification efficiency of Random Forest algorithm is found to be 88.69%. The result in the form of confusion
matrix is shown in Table 2(b). In the confusion matrix a means Lack of fusion condition, b is the good weld
condition and c denotes burn through condition.
Table.2 (a): Confusion Matrix (J48 algorithm).

Table: 2(b): Confusion Matrix (Random forest algorithm).

Classified as

Classified as

268

268

43

a=LF

524

67

41

a=LF

34

779

51

b=Good

18

761

26

b=Good

15

150

315

c=BT

31

34

418

c=BT

4. RESULTS AND DISCUSSION

It is observed from the Table 1 that amperage and overall heat input are the major parameters that influence the
weld quality. To obtain a good weld it is necessary that welding has to be done with optimum current and heat input.
Too low current or too law heat input results in lack of fusion at times associated with the incomplete penetration, as
proved in the sample numbers 1,3, 5, 9, 11, and 18. Whereas very high current or veryhigh heat input results in burnthrough as in the case of samples numbers 12, 13, 14, 15 and 17.
The sample shown in the Figure 3 is that of weld joint without any defects. Figure 4 is that of sample 3 which is
having a lack of fusion. From the Table 1, it can be seen that both current and heat input for the sample is low compared
to that of sample shown in Figure 4. The overall heat input is very less when compared to the one having good weld.
This is could be the reason for lack of fusion, since the root gap for both the joints were kept same and welded under
similar condition. The figure 5 shows the joint with a burn through (Sample No.14). From the Table 1it can be seen
that both current and overall heat input is very high for this joint when compared to that of Good weld (Sample number
7).

A. Sumesh et al. / Procedia Computer Science 50 (2015) 316 322

Figure. 3: (a) Good Weld Specimen

Figure 4(a) : Joint with Lack of Fusion

Figure 3 (b):Good weld Sound signal

Figure 4( b) : Lack of fusion Sound Signal

Figure 5( a): Joint with Burn Through Figure 5(b) : Burn Through Sound Signal.

Statistical features representing good weld, weld with burn through and lack of fusion which are extracted from
the sound signals were given as an input to J48 and Random Forest algorithms. Results were given in the form of
confusion matrix as shown in Table 2 (a) and (b). Classification efficiency of J48 algorithm and Random Forest
algorithm is shown in Table 3. The comparison indicates the better performance of the Random Forest Algorithm.
Table 3: Classification Efficiency of algorithms.
Algorithm
J48
Random Forest Algorithm

Classification Efficiency
70.78
88.69

321

322

A. Sumesh et al. / Procedia Computer Science 50 (2015) 316 322

5. CONCLUSION
An experimental set up has been established to carry out SMA Welding of CS plates with an objective of correlating
arc sound with good weld and weld with lack of fusion and burn through. The input weld parameters chosen in this
study are current, voltage and travel speed. Based on Central Composite Design, twenty experiments were conducted.
Arc sound signal was recorded while welding and the raw data of sound signal was generated using Matlab software.
The statistical features extracted from the raw data were given as an input to the classifier for classifying the features
as good weld and the weld with burn through and lack of fusion. Two different classifier algorithms, J48 and Random
forest were used for classification and the results were compared. The efficiency of Random forest is 88.69% when
compared to 70.78% of J48 algorithm. From this study, it can be concluded that sound signature of welds in SMAW
gives good insight to the condition of welding and can be effectively used for quality monitoring. The performance of
the algorithms may further be improved by suitably filtering the unwanted noises generated during welding other than
arc sound. The results of this research shows that arc sound can be used as an effective signature for weld quality
monitoring.
References
Julio Molleda, Juan L. Caru s, RubenUsamentiaga, Daniel F. Garca, Juan C. Granda, Jose L. Rendueles,
A fast and robust decision support system for in-line quality assessment of resistance seam welds in the
steelmaking industry, J. Computers in Industry. 2012; 63:222230.
2. J. Mirapeix, P.B. Garca-Allende, A. Cobo, O.M. Conde, J.M. Lo pez-Higuera, Real-time arc-welding
defect detection and classification with principal component analysis and artificial neural networks, J.
NDT&E International. 2007; 40: 315323.
3. Li Zhiyong,WangBao, Ding Jingbin, Detection of GTA welding quality and disturbance factors with spectral
signal of arc light , Journal of Materials Processing Technology 209 (2009) 48674873.
4. Sukhomay Pal, Surjya K. Pal , Arun K. Samantaray, Artificial neural network modeling of weld joint strength
prediction of a pulsed metal inert gas welding process using arc signals, J. Materials processing technology.
2008; 2 02: 464474.
5. BappaAcherjee, SubrataMondal, BipanTudu, DiptenMisra, Application of artificial neural network for
predicting weld quality in laser transmission welding of thermoplastics, J. Applied Soft Computing. 2011;
11: 25482555.
6. Kamal Pal, Sandip Bhattacharya, Surjya K. Pal, Investigation on arc sound and metal transfer modes for online monitoring in pulsed gas metal arc welding, Journal of Materials Processing Technology. 2010; 210:
13971410.
7. EberHuancaCayo and Sadek Crisostomo Absi Alfaro, A Non-Intrusive GMA Welding Process Quality
Monitoring System Using Acoustic Sensing, J. Sensors. 2009; 9: 7150-7166.
8. D. Saini and S Floyd, An investigation of Gas Metal Arc Welding Sound signature for online quality
monitoring, AWS, WJ Supplement. 1998; 172-179.
9. Wikle, H. III, R, Zee, B. Chin, Infrared Sensing techniques for penetration depth control of the submerged
arc Welding Process, Journal of Materials Processing Technology. 2010; 113: 228-233.
10. U. Sreedhar, C.V. Krishnamurthy, Krishnan Balasubramaniam, V.D. Raghupathy, S. Ravisankar, Automatic
defect identification using thermal image analysis for online weld quality monitoring, Journal of Materials
Processing Technology. 2012; 212: 1557 1566.
11. Shih-Fu Ling,Li-XueWan, Yoke-RungWongn, Dong-NengLi, Input electrical impedance as quality
monitoring signature for characterizing resistance spot welding, J. NDT&E International.2010; 43:200205.
12. Shih-Fu Ling, Jingen Luan, Xiangchao Li, Wendy Lee Yong Ang, Input electrical impedance as signature
for nondestructive evaluation of weld quality during ultrasonic welding of plastics, J.NDT&E International.
2006; 39:
1.