Materials Today: Proceedings xxx (xxxx) xxx

Contents lists available at ScienceDirect

Materials Today: Proceedings

journal homepage: www.elsevier.com/locate/matpr

Development of a machine learning model to predict the friction stir

welding tool condition
K. Balachandar a,b, K.S. Salamon Arockiaraj a, G. Sriraman a, R. Jegadeeshwaran b, G. Sakthivel b,
J. Lakshmipathi b,⇑
a
School of Mechanical Engineering, Sathyabama Institute of Science and Technology, Chennai, Tamilnadu 600119, India
b
School of Mechanical Engineering, Vellore Institute of Technology, Chennai, Tamilnadu 600127, India

a r t i c l e i n f o a b s t r a c t

Article history: Tool damage results in surface polishing and dimensional accuracy losses for machined parts, as well as
Available online xxxx potential damage to a workpiece or machine. In the industrial business, tool condition monitoring (TCM)
is significant. As a result, it is essential for production efficiency and profitability. In order to prevent
Keywords: damages from occurring, also known as predictive maintenance, we may use the TCMS to identify tool
Friction Stir Welding wear and damage that eventually impacts the workpiece. Machine learning models play a significant part
Condition Monitoring in TCMS by way of a GUI that identifies the tool’s failure state. This method can be applied in any kind of
Feature Extraction
system which is used in the industry to predict the damage. This TCMS is applied in a friction stir welding
Feature Selection
Feature Classifiers
machine (FSW) to predict its tool wear. Friction stir welding, an eco-friendly solid state welding method,
GUI is used to join two metals that are difficult to fuse together using conventional fusion welding. FSW is
largely recognized as the most important innovation in metal joining technology in the previous decade.
This FSW has been widely employed in numerous sectors such as shipbuilding, computer covers, vehicles,
and so on. Developing a GUI for the prediction of tool condition in FSW with AZ31B(Mg) as a workpiece is
a novel approach. In this condition monitoring technique, the real time vibrational data is collected with
the help of an accelerometer. These data’s are being processed and features are extracted and a proper
classifier is used to predict the faults. This paper aims to develop a part of TCMS through vibrational data.
GUI and ML models were developed to predict the condition of FSW tools. The ML model includes
Decision Tree, Random Forest, Light Gradient Boosted Machine Classifiers. Among all these Light
Gradient Boosted Machine Classifiers perform better. Using this Classifier a GUI is built to predict the
condition.
Copyright Ó 2023 Elsevier Ltd. All rights reserved.
Selection and peer-review under responsibility of the scientific committee of the International Confer-
ence on ‘‘Application in Computational Engineering .

1. Introduction cial sectors. One such alloy, AZ31B, is a wrought magnesium alloy
that exhibits good room-temperature ductility and strength, corro-
The lightest structural metal, Magnesium Alloy (AZ31B), is often sion resistance, and weldability. AZ31B is utilized in various appli-
combined with other metals to enhance their physical properties. cations, including but not limited to speaker cones, concrete tools,
Some of the elements used for this purpose include manganese, mobile phone and laptop cases, and airplane fuselages. At high
aluminum, zinc, silicon, copper, zirconium, and rare-earth metals temperatures, AZ31B can be super formed to create intricate com-
[1]. Magnesium alloys, due to their light weight, high strength- ponents for use in automobiles [2]. The process of FSW takes place
to-weight ratio, stiffness-to-weight ratio, castability, machinabil- in the solid phase, at temperatures below the melting point of the
ity, and exceptional damping, are suitable for a variety of structural material being welded. This helps to avoid the problems that can
applications in the automotive, aerospace, industrial, and commer- arise during resolidification, such as the formation of second
phases, porosity, embrittlement, and cracking. Moreover, the lower
temperature of the FSW process enables welding with less distor-
⇑ Corresponding author. tion and residual strains, making it a preferable method for joining
E-mail address: [email protected] (J. Lakshmipathi).
materials [3]. The operating principle of the system is simple. A

https://doi.org/10.1016/j.matpr.2023.05.400
2214-7853/Copyright Ó 2023 Elsevier Ltd. All rights reserved.
Selection and peer-review under responsibility of the scientific committee of the International Conference on ‘‘Application in Computational Engineering .

Please cite this article as: K. Balachandar, K.S. Salamon Arockiaraj, G. Sriraman et al., Development of a machine learning model to predict the friction stir
welding tool condition, Materials Today: Proceedings, https://doi.org/10.1016/j.matpr.2023.05.400
K. Balachandar, K.S. Salamon Arockiaraj, G. Sriraman et al. Materials Today: Proceedings xxx (xxxx) xxx

ture classification, and uses the Gradient boost method (lgbm)

and tree-based models (Random Forest and Decision Tree) to
detect faults in the vibration characteristics of the tool. A graphical
user interface (GUI) has also been developed using PyQt-5 to allow
for real-time prediction of the tool’s condition. This approach offers
a reliable and effective way to predict the condition of the FSW tool
and avoid potential welding defects.

1.1. Methodology

The study involved conducting full tests on nonferrous metal

such as Magnesium alloy (AZ31B), using H13 tool and a healthy
friction stir welding equipment to collect the vibration signals for
tool health condition monitoring. Digital signal processing was
identified as the key factor in machine condition monitoring, with
Fig. 1. FSW process.
a piezoelectric transducer sensor used to record signals from the
vibrating component. The recorded acoustic vibration signal was
converted into a digital signal using NI’s 4-channel DAQ device
non-consumable spinning tool is located inside the welding line and an ADC, and the signal was processed using a tool called Lab-
(as shown in Fig. 1). The tool’s shoulder generates heat, which soft- VIEW, which records signals in digital form. Feature extraction,
ens the material and causes it to flow and stir. Simultaneously, the selection, and classification components of machine learning were
spinning shoulder comes into contact with the surface of the used to enable the machines to learn from the past and make
plates, creating frictional heat that prevents any material from future corrections. Developing a GUI, for tool condition monitoring
leaking out of the seam and resulting in a smooth surface [4]. is to display real-time data (health of the tool). This data can be
The microstructure of the central zone of the FSW joint is signifi- displayed graphically on the GUI, allowing operators to easily visu-
cantly altered by the subsequent dynamic recrystallization of the alize any trends or patterns that may indicate a problem. The
material, resulting in the creation of even finer axial equivalent workflow of the project is depicted in Fig. 2.
grains by an order of magnitude. This event leads to a reduction
in the mechanical properties of the joint. To produce high-quality
welds, it is important to choose the appropriate technical and geo-
metrical welding settings. While optimizing welding parameters
such as rotation speed and tool feed (welding speed) is essential,
optimizing the tool’s shape and material is even more critical [5].
Achieving acceptable mechanical properties and defect-free joints
is possible through careful selection of the FSW tool profile, tool
dimension, and process parameters, which have the most signifi-
cant impact on the material flow behavior [6]. Faults are often
observed in the weldment due to the condition of the tool. Incor-
rect welding methods and the use of a faulty tool are the primary
causes of welding errors. As every tool undergoes wear and tear
after a certain duration of welding, which is known as slow or con- Fig. 2. Methodology.
tinuous tool wear, a damaged tool can affect the heating rate. Inap-
propriate heating rates can lead to the production of various FSW
flaws, such as a lack of tunnels, cavities, surface grooves, excessive
flash, surface galling, nugget collapse, and kissing bonds, which are
all indications of fusion failure. As a result, the overall cost of the
welding process increases [7]. To prevent such failures and dam-
ages, condition monitoring systems, also known as predictive
maintenance, are employed. Condition monitoring involves moni-
toring the characteristics of a tool or equipment to prevent harm.
By continuously monitoring its operational parameters, it is possi-
ble to detect a machine’s need for repair before it deteriorates or
malfunctions. Unlike conventional preventative maintenance prac-
tices, Condition-Based Maintenance (CBM) focuses on the current
condition of the equipment rather than a predetermined timetable
[8]. Vibration analysis is considered to be one of the most effective
non-destructive techniques available as it provides accurate infor-
mation about a component’s condition and helps to predict its
lifespan [9]. To ensure optimal welding properties, it is essential
to maintain the FSW tool in the correct position and condition.
However, it can be difficult to determine the quality of the weld
during the welding process. To address this challenge, a
vibration-based condition monitoring tool has been proposed
using a machine learning-based approach. This method involves
three main phases: feature extraction, feature selection, and fea-
Fig. 3. FSW Machine Setup.

2
K. Balachandar, K.S. Salamon Arockiaraj, G. Sriraman et al. Materials Today: Proceedings xxx (xxxx) xxx

cally automated FSW machine. A PLC-controlled FSW equipment

was rented for conducting the testing, as shown in Fig. 3.
For the experiment, 25 mm  25 mm  8 mm magnesium alloy
(AZ31B) plates were used, and a single-pass welding technique
was utilized to create the joints. The entire process was carried
out using a threaded cylindrical pin profiled, non-consumable
H13 tool steel (as shown in Fig. 4).
Magnesium alloy AZ31B grade metal with dimensions of
25 mm  25 mm  8 mm was joined utilizing FSW, also known
as solid-state joining, using a steel tool mounted on a machine with
a 5.75 mm depth pin. 30 mm/min of feed rate and a constant spin-
dle speed of 1200 rpm were employed. The DAQ instrument
(NI9234) was used to collect vibration signals in this setup. A
piezoelectric accelerometer sensor (Dytron Make, 500 g,
10.26 mV/g sensitivity) was mounted on the FSW machine to
record the signals (Fig. 5). The analogue signals from the sensor
were converted into digital signals by the DAQ module, enabling
them to be used for decision-making. The vibration was captured
Fig. 4. FSW Tool.
using NI LabVIEW, a visual tool.
Both the FSW weld’s excellent and bad conditions were used to
collect the vibration data for the investigation. Using a brand-new
instrument, the vibration signals in good shape were collected. For
the study, five flawed circumstances were taken into account.
When Magnesium sticks to the taper threaded pin, the tool is said
to be in terrible condition. The tool becomes broken when it is han-
dled incorrectly or when the stated rpm is not appropriate for the
given material. The Fig. displays a selection of raw vibration data
that were collected under various simulated fault scenarios. To col-
lect the vibration signals, a piezoelectric accelerometer was
attached to the tool head using an adhesive method. The wireless
4-channel DAQ device (NI 9234, 51.2k Samples/sec) was utilized
to capture the signal from the tool head, as shown in Fig. 6(a&b).
The vibration signal was recorded in digital form using the cDAQ
chassis NI9191 data acquisition system through a graphical NI Lab-
VIEW application. The raw data from various conditions are pre-
sented in Fig. 7a–e.

3. Feature extraction

Fig. 5. Piezoelectric Accelerometer.

Feature extraction refers to the process of deriving parameter
data from raw vibration signals. Raw vibration may be used to gen-
erate several features, including wavelets [10], statistics [11], and
2. Experimental studies histograms [12]. In this study, statistical characteristics were uti-
lized for defect prediction, and various parameters such as Mean,
The experimental setup for monitoring the tool condition of Standard Deviation, Variance, Kurtosis, Median, Sum, Skewness,
FSW involved the use of a high-quality, fully pressurized, hydrauli- Maximum, Minimum, Range, Count, RMS, Standard Error, and

Fig. 6. A. ni daq 9234. b. cdaqchasis 9191.

3
K. Balachandar, K.S. Salamon Arockiaraj, G. Sriraman et al. Materials Today: Proceedings xxx (xxxx) xxx

Fig. 7. A. good condition, b. air gap condition, c). notch condition. d). broken condition. e). misalignment condition.

Mode [13] were extracted from the raw vibration signals using a ing. The data is trained using a model, which is then tested on
python script [Fig. 8]. However, not all of these features are untrained data according to the training. To develop the model,
required for the classifier analysis, and some of them may not pro- an algorithmic technique is necessary, and in the following subsec-
vide significant information. Therefore, the next step was to per- tions, three such algorithmic models are examined.
form feature selection, which involves eliminating the irrelevant
or redundant features from the set of extracted features. For 4.1. Decision Tree algorithm
instance, the ’count’ feature was removed from the feature set.
Fig. 9.Fig. 10.Fig. 11.Fig. 12.Fig. 13.Fig. 14.. The Decision Tree (DT) is one of the decision-based models uti-
lized to categorize data. It uses all of the selected features as input
to the classifiers. During training, the model learns from the input
4. Feature classification data and generates a graphical tree with nodes and branches, rep-
resenting some relational output. Decision trees are capable of
The next step is Classification. In this stage, we need to create an handling dimensional data, where the selected features are used
ML for the GUI which will predict the data. In machine learning, to determine the best way to divide the dataset into distinct classes
the given data is categorized through a process of training and test- during tree building. The J48 is a widely used decision tree classi-
4
K. Balachandar, K.S. Salamon Arockiaraj, G. Sriraman et al. Materials Today: Proceedings xxx (xxxx) xxx

Fig. 8. Feature extraction Script.

Fig. 13. Accuracy Graph.

Fig. 9. Decision Tree using python.

Fig. 14. Confusion Matrix of Random Forest.

Fig. 10. Random Forest using python.

Fig. 11. LGBM using python.

Fig. 15. Confusion Matrix of Decision Tree.

Fig. 12. PyQt Designern python.

fier for classification purposes. When building the tree, J48 skips
hidden values, which are values that may potentially predict a
specific item based on information about the attribute values for
future entries. The primary objective is to divide the data into Fig. 16. Confusion Matrix of Light Gradient Boost Machine.

5
K. Balachandar, K.S. Salamon Arockiaraj, G. Sriraman et al. Materials Today: Proceedings xxx (xxxx) xxx

ranges by considering the attribute values for that item that can be
discovered in the training data. J48 allows for categorization using
decision trees (DT) or rules generated from them [14]. In this study,
DT model created using python (Jupyter Notebook IDE).
Fig. 19. Gradient Boosted Machine Model.
4.2. Random Forest algorithm

One strategy that may measure accuracy rate in the best

method is Random Forest (RF). In comparison to other methods,
effectiveness in massive data computations offers consistency in
accuracy, RF estimate of hidden clusters is the most straightfor-
ward. RF can estimate hidden data in bulk as well. It is possible
to identify key categorization characteristics [15]. In this study,
RF model created using python (Jupyter Notebook IDE).

4.3. Light Gradient Boosted machine algorithm

LightGBM is an open-source toolkit that implements the gradi-

ent boosting technique in a highly efficient and effective way. It
stands for Light Gradient Boosted Machine and is designed to
improve upon the gradient boosting technique by incorporating a
form of autonomous feature selection and focusing on improving
cases with higher gradients. By doing so, it can significantly
increase the speed of training and enhance the performance of pre- Fig. 20. Start page of GUI.
diction [16]. A gradient boosting technique achieves a very com-
petitive and reliable accuracy for both classification and such as primary notation, instead of text-based user interfaces,
regression [17].In this study, LGBM model created in python (Jupy- written command labels, or text navigation. This is in contrast to
ter Notebook IDE). command-line interfaces (CLIs), which require users to input
instructions using a computer keyboard and are often considered
5. Graphical user interface (GUI) to have a steep learning curve [18]. There is no literature review
or any work for GUI is found. This is also another novel idea in this
The graphical user interface (GUI) allows people to interact with study. GUI has been created using PyQt5. PyQt5 is the latest ver-
electronic devices using graphical icons and auditory indicators, sion of Riverbank Computing’s GUI widget toolkit, which is based

Fig. 17. Decision Tree model.

Fig. 18. Random Forest Model.

6
K. Balachandar, K.S. Salamon Arockiaraj, G. Sriraman et al. Materials Today: Proceedings xxx (xxxx) xxx

Fig. 24. Graph for Notch.

Fig. 21. RPM, Classifier selection Window.

Fig. 25. Graph for Misalignment.

Fig. 22. Result window (Good Condition).

Fig. 26. Graph for Broken.

Boosted Machine Classifiers are employed. These algorithms use

statistical characteristics extracted from the raw data. The
expected results are discussed in the following sections.

6.1. Accuracy and confusion matrix

d For the 5 conditions the statistical features are extracted and

the 80 percent data have been fed to train the model and
remaining for testing.
Fig. 23. Graph for Airgap. d The Model accuracy percentage are Decision Tree(91.55 %), Ran-
dom Forest(93.55 %), and Light Gradient Boosted Machine
on the Qt library - one of the most popular and powerful cross- (94.80%)
platform GUI libraries. PyQt5 provides a Python interface to the d Therefore, we can conclude that the Light Gradient Boosted
Qt library, allowing the combination of the Python programming Machine performs better than all others with 94.80% accuracy.
language and the Qt library to create robust and efficient graphical Fig. 15 gives better visualization. Fig. 18
user interfaces [19]. In order to predict when maintenance or d Figs. 16, 17 are the confusion matrix of the models and Figs. 19,
replacement of a tool is necessary, GUIs can assist operators in 20, 21 are the tree generated by the models. Fig. 22.Fig. 23.
instantly identifying patterns and trends in tool condition data Fig. 24.Fig. 25.Fig. 26.
[20].
6.2. GUI and prediction graph
6. Results and discussion
 The Graphical User Interface (GUI) enables people to interact
with electronic devices using graphical icons and auditory indi-
To identify defects in the friction stir welding machine, ML algo-
cators like primary notation.
rithms including Decision Tree, Random Forest, and Light Gradient
7
K. Balachandar, K.S. Salamon Arockiaraj, G. Sriraman et al. Materials Today: Proceedings xxx (xxxx) xxx

 There is no literature review or GUI work discovered. This is also automation, and control, J. Manuf. Process. 16 (1) (2014) 56–73, https://doi.
org/10.1016/j.jmapro.2013.04.002.
a fresh concept in our study. We used PyQt5 to construct a
[4] S. Eslami, T. Ramos, P. Tavares, P. Moreira, Shoulder design developments for
graphical user interface. FSW lap joints of dissimilar polymers, J. Manuf. Process. 20 (2015) 15–23,
 For the five conditions, the GUI results has been attached below https://doi.org/10.1016/j.jmapro.2015.09.013.
and it gives a better idea on Tool Condition. [5] A. Dorbane, G. Ayoub, B. Mansoor, R.F. Hamade, G. Kridli, R. Shabadi, A. Imad,
Microstructural observations and tensile fracture behavior of FSW twin roll
 While training the ML mode, the good condition is labeled as 1, cast AZ31 Mg sheets, Mater. Sci. Eng.: A 649 (2016) 190–200, https://doi.org/
misalignment is labeled as 2, air gap is labeled as 3, notch is 10.1016/j.msea.2015.09.097.
labeled as 4, broken is labeled as 0. [6] G. Rambabu, D. Balaji Naik, C.H. Venkata Rao, K. Srinivasa Rao, G. Madhusudan
Reddy, Optimization of friction stir welding parameters for improved corrosion
resistance of AA2219 aluminum alloy joints, Defence Technology 11 (4) (2015)
7. Conclusion 330–337, https://doi.org/10.1016/j.dt.2015.05.003.
[7] A. Albannai, Review the common defects in friction stir welding, Int. J. Sci.
Technol. Res. 9 (2020) 318–329.
FSW is a complex process that requires continuous monitoring [8] J. Dai, C. L. P. Chen, Xiao-Yan Xu and Peng Hu, Condition monitoring on
to prevent tool failure. An ML model can be developed to predict complex machinery for predictive maintenance and process control, in: 2008
the tool’s condition using data from sensors placed on the tool. IEEE international conference on systems, man and cybernetics, Singapore,
2008, pp. 3595-3600, doi: 10.1109/ICSMC.2008.4811856.
These sensors gather data on variables such as temperature, vibra- [9] Manish Vishwakarma, Rajesh Purohit, V. Harshlata, P. Rajput, Vibration
tion, and acoustic emission signals. The ML model can predict analysis & condition monitoring for rotating machines: A review, J. Mater.
when the tool is nearing its end-of-life and identify areas of Today: Proc. 4 (Issue 2, Part A) (2017) 2659–2664, https://doi.org/10.1016/
j.matpr.2017.02.140.
improvement in the welding process. Optimizing hyperparameters
[10] T.M. Alamelu Manghai, R. Jegadeeshwaran, Vibration based brake health
is crucial for improving the accuracy of the ML model, and grid monitoring using wavelet features: A machine learning approach, J. Vibrat.
search, random search, and Bayesian optimization can be used Control. 25 (18) (2019) 2534–2550, https://doi.org/10.1177/
for this purpose. A GUI can be developed with real-time monitor- 1077546319859704.
[11] M. Alamelu, Vibration based condition monitoring of a brake system using
ing, data visualization, and alert notifications to assist in tool con- statistical features with logit boost and simple logistic algorithm, Int. J.
dition monitoring. The GUI can display the predicted tool condition Perform. Eng. (2018) 14, https://doi.org/10.23940/ijpe.18.01.p1.18.
and extract statistical features for further analysis and optimiza- [12] R. Jegadeeshwaran, V. Sugumaran, Method and apparatus for fault diagnosis of
automobile brake system using vibration signals, Rec. Patents Signal Process 3
tion of the ML model. The development of a semi-automatic sys- (1) (2013) 2–11, https://doi.org/10.2174/2210686311303010002.
tem for tool condition monitoring can increase the effectiveness [13] K. Balachandar, R. Jegadeeshwaran, Friction stir welding tool condition
and dependability of the welding process, and the GUI develop- monitoring using vibration signals and Random forest algorithm – A
Machine learning approach, Mater. Today: Proc. 46 (Part 2) (2021) 1174–
ment scripts can be automated to produce an onboard diagnostic 1180, https://doi.org/10.1016/j.matpr.2021.02.061.
tool that doesn’t require human participation. [14] W. Sun, J. Chen, J. Li, Decision tree and PCA-based fault diagnosis of rotating
machinery, Mech. Syst. Signal Process. 21 (3) (2007) 1300–1317, https://doi.
org/10.1016/j.ymssp.2006.06.010.
Data availability [15] Shiv Virendra Dahe, G. Sai Manikandan, R. Jegadeeshwaran, G. Sakthivel, J.
Lakshmipathi, Tool condition monitoring using Random forest and FURIA
Data will be made available on request. through statistical learning, Mater. Today: Proc. 46 (Part 2) (2021) 1161–1166,
https://doi.org/10.1016/j.matpr.2021.02.059.
[16] M. Wang, K. Shen, C. Tai, Q. Zhang, Z. Yang, et al., Research on fault diagnosis
Declaration of Competing Interest system for belt conveyor based on internet of things and the LightGBM model,
PLoS One 18 (3) (2023), https://doi.org/10.1371/journal.pone.0277352.
[17] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma,
The authors declare the following financial interests/personal Qiwei Ye, and Tie-Yan Liu, LightGBM: a highly efficient gradient boosting
relationships which may be considered as potential competing decision tree, in: Proceedings of the 31st International conference on neural
interests: [Balachandar K reports equipment, drugs, or supplies information processing systems (NIPS’17). Curran Associates Inc., Red Hook,
NY, USA, 2017, pp. 3149–3157.
was provided by Vellore Institute of Technology - Chennai Campus]. [18] I. Banerjee, B. Nguyen, V. Garousi, A. Memon, Graphical user interface (GUI)
testing: Systematic mapping and repository, Information and Software
References Technology, 2013, p. 55.
[19] Mark Summerfield, Rapid GUI programming with Python and Qt: the
definitive guide to PyQt programming, 2007.
[1] V. Jaiganesh, Impact of tool profile on mechanical properties of AZ31B Mg alloy
[20] G. Terrazas, G. Martínez-Arellano, P. Benardos, S. Ratchev, Online tool wear
during FSW using optimized parameters, FME Trans. 44 (2016) 43–49, https://
classification during dry machining using real time cutting force
doi.org/10.5937/fmet1601043J.
measurements and a CNN approach, J. Manuf. Mater. Process. 2 (2018) 72,
[2] P. Myśliwiec, R.E. Sliwa, Friction stir welding of thin sheets of magnesium alloy
https://doi.org/10.3390/jmmp2040072.
az31b, Arch. Metall. Mater. 63 (2018) 45–54, https://doi.org/10.24425/118907.
[3] B.T. Gibson, D.H. Lammlein, T.J. Prater, W.R. Longhurst, C.D. Cox, M.C. Ballun, K.
J. Dharmaraj, G.E. Cook, A.M. Strauss, Friction stir welding: Process,