1 s2.0 S2213846322000207 Main

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

Manufacturing Letters 32 (2022) 83–86

Contents lists available at ScienceDirect

Manufacturing Letters
journal homepage: www.elsevier.com/locate/mfglet

Letters

Determination of the Ultimate Tensile Strength (UTS) of friction stir


welded similar AA6061 joints by using supervised machine learning
based algorithms
Akshansh Mishra a,⇑, Rakesh Morisetty b
a
Department of Chemistry, Materials, and Chemical Engineering ‘‘Giulio Natta”, Politecnico di Milano, Milan, Italy
b
DevOps Consultant, Azatec, Milan, Italy

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

Article history: The Ultimate Tensile Strength (MPa) of identical FSWed AA 6061 joints is determined using five
Received 27 January 2022 Supervised Machine Learning regression-based and classification based techniques in this study. Tool
Received in revised form 27 March 2022 Traverse Speed (mm/min), Tool Rotational Speed (RPM), and Axial Force are the three input parameters
Accepted 25 April 2022
in the dataset (KN). The findings revealed that the Axial Force (KN) has a greater impact on the output,
Available online 7 May 2022
namely the Ultimate Tensile Strength (MPa). The algorithms’ performance was compared using measures
such as the Mean Absolute Error (MAE), Mean Square Error (MSE), Coefficient of Determination (R2) value
Keywords:
and F1- Score for ML based algorithms.
Machine learning
Friction stir welding
Ó 2022 Society of Manufacturing Engineers (SME). Published by Elsevier Ltd. All rights reserved.
Supervised learning
Ultimate tensile strength
Python computing

1. Introduction joints of AA6061 and exposed them to five supervised machine


learning regression and classification based methods [8–26].
Machine Learning is a subset of Artificial Intelligence that solves
a problem by automatically improving the provided answer based 2. Experimental procedure
on previous experience [1–3]. Supervised Learning, Unsupervised
Learning, and Reinforcement Learning are the three types of The dataset consists of input parameters i.e. Axial Force (KN), Tool
machine learning. Supervised Learning is a task-driven algorithm Traverse Speed (mm/min), and Tool Rotational Speed (RPM) with
that learns how to predict the output label using the labeled data- output parameters i.e. Ultimate Tensile Strength which is prepared
set as an example, then improves its accuracy by learning from a in the excel sheet format and further is converted to Comma Sepa-
larger dataset [4–5]. rated Values (CSV) file. The CSV file is imported to the Google Colab-
The Friction Stir Welding technique, for example, uses Machine oratory platform for subjecting it to Python programming developed
Learning-based algorithms to improve mechanical parameters such supervised machine learning regression and classification-based
as Ultimate Tensile Strength, Elongation, Hardness, and Fracture algorithms. The important Python libraries such as scikit-learn,
Strength. For the prediction of tensile characteristics of friction stir numpy, pandas, seaborn, and matplotlib.pyplot were imported.
welded AA 7039 joints, Verma et al. [6] employed machine learning sys- Fig. 1a) shows the necessary flowchart of the procedure. As indicated
tems such as Gaussian Process Regression (GPR), Linear Regression (LR), in Table 1, statistical analysis is used to calculate the standard devia-
and Artificial Neural Network (ANN). In compared to other algorithms, tion value, minimum value, mean value, maximum value, and so on.
the results showed that the ANN model produced more accurate results. Feature importance was calculated to determine which input
Sandeep et al. [7] developed machine learning algorithms for forecast- parameters affect the output parameter i.e. UTS (MPa) most. For
ing the maximum temperature value of AA 7075 and PPS polymer prediction purpose, feature importance allocates values to input
hybrid joints that were friction stir lap welded. The Support Vector variables to a forecasting model that reflects the relative value of
Machine (SVM) algorithm was shown to produce more accurate results. each feature. In our present study, the Decision Tree which is based
The current study combined 85 data from research works based on the CART technique is used to calculate the feature importance
on the assessment of tensile properties of Friction Stir welding of each input feature. From Fig. 1 b) it is observed that the Axial
Force (KN) has higher feature importance value in comparison to
⇑ Corresponding author. other input parameters. The dataset is split into an 80–20 ratio

https://doi.org/10.1016/j.mfglet.2022.04.003
2213-8463/Ó 2022 Society of Manufacturing Engineers (SME). Published by Elsevier Ltd. All rights reserved.
A. Mishra and R. Morisetty Manufacturing Letters 32 (2022) 83–86

Fig. 1. a) Flowchart of the implementation of Supervised Machine Learning algorithms b) Calculation of Feature importance of the input parameters.

Table 1
Statistical analysis of Friction Stir Welded AA 6061 dataset.

index Tool Rotational Speed (RPM) Tool Travel Speed (mm/min) Axial Force (KN) Ultimate Tensile Strength (Mpa)
count 85.0 85.0 85.0 85.0
mean 1143.5129411764706 64.25870588235296 7.086 155.2383529411765
std 254.37163091871412 26.49786369593178 1.4541125586877148 28.749996929484674
min 700.0 12.43 2.5 77.83
25% 1000.0 42.5 6.0 137.0
50% 1100.0 60.0 7.0 155.0
75% 1250.0 85.1 8.0 172.0
max 2000.0 125.0 10.37 208.0

after the feature importance is calculated. The training is carried put variable and the input variables. It is observed that the more
out on 80% of the available data, while the testing is carried out the light pink then more is the positive correlation and purple col-
on the remaining 20%. Feature Scaling is done to standardize the our indicates negative correlation.
features by removing the mean and further scaling it to unit vari- The performance of the algorithms was evaluated using metrics
ance. After these procedures, the dataset is finally subjected to the such as MSE, MAE, and R2. As mentioned in equation (1), MAE is
machine learning regression-based algorithms i.e. XG Boost, derived by averaging the difference between the actual and pro-
Random Forest, Catboost, Decision Trees and AdaBoost algorithms. jected values of Ultimate Tensile Strength (MPa). As indicated in
In the case of the classification-based approach, the output param- equation (2), MSE is calculated by taking the average of the square
eter is categorized into two categories i.e. 0 (UTS < 70 % of the base of the difference between the actual and anticipated values. As indi-
metal) and 1 (UTS > 70 % of the base metal). cated in equation (3), the R2 value is derived by dividing the sum of
squared error regression by the sum of squared total variation.
3. Results and discussion
1X N 
yj  b

MeanAbsoluteError ¼ yj ð1Þ
The main characteristics of the dataset are analyzed and sum- N j¼1
marized by using exploratory data analysis as shown in Fig. 2 a)
and further visualized in a two-dimensional format in the form
1 XN
 2
of a heatmap as shown in Fig. 2 b). Fig. 2 a) represents the statisti- MeanSquareError ¼ y b
yj ð2Þ
N j1 j
cal analysis of the data. Fig. 2 b) represents the seaborn heatmap
which is used to find the correlation information between the out-
84
A. Mishra and R. Morisetty Manufacturing Letters 32 (2022) 83–86

Fig. 2. a) Exploratory data analysis plot b) Exploratory data analysis heatmap visualization c) Confusion Matrix results for Supervised Machine Learning Classification based
Algorithms d) ROC Curve Characteristics.
85
A. Mishra and R. Morisetty Manufacturing Letters 32 (2022) 83–86

Table 2a Declaration of Competing Interest


Metrics features for measuring the performance of the regression based algorithms.

Algorithm MAE MSE R2 Training Testing The authors declare that they have no known competing finan-
Value Value Value Accuracy in % Accuracy in % cial interests or personal relationships that could have appeared
Decision 16.45 506.08 0.49 93.83 53.44 to influence the work reported in this paper.
Tree
Random 14.90 479.54 0.52 87.54 52.49 References
Forest
XG Boost 13.39 433.70 0.57 87.99 57.03 [1] Jordan MI, Mitchell TM. Machine learning: Trends, perspectives, and prospects.
Cat Boost 15.73 465.33 0.53 93.30 53.90 Science 2015;349(6245):255–60.
Ada Boost 11.36 249.98 0.75 79.96 75.23 [2] Goodfellow I, Bengio Y, Courville A. Machine learning basics. Deep Learn
2016;1(7):98–164.
[3] Zhang X-D. Machine learning. In: Zhang X-D, editor. A Matrix Algebra
Table 2b Approach to Artificial Intelligence. Singapore: Springer Singapore; 2020. p.
Performance evaluation of the classification based algorithms. 223–440.
[4] Shalev-Shwartz S, Ben-David S. Understanding machine learning: From theory
Algorithm F1-Score of ‘00 F1-Score of ‘10 Overall F1-Score to algorithms. Cambridge University Press; 2014.
[5] Wagstaff, K., 2012. Machine learning that matters. arXiv preprint
Decision Tree 0.86 0.33 0.7647 arXiv:1206.4656.
Random Forest 0.90 0.40 0.8235 [6] Verma S, Misra JP, Singh J, Batra U, Kumar Y. Prediction of tensile behavior of FS
XG Boost 0.93 0.67 0.8824 welded AA7039 using machine learning. Mater. Today Commun
Cat Boost 0.80 0.49 0.7520 2021;26:101933.
Ada Boost 0.85 0.50 0.7631 [7] Sandeep R, Natarajan A. Prediction of peak temperature value in friction lap
welding of aluminium alloy 7475 and PPS polymer hybrid joint using machine
P 
 2
learning approaches. Mater Lett 2022;308:131253.
j
b
yj  y [8] Pothur Hema (October 24th 2019). Experimental Investigations on AA 6061
2
R ¼P  ð3Þ Alloy Welded Joints by Friction Stir Welding, Aluminium Alloys and
j yj  yÞ 2 Composites, Kavian Omar Cooke, IntechOpen, DOI: 10.5772/
intechopen.89797. Available from: https://www.intechopen.com/books/
Table 2a) shows the obtained metrics features of the regression aluminium-alloys-and-composites/experimental-investigations-on-aa-6061-
alloy-welded-joints-by-friction-stir-welding.
based algorithms while Table 2b) shows the F1-Score obtained by [9] Elatharasan G, Kumar VS. An experimental analysis and optimization of
the classification based algorithms. The results show that the process parameter on friction stir welding of AA 6061–T6 aluminum alloy
AdaBoost algorithm resulted in the less MSE and MAE values in using RSM. Procedia Eng 2013;64:1227–34.
[10] Ghiasvand A, Hassanifard S, Saadi S, Varvani-Farahani A. Tensile properties and
comparison to the other algorithms and also resulted in high R2 microstructural features of friction stir welded Al 6061 joints fabricated by
value which determines the best curve fitting in comparison to various dual-pin tool shapes. Sci Technol Weld Joining 2021;26(6):493–502.
the other algorithms. [11] Fuse K, Badheka V. Hybrid self-reacting friction stir welding of AA 6061–T6
aluminium alloy with cooling assisted approach. Metals 2021;11(1):16.
The AdaBoost algorithm, as shown in Table 2a), can be used to [12] Banik A, Saha A, Deb Barma J, Acharya U, Saha SC. Determination of best tool
improve the performance of a machine learning algorithm and is geometry for friction stir welding of AA 6061–T6 using hybrid PCA-TOPSIS
also optimal for weak learners because it minimizes the loss func- optimization method. Measurement 2021;173:108573.
[13] Alinaghian I, Amini S, Honarpisheh M. Residual stress, tensile strength, and
tion. AdaBoost performs a succession of t = 1,. . .,T classifiers, con-
macrostructure investigations on ultrasonic assisted friction stir welding of AA
tinually calling the weak classifiers. The ‘‘weight” calculated by 6061–T6. J Strain Anal Eng Design 2018;53(7):494–503.
erroneously classified instances increases with each execution [14] Maneiah D, Mishra D, Rao KP, Raju KB. Process parameters optimization of
(or, alternatively, weights of each correctly classified examples friction stir welding for optimum tensile strength in Al 6061–T6 alloy butt
welded joints. Mater Today: Proc 2020;27:904–8.
decreases). New classifiers are forced to concentrate on cases that [15] Asmare A, Al-Sabur R, Messele E. Experimental investigation of friction stir
prior classifiers improperly classified. From Table 2b) it is observed welding on 6061–T6 aluminum alloy using taguchi-based GRA. Metals
that the XG Boost algorithm yields better results in comparison to 2020;10(11):1480.
[16] Fuse K, Badheka V. Effect of shoulder diameter on bobbin tool friction stir
other classification based algorithms. Fig. 2 c) represents the con- welding of AA 6061–T6 alloy. Mater Today: Proc 2021;42:810–5.
fusion matrix of each algorithm and Fig. 2 d) represents the ROC [17] Gomathisankar M, Gangatharan M, Pitchipoo P. A novel optimization of
curve of each algorithms. It is observed that the XG Boost performs friction stir welding process parameters on aluminum alloy 6061–T6. Mater
Today: Proc 2018;5(6):14397–404.
better than the other algorithms. [18] Napitupulu RA, Simanjuntak SL, Manurung C, Hutabarat C, Sitompul S, Aoh JN.
Friction stir welding of aluminium alloy 6061-t651. IOP Conference Series:
Materials Science and Engineering. IOP Publishing; 2019. Vol. 508, No. 1, p.
012064.
4. Conclusion
[19] Hema P. Experimental investigations on AA 6061 alloy welded joints by
friction stir welding. Aluminium Alloys and Composites. IntechOpen; 2019.
In the present study, Python computing was successfully used [20] Liu F, Fu L, Chen H. High speed friction stir welding of ultra-thin AA6061-T6
for developing and executing the supervised machine learning- sheets using different backing plates. J Manuf Processes 2018;33:219–27.
[21] Rao RV, Kumar MS. Experimental investigation on effect of welding
regression and classification based algorithms for the determina- parameters on the friction stir welding of AA 6061. Mater Today: Proc
tion of the UTS (MPa) of similar FSWed AA6061 joints. It was 2018;5(5):12265–72.
observed that out of three input parameters, Axial Force (KN) has [22] Emamian S, Awang M, Yusof F, Hussain P, Meyghani B, Zafar A. The effect of pin
profiles and process parameters on temperature and tensile strength in
more influence on the output parameter i.e. Ultimate Tensile friction stir welding of AL6061 alloy. In: The Advances in Joining
Strength (MPa). It is also noticed that in regression based apprach Technology. Singapore: Springer; 2019. p. 15–37.
the Adaptive Boosting algorithm (AdaBoost) outperforms the other [23] Abioye TE, Zuhailawati H, Anasyida AS, Ayodeji SP, Oke PK. Effects of
particulate reinforcements on the hardness, impact and tensile strengths of
four algorithms by resulting in less MSE and MAE values as well as AA 6061–T6 friction stir weldments. Proc Inst Mech Eng, Part L: J Mater:
a high coefficient of determination value. While in case of the clas- Design Appl 2021;235(6):1500–6.
sification based algorithms resulted in the highest F1-Score in [24] Setiawan W, Darmadi DB, Suprapto W, Soenoko R, Tri SG. Varied corner joint
design alumunium 6061 uing friction stir welding Vol. 494, No. 1, p. 012106.
comparison to other algorithms. XG Boost is a fast adaptation of IOP Conference Series: Materials Science and Engineering. IOP Publishing;
the stochastic gradient boosting algorithm with a number of 2019.
hyperparameters that let you fine-tune the model training process. [25] Rathinasuriyan C, Kumar VS. Optimisation of submerged friction stir welding
parameters of aluminium alloy using RSM and GRA. Adv. Mater. Process.
The future scope of the work will be to implement the nature based
Technol. 2021;7(4):696–709.
optimization algorithms and further compare the results with the [26] Pasha, M.A., Aditya, T.N., Chandrakanth, A. and Gurubrahmam, K., Friction Stir
Machine Learning algorithms. Welded Joint Efficiency of Aluminium Alloy 6061-T6.

86

You might also like