Machine Learning for Friction Stir

Welding Process
Akshansh Mishra,

Computing Scientist, Centre for Artificial Intelligent Manufacturing Systems,

Stir Research Technologies, India

This work is licensed under a Creative Commons AttributionNonCommercial

4.0 International License.

PREFACE

The Machine Learning process is revolutionizing many sectors in engineering

including manufacturing sectors such as the Friction Stir Welding Process.
The implementation of machine learning algorithms in Friction Stir Welding
process not only reduced the time of the experiment but also reduces the cost
of the experiment.

This book highlights the various application of machine learning algorithms

in the Friction Stir Welding process such as prediction of the mechanical
properties, detection of surface defects etc. The book is written in a simple
language so that the readers will understand the concepts clearly while going
through this book.

This page is left intentionally blank.

ACKNOWLEDGEMENT

The author would like to thanks his close friends Rajat Sarawagi, Upendra
Srivastava, and Vijay Patel. The author is deeply grateful to his mother
Sucheta Mishra and his brother Dewansh Mishra for motivating him in his
whole career path towards Machine Learning. The author would also like to
thank Amrendra Kumar Sharma, the best faculty and motivator he knows to
date.

TABLE OF CONTENTS
1. Machine Learning in Friction Stir Welding
2. Image Processing in Friction Stir Welding Process
3. Supervised Machine Learning in Friction Stir Welding
4. Discrete Wavelet Transformation in Friction Stir Welding Process
5. Grain Size Distribution Analysis in Friction Stir Welding Process
6. Prediction of the Mechanical Property of Friction Stir Welded Joint
Chapter 1
Machine Learning in Friction Stir Welding
1. Introduction

Machine learning process drastically decreases the time it takes to develop

stronger, lighter materials. This is important to the aerospace, automotive and
manufacturing sectors. Machine Learning techniques like Artificial Neural
Networks and Image processing are used in Friction Stir Welding process for
the optimization of mechanical properties like Ultimate Tensile Strength,
Fracture Strength and elongation % and microstructure properties like grain
size and understanding defects formation. In the recent chapter, application of
Machine Learning technique in Friction Stir Welding technology will be
discussed.

2. Understanding Human Nervous System

The human nervous system can be classified as a three stage system which is
shown in the block diagram of Figure 1. The brain is in central position to the
nervous system which is represented by the neural net. The function of the
neural net is to receive the information continuously, perceive it and then to
make appropriate decision.
The green arrows which are pointing from left to right represent the forward
transmission of information carrying signals through the nervous system. The
red arrows in the diagram show the presence of feedback in the nervous
system. Stimuli from the external environment or the human body are
converted to electrical impulses by receptors which further convey the
information to the brain (neural net). Electrical impulses which are generated
by the receptors are further converted to a distinguishable response as a
nervous system outputs by the effectors [1]. Figure 2 represents the schematic
diagram of the human nerve cell.
Fig.2. Representation of human nerve cell

Neuron is the fundamental unit of the neural network architecture. Dendrites,

soma and axon constitute the neuron structure. Dentrites are tree-like
structure that basically receives the signal from surrounding neurons, where
each line is connected to one neuron. Axon is a thin cylinder that transmits
the signal from one neuron to others. The contact to the dendrites is made
through a synapse, at the end of axon. The inter-neuronal signal at the
synapse is usually chemical diffusion but sometimes electrical impulses.
Only if certain condition is met, the neuron fires an electrical impulse [2]. A
neuron fires only if the total weight of the synapses that receive impulses in
the period of latent summation exceeds the threshold [3].

3. Modeling Artificial Neural Network

The information processing unit of the basic neural network architecture is a

neuron. Non-linear model of a neuron is shown in the Figure 3.
The given neural model has three basic elements i.e. set of synapses, an adder
and an activation function. The characterization of the set of synapses or
connecting links is done by its own weight or

strength. Signal
at the input of synapse j which is connected to the neuron k is further
multiplied by the synaptic weight
. The

synaptic weight of the artificial neuron lies between the both positive and
negative values. An adder performs the operation of the linear combination. It
sums up the input signals which are weighted by the synaptic strengths. The
main application of the activation function is to limit the amplitude of the
output of neuron. The bias which is shown in the Figure 3, is applied
externally and its main function is to increase or decrease the net input of the
activation function. Mathematically, we can define the the neuron k shown in
the Figure 3 with the help of the pair of equations:

(1)
and,
 (2)
4. Machine Learning and its classification

Machine learning technique generally focuses on the development and

execution of the computer programs that can access data and use it learn for
themselves. It’s just like before solving calculus exercise numerical
problems, you go through the various examples given in that particular
chapter i.e. you are training your mind and when you solve new question it is
called testing your mind and the result you get for that particular question
measures the accuracy of your performance. So we can define Machine
learning as an application of artificial intelligence (AI) that provides the
given systems or state the ability to automatically learn and improve from
experience without being explicitly programmed [4-8]. Machine learning
classified into three main types i.e. supervised learning, unsupervised
learning and reinforcement learning which will be discussed one by one in
the sub sections.

4.1 Supervised Learning

Supervised learning can be referred to as the situation of learning with a

teacher. The working of supervised machine learning is shown in the Figure
4. We may assume that the teacher possess the knowledge of the given
environment which is unknown to the neural network architecture. The
representation of the knowledge is done by a set of input-output examples.
Figure 4 shows that a training vector which is drawn from the same
environment is exposed to the neural network and the teacher. So, for that
training vector, a desired response is given by the teacher with the help of the
possessed knowledge. Error signal shown in the Figure 4 is the difference of
the desired response and the actual response. Under the combined influence
of the error signal and the training vector, network parameters are adjusted.
This step by step adjustment is carried out further which makes proper
transfer of knowledge of the environment possessed by the teacher to the
neural network.

4.2 Unsupervised Learning

Unsupervised machine learning is a self-organized learning process in which

there is no external person or teacher to look after or observe the learning
process. The working of the unsupervised machine learning process is shown
in the Figure 5. In order to carry out unsupervised machine learning task, a
competitive learning rule is used. Just take an example of a neural network
consisting of two layers i.e. an input layer and a competitive layer. The
available data is received by the neural network. If we talk about the
competitive layer, it consists of various neurons which compete against each
other for the opportunity for responding the features contained in the input
data as per the given learning rule.

4.3
Reinforcement Learning

Reinforcement learning is a semi-supervised machine learning method which

trains the models to make a sequence of decisions. The working of the
Reinforcement learning is shown in the Figure 6. In an uncertain, potentially
complex environment agent learns to achieve a particular goal. An artificial
intelligence is subjected to a game-like situation where the computer employs
trial and error method to come up with a solution to the given problem. The
main objective of the Reinforcement Learning is the maximization of the
total reward. Reinforcement Learning works on the rewards and penalties
scheme. The Reinforcement Learning model tries to figure out the method to
perform the given task in order to maximize the reward, starting from totally
random trials and finishing with advanced tactics and superhuman skills.
 5.
Literature Survey

Verma et al. [9] used various sophisticated machine learning approaches like
. Gaussian process (GP) regression, support vector machining (SVM) and
multi-linear regression (MLR) for evaluating the friction stir welding process.
Fig.7. Actual (x-axis) vs predicted values (y-axis) of Ultimate Tensile Strength (UTS) using GP, SVM,
and MLR using training data [9].

As shown in Figure 7 it is observed that the GPR approach works better than
the SVM and MLR techniques. Therefore, GPR approach is generally used
successfully for prediction of the UTS of Friction Stir welded joints.

Debroy et al. [10] studied the conditions for void formation using a decision
tree and a Bayesian neural network. Schematic representation of the research
is shown in the Figure 8.
Fig.8. The components are FSW process, mechanistic models, and machine learning methods (neural
network and decision tree) [10].

The study showed that the neural network and the decision tree predicted
void formation with 96.6% accuracy by computing the causative variables
like temperature, strain rate, torque, and maximum shear stress on the tool
pin.

Celik et al. [11] investigated the correlation between the friction welding
parameters and tensile strength of both AISI 316 austeniticstainless steel and
Ck 45 steel by developing an Artificial Neural Network (ANN) model.
Figure 9 represents the artificial neural network architecture used in the
study.
Fig.9. Four layered neural network architecture [11]

Figure 10 shows that a good correlation was obtained between the Artificial
Neural Network predicted values and the experimental values.

Fig.10.
Comparison of measured and predicted outputs for tensile strength [11]

Maleki et al. [12] developed Artificial Neural Network based on backward

propagation algorithm for predicting the yield strength, tensile strength,
notch-tensile strength and hardness of friction stir welded AA 7075-T6 joints.
The Neural Network architecture used in the study is shown in the Figure 11.
Fig.11.Schematic representation of Neural Network architecture used by Maleki et al. [12]

Figure 12 shows the comparison of the predicted and the experimental

results. The results show that the if neural networks are adjusted carefully
then it can be used for modeling of various Friction Stir Welding parameters.
Fig.12. Predicted values in comparison to experimental values [12]

Mishra et al. [13] implemented the Convolutional Neural Network for

identification of the texture of Friction Stir Welded joints and Conventional
Welded joints.

Macias et al. [14] established a correlation between the acoustic emission

signals and the various main parameters of friction stir welding process based
on artificial neural networks (ANNs) trained on Levenberg-Marquardt
algorithm. Figure 13 and 14 shows the methodology and the development of
Artificial Neural Network architecture respectively.
Fig.13. Artificial Neural Network development methodology [14]
Fig.14. Artificial Neural Network architecture [14]

Figure 15 shows the predicted and measured results on the given dataset. It is
clearly observed that the results obtained from the new model obtained, based
on Neural Network architecture is an effective technique for the prediction of
Friction Stir Welding process parameters and tensile strength of joint.
Fig.15. Predicted and Measured Values of the experimental dataset [14].
5. Designing a simple Artificial Neural Network on Google Colab by
using Python coding

In this section, development of an Artificial Neural Network model by using

Python Programming on Google Colaboratory platform as shown in Figure
16 will be discussed. Table 1 shows the experimental datset which shows the
Ultimate Tensile Strength (MPa) value of Friction Stir Welded joints against
the tool rotational speed of the tool (rpm).
Table 1. Experimental Dataset

Tool Rotational Speed (RPM) UTS (MPa)

425 465
575 444
350 440.6
650 415
500 448
Fig.16. Google Colaboratory Platform

First we will start with our imports. Here we are importing TensorFlow and
calling it tf for ease of use. We then import a library called numpy, which
helps us to represent our data as lists easily and quickly. The framework for
defining a neural network as a set of Sequential layers is called keras, so we
import that too as shown in the Figure 17.

Fig.17. Importing process

Next we will create the simplest possible neural network. It has 1 layer, and
that layer has 1 neuron, and the input shape to it is just 1 value as shown in
the Figure 18.
Fig.18. Defining the Neural Network

Now we compile our Neural Network as shown in Figure 19. When we do so,
we have to specify 2 functions, a loss and an optimizer. The LOSS function
measures the guessed answers against the known correct answers and
measures how well or how badly it did. It then uses the OPTIMIZER
function to make another guess. Based on how the loss function went, it will
try to minimize the loss.

It will repeat this for the number of EPOCHS which you will see shortly. But
first, here's how we tell it to use 'MEAN SQUARED ERROR' for the loss
and 'STOCHASTIC GRADIENT DESCENT' for the optimizer.
Fig.19. Compiling the Neural Network

Next up we'll feed in some data as shown in the Figure 20. A python library
called 'Numpy' provides lots of array type data structures that are a defacto
standard way of doing it. We declare that we want to use these by specifying
the values as an np.array[].
Fig.20.Providing the data

The process of training the neural network, where it 'learns' the relationship
between the Xs and Ys is in the model.fit call which is shown in Figure 21.
This is where it will go through the loop we spoke about above, making a
guess, measuring how good or bad it is (aka the loss), using the opimizer to
make another guess etc. It will do it for the number of epochs you specify.
When you run this code, you'll see the loss on the right hand side.
Fig.21. Training and Testing the Neural Network
6. Conclusion

There is a loss of time and materials if the optimization of the Friction Stir
Welding parameters is done through experimental studies which further leads
to increase in the cost of the experiment. Machine Learning approach like
Artificial Neural Network and image processing overcome these issues. So, it
can be concluded that the mechanical and microstructure properties can be
predicted and also the defects formation can also be observed by the
implementation of various Machine Learning tools in the Friction Stir
Welding process.

References

1. Haykin, S.S., 2009. Neural networks and learning

machines/Simon Haykin.
2. Zurada, Jacek M.``Introduction to Artificial Neural System''. West
Publishing Company, St. Paul, MN, 1992
3. Arbib, Michael A. ``Brains, Machines, and Mathematics: Second Edition''.
Springer-Verlag, New York, NY, 1987.
4. Alpaydin, E., 2020. Introduction to machine learning. MIT press.
5. Mitchell, T.M., 1997. Machine learning.
6. Marsland, S., 2015. Machine learning: an algorithmic perspective. CRC
press.
7. Michie, D., Spiegelhalter, D.J. and Taylor, C.C., 1994. Machine learning.
Neural and Statistical Classification, 13(1994), pp.1-298.
8. Quinlan, J.R., 2014. C4. 5: programs for machine learning. Elsevier.
9. Verma, S., Gupta, M. and Misra, J.P., 2018. Performance evaluation of
friction stir welding using machine learning approaches. MethodsX, 5,
pp.1048-1058.
10. Du, Y., Mukherjee, T. and DebRoy, T., 2019. Conditions for void
formation in friction stir welding from machine learning. npj Computational
Materials, 5(1), pp.1-8.
11. Ersozlu, I. and Celik, S., 2019. Artificial Neural Network application to
the friction welding of AISI 316 and Ck 45 steels. KOVOVE MATERIALY-
METALLIC MATERIALS, 57(3), pp.199-205.
12. Maleki, E., 2015. Artificial neural networks application for modeling of
friction stir welding effects on mechanical properties of 7075-T6 aluminum
alloy. In IOP Conference Series: Materials Science and Engineering (Vol.
103, No. 1, p. 012034). IOP Publishing.
13. Mishra, A. and Nagpal, K., 2019. Convolutional Neural Network for
Image Processing of Friction Stir Welded and Conventional Welded Joints
Texture. Int. J. Hum. Comp. Inter. Data Min, 2(1&2), pp.5-9.
14. Jiménez-Macías, E., Sánchez-Roca, A., Carvajal-Fals, H., Blanco-
Fernández, J. and Martínez-Cámara, E., 2014. Wavelets application in
prediction of friction stir welding parameters of alloy joints from
vibroacoustic ANN-based model. In Abstract and Applied Analysis (Vol.
2014). Hindawi.

Chapter 2
Image Processing in Friction Stir Welding Process
1. Introduction

Image processing can be a breakthrough in aerospace and manufacturing

industries. In this research study, Friction Stir Welding of 6060 T5 Alloy is
carried out. The image obtained of the Friction Stir Welded joint is subjected
to basic image processing techniques using the OpenCV computer vision
library and some standard data analysis libraries in Python.

Image Processing is a machine learning technique which is basically used to

perform various operations on an image, in order to get an used to perform
various operations on an image, in order to get an 4].

Image Processing has been used by various industries for detection of surface
defects. Elbehiery et al. [5] used an image processing method for the
detection of surface defects in ceramic tiles. Vasilic et al. [6] used the edge
detecting image processing method for ceramic tiles defects detection on the
basis of color and surface texture. Alegre et al. [7] carried out study on the
eggshell defect detection by using color processing method. Liu et al. [8]
used a rapid real-time defect detection image processing algorithm based on
the pixels brightness for defects detection on rail tracks. In the present work,
image processing is used to extract various features like generating
histograms form color and grayscale image, histogram equalization, and
generating Fourier transform of the grayscale image of the friction stir
welded joint. Friction Stir Welding process is a novel solid state joining
technique developed by The Welding Institute in 1991. Friction Stir Welding
process overcomes the drawbacks of various conventional welding process
[9-11]. Friction Stir Welding is generally used to join aluminum alloys,
magnesium alloys and as well as various titanium alloys which find
application in aerospace industries [12-15]. In this case study, AA 6060-T5
alloy is used for Friction Stir Welding purpose.

2. Materials and Methods

Aluminum alloy plate 6060 T5 of the dimension 150 mm X 100mm X 6mm
was joined by Friction Stir Welding process. The plate was fixed on the
fixture mounted on the CNC bed and tool fabricated from H13 steel is used
for joining the plates. The input parameters which should be properly
monitored in Friction Sir Welding process for obtaining sound joints are Tool
Rotational Speed (rpm), Tool Traverse Speed (mm/min) and Axial Force. In
this case study Tool Rotational Speed of 1500 rpm, Tool Traverse speed of
400 mm/min and an axial force of 2.5KN were used. With the help of a
digital camera, image of the welded joint was captured and further cropping
process is done to the size of 814 X 2559 pixels for the image processing task
as shown in the Figure 1.

Jupyter notebook using Python 3.0 programming language on the Google

Colaboratory platform is used for developing image processing algorithm
while using opencv, numpy and matplotlib libraries. Image processing
method is implemented in three steps as shown in the Figure 2.
 Fig.1 Cropped Image of Friction Stir Welded
Joint
Importing the image via Image acquisition tools
Analyzing and manipulating the images
Output in the form of altered image
Fig.2. Image Processing steps carried out via Friction Stir Welded image
3. Results and Discussions

In the first step image is read from urls, and display it by using openCV,
please note the difference when reading image in RGB and BGR format. The
default input color channels are in BGR format for openCV . The obtained
result is shown in the Figure 3.
Sometimes it is desired to enhance the contrast in a particular image or
expand the contrast in a particular region while sacrificing the detail in colors
that don’t vary much, or don’t matter. A good tool to find interesting regions
is the histogram shown in the Figure 4. To create a histogram of our image
data, we use the matplot.pylab hist() function.

Figure 5 display the histogram of R, G, B channel. It is observed that the

green channel has many pixels in 255, which represents the white patch in the
image.
Figure 6 represents the grayscale image. Histogram of the grayscale image is
shown in the Figure 7. It is observed that the frequency of the image
histogram has decreased ~ 1/3 of the histogram of color image.

Figure 8 shows the obtained contour by using matplotlib.contour.

Fig.3 Read and Display

Fig.4. Histogram of all the pixels in the

color image Fig.5.Histogram of R,G, B
Channel

Fig.6.Grayscale Image
 Fig.7.Histogram of the grayscale image

Fig.8.Contour of the friction stir welded

joint image

Grayscale transformation and histogram equalization was performed on the

given image. Grayscale transformation is an inverse operation of the
grayscale image, it is seen from Figure 9 that the bright pixels become dark,
and the dark pixels become bright.
Fig.9.Grayscale Transformation of the
image Fig.10.Haze effect in image

Another transform of the image shown in the Figure 10, after adding a
constant, all the pixels become brighter and a hazing-like effect of the image
is generated.

Histogram Equalization transform flattens the gray-level histogram so that all

intensities are as equally common as possible which is compared in Figure 11
and Figure 12. The transform function is a cumulative distribution function
(cdf) of the pixel values in the image (normalized to map the range of pixel
values to the desired range).
 Fig.11. Before Histogram Equalization

Fig.12.After Histogram Equalization

A Fourier transform is used to find the frequency domain of an image. You

can consider an image as a signal which is sampled in two directions as
shown in the Figure 13. So taking a Fourier transformation in both X and Y
direction gives you the frequency representation of image. For the sinusoidal
signal, if the amplitude varies so fast in short time, you can say it is a high
frequency signal. If it varies slowly, it is a low frequency signal. Edges and
noises are high frequency contents in an image because they change
drastically in images.
 Fig.13.Fourier transform of the
grayscale image

Figure 14 demonstrates conducting a high pass filter to remove the low

frequency component, resulting in a sharpened image which contains the
edges.

Fig.14.Conduction of High Pass

filter

Image processing is successfully implemented on the image of Friction Stir

Welded joint for extracting the necessary features from the image. Improper
selection of the input parameters like Tool Rotational Speed, tool traverse
speed and axial force during friction stir welding process may result in flash
formation, groovy edges formation and other various types of surface defects.
So, image processing techniques used in this research work can be also used
for the detection of any types of irregularities or surface defects present on
the Friction Stir Welded joint. Hence, it can be used as an application for
monitoring of the quality of the welded joints.

References
1. Russ, J.C., 1990. Image processing. In Computer-assisted microscopy (pp.
33-69). Springer, Boston, MA.
2. Lim, J.S., 1990. Two-dimensional signal and image processing. ph.
3. Abràmoff, M.D., Magalhães, P.J. and Ram, S.J., 2004. Image processing
with ImageJ. Biophotonics international, 11(7), pp.36-42.
4. Pavlidis, T., 2012. Algorithms for graphics and image processing. Springer
Science & Business Media.

5. Elbehiery, H., Hefnawy, A. and Elewa, M., 2005. Surface defects detection
for ceramic tiles using image processing and morphological techniques.

6. Vasilic, S. and Hocenski, Z., 2006, July. The edge detecting methods in
ceramic tiles defects detection. In 2006 IEEE methods in ceramic tiles defects
detection. In 2006 IEEE 472). IEEE.

7. García-Alegre, M.C., Ribeiro, A., Guinea, D. and Cristóbal, G.,

2000, March. Eggshell defects detection based on color processing. In
Machine Vision Applications in Industrial Inspection VIII (Vol.
3966, pp. 280-287). International Society for Optics and Photonics.

8. LIU, Y., LIU, T., WANG, Q. and LUO, S., 2007. Rail Surface Defects
Detection Algorithm Based on Image Processing [J]. Computer Engineering,
11.

9. Mishra, R.S. and Ma, Z.Y., 2005. Friction stir welding and processing.
Materials science and engineering: R: reports, 50(1-2), pp.1-78.

10. Thomas, W.M. and Nicholas, E.D., 1997. Friction stir welding for the
transportation industries. Materials & design, 18(4-6), pp.269-273.
11. Guerra, M., Schmidt, C., McClure, J.C., Murr, L.E. and Nunes, A.C.,
2002. Flow patterns during friction stir welding. Materials characterization,
49(2), pp.95-101.
12. Threadgill, P.L., Leonard, A.J., Shercliff, H.R. and Withers, P.J.,
2009. Friction stir welding of aluminium alloys. International Materials
Reviews, 54(2), pp.49-93.

13. Kwon, Y.J., Shigematsu, I. and Saito, N., 2008. Dissimilar friction stir
welding between magnesium and aluminum alloys. Materials Letters, 62(23),
pp.3827-3829.

14. Dressler, U., Biallas, G. and Mercado, U.A., 2009. Friction stir welding
of titanium alloy TiAl6V4 to aluminium alloy AA2024-T3. Materials Science
and Engineering: A, 526(1-2), pp.113-117.

15. Mishra, A. and Dixit, D., 2018. Friction Stir Welding of

15. Mishra, A. and Dixit, D., 2018. Friction Stir Welding of

46.

Chapter 3
Supervised Machine Learning in Friction Stir Welding
1. Introduction

In this study, two supervised machine learning algorithms i.e. supervised

machine learning regression and supervised machine learning classification
model have been used for predicting the Ultimate Tensile Strength (MPa) and
Weld joint efficiency of Friction Stir Welded joint. The results showed that
Polynomial regression model yields better result than other supervised
machine learning regression model and Decision Tree (Both Gini Index and
Information Gain criterion), Artificial Neural Network classification models
gave better classification result than K-Nearest Neighbor classification
model.

Machine Learning can be considered as a subset of artificial intelligence

where available information and data are being learned by some algorithms.
The history of the development of machine learning dates back to 1950 when
the Turing test was created by Alan Turing in order to verify the intelligence
of the computer system [1]. In 1952 the first learning-based computer
program was developed by Arthur Samuel [2]. The first neural network i.e.
the perception model was created for computers by Frank Rosenblatt in 1957
with the main objective of simulating the thought process of humans. In
1967, basic pattern recognition activities for computers were carried out by a
designed nearest neighbor algorithm. Machine learning algorithms are
divided into four categories i.e. supervised machine learning algorithm,
unsupervised machine learning algorithm, semi-supervised machine learning
algorithm, and reinforcement machine learning algorithm [3-5].
Supervised machine learning algorithm finds application in predicting the
future, unseen or unavailable data based on the available dataset [6-7]. The
supervised machine learning algorithm is further subdivided into two types
i.e. regression model and classification model [8-9].

Supervised machine learning algorithms are finding various applications in

the manufacturing and materials industries. Supervised machine learning
algorithms are being used for predicting the material properties such as
fracture strength, tensile strength, elongation percentage, and also the
hardness of the given material [10-13].

The present research work focuses on the implementation of supervised

machine learning regression algorithms such as polynomial regression,
support vector regression, decision tree regression, random forest regression
as well as an artificial neural network is used for predicting the Ultimate
Tensile Strength of friction stir welded similar AA2024-T3 aluminum alloy
and supervised machine learning classification algorithms such as Knearest
neighbors, decision tree classifier with Gini Index and Information Gain as
criteria for classifying the joint having weld strength efficiency greater 70
percent of the base metal.

2. Materials and Methods

Similar AA2024-T3 aluminum alloy plates of dimension 200 X 100 X

3.5mm were mounted on a vertical milling machine in order to be butt
welded by the Friction Stir Welding process by using a tapered pin profile
tool steel (HSS) [14]. The experimental dataset consists of 27 data which is
fetched from the work of Hussein et al. [14] and is further converted to CSV
(Comma Separated Values) file for importing purposes during the execution
of Python codes. The dataset consists of Tool Shoulder Diameter (mm), Tool
Rotational Speed (rpm), Tool Traverse Speed (mm/min), Number of weld
passes, and Tool Tilt Angle as input parameters while the Ultimate Tensile
Strength (UTS) is an output parameter.
The Python libraries which are imported for constructing and executing the
Machine Learning algorithms were Numpy, Matplotlib, Seaborn, Pandas,
Tensorflow, and Keras. Figure 1 shows the hierarchy of the experimental
procedure subjected to the CSV dataset.

Figure 1: Various operations subjected to the imported dataset

The dataset is subjected to Machine Learning Classification Models. For the

Machine Learning classification modeling, the output of the given dataset is
classified as 0 if the UTS is less than 70% of the reference value and as 1 if
UTS is greater than or equal to 70 % of the reference value and at last it is
subjected to three classifier models i.e. K- Nearest Neighbors, Decision Tree
with Gini Index and Information Gain as criteria and Neural Network
classification model.

3. Results and Discussions

3.1 Exploratory Data Analysis

At this stage, we have explored the relation the features shared with the target
variable. Accordingly, the features were dropped which have no relation with
the target variable. From Table 1 we can see the data's distribution and judge
whether we need to normalize our data or not. We also get other statistics
using the table.
Table 1: Statistical model of the experimental dataset

3.2 Checking Null Values

The check_null( ) function is used to check the number of null values in the
dataset. The null values are replaced by mean.
3.3 Plotting Graph of p-Value Function and Contour Plot

The plot_graph_pvalue ( ) function plots a line plot between given variables

and prints the p-value and Pearson values. The
contour_plot ( ) function plots a contour plot for the given variables. Figure 2
shows the plot against the shoulder diameter (mm) and UTS (MPa). The
obtained p-value and Pearson value for the given parameters is 0.50151 and
0.135 respectively.
Figure 2. Relationship between UTS (MPa) and Shoulder Diameter (mm)

From the p-value and Pearson value, we can clearly interpret that shoulder
diameter is highly correlated with UTS. From the graph, we can see the
Pearson's predictions come to life as we see the UTS values start to drop with
increasing shoulder diameter after 14 mm. Figure 3 shows the contour plot of
shoulder diameter and UTS.
Figure 3. Contour Plot between UTS (MPa) and Shoulder diameter (mm)

Figure 4 shows the plot against the Tool Rotational Speed (rpm) and UTS
(MPa). The obtained p-value and Pearson value for the given parameters is
0.00034 and 0.638 respectively. Figure 5 shows the contour plot between
Tool Rotational Speed (rpm) and UTS (MPa).
Figure 4. Relationship between UTS (MPa) and Tool Rotational Speed (rpm)
Figure 5. Contour Plot between UTS (MPa) and Tool Rotational Speed (rpm)

From the graph depicted in Figure 4 it is observed that the UTS increases
after tool rotational speed of 1000 rpm. Figure 6 shows the plot against the
Tool Travel Speed (mm/min) and UTS (MPa). The obtained p-value and
Pearson value for the given parameters is 0.00045 and 0.628 respectively.
Figure 7 shows the contour plot between Tool Traverse Speed (mm/min) and
UTS (MPa).
Figure 6. Relationship between UTS (MPa) and Tool Traverse Speed (mm/min)
Figure 7, Contour Plot between UTS (MPa) and Tool Traverse Speed (mm/min)

From the graph shown in Figure 6 it is observed that the UTS continue to
increase after traverse speed greater than 55 mm/min. Figure 8 shows the plot
against the Tilt Angle and UTS (MPa). The obtained p-value and Pearson
value for the given parameters is 0.34943 and 0.187 respectively. Figure 9
shows the contour plot between Tilt Angle and UTS (MPa).
Figure 8, Relationship between UTS (MPa) and Tool Tilt Angle
Figure 9. Contour Plot between UTS (MPa) and Tool Tilt Angle

From the graph shown in Figure 8 it is observed that UTS continues to

decrease with increase in Tool Tile angle upto 2 degree but after 2 degree,
UTS starts increasing. . Figure 10 shows the plot against the number of weld
passes and UTS (MPa). The obtained p-value and Pearson value for the given
parameters is 0.63737 and 0.095 respectively. Figure 11 shows the contour
plot between number of weld passes and UTS (MPa).
Figure 10. Relationship between UTS (MPa) and Number of passes
Figure 11. Contour Plot between UTS (MPa) and Number of passes
Figure 12 shows the correlation heat map.
Figure 12. Correlation Heat map

Figure 13 shows the correlation plot with response variable using bar graph.
From the plot analysis, we conclude that all input parameters have some
relation with UTS. So, we will not be dropping them from the dataset.
Figure 13. Response variable plot
3.4 Implementation of Supervised Machine Learning Regression Models

Firstly, the dataset is split after that we decide whether to normalize or not.
From the statistics shown in the Table 2, we will decide whether or not to
normalize the dataset.
Table 2: Statistical Data Analysis
Since the range of Travel speed is larger than that of other variables, we will
be normalizing the data.

The get_score_regression function will give Mean Absolute Error, Mean

Squared Error, and R squared error to analyze the model's performance. The
Polynomial regression, Decision Tree regression, Random Forest regression,
Support Vector Regression and Neural Network regression models were
implemented and the one which performs the best was selected. Table 3
shows the model analysis of the regression models implemented on the
dataset.
Table 3. Mean Absolute Error, Mean Square Error and Coefficient of determination
Figure 14 shows the model performance of the Neural Network Regressor
model.
Figure 14. Variation of loss function with number of epochs
Figure 15 and Figure 16 shows the model performance in terms of Mean
Square Error (MSE) and Mean Absolute Error (MAE).
Figure 15. Representation of MSE of each Machine Learning regression models
Figure 16. Representation of Mean Absolute Error of each Machine Learning Regression Models

From Figure 15 and 16 it can be clearly interpreted that the Polynomial

Regression model and Decision Tree are a better fit than other models while
on the basis of Mean absolute error, we can see that the Polynomial
Regression model outperforms all other models.

3.5 Implementation of Supervised Machine Learning Classification

Models

K-nearest Neighbors, Decision Tree Classifier (with Gini Index and

Information Gain as Criterions), and Neural Network classification models
were subjected to the dataset. The accuracy of the models for classification is
shown in the Table 4. It is observed that Artificial Neural Network, Decision
Tree classifier (Gini Index) and Decision tree classifier (Information gain as
criterion) results better accuracy.
Table 4. F1 Accuracy of Machine Learning Classification models

Figure 17 shows the performance of training and testing set while

implementing Neural Network classification model.
Figure 17. Training and testing set model performance
Figure 18 shows the plot of F1-accuracy of each classification models.
Figure 18. F1-accuracy of Machine learning classification models
4. Conclusion

From the obtained results, we can see that Decision Tree Regressor and
Polynomial Regressor outperform all other regressor models., As a classifier,
Artificial Neural Network, Decision Tree (Gini Index) and Decision Tree
(Information Gain) outperform the KNearest Neighbor model. In both the
case, the Deep Learning model could have given a much better performance
had there not been lack of data.

References
1. A. M. Turing, “Computing machinery and intelligence,” Mind LIX, 433–
460 (1950). https://doi.org/10.1093/mind/lix.236.433
2. A. L. Samuel, “Some studies in machine learning using the game of
checkers,” IBM J. Res. Dev. 3, 210–229 (1959).
3. Alpaydin, Ethem. Introduction to machine learning. MIT press, 2020.

4. Molnar, C., 2020. Interpretable Machine Learning. Lulu. com. 5. Zhang,

X.D., 2020. Machine learning. In A Matrix Algebra Approach to Artificial
Intelligence (pp. 223-440). Springer, Singapore.

6. Di Noia, A., Martino, A., Montanari, P. and Rizzi, A., 2020. Supervised
machine learning techniques and genetic optimization for occupational
diseases risk prediction. Soft Computing, 24(6), pp.4393-4406.

7. Chen, J., Lange, T., Andjelkovic, M., Simevski, A. and Krstic, M.,
2020. Prediction of solar particle events with SRAM-based soft error rate
monitor and supervised machine learning. Microelectronics Reliability,
p.113799.

8. Sen, P.C., Hajra, M. and Ghosh, M., 2020. Supervised classification

algorithms in machine learning: A survey and review. In Emerging
Technology in Modelling and Graphics (pp. 99-111). Springer, Singapore.

9. Gan, J., Wen, G., Yu, H., Zheng, W. and Lei, C., 2020. Supervised feature
selection by self-paced learning regression. Pattern Recognition Letters, 132,
pp.30-37.

10. Marani, A. and Nehdi, M.L., 2020. Machine learning prediction of

compressive strength for phase change materials integrated cementitious
composites. Construction and Building Materials, 265, p.120286.

11. Tawfik, S.A., Isayev, O., Spencer, M.J. and Winkler, D.A., 2020.
Predicting thermal properties of crystals using machine learning. Advanced
Theory and Simulations, 3(2), p.1900208.

12. Qin, Z., Wang, Z., Wang, Y., Zhang, L., Li, W., Liu, J., Wang, Z., Li, Z.,
Pan, J., Zhao, L. and Liu, F., 2020. Phase prediction of Nibase superalloys
via high-throughput experiments and machine learning. Materials Research
Letters, 9(1), pp.32-40.

13. Chapman, J., Batra, R. and Ramprasad, R., 2020. Machine learning
models for the prediction of energy, forces, and stresses for platinum.
Computational Materials Science, 174, p.109483. 14. Hussein, A.K., Abbas,
L.K. and Seger, A.A., 2020. MultiObjective Optimization of Friction Stir
Welding for Aluminum Alloy (2024-T3). Engineering and Technology
Journal, 38(2 part (A) Engineering), pp.185-198.
Chapter 4
Discrete Wavelet Transformation in Friction Stir Welding Process
1. Introduction

Friction Stir Welding joint quality depends on input parameters such as tool
rotational speed, tool traverse speed, tool tilt angle and an axial force. Surface
defects formation occurs when these input parameters are not selected
properly. The main objective of the recent paper is to develop Discrete
Wavelet Transform algorithm by using Python programming and further
subject it to the Friction Stir Welded samples for the identification of various
external surface defects present.

Friction Stir Welding is a solid-state joining process that generally finds

application in the joining of alloys which are difficult to weld by a
conventional welding process. The important input parameters which govern
the quality of weld obtained from the Friction Stir Welding process are tool
rotational speed (rpm), tool traverse speed (mm/min), axial force (KN), and
tool tilt angle. Improper selection of these input parameters during the
Friction Stir Welding process results in the formation of intermetallic
compounds which are responsible for the initiation of crack nucleation and
also there is the formation of various external and internal defects such as
surface grooves, tunnel formation, flash formation and void formation which
are responsible for stress concentration [1-4].

Nowadays, Machine Learning is being radically used in various materials and

manufacturing sectors for the optimization of mechanical and microstructure
properties like Ultimate Tensile Strength, Fracture Strength, Elongation
percentage, Grain size, etc. Unsupervised machine learning classifiers were
used by Kolokas et al. [5] for fault prognosis and forecasting in industrial
equipment related to plastic and aluminium production. The results
concluded that the used machine learning models were capable of predicting
the faults before their occurrence.
Aimiyeakagbon et al. [6 ] used machine learning time series forecasting
approach for prediction of crack length in the riveted aluminium plates.
Mongan et al. [7] combined genetic algorithm (GA) with Artificial Neural
Network (ANN) for predicting the strength of ultrasonically welded joints.
The model resulted high accuracy with 0.9827 as a correlation coefficient.
Likely these applications machine learning is also being used in Friction Stir
Welding process. Dutt et al. [8 ] developed artificial neural network model
for studying about the correlation between the Friction Stir Welding input
parameters such as a rotation rate and traverse rate with the mechanical
property of friction stir welded precipitation strengthened AA7050 aluminum
alloys.

Hartl et al. [9] used Bayesian optimization and reinforcement learning

method to improve the surface quality of friction stir welded joints. The
present research focuses on the implementation of Discrete Wavelet
Transformation for the detection of surface defects present on Friction Stir
Welded joints. Selim et al. [10] applied a wavelet transformation algorithm
for the detection of internal defects in a given aluminum component. Figure 1
shows the wavelet contour map of the three scan points on the metallic
component.
Fig 1. Contour maps of wavelet transform at a) Scan point 1 b) Scan point 2 c) Scan Point 3 [10].

Vermaak et al. [11] used the Dual-Tree Complex Wavelet Transform

algorithm for improving fabric defect detection. Guminiak et al. [12] applied
a discrete wavelet transform in the truss structures with rigidly connected
bars for detecting the defects.

2. Material and Methods

In the present work, Aluminium alloy 6060 T5 plates were joined. The
chemical composition of the base alloy plate is shown in Table 1. The base
alloy plate of the dimensions 150 mm X 100 mm X 6 mm was mounted
tightly on the CNC bed with the help of a fixture. The main purpose of the
fixture is to help both workpieces in a proper grip so that they do not
dislocate from their original position while carrying out the Friction Stir
Welding Process. The tool material for joining the plates is H13.
Table 1: Chemical Composition of 6060-T5 Al alloy in wt %

Alloy Si Fe Cu Mg Mn Cr Zn Ti Ti 0.42 0.17 0.002 0.02 0.45 0.0006 0.0002 0.01 T5

Five Friction Stir Welded samples were prepared at particular input

parameters shown in Table 2. The digital images captured of the welded
samples are shown in Figure 2 – Figure 6.
Table 2: Parameters selected for Friction Stir Welding process

Sample Number Tool Rotational Speed (rpm) 1 2000

2 1000
3 1500
4 1500
5 2000

Tool Traverse Axial Force Speed (mm/min) (KN)

400 2.5
200 1.5
400 1.5
400 2.5
400 1.5
Fig 2. Digital Cropped image of Sample 1.
Fig 6. Digital Cropped image of Sample 5

These digitally captured cropped images were imported to Google

Colaboratory platform for the im-plementation of Discrete Wavelet

Transform algorithm developed by using Python programming language.

3. Results and Discussions

The digital captured cropped image can be considered as a twodimensional

signal s(n) where n is the samples of a given signal and n=0,1,2,…..M-1.
Discrete time signal is obtained by summing up scaling function term
represented by Equation 1 and wavelet function term represented by Equation
2.

(1)
(2)
In Equation 1,
is a scaling function, is a scaling parameter and is a normalizing term
used for converting a spacial domain s(n) to . In Equation 2 it should
be noted that j
. Equation 1 and

2 constitutes Forward discrete wavelet transformation while Equation 3

represents an expression for an inverse discrete wavelet transformation.

s(n) =
(n) (3)
Fig 7. Friction Stir Welded image subjected to Discrete Wavelet Transform algorithm

It is a common fact that filters are one-dimensional in nature while images

are two-dimensional in nature. So, in order to apply onedimensional filter to
the two-dimensional images, we have to apply one-dimensional filter along
rows of images and then along column of images as shown in Figure 7. High
pass filter extract the edges and low pass filter does the approximation. I_LH
extracts the information from the input image and has passed through the
high pass filter which acts on the row of an input image and finally results
horizontal edges. I_LL has passed through two low pass filters and results an
approximation image. I_HL results vertical edges while I_HH extracts the
vertical feature of the image and has passed through high pass filter along the
column of the image. The high pass filter operating along the row of the input
image, so, I_HH emphasizes the edges along the diagonal of the image. The
results obtained as depicted in Figure 8-12.

Fig 8 Result for Sample 1

Fig 9 Result for Sample 2

Fig 10 Result for Sample 3

Fig 11 Result for Sample 4

Fig 12 Result for Sample 5

It is observed from the results that inhomogeneous pixel present in the

obtained transforms depicts external surface defects like flash formation and
groovy edges.

4. Conclusion

The present study implemented the Discrete Wavelet Transform algorithm to

the five friction stir welded samples for detecting the presence of surface
defects such as flash formation, groovy edges etc. as a main objective. The
results showed that the Discrete Wavelet Transform is able to capture and
extract the minute details present on the surface of the weld and further can
be used for defects detection purpose. The future study which be further
carried out on this work is to implement it on real-time monitoring of defects
formation.

References

1. Mishra, R.S. and Ma, Z.Y., 2005. Friction stir welding and processing.
Materials science and engineering: R: reports, 50(1-2), pp.1-78.

2. Thomas, W.M. and Nicholas, E.D., 1997. Friction stir welding for

2. Thomas, W.M. and Nicholas, E.D., 1997. Friction stir welding for
273.

3. Lohwasser, D. and Chen, Z. eds., 2009. Friction stir welding: From basics
to applications. Elsevier.

4. Akinlabi, E.T. and Mahamood, R.M., 2020. Introduction to Friction

Welding, Friction Stir Welding and Friction Stir Processing. In Solid-State
Welding: Friction and Friction Stir Welding Processes (pp. 1-12). Springer,
Cham.

5.Kolokas, N., Vafeiadis, T., Ioannidis, D. and Tzovaras, D., 2020. Fault
Prognostics in Industrial Domains using Unsupervised Machine Learning
Classifiers. Simulation Modelling Practice and Theory, p.102109.

6. Aimiyekagbon, O.K., Bender, A. and Sextro, W., 2020, July. Evaluation of

time series forecasting approaches for the reliable crack length prediction of
riveted aluminium plates given insufficient data. In PHM Society European
Conference (Vol. 5, No. 1, pp. 11-11).

7. Mongan, P.G., Hinchy, E.P., O’Dowd, N.P. and McCarthy, C.T.,

2020. Optimisation of Ultrasonically Welded Joints through Machine
Learning. Procedia CIRP, 93, pp.527-531.

8. Dutt A.K., Sindhuja K., Reddy S.V.N., Kumar P. (2021) Application of

Artificial Neural Network to Friction Stir Welding Process of AA7050
Aluminum Alloy. In: Arockiarajan A., Duraiselvam M., Raju R. (eds)
Advances in Industrial Automation and Smart Manufacturing. Lecture Notes
in Mechanical and Smart Manufacturing. Lecture Notes in Mechanical 15-
4739-3_34
9. Hartl, R., Hansjakob, J. & Zaeh, M.F. Improving the surface

quality of friction stir welds using reinforcement learning and Bayesian

optimization. Int J Adv Manuf Technol 110, 3145–3167 (2020).
https://doi.org/10.1007/s00170-020-05696-x

10. Hossam Selim, Fernando Piñal Moctezuma, Miguel Delgado Prieto, José
Francisco Trull, Luis Romeral Martínez and Crina Cojocaru (April 3rd 2019).
Wavelet Transform Applied to Internal Defect Detection by Means of Laser
Ultrasound, Wavelet Transform and Complexity, Dumitru Baleanu,
IntechOpen, DOI:
10.5772/intechopen.84964. Available from:
https://www.intechopen.com/books/wavelet-transform-
andcomplexity/wavelet-transform-applied-to-internal-defectdetection-
by-means-of-laser-ultrasound

11. Vermaak, H., Nsengiyumva, P. and Luwes, N., 2016. Using the dual-tree
complex wavelet transform for improved fabric defect detection. Journal of
Sensors, 2016.

12 KNITTERPIĄTKOWSKA, Anna; GUMINIAK, Michał Jan;

PRZYCHODZKI, Maciej. Application of Discrete Wavelet Transformation
to Defect Detection in Truss Structures with Rigidly Connected Bars.
Engineering Transactions, [S.l.], v. 64, n. 2, p. 157– 170, feb. 2016. ISSN
2450-8071. Available at:
http://www.entra.put.poznan.pl/index.php/et/article/view/319. Date accessed:
01 Nov. 2020.

Chapter 5
Grain Size Distribution Analysis in Friction Stir Welding Process
1. Introduction

Machine learning has widely spread in the areas of pattern recognition, pre-
diction or forecasting, cognitive game theory and in bioinformatics. In recent
days, machine learning is being introduced into manufacturing and material
industries for the development of new materials and simulating the
manufacturing of the required products. In the recent chapter, machine
learning algorithm is developed by using Python programming for the
determination of grain size distribution in the microstructure of stir zone
seam of Friction Stir Welded magnesium AZ31B alloy plate The grain size
parameters such as an equivalent diameter, perimeter, area, orientation etc.
were determined. The results showed that the developed algorithm is able to
determine various grain size parameters accurately.

The development of various Artificial Intelligence applications is made

possible by the usage of Machine Learning algorithms. In the available set of
a massive amount of data, the Machine learning model uses a statistical
technique to derive the required pattern. The Machine Learning model
generally feeds on digitally available data which is composed of numbers,
images, videos, or texts. Machine Learning algo-rithms generally operate on
low-level tasks in which raw data is processed and is further acquired by an
Artificial Intelligence-based system.
The Machine Learning model generally feeds on digitally available data
which is composed of numbers, images, videos, or texts. Machine Learning
algorithms generally op-erate on low-level tasks in which raw data is
processed and is further acquired by an Artificial Intelli-gence-based system.
The Machine Learning algorithm's working principle is based on three points.
Firstly, a hypothesis of a certain phenomenon is formulated. Secondly, to test
the hypothesis i.e. for the validation of the chosen model, a dataset is
collected. Thirdly, the given hypothesis is refined by iterating it.

Traditional manufacturing processes are being spiced up by the introduction

of various Machine Learning algorithms in their applications. For example,
GE research is combining artificial intelligence and machine learning models
with "physics" to inspection, processing, and advanced manufacturing de-
sign to obtain the real-world product [1].

Li et al used a data-driven machine learning model approach for predicting

the surface roughness of an additively manufactured part. It was observed
that the developed machine learning model was able to predict the surface
roughness with good accuracy [2]. Moreno et al. proposed an image
processing and machine learning model based on the Random Forest
algorithm for automatic classification of pores present in Al-alloys by laser
melt deposition. The proposed model resulted in an accuracy of 94.41 % [3].
Kopper et al. predicted the mechanical property i.e. Ultimate Tensile Strength
(UTS) of highpressure die-cast Al-alloy component by using various
regression machine learning models like Support Vector Machines (SVM),
XG Boost, and Random Forest algorithm [4].
Machine Learning algorithms are also finding applications in solidstate
welding processes like Friction Stir Welding (FSW). Unlike other
conventional welding processes, Friction Stir Welding does not use any filler
material for joining purposes. Friction Stir Welding tool which is harder than
the base material to be joined is used. The schematic representation of the
Friction Stir Welding process is shown in Figure 1 [5]. Verma et al. [6] used
three machine learning models i.e. Artificial Neural Network, Random
Forest, and M5P tree regression for analyzing and predicting the tensile
behavior of aviation-grade aluminum alloy friction stir welded joint. It was
observed from results that the Random Forest algorithm model outperforms
the other two models in terms of predicting the tensile behavior accurately.

Hartl et al. [7] used three Artificial Neural Network models i.e. Convolutional
Neural Networks, Feedforward fully con-nected Neural Networks, and
Recurrent Neural Networks for predicting the surface quality of friction stir
welded EN AW-6082 T6 sheets. Out of these three used Artificial Neural
Networks, Convolutional Neural Networks achieved the highest classification
accuracy. Srinivasan et al. [8] improved the degraded mi-crostructure of the
friction stir welded joint by using a hybrid sparsity machine learning model.
Figure 1. Working Mechanism of Friction Stir Welding process [5].

In the recent study, a machine learning algorithm has been developed for the
study of grain size distri-bution in the microstructure of stir zone seam of
Friction Stir Welded magnesium AZ31B obtained by Subramani et al [9].

2. Materials and Methods

In the experiment, AZ31 B Magnesium alloy plates (commercial) of 6 mm

thickness were joined by Friction Stir Welding process [9]. The weld plates
were in the dimensions of 100 mm length and 75 mm width. SiC
nanoparticles were incorporated and compacted in the grooves of width 0.3,
0.6, 0.9, 1.2, and 1.5 mm, of the corresponding volume percentages were
indicated as V4, V8, V12, V16, and V20, respec-tively. The schematic
diagram of volume fraction sample is shown in the Figure 2.

Figure 2 : Volume fraction sample [9].

The process parameter of the experimentation is shown in the Table 1.
Table 1. Friction stir welding (FSW) process parameters used in this study.
Control Parameters Values

Tool rotational speed, N (rpm) 1250 rpm

Tool travel rate, F (mm/min) 25 mm/min
Volume percentage V (%) 4 (V4), 8 (V8), 12 (V12), 16 (V16), 20 (V20)
Groove width (W) mm 3 and 5.7 mm Tool Pin profile
material
0.3, 0.6, 0.9, 1.2, 1.5 D/d ratio and Pin height in

Cylindrical threaded tool pin profile Tool H13 tool steel

The optical microstructure of the matrix alloy (Figure3a) and base metal
welded joint without nano-SiC (Figure3b) is shown in Figure 3.

Figure 3. Optical microstructure: (a) as-received matrix alloy (b) FSW AZ31B joint (without rein-
forcement) [9].

Grain size distribution analysis is done by loading the above microstructures

which have a whole bunch of grains and extracting some statistical features
out of it. The given RGB image is converted to a grayscale image. It is
important to define a particular scale in this case in order to track the size of
the pixel while working on a microscopic image.

The first step is to read an image and define pixel size state(if needed to
convert results into microns, no pixels). The second step involves the
denoising process if required and threshold image to separate grain from
boundaries. Thirdly, image clean up is performed if needed and create a mask
for observed grains in the microstructure. The fourth step involves the
labeling of grains in the masked image. At the fifth stage, the measurement of
the properties of each grain in the microstructure is performed. At the sixth
and last stage, output results are obtained in the form of a CSV (Comma
Seperated Values) file.

3. Results and Discussions

The important Python programming libraries which were important for

coding purpose are cv2, numpy, pyplot from matplotlib, ndimage, io, color to
show images in the color library so we can assign a different color to grains,
measure library to measure the grain size distribution. The image is read by
indi-cating the file path, then the scale is defined which is 1 pixel=0.5μm
which is further multiplied to every measurements as shown in the Figure 4
and Figure 5. If the image has a scale bar in the microscope then the image
needed to be cropped by slicing the dataset of an image array.

Figure 4. Loaded image of optical microstructure of received matrix alloy

Figure 5. Loaded image of optical microstructure of FSW AZ31B joint (without reinforcement).

Step 1 of the thresholding is to look at the histogram itself as shown in Figure

6 and Figure 7. The loaded images are in form of a 2D array, so in order to
obtain a histogram we need 1D array. So we need to flat-ten the image which
takes a 2D image and flatten it to a 1D array of bins=100 and range = (0,255).
From Figure 6 it is observed that a bunch of pixels lies between 130–200 and
From Figure 7 it is observed that a bunch of pixels lies between 60–200.
Figure 6. Histrogram distribution of optical microstructure of received matrix alloy
Figure 7. Histogram distribution of the optical microstructure of FSW AZ31B joint (without reinforce-
ment).

The two types of thresholding operation can be done i.e. either manual
thresholding or auto thresholding. The threshold value obtained for the
optical microstructure of received matrix alloy is 130 while thresholded value
obtained for the optical microstrucutre of FSW AZ31B joint (without
reinforcement) is 128. The obtained thresholded image for the optical
microstrucutre of received matrix alloy and FSW AZ31B joint (without
reinforcement) is shown in the Figure 8 and Figure 9.
Figure 8. Thresholded image of the optical microstructure of received matrix alloy
Figure 9. Thresholded image of the optical microstructure of FSW AZ31B joint (without rein-
forcement).

It is observed that all pixels corresponding to the grain will have a value of
255 and all pixels correspond-ing to the grain boundary have a value 0. So
this is only a threshold image, not a binary image. So, we need to convert this
thresholded image to a binary image. The thresholded image shows some
missing pixels, we can close these areas of the grain by eroding and dilating
process. When we use the eroding process then the grain will shrink by 1
pixel and when we use dilating process they go up by 1 pixel. In order to
execute eroding and dilating operation, we created kernel size of (3,3) of type
int 8. Eroding and dilating operation is carried out on thresholded image. The
thresholded image is nothing but 8-bit integer with all values of 255 and 0's.
It is a binary image but the system doesn't know its a binary image. So it is
converted to a binary image with the help of masking as shown in the Figure
10 and Figure 11.
Figure 10. Masked image of the optical microstructure of received matrix alloy.
Figure 11. Masked image of optical microstructure of FSW AZ31B joint (without reinforcement).

Each grain is labeled in the masked image as shown in Figure 12 and Figure
13. Structure factor of [[1,1,1], [1,1,1],[1,1,1]] is used to define the nature of
pixel connection i.e. whether connected or discon-nected. In the ndimage
library, there is a function called label which labels the unconnected grains.
Label assignment is done through the masked image and assigns a value to
the all unconnected objects.
Figure 12. Labeled image of the optical microstructure of received matrix alloy
Figure 13. Labeled image of the optical microstructure of FSW AZ31B joint (without reinforcement).

Measuring the grain property is done by the extraction of property from each
labeled image. The obtained measurement of grain size distribution of the
optical microstructure of the received matrix alloy and FSW AZ31B joint
(without reinforcement) is shown in the Table 2 and Table 3. Table 2
obtained as a CSV file has 1296 grain distribution properties of the optical
microstructure of received matrix alloy out of which only 10 is shown in the
Table 2 (rest can be further found in supplementary file) while Table 3 has
696 grain distribution properties out which only 10 is shown in the Table 3
(rest in supplementary file).
Table 2: Grain Size distribution Property of received metal matrix alloy
Table 3: Grain Size distribution Property of FSW AZ31B joint (without reinforcement).
4. Conclusion

The research work focuses on the development of machine learning

algorithm for estimating the grain size distribution property of received
matrix alloy and FSW AZ31B joint (without reinforcement) by us-ing Python
Programming. The various grain size parameters such as equivalent diameter,
orientation, major axis, minor axis, perimeter, area, minimum intensity,
maximum intensity and mean intensity were successfully determined. It is
further observed from the Table 2 and Table 3 that the area, perimeter and
equivalent diameter of the grain decreases after carrying out Friction Stir
Welding process due to the refinement.

So, it can be concluded that the machine learning approach can be easily
incorporated in material science and manufacturing domain in order to reduce
the cost and time of the experiment. It is also observed that there is loss of
information by the observation of two and three dimensional metrics distri-
bution such as aspect ratio, number density, total curvature, interface line and
area density when conventional method is used for microstructure
characterization but the application of machine learning overcomes these
problems.

The future work will deal with implementing this algorithm to the
microstructure obtained by other characterization techniques also some work
will be carried out on the development of a unique microstructure identifier
known as microstructural fingerprint in order to classify micrographs.
Machine learning can be further combined with the advanced microstructural
metrics for improving the classification accuracy.

References

1. Aggour, K.S., Gupta, V.K., Ruscitto, D., Ajdelsztajn, L., Bian, X.,
Brosnan, K.H., Kumar, N.C., Dheeradhada, V., Hanlon, T., Iyer, N. and
Karandikar, J., 2019. Artificial intelligence/machine learning in
manufacturing and inspection: A GE perspective. MRS Bulletin, 44(7),
pp.545-558.

2. Li, Z., Zhang, Z., Shi, J. and Wu, D., 2019. Prediction of surface roughness
in extrusion-based additive manufacturing with machine learning. Robotics
and Computer-Integrated Manufacturing, 57, pp.488-495.

3. García-Moreno, A.I., Alvarado-Orozco, J.M., Ibarra-Medina, J. and

Martínez-Franco, E., 2020. Image-based porosity classification in Al-alloys
by laser metal deposition using random forests. The in Al-alloys by laser
metal deposition using random forests. The 19.

4. Kopper, A., Karkare, R., Paffenroth, R.C. and Apelian, D., 2020. Model
Selection and Evaluation for Machine Learning: Deep Learning in Materials
Processing. Integrating Materials and Manufacturing Innovation, pp.1-14.

5. Pothur Hema (October 24th 2019). Experimental Investigations on AA

6061 Alloy Welded Joints by Friction Stir Welding, Aluminium Alloys and
Composites, Kavian Omar Cooke, IntechOpen, DOI:
10.5772/intechopen.89797. Available from:
https://www.intechopen.com/books/aluminium-alloys-
andcomposites/experimental-investigations-on-aa-6061-alloywelded-
joints-by-friction-stir-welding

6. Verma, S., Misra, J.P. and Popli, D., 2020. Modeling of friction stir
welding of aviation grade aluminium alloy using machine learning
approaches. International Journal of Modelling and Simulation, pp.1-8.
7. Hartl, R., Praehofer, B. and Zaeh, M.F., 2020. Prediction of the surface
quality of friction stir welds by the analysis of process data using Artificial
Neural Networks. Proceedings of the Institution of Mechanical Engineers,
Part L: Journal of Materials: Design and Applications, 234(5), pp.732-751.

8. Srinivasan, K., Deepa, N. and PM, D.R.V., 2020. Realizing the Resolution
Enhancement of Tube-to-Tube Plate Friction Welding Microstructure Images
Via Hybrid Sparsity Model for Improved Weld Interface Defects Diagnosis.
Journal of Internet Technology, 21(1), pp.61-72.

9. Subramani, V., Jayavel, B., Sengottuvelu, R. and Lazar, P.J.L.,

2019. Assessment of microstructure and mechanical properties of stir zone
seam of friction stir welded magnesium AZ31B through nano-SiC. Materials,
12(7), p.1044.

Chapter 6
Prediction of the Mechanical Property of Friction Stir Welded Joint
1. Introduction

In modern computational science, the interplay existing between machine

learning and optimization process marks the most vital developments.
Optimization plays an important role in mechanical industries because it
leads to reduce in material cost, time consumption and increase in production
rate. The recent work focuses on performing the optimization task on Friction
Stir Welding process for obtaining the maximum Ultimate Tensile Strength
(UTS) of the friction stir welded joints. Two machine learning algorithms i.e.
Artificial Neural Network (ANN) and Decision Trees regression model are
selected for the purpose.

The input variables are Tool Rotational Speed (RPM), Tool Traverse Speed
(mm/min) and Axial Force (KN) while the output variable is Ultimate Tensile
Strength (MPa). It is observed that in case of the Artificial Neural Networks
the Root Mean Square Errors for training and testing sets are 0.842 and 0.808
respectively while in case of Decision Trees regression model, the training
and testing sets result Root Mean Square Errors of 11.72 and 14.61. So, it can
be concluded that ANN algorithm gives better and accurate result than
Decision Tree regression algorithm.
According to Andrew Ng who is the co-founder of Google Brain, Artificial
Intelligence (AI) is going to revolutionise the
manufacturing and materials industries. AI finds application in quality
checks, prediction of failure modes, generative design and forecasting output
for the available dataset [1–2]. AI can be defined as a simulation of human
intelligence which possesses the skills of identifying, observing, learning,
performing the task and if needed, it can self-correct itself [3-4]. AI can be
achieved by various numbers of ways as shown in the Fig.1. It can be seen
that areas like machine learning, machine vision, natural language processing,
robotics etc. contributes to the formation of AI [5]. From Fig.2 it is observed
that machine learning field makes an important contribution for the
development of AI.

Fig. 1. Representation
of number of ways AI can be achieved [6]
 Fig. 2. Classification
architecture of AI [6]

Machine learning aims to design, understand, and apply computer programs

that learn from the past experience i.e. data for the purpose of modelling,
prediction or control. Machine learning finds application in the development
of new type of material, also it involves high-throughput density functional
theory for the calculations of materials properties [7].
Lu et al. [8] worked on the development of a general machine learning
framework for the extraction of elastoplastic properties of engineering alloys
with strikingly improved training efficiency and accuracy. Lilienfeld et al. [9]
introduced machine learning models for a sample prediction for proton and
forces on atoms, excitations at atomic core level and carbon nuclear chemical
shifts.
Machine Learning finds application in Friction Stir Welding process which is
a solid-state welding technology. Friction Stir Welding process consists of
base alloy plates to be joined, welding tool whose material is harder than the
alloy plates to be joined and a fixture design for holding the work piece
together. In the present work, two machine learning algorithms i.e. Artificial
Neural Network and Decision Tree regression model have been used for the
prediction of Decision Tree regression model have been used for the
prediction of T6 aluminium alloy joints.

2. Working principle of Artificial Neural Network (ANN) and Decision

Tree Regression

Perceptron is the simplest neuron which consists of an output layer and an

input layer. This neuron contains two types of functions i.e. summation
function and activation function. It is observed from the Fig. 3 shows that the
summation function processes the inputs incorporated into the perceptron and
then it is further subjected to the activation function in order to get the output.

Fig. 3. Design of Simple Perceptron model [10]

The main disadvantage of simple perceptron model is its incapability to

process huge datasets and multiple inputs. In order to overcome this issue, we
should keep increasing the number of neurons. Fig. 4. shows a basic artificial
neural network architecture consisting of an output layer, hidden layers and
an input layer. It should be noted that the artificial neural network can have
multiple hidden layers, but it should have single output layer and input layer.
The main objective of the activation function is to achieve an output by the
conversion of the weighted sum of input signals.
 Fig.4.
Architecture of Artificial Neural Network [10]

The working principle of an Artificial Neural Network is generally based on

the two types of propagation methods i.e. forward propagation and backward
propagation. Fig. 5. explains the working principle of forward propagation
method.

Fig. 5. Working principle of Forward Propagation method

are the features of a sample in a given dataset. In order to predict the
outcome, these features are subjected to some
mathematical operations. are the weights which are associated with each
features. These weights together serve the neuron as an input. At the
summation stage, all available features are multiplied by their alotted weights
and a bias is added further as shown in the equation 1.

Y = (1)

Activation function is further subjected to the above summed up function.

Weight W3 is multiplied with the output of the former neuron and it further
behaves as the input for an output layer. The backpropagation method is used
to update the weights after each initiation.
Fig. 6. shows the architecture of the backpropagation method. The main
objective is to make the error minimum which can be achieved

by finding the changing in weights and also calculating and .

Fig. 6. Working principle of backward propagation method

Summation (S) = (2)
Activation = sigmoid = (3)
By using chain rule, we can find change in weights,
= (4)

= (5) Senthilnathan et al. [11] used Artificial Neural Network for predicting
the micro hardness and tensile strength of friction stir predicting the micro
hardness and tensile strength of friction stir T6. Mishra et al. [12] predicted
the corrosion potential of friction stir welded joints by using back propagated
artificial neural network. Hartl et al. [13] used artificial neural networks for
the surface quality of friction stir welded joints. Rehim et al. [14] predicted
and simulated the Vickers hardness of friction stir welded joints by using
artificial neural network. From these researches it is observed that Artificial
Neural Network has a wide application in Friction Stir Welding process.

Decision Tree regression algorithm is a non-continuous model which falls

into non-linear regression category. It can be thought of as a supervised
learning method which is non-parametric in nature and can be both used for
regression and classification. Decision tree operates on if-then- else decision
rules which learn from the given dataset and do better approximation. It
should be noted that deeper the decision tree is the more fit model we get.
The regression models in the decision tree are in the form of a tree structure.
Fig. 7 shows the working criteria of decision tree algorithm.

Fig. 7.
Schematic representation of decision tree [15]

Bozkurt et al. [16] used decision tree regression algorithm for the prediction
of the tensile property of friction stir welded
AA2124/SiC/25p metal matrix composite (MMC) plates. It was observed that
the decision tree algorithm can be applied with a greater accuracy. Du et al.
[17] investigated about the condition of void formation using decision tree
classification model and artificial neural network. It was concluded that both
Artificial Neural Network and decision tree model were able to predict the
void formation with 96.6 % accuracy. In the next section, the experimental
procedure dealing with the present work will be discussed.

3. Experimental Procedure
Firstly, 85 datasets were prepared from the previous research carried out on
the friction stir welded 6061-T6 aluminium alloys [18–21]. In the prepared
dataset, tool rotational speed (rpm), tool traverse speed (mm/min) and axial
force (KN) are the input variables while the Ultimate Tensile Strength (MPa)
is the output variable. Secondly, the Python programming is used for
developing the code for the execution of Artificial Neural Network and
Decision Tree regression algorithm. Fig. 8. shows the methodology for the
Artificial Neural Network.

Fig. 8. Dataset
subjected to Artificial Neural Network algorithm

Tool rotational speed ranged between 500 rpm to 2000 rpm and on other
hand tool traverse speed ranged between 20 mm/min to 150 mm/min. Fig. 9
shows the Artificial Neural Network architecture used in the present study.
The input layer consisting of tool rotational speed (rpm), tool traverse speed
(mm/min) and axial force (KN) constitute the input nodes. The first hidden
layer has 4 nodes which is subjected to ReLu activation function, the second
hidden layer has 3 nodes which is also subjected to ReLu activation function
and the output layer has one node i.e. Ultimate Tensile Strength which is
subjected to sigmoid activation function. The training was performed on 63
datasets and test was performed on 22 sample datasets.
Fig. 9. Artificial Neural Network architecture for the prediction of UTS of friction stir welded alloy

For subjecting the dataset to decision tree regression, firstly the python
libraries like Numpy, Matplotlib and Pandas were imported. Secondly, the
datasets were imported and the whole dataset was trained on decision tree
regression model. Thirdly, the new result was predicted, and the results were
visualised.

4. Results and Discussions

The Artificial Neural Network (ANN) model was fit with stochastic gradient
descent with a learning rate of 0.01 and momentum of 0.9. Dataset training
was performed for 800 epochs and test set was evaluated at the end of each
epoch.

Fig. 10 shows the plot of loss function for both training and testing dataset. It
is observed that the loss function for both testing and training set decreases
with increase in number of epochs. Fig. 11 shows the contour plot obtained
between the input variables i.e. Tool traverse speed (mm/min), the tool
rotational speed (rpm) and the output variable i.e. Ultimate Tensile Strength
(MPa). Fig. 12 shows the contour plot obtained between the input variables
i.e. axial force (KN), tool rotational speed (rpm) and the output variable i.e.
Ultimate Tensile Strength (MPa).

Fig. 13 shows the contour plot obtained between in the input variables i.e.
axial force (KN), tool traverse speed (mm/min) and the output variable i.e.
Ultimate Tensile Strength (MPa). After the end of epoch, it is observed that
for the Artificial Neural Networks, the Root Mean Square Errors for training
and testing sets are 0.842 and 0.808.

Fig. 10. Plot of loss function wrt epoch

Fig. 11. Contour plot of UTS wrt Tool Traverse Speed (mm/min) and Tool Rotational Speed (rpm)
Fig. 12. Contour plot of UTS wrt Axial Force (KN) and Tool Rotational Speed (rpm)

Fig. 13. Contour plot of UTS wrt Axial Force (KN) and Tool Traverse Speed (mm/min)

Decision trees regression model are bit more complex than classification tree
model. Once we execute the regression tree algorithm, the scatter plot
obtained will be split into many segments as shown in the Fig. 14 and Fig.
15.

An algorithm called Information Entropy governs how and where these splits
are conducted. So, it basically means that when the split is performed it is
checked whether the split is increasing the amount of information that we
have about our points or is it actually it is adding some value we want to
group our points. The algorithm can well handle the optimal splits of the
dataset into these leaves and the final leaves are called terminal leaves.
 Fig. 14. Decision
Tress Regression plot between UTS and Tool traverse speed (yaxis), Rotational Speed (x-axis)

Fig.15. Decision
Tress Regression plot between UTS and Tool traverse speed (xaxis), Rotational Speed (y-axis)

It is observed that the training and testing sets in decision tree regression
algorithm result Root Mean Square Errors of 11.72 and 14.61.

5. Conclusion
Machine Learning applications in materials and manufacturing industries are
found to be ground breaking. The following conclusions can be drawn from
the present work:

• Artificial algorithms i.e. Artificial Neural Network (ANN) and decision tree
regression models are successfully implemented in the present work.

• It is observed that in ANN, loss function decreases continuously with the

increase in number of epochs.

• It is also observed that the root means square error obtained in ANN is far
more less than those obtained by implementing decision tree regression
model.

• The accuracy of both of the algorithms can be achieved if the number of

data set is increased.
• The cost and time of the experiment is further reduced by implementing
machine learning techniques.
References

1. Russell S.J., Norvig P., Artificial Intelligence: A Modern

ApproachRussell, S. J., & Norvig, P. (2010). Artificial Intelligence: A
Modern Approach. Artificial Intelligence.
https://doi.org/10.1017/S0269888900007724.Artificial Intelligence. 2010.

2. Fogel D.B., Defining Artificial Intelligence.In: Evolutionary

Computation, 2006.
3. Ramos C., Augusto J.C., Shapiro D., Ambient intelligencethe next step for
artificial intelligence. IEEE Intelligent Systems, 2008, Vol. 23(2), 15–8.
https://doi.org/10.1109/MIS.2008.19

4. Ghahramani Z., Probabilistic machine learning and artificial intelligence.

Nature, 2015, Vol. 521, 452–9.
https://doi.org/10.1038/nature14541

5. Ng A., What artificial intelligence can and can’t do right now. Harvard
Business Review Digital Articles, 2016,(2–4).
6. Chethan K.G., Artificial Intelligence: Definition, Types,
Examples,Technologies.https://doi.org/https://medium.com/@chetha
nkumargn/artificial-intelligence-definition-types-examplestechnologies-
962ea75c7b9b

7. Meredig B., Five High-Impact Research Areas in Machine Learning for

Materials Science. Chem Mater, 2019, Vol. 31(23),
9579–81. https://doi.org/10.1021/acs.chemmater.9b04078

8. Lu L., Dao M., Kumar P., Ramamurty U., Karniadakis G.E., Suresh S.,
Extraction of mechanical properties of materials through deep learning from
instrumented indentation. Proceedings of the National Academy of Science
of the USA, 2020, Vol. 117(13), 7052–62.
https://doi.org/10.1073/pnas.1922210117

9. Rupp M., Ramakrishnan R., Von Lilienfeld O.A., Machine Learning for
Quantum Mechanical Properties of Atoms in Molecules. Journal of Physical
Chemistry Letters, 2015, Vol. 6(16), 3309–13.
https://doi.org/10.1021/acs.jpclett.5b01456

10. Gajawada S.K., The Math behind Artificial Neural Networks.2019.

https://doi.org/https://towardsdatascience.com/the-heart-ofartificial-
neural-networks-26627e8c03ba

11. Senthilnathan T., Sujay Aadithya B., Balachandar K., Prediction of

mechanical properties and optimization of process parameters in friction-stir-
welded dissimilar aluminium alloys. World Journal of Engineering, 2020,
Vol. 17(4), 519–26.
12. Mishra A., Tiwari A., Dubey N.K., Machine Learning Approach to
Determine Corrosion Potential of Friction Stir Welded Joints. Journal of
Image Processing & Pattern Recognition Progress, 2020, Vol. 7(1), 5–17.

13. Hartl R., Praehofer B., Zaeh M.F., Prediction of the surface quality of
friction stir welds by the analysis of process data using Artificial Neural
Networks. Proceedings of the Institution of Mechanical Engineers, Part L:
Journal of Materials: Design and Applications, 2020, Vol. 234(5), 732–51.
https://doi.org/10.1177/1464420719899685
14. Abd El-Rehim A.F., Zahran H.Y., Habashy D.M., Al-Masoud H.M.,
Simulation and prediction of the vickers hardness of AZ91 magnesium alloy
using artificial neural network model. Crystals, 2020, Vol. 10(4), 290.
https://doi.org/10.3390/cryst10040290

15. Decision Tree Introduction with example.

https://doi.org/https://www.geeksforgeeks.org/decision-treeintroduction-
example/

16. Bozkurt Y., Kentli A., Uzun H., Salman S., Experimental Investigation
and Prediction of Mechanical Properties of Friction Stir Welded Aluminium
Metal Matrix Composite Plates. MATERIALS SCIENCE
(MEDŽIAGOTYRA), 2012, Vol. 18(4), 336–40.
https://doi.org/10.5755/j01.ms.18.4.3092

17. Du Y., Mukherjee T., DebRoy T., Conditions for void formation in
friction stir welding from machine learning. npj Comput Mater,
2019, Vol. 5, 68. https://doi.org/10.1038/s41524-019-0207-y

18. Hema P., Experimental Investigations on AA 6061 Alloy Welded Joints

by Friction Stir Welding. In: Cooke KO, editor. Aluminium Alloys and
Composites, IntechOpen Limited, 2019.
https://doi.org/10.5772/intechopen.89797

19. Elatharasan G., Kumar V.S.S., An experimental analysis and optimization

of process parameter on friction stir welding of AA
6061-T6 aluminum alloy using RSM. Procedia Engineering, 2013, Vol. 64,
1227–34. https://doi.org/10.1016/j.proeng.2013.09.202 20. Netto N.,
Tiryakioglu M., Eason P., Characterization of Tool Degradation during
Friction Stir Processing of 6061-T6 Aluminum Alloy Extrusions. Preprints,
2018, 2018080286.
https://doi.org/10.20944/preprints201808.0286.v1

21. Amirabadi H., Bani Mostafa Arab N., Safi S.V., A Novel Approach for
Improving Mechanical Properties in Friction Stir Butt Welded AA6061
Aluminum Plates by Using Preheat. Preprints, 2020, 2020020183.