Professional Documents
Culture Documents
Machine Learning For Friction Stir Welding Process
Machine Learning For Friction Stir Welding Process
Welding Process
Akshansh Mishra,
PREFACE
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
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
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
4.3
Reinforcement Learning
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]
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
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.
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
Chapter 2
Image Processing in Friction Stir Welding Process
1. Introduction
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.
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.
Fig.6.Grayscale Image
Fig.7.Histogram of the grayscale 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.
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.
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.
Chapter 3
Supervised Machine Learning in Friction Stir Welding
1. Introduction
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
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
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.
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.
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.
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.
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.
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 %
(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
s(n) =
(n) (3)
Fig 7. Friction Stir Welded image subjected to Discrete Wavelet Transform algorithm
4. Conclusion
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.
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.
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.
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.
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].
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].
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.
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
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.
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.
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.
Chapter 6
Prediction of the Mechanical Property of Friction Stir Welded Joint
1. Introduction
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]
Y = (1)
= (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.
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.
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. 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 also observed that the root means square error obtained in ANN is far
more less than those obtained by implementing decision tree regression
model.
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
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
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
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
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
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.