Research Progress in Mechanical and Manufacturing Engineering Vol. 3 No.

1 (2022) 875-883
© Universiti Tun Hussein Onn Malaysia Publisher’s Office

RPMME
Homepage: http://publisher.uthm.edu.my/periodicals/index.php/rpmme
e-ISSN : 2773-4765

Application of Machine Learning Algorithm to

Prediction of Thermal Spring Back of Hot Press
Forming
Mah Wei Chun1, Ng Chuan Huat2*, Ong Pauline1
1
Faculty of Mechanical and Manufacturing Engineering
Universiti Tun Hussein Onn Malaysia (UTHM), Parit Raja, 86400, MALAYSIA

*Corresponding Author Designation

DOI: https://doi.org/10.30880/rpmme.2022.03.01.093
Received 01 Dec 2021; Accepted 01 April 2022; Available online 30 July 2022

Abstract: Some steels are very difficult to fabricate using cold forming, which is a
conventional sheet metal forming method. As a result, hot stamping is one of the ways
utilized to manufacture components made of advanced high strength steel (AHSS).
Although hot press sheet can form the high-strength steels, but it also can cause to
thermal springback defects. In this paper, thermal spring back simulation data is used
to examine and predict the springback condition by using machine learning algorithm.
The focus of this research is to determine which machine learning model performs
best for thermal springback predictions. To forecast thermal springback, three
machine learning techniques were used in this paper: the KNN algorithm, the DT
algorithm, and the SVR algorithm. Furthermore, the predicting errors of these three
models are compared. The compared results indicate that the Decision Trees model
can properly forecast and capture thermal springback variation trends.

Keywords: Hot Press Forming, Machine Learning, Thermal Springback,

Supervised

1. Introduction
The Industry 4.0 is an important transformation by synthesize the digital and internet technologies
with conventional industry. Increase the flexibility and resource efficiency through digitization is the
aim of an industry to develop products faster. Cyber-physical systems monitor actual processes, create
a digital replica of the actual environment, and make decentralised decisions in intelligent factories
enabled by Industry 4.0. Also, they cooperate in real time with human resources and each other by using
Internet of Things (IoT). The algorithms integrated with users monitored the Cyber-physical systems
via Internet [1]. The transition to the Fourth Industrial Revolution, sometimes known as "Industry 4.0"
is pervasive today and has had a significant impact on the manufacturing industry [2].
Hot stamping forming is a sheet metal forming technique that involves heating a blank before
producing it. The young’s modulus and yield limit are decreased by the elevated temperature while

*Corresponding author: [email protected]

2022 UTHM Publisher. All rights reserved.
publisher.uthm.edu.my/periodicals/index.php/rpmme
Mah Wei Chun et al., Research Progress in Mechanical and Manufacturing Engineering Vol. 3 No. 1 (2022) p. 875-883

enhance ductility in comparison to the material's cold condition. The blank is heated to 900 °C during
the heating state and held at this degree until it is totally austenite. For the utilised sheet thickness, it
took at least 5 minutes at 900 °C. The Figure 1.1 shown that the blank is then moved and formed by a
hydraulic press, which employs a die, a blank holder, and a punch. When the forming is completed, the
component is hardened using contact pressure from cooler tools and air cool to room temperature [3].

Figure 1.1: Hot stamping process

1.1 Problem Statement

Some steels are extremely hard to construct using cold forming, which is a standard procedure for
sheet metal forming. Therefore, hot stamping is one of the methods used for producing parts of
advanced high strength steel (AHSS). Although the hot stamping only has very small spring back.
However, it can also be seen those non-symmetric sections with specialised tempering, such as A-pillars
on an automobile, have a substantially bigger spring back during the cooling stage [3]. Because there is
more springback following hot press sheet forming of high-strength steels, the evolution of sheet metal
forming technologies is required in the future [4].

1.2 Objectives

1. To investigate the thermal spring back behaviour between the blank and dies under
forming conditions.
2. To develop several machine learning model by using design of experiments with the
forming parameters for extracts the thermal spring back image.
3. To compare the performance of machine learning models for monitors the thermal spring
back in hot press forming process.

2. Materials and Methods

The materials and methods section, often known as methodology, covers all of the information
required to produce the study's results.

2.1 Simulation
Deform software is used to create a simulation of hot press forming. The parameters utilised to run the
simulation include preheating temperature, blank thickness, punch and die temperature, cooling water
temperature, punch velocity, displacement between punch and blank, forming time, die holding time,
and punch lifting time.

2.2 Data Preparation

This section will be focusing on handling the data quality issues, data filtration, feature extraction and
etc, before they are ready to be feed into the machine. This is to make the quality of the data as high as

876
 Mah Wei Chun et al., Research Progress in Mechanical and Manufacturing Engineering Vol. 3 No. 1 (2022) p. 875-883

possible and minimize the errors or inaccuracies that might occur later. The dataset after EDA is shown
in Figure 2.1.

Figure 2.1: Dataset after EDA

877
Mah Wei Chun et al., Research Progress in Mechanical and Manufacturing Engineering Vol. 3 No. 1 (2022) p. 875-883

2.3 Machine Learning Models

In addition to the project implementation, the technique used for the three selected machine learning
models will be described.

2.3.1 Model 1 (K-Nearest Neighbors)

The KNN algorithm presumes that comparable objects exist nearby. In other words, comparable objects
are close together. KNN algorithm captures the concept of similarity (also known as distance,
proximity, or closeness) with certain mathematics we may have learned in primary school, such as
calculating the distance between points on a graph. There are various methods for determining distance,
and depending on the task at hand, one method may be preferred. However, the straight-line distance
(also known as the Euclidean distance) is a famous and well-known option.

2.3.2 Model 2 (Decision Trees)

Decision Tree is a Supervised learning approach that is being used to solve both classification and
regression issues, however it is most commonly used to solve classification problems. Decision Trees
categories examples by sorting them along the tree from the root to some leaf node, with the leaf node
providing the example's categorization. This algorithm checks the values of the root property with the
values of the record (actual dataset) attribute and then follows the branch and jumps to the next node
depending on the comparison. After that, the algorithm will repeat the comparison of the attribute value
and other sub-node and move further. This step will continue until the algorithm meets the leaf node
(results).

2.3.3 Model 3 (Support Vector Regression)

The SVR operates on the same basis as the SVM. The aim of SVR is to locate the best fit line among
all the provided data, or more specifically, the hyperplane with the greatest number of data points.
Unlike other regression models that seek to minimize the difference between the real and predicted
values, SVR simply tries to fit the best line within a threshold value, which is effectively the distance
between the hyperplane and the boundary line. The complexity of SVR's fit time rises more than
quadratically with the number of samples, making it difficult to scale to datasets with more than a few
tens of thousands of samples.

3. Results and Discussion

3.1 K-Nearest Nrighbors model

MATLAB built-in function (fitcknn) is utilized. The optional parameters of command (fitcknn)
"NumNeighbors" are then set to 3, and the variable 'cond' (condition of springback) in dataset is set as
the output variable. Figure 3.1 shows the detail information of K-NN model. The accuracy of the KNN
model is 90%, training error is 0.26923, and testing error is 0.096154 as shown in Figure 3.2.

878
 Mah Wei Chun et al., Research Progress in Mechanical and Manufacturing Engineering Vol. 3 No. 1 (2022) p. 875-883

Figure 3.1: Detail information of KNN model

Figure 3.2: Accuracy, Training error and Testing error of KNN model

3.2 Decision Trees model

The decision tree model is created in MATLAB using the built-in command 'fitctree'. We don't use any
parameters in this model, simply input and output variables to train it. The coding and detail information
are shown in Figure 3.3. As shown in Figure 3.4, the DT model's accuracy is 90%, its training error is
0.076923, and its testing error is 0.086957. For Decision Trees algorithm, we can view the Decision
Tree diagram as shown Figure 3.6 in by using the coding in Figure 3.5. From the Tree Diagram, we can
see that how the Decision Trees algorithm make decision and predict the outputs.

Figure 3.3: Detail information of Decision Trees model

879
Mah Wei Chun et al., Research Progress in Mechanical and Manufacturing Engineering Vol. 3 No. 1 (2022) p. 875-883

Figure 3.4: Accuracy, Training error and Testing error of Decision Trees model

Figure 3.5: Coding to show the Decision Tree diagram in MATLAB

Figure 3.6: Tree diagram to predict the springback condition

880
 Mah Wei Chun et al., Research Progress in Mechanical and Manufacturing Engineering Vol. 3 No. 1 (2022) p. 875-883

3.3 Support Vector Regression model

In MATLAB, the SVR model is developed using the built-in programme 'fitcecoc.' The parameters of
the SVR algorithm cannot modify directly. To do so, we must first construct a template with the updated
parameters. Then, using the built-in parameters 'Learners,' assign the template just created as shown in
Figure 3.7. The detail information of SVR model is shown in Figure 3.8. The SVR model's accuracy is
80%, its training error is 0, and its testing error is 0.21053 as shown in Figure 3.9.

Figure 3.7: Coding to develop SVR model

Figure 3.8: Detail information of SVR model

Figure 3.9: Accuracy, Training error and Testing error of SVR model

3.4 Comparison between models

After finishing all three machine learning models, they were compared to see how accurate they were.
Based on the results of the accuracy tests, one of the models must be chosen as the best match with the
most dependable projected responses. The evaluation matrix for all three machine learning models is
shown in Table 3.1.

Table 3.1: Evaluation matrix of all machine learning models

Name of Model Accuracy (%) Training Error Testing Error Rating

881
Mah Wei Chun et al., Research Progress in Mechanical and Manufacturing Engineering Vol. 3 No. 1 (2022) p. 875-883

KNN 90 0.26923 0.09615 2

DT 90 0.07692 0.08696 1
SVR 80 0 0.21053 3

By observing Table 3.1, Decision Trees model (DT) was discovered to have the least level of
Training and Testing error, followed by K-Nearest Neighbors model (KNN) to the Support Vector
Regression model (SVR). As a consequence, the Decision Trees algorithm is regarded as the best model
among all since it can produce the best predicted results that are the closest to the actual results.

4. Conclusion
This research aims at the use of machine learning in predicting springback in hot sheet metal
forming. According to the research, Decision Trees might be a viable alternative to the K-NN and SVR
algorithms for predicting springback errors. In the experiment, the Decision Trees technique predicted
springback substantially better than the SVR algorithm and somewhat better than the KNN algorithm.
The superior performance of Decision Trees over KNN and SVR can be due to the following factors:

1. When compared to the other two algorithms, the Decision Trees approach can be trained
effectively even with a small amount of dataset. For Decision Trees algorithm, Once the
variables have been created, there is less data cleansing necessary. So that, missing values and
outliers have less impact on the data in the decision tree.

2. When compared to the K-NN and SVR algorithms, the Decision Trees algorithm has less
parameters and even doesn't require them. The accuracy of the Decision Trees model is
substantially better than the SVR model, as shown in the research, despite the fact that the
SVR model contains more modified parameters.

3. The use of a training dataset to proceed with KNN and SVR training necessitates a high level
of knowledge and care. As observed in the results, even though the SVR model's training
error is zero, the predicted outcome is fully 'Normal,' showing that it is over-fitting to the
training dataset.

We may deduce from the three factors mentioned above that incorrect parameter selection might
result in either over-fitting or under-fitting of the training phase. Because the Decision Trees method
has better ability of prediction of springback defects over the KNN and SVR algorithms, future research
will look at improving the Decision Trees algorithm to obtain better generalization performance.

Acknowledgement
The authors would like to thank the Faculty of Mechanical and Manufacturing Engineering,
Universiti Tun Hussein Onn Malaysia for the support in accomplishing this research.

References

[1] Stăncioiu, A., The fourth industrial revolution „Industry 4.0”. Fiabilitate Şi
Durabilitate, 2017. 1(19): p. 74-78.

882
 Mah Wei Chun et al., Research Progress in Mechanical and Manufacturing Engineering Vol. 3 No. 1 (2022) p. 875-883

[2] Ong, P., W.K. Lee, and R.J.H. Lau, Tool condition monitoring in CNC end milling
using wavelet neural network based on machine vision. The International Journal of
Advanced Manufacturing Technology, 2019. 104(1): p. 1369-1379.

[3] Lugnberg, M. and T. Netz, Investigation of thermal spring back of a hot formed
22MnB5 A-pillar with tailored properties. 2016.

[4] Yanagimoto, J., K. Oyamada, and T. Nakagawa, Springback of High-Strength Steel

after Hot and Warm Sheet Formings. CIRP Annals, 2005. 54(1): p. 213-216.

883