The 9th Asian Symposium on Visualization

Hong Kong, 4-8 June, 2007

A Prediction on Rubber Product Shrinkage

in Compression Moulding Using Neural Network
W. Chookaew 1*, S. Rodkwan 1,2 and N. Ranong 3
1. Center of Excellence in Rubber Mould, Research and Development Institute of Industrial Production
Technology, Faculty of Engineering, Kasetsart University, Bangkok, Thailand. *E-mail: fengwpc@ku.ac.t h
2. Department of Mechanical Engineering, Faculty of Engineering, Kasetsart University,
Bangkok, Thailand. E-mail: fengssr@ku.ac.th
3. Postharvest and Processing Research and Development Office, Department of Agriculture,
Ministry of Agriculture and Cooperative, Thailand. E-mail: nuchanatn@yahoo.com

Corresponding author W. Chookaew

Abstract

One of the common problems found in rubber compression moulding is the shrinkage of rubber
products leading to the loss of shape. In this research, an application of artificial neural network in
prediction of the rubber product shrinkage in compression moulding is presented. A back propagation
neural network was developed to determine the shrinkage based on the variables of the rubber
compound, processing variables like mould temperature and mould sizing variables. The neural
network prediction for the inside diameter shrinkage and the cross sectional diameter shrinkage
indicate that architectures 5-11-21-1 and 5-11-16-1 provide a prediction within 95.9% and 96.1%
accuracy, respectively.

Keyword: Neural Network, Shrinkage, Rubber Compression Moulding

1. Introduction

Generally, rubber products can be manufactured through various processes such as moulding,
calendaring and extrusion. For rubber moulding, it can be further categorized as compression, transfer
and injection mouldings. Among those processes, a compression moulding is a simplest, most
economical and widely used technique to produce a variety of parts such as rubber bushes, o-rings for
automotive, electrical, and consumer applications. Nevertheless, there are several deficiencies related
to the quality of finished rubber moulded products obtained from compression moulding process. It
has been found that the shrinkage of rubber compression moulded products often causes the loss of
their shapes and dimensions. Consequently, in this research, an application of neural network in
prediction of the rubber product shrinkage in compression moulding is investigated in order to reduce
the product defects due to the shrinkage.

2. Neural Network Models

A neural network is an attempt to model the function of biological neural systems to simulate the
highly complicated process and to predict a nonlinear behavior of any physical systems. The back
propagation can be used to train a multi-layer network with and Levenberg-Marqurdt algorithm due to
its simplification and less time [1]. The input of the neural network should be between 0 and 1. Thus,
before conducting data to train the network, the data is needed to be normalized by dividing by two
maximum absolutes [2]. Since there is no certain rule for determining the architecture of the neural
network for each application, various architectures should be examined to determine the proper
selection. A network with only one hidden layer which has numbers of neurons in hidden layer no
greater than half of the sum of neurons in the input and output layers is examined first [3]. In addition,
the number of neurons should not be greater than a double of number of neurons in input layer plus
one [2]. If the network does not converge to the acceptable error, more complicated network can be

ASV0056-002 ASV0056-002-1
 W. Chookaew, S. Rodkwan and N. Ranong

selected. Four layers or more are not widely used to trained network [1]. Training begins with arbitrary
values of weights, and proceeds iteratively. Each iteration is called an epoch. In each epoch, the
network adjusts the weight in the direction that reduces error. As the examples are processed, each
output neurons and each hidden neurons accumulates the net desire change in each of its own weights.
At the end of epoch, all the hidden neurons and the output neurons change each of their weights by the
appropriate amount. This process is repeated iteratively until the weights reach their optimal values.
The neural network architecture is shown in Figure 1. This research was carried out using the neural
network model to predict the product shrinkage; o-rings in this case. Figure 2. shows the compression
mould for rubber o-rings. The input used in this model can be classified as the additives for rubber
compound, the processing condition of moulding, the product shape and dimension [4, 5, 6]. The
outputs of the network in this work are an inside diameter shrinkage and a cross sectional diameter
shrinkage as shown in Figure 3.

Fig. 1. A neural network architecture.

Fig. 2. A compression mould for rubber o-rings.

Fig. 3. A geometry definition of a rubber o-ring.

9th Asian Symposium on Visualization, Hong Kong SAR, China, 2007. 5-2
 A Prediction on Rubber Product Shrinkage in Compression Moulding using Neural Network

2. Materials and Methods

Rubber o-rings were used as a case study of rubber compression molded product. Five independent
variables leading to the product shrinkage were chosen. These variables are the amount of sulphur, the
amount of carbon black, the mould temperature, an inside diameter and a cross sectional diameter. The
experiments were carried out in three steps. In the first step, the operating range for each variable was
determined by varying its value; maximum and minimum values. The operating range for each
variable is provided in Table 1. The two level factorial design (2k factorial design) was used in this
work. The random order method was used in the test to avoid any biases in the collected data. For the
second step, all five independent variables which test with a two level design of experiment were then
analyzed statistically to determine the main and interaction effects. In the third step, all variables were
tested again with three levels as shown in Table 2. The data set consisting of 243 cases were used in
training of a multi-layer perceptron neural network. The other 49 cases were used in testing the
network. In this research, two networks are used separately to predict the inside diameter shrinkage
and the cross sectional diameter shrinkage, respectively.

Table 1. Operating range of control parameters used in two level factorial design.
parameter sulphur, carbon mould inside dia., cross section
phr black, temperature, mm dia.,
o
range phr c mm
low 0.5 5 140 5 2
high 2.5 50 160 45 8

Table 2. Operating range of control parameters used in neural network training.

parameter sulphur, carbon mould inside dia., cross section
phr black, temperature, mm dia.,
o
range phr c mm
low 0.5 5 140 5 2
medium 1.5 27.5 150 25 4
high 2.5 50 160 45 8

3. Results and Discussions

Two level factorial design was firstly used and it is found that the amount of sulphur, the amount of
carbon black and the mould temperature have effects on the shrinkage of the o-rings as shown as the
main effect relations in Figure 4. As the amounts of sulphur and the mould temperature are increased,
the shrinkage ability is high. In contrast, as the amounts of carbon black are risen, the shrinkage ability
is low. Generally, sulphur involves in the complicated chemical reaction of vulcanization process
which the cross link are created during the process of reaction resulting in various molecular types in
rubber products such as Mono-, Di-, Tri- or Poly-Sulphide. As the content of sulphur is added in the
compound, the force in the molecular chain is risen. As a result, the shrinkage of the rubber product
can occur. Since the coefficient expansion of carbon black is lower than one in rubber, when the
amount of carbon black is increased, the rubber particle will be replaced by additional carbon black
with lower expansion coefficient. Therefore, more carbon black in the rubber part can lead to the
possibility of the shrinkage. When the mould temperature is increased, the difference temperature
between mould and rubber compound is larger. Therefore, the shrinkage of rubber product can be
found. For results on mould sizing variables, it is shown in Figure 5 that an inside diameter has an
effect on the shrinkage but the cross sectional diameter does not play a major role on product
shrinkage. In addition, it is found that the shrinkage varies inversely on the mould sizes. Subsequently,
a one hidden layer model of neural network reveals that the network 5-11-1 provides the most
accuracy of o-ring shrinkage prediction as shown in Table 3. For a two hidden layer model, Table 4, 5
show that the network 5-11-21-1 gives the best prediction for an inside diameter shrinkage while the
network 5-11-16-1 leads to the optimum prediction for a cross sectional diameter shrinkage of o-rings.

9th Asian Symposium on Visualization, Hong Kong SAR, China, 2007. 5-3
 W. Chookaew, S. Rodkwan and N. Ranong

Main Effects Plot (data means) for Shrinkage (%)

Sulphur Carbon Black

3.4

3.2

Mean of Shrinkage (%)

3.0

0.5 2.5 5 50
Mould Temperature

3.4

3.2

3.0

140 160

Fig. 4. The main effect of an experiment

for determining compound variables and process variables.
Main Effects Plot (data means) for Shrinkage (%)
Inside Diameter Cross Section Diameter
6.0

5.5
5.0
Mean of Shrinkage (%)

4.5
4.0

3.5
3.0
2.5

2.0
1.5

1.0

5 45 2 8

Fig. 5. The main effect of an experiment

for determining mould sizing variables.

Table 3. Average accuracy of one hidden layer model for o-ring shrinkage prediction.
Average accuracy, %
Architectural network
Inside Diameter Cross Sectional Diameter
5-3-1 93.1 81.6
5-4-1 94.6 84.7
5-5-1 94.6 85.5
5-6-1 94.7 86.7
5-7-1 94.5 89.0
5-8-1 94.4 89.3
5-9-1 94.5 88.8
5-10-1 94.6 90.5
5-11-1 95.2 91.3

Table 4. Average accuracy of two hidden layer model for inside diameter shrinkage prediction.
Average accuracy, %
Architectural network Standard Deviation
in direction of inside diameter
5-11-15-1 95.9 0.038
5-11-16-1 95.9 0.020
5-11-18-1 95.9 0.024
5-11-19-1 95.9 0.040
5-11-20-1 95.9 0.037
5-11-21-1 95.9 0.008
5-11-22-1 95.9 0.023
5-11-23-1 95.9 0.011
5-11-15-1 95.9 0.038

9th Asian Symposium on Visualization, Hong Kong SAR, China, 2007. 5-4
 A Prediction on Rubber Product Shrinkage in Compression Moulding using Neural Network

Table 5. Average accuracy of two hidden layer model for cross section diameter shrinkage
prediction.
Average accuracy, %
Architectural network in direction of cross section Standard Deviation
diameter
5-11-15-1 95.3 1.042
5-11-16-1 96.1 0.183
5-11-17-1 96.0 0.188
5-11-18-1 96.0 0.223
5-11-19-1 96.1 0.182
5-11-20-1 95.8 0.513
5-11-21-1 96.0 0.167
5-11-22-1 95.8 0.517
5-11-23-1 96.1 0.185

4. Conclusions

In this research, an application of neural network model in prediction of the rubber o-ring shrinkage in
compression moulding is performed. A back propagation neural network was used to determine the
shrinkage based on amount of sulphur, the amount of carbon black, the mould temperature, an inside
diameter and a cross sectional diameter. The neural network prediction for an inside diameter
shrinkage and a cross sectional diameter shrinkage indicate that architectures 5-11-21-1 and 5-11-16-1
provide an optimized prediction within 95.9% and 96.1% accuracy, respectively.

References

[1] Hagan, M. T., Demuth, H. B. and Beals, M. “Neural Network Design”, PWS Publishing Company, U.S.A.
(1996).
[2] Chungchoo, C. “A New Nueral Network for Tool Wear Classification”, The 20th Conference of Mechanical
Engineering Network of Thailand. (2002).
[3] Abu-Zahra, N. H. “Neural Network Approach to Density Control of Rigid PVC Foam in Extrusion
Process”. International Polymer Processing. (2003).
[4] Davis, B.A., Gramanm, P.J., Osswald,T.A. and Rios, A.C. “Compression Moulding”. Carl Hanser Verlag,
Munich. (2003).
[5] Deng, J-C. “Basic Study of Injection Moulding of Rubber Compounds”, Ph.D. Thesis, The University of
Akron. (1990).
[6] Sommer, J.G. “Elastomer Molding Technology”. Bookmasters, Inc., Ashland. (2003).

9th Asian Symposium on Visualization, Hong Kong SAR, China, 2007. 5-5