You are on page 1of 10

A Study of Using Back-Propagation Neural Network for the Sales

Forecasting of the Thin Film Sputtering Process Material


Tian-Syung Lan, Pin-Chang Chen, Chun-Hsiung Lan, Kai-Chi Chuang

A Study of Using Back-Propagation Neural Network for the Sales


Forecasting of the Thin Film Sputtering Process Material
1
Tian-Syung Lan, 2 Pin-Chang Chen, 3Chun-Hsiung Lan, 4Kai-Chi Chuang
1
Department of Information Management, Yu Da University, (tslan@ydu.edu.tw)
*2
Department of Information Management, Yu Da University, (chenpc@ydu.edu.tw)
3
Department of Business Administration, Hsing Wu Institute of Technology,
(100001@mail.hwc.edu.tw)
4
Department of Information Management, Yu Da University, (99310024@ydu.edu.tw)

Abstract
Sales forecasting is the most important aspect in managing company operations. The forecast
accuracy becomes an important goal for companies. This study introduces the target industry (an
optoelectronic material) as an example for case study which the annual sales amounts from year 2004
to year 2010 are provided by the company. By using the back-propagation neural network with the
advantages of high-speed computing power, fast recall speed, high learning accuracy, and fault
tolerance, the neural network uses 12 neurons in the input layer, 8 neurons in the hidden layer, and 1
neuron in the output layer. Through a learning cycle of 20,000 times and a learning rate of 0.1, the
forecast values after conducting training and testing are received.
It is shown that the forecasted output values by the back-propagation neural network are extremely
related to the actual sales values. Additionally, the correlation and the R-squared values are 99.3967%
and 0.975987, respectively. It indicates that the future case companys chromium target sales can be
satisfactorily forecasted within certain accuracy. The result of this study demonstrates that the case
company can surely utilize this proposed forecasting method to not only prevent increased inventory
cost caused by excessive importing, but also reduce loss of orders caused by material shortages.
With the result from this study, the advantages of the back-propagation neural network can be
furthermore applied to other industries in enabling companies to identify the range of possible
influences by reviewing the overall forecast data. The back-propagation neural network can also
provide managers with estimates of future demand to develop coping strategies in advance and
manage the transmission of company knowledge and experience to reduce company losses and
strengthen company competitiveness through perpetual operation.

Keywords: Target Industry, Back-Propagation, Neural Network, Sales Forecasting

1. Introduction

The target industry is greatly influenced by semiconductor and panel manufacturers. Additionally,
its industry characteristics are also included within very uncertain market demand. Under the shocks
from the international financial crisis, the industry is frequently in an over supplied situation. Because
the range of electronics products has declined, the sales of small and medium sized panels have
increased annually which is shown in Figure 1. Since the field of coating has expanded, the thin film
sputtering process becomes the key to producing light and thin products.

International Journal of Advancements in Computing Technology(IJACT) 118


Volume4, Number8, May 2012
doi: 10.4156/ijact.vol4.issue8.15
A Study of Using Back-Propagation Neural Network for the Sales
Forecasting of the Thin Film Sputtering Process Material
Tian-Syung Lan, Pin-Chang Chen, Chun-Hsiung Lan, Kai-Chi Chuang

Source: MIC (Institute for Information Industry) compiled by TISC (2010)

Figure 1. Global sales of small- and medium-sized panels

The main material of the thin film sputtering process is the target. The quality demands for the
target are very strict and substantial. Targets are typically imported and then put onto the bonding
process in Taiwan. Currently, targets are primarily used for high concentration of pure metal in hot
isostatic pressing (HIP) [1]. However, the cost is significantly influenced by the exchange rate because
of the lengthy production period caused by the HIP and the primarily imported targets. Additionally,
because most targets are imported, there is much cost influenced by the exchange rate. There are many
factors that make it impossible to maintain high inventory, such as backing bonding by customers, the
oxidation of finished products stored, and a limited delivery time. Therefore, sales forecasting is
extremely important to the target industry. Since it is hard to find sales forecasting for the target
industry in the existing researches, this paper is motivated for the industry.
The back-propagation neural network has been applied to numerous fields. Hashim et al. (2004)
used the back-propagation neural network to build a module with flexible architecture for determining
errors when correcting nonlinear characteristics. After continuous repetition, the module was then
developed by Hashim et al. to correct errors consistently. This simulation result has proved that using
the back-propagation neural network was superior to the traditional method of linear module correction
[2]. Purna et al. (2006) used an artificial neural network to forecast the underground water level in the
coastal plains of India. The results showed that the artificial neural network model could accurately
forecast the underground water level for the following few months [3]. Pai et al. (2007 to 2009)
conducted forecasts on the water quality of waste water floating from hospitals and industrial areas
[4],[5],[6]. The results showed that the artificial neural network model could provide good forecasts of
the chemical oxygen demand (COD) and suspended solids of water.
The results of these studies indicated that the back-propagation neural network is effective in
forecasting. In this study, the chromium target sales records of the case study company is obtained to
training the proposed back-propagation neural network for further analysis and discussion of the sales
forecast for the next three years.

2. Literature Review

2.1 Target production process

The production process of the target can be expressed as follows and is shown in Figure 2.
a. Pre-processing
Chromium metal powder is mixed with an additive and poured into a mold at 1600 C. Using
high-pressure sintering, the powder transforms into a solid state. The sintered target is then cut and
polished according to the size specified by the customer. Finally, the target undergoes the target

119
A Study of Using Back-Propagation Neural Network for the Sales
Forecasting of the Thin Film Sputtering Process Material
Tian-Syung Lan, Pin-Chang Chen, Chun-Hsiung Lan, Kai-Chi Chuang

concentration test, ultrasonic washing, drying, packing, and air shipping. The production process in this
stage is conducted by foreign manufacturers.
b. Post-processing
The spent target and backings are heated and separated. The smoothness of the backing is confirmed
within tolerance. Molten indium is poured into the backing bonding area to perform target bonding.
After bonding, pressing is used to smooth the target out, and the warp is measured to determine if it is
within tolerance. After the bond cools, the finished targets undergo ultrasonic testing to verify that the
target and the backing are completely sealed in preventing the loss of the target during the sputtering
process. Loss of the target may cause the customer substantial losses and effect the repetition of the
company.
For the finished targets that have passed ultrasonic testing, the oxides from the two sides of the
copper backing are removed. The target surface is then polished using pneumatic tools and sandpaper
and cleaned using alcohol wipes. The targets surface roughness is measured to ensure compliance with
the customers requirements. Finally, the finished products are vacuum packed in two layers to prevent
any contact to the atmosphere in preventing from oxidation. This part of the production process is
conducted by the manufacturers in Taiwan [1].

Power metallurge

Pre-processing
Appearance and size processing

Backing separation testing Imported target testing

Bonding
Post-processing

Ultrasonic testing

Polishing

Export verification

Figure 2. Target production process

2.2 Back-propagation neural network

The development of an artificial neural network started when McCulloch proposed a mathematic
model for a neural network in 1943, which has since been developed into dozens of different models.
Artificial neural network was inspired by biology; its pattern is similar to the neural tissues of an
organism connected by neurons to simulate the information processing ability of the neural network of
organisms. Artificial neural network simulates how the brains of organisms in the nature receive
external stimuli and learn, during which they gradually adapt to the environmental rules by constantly
learning and revising. Artificial neural network uses a significant amount of simple connected artificial
neurons to imitate the ability of the neural network of organisms. Artificial neurons are simple
simulations of organisms neurons. They receive information from an external environment or other

120
A Study of Using Back-Propagation Neural Network for the Sales
Forecasting of the Thin Film Sputtering Process Material
Tian-Syung Lan, Pin-Chang Chen, Chun-Hsiung Lan, Kai-Chi Chuang

artificial neurons, calculate it, and output the results to the external environment or other artificial
neurons [7].
Since the artificial neural network is a parallel distributed processing model with a highly parallel
and distributed association memory, mistake tolerability, adaptability, and ability to learn from the
environment, it is widely used in the fields of figure identification, voice identification and composition,
signal processing, image compression, construction of expert systems, and decision analysis.
Regarding the development of the artificial neural network, Hopfields study in 1982 was an important
breakthrough in the history of artificial neurons, bringing studies related to the artificial neural network
to the attention of the entire scientific community again [8].
The back-propagation neural network is currently the most representative and widely applied neural
network learning model. The concept of the back-propagation neural network was discussed by
Werbos in 1974 and Parker in 1982. However, this theory and algorithm were only clearly defined in
1985 when Rumelhart et al. (1986) introduced the Propagation Learning Rule at Stanford University
[9]. The architecture of the back-propagation neural network is shown in Figure 3.
Normalized input

Input Layer Hidden Layer Output Layer


Xi Wj Yj
Figure 3. Back-propagation neural network

The basic principle of the back-propagation neural network is using the gradient steepest decent
method to minimize the error function and infer the Delta Rule. The back-propagation neural network
processing is divided into forward pass and reverse pass stages. Errors can be effectively reduced by
using these two stages, thereby achieving the expected learning results. Besides, the back-propagation
neural network is divided into two sections. One is called the learning process, where the learning
model is adopted to obtain a weighted value and a threshold value. The other is the recall process,
where the weighted value and threshold value are entered to provide the model forecasting power. The
neurons repeatedly connect formula (1) and active formula (2) to update the neurons activation
value a .
i
N (1)
ui (k 1) Wij (k )a j ( k ) ( k )
j 1

ai (k 1) f ui (k 1), ui (k ), ai (k ) (2)

After a new activation value is generated, it is sent to other neurons. The two preceding formulae
are collectively called the system dynamic formula, where
k indicates the number of updates
u represents the input value of the neuron number i
i

Wij indicates the weighted value between neuron number i and neuron number j
i denotes the internal threshold value of neuron number i .
N indicates the number of neurons connected to neuron number i

121
A Study of Using Back-Propagation Neural Network for the Sales
Forecasting of the Thin Film Sputtering Process Material
Tian-Syung Lan, Pin-Chang Chen, Chun-Hsiung Lan, Kai-Chi Chuang

f describes the activation function derived from the simplified bio-effect in the neural cells
The back-propagation neural network processing steps are repeated until the error value no longer
changes significantly and the convergence is achieved. When learning is completed, the network stores
the weighted and threshold values, which represent the special characteristics of the neural network.
The weighted value matrix and the threshold value matrix are stored for further use or recall by the
network.
The learning rate, which influences the convergence speed of the artificial neural network, is an
essential parameter of the artificial neural network learning process. If the learning rate is higher, the
artificial neural network convergence speed is increasing. On the contrast, a lower learning rate slows
the convergence speed in the artificial neural network. Selecting either excessively high or low learning
rate can have an adverse effect during the artificial neural network training [10].

3. Research Methodology

3.1 Data collection and data preprocess

This study introduces the target industry (an optoelectronic material) as an example for case study
which the annual sales amounts from year 2004 to year 2010 are provided by the Polema S.A. Taiwan
Branch. This study is intended to forecast the annual sales amounts, which uses year 2004 to year 2008
data as the training data, year 2009 data as the validation data, and year 2010 data as the testing and
evaluation data. In addition, data preprocess is applied to make sure all input data has been transformed
into the scale between [-1, 1]. It is intended to transform the original disorderly raw data into a
dimensionless series in order to obtain an appropriate fundamental for the accurate mathematical
relations.

3.2 Neural network design

After data collection and data preprocess, the neural network design is then conducted. There are
many topologies possible for artificial neural network. The back-propagation algorithm is one of the
most commonly used neural network paradigms. The back-propagation algorithm is a general purpose
learning algorithm, which consists of three basic steps: input signals go forward, errors are calculated
from expected outputs, and these errors are sent to previous layers for weight adjustments. The
back-propagation algorithm is used to develop the artificial neural network model in this study.
As adjusting the learning rate and learning cycles, the artificial neural networks synaptic weights
are gradually adjusted so that the error between the target value and the artificial neural network output
becomes minimized. When the error basically does not change, the convergence of artificial neural
network is then achieved, and the training process is thus completed. It is intended to validate the
accuracy and reliability of this neural network model, as well as the optimal weight values and bias for
the proposed artificial neural network simulation.

3.3 Training and testing

After neural network design, the input and output layer of the model is primarily determined. The
sales data provided by the case company from year 2004 to year 2010 is entered to the artificial neural
network software and the parameters of the proposed neural network are set as well. After training and
testing the input data, the forecasting values obtained from this artificial neural network model is
considered as the results of this study.

3.4 Evaluation

After training and testing, the forecasting accuracy of the model is moreover testified. To examine
the performance of forecasting, it is necessary to evaluate the previously unseen data. It is likely to be
the closest to a true forecasting by using the direct comparison of out-of-sample data. Typically,
forecasting models are optimized using a mathematical criterion, and subsequently analyzed using

122
A Study of Using Back-Propagation Neural Network for the Sales
Forecasting of the Thin Film Sputtering Process Material
Tian-Syung Lan, Pin-Chang Chen, Chun-Hsiung Lan, Kai-Chi Chuang

statistical measures. This study uses forecasting accuracy via statistical measures such as correlation
and R-squared values to evaluate the forecasting performance. The result of this study is intended to
show that the proposed model is capable of forecasting the annual sales amounts with accuracy and
reliability.

The step by step research methodology of this study is shown in Figure 4.

Data Collection

Data Preprocess

Neural Network Design

Training

Testing

Evaluation

Figure 4. Research methodology

4. Data analysis and model implementation

4.1 Data collection and data preprocess

This study uses the Alyuda NeuroIntelligence 2.1 software to develop a sales forecasting model.
First, the annual sales data are used as input data which is shown in Table 1. The data collection and
data preprocess by using the proposed forecasting model is shown in Figure 5.

Table 1. Input data


1 2 3 4 5 6 7 8 9 10 11 12 Total
2004 1.26 0.84 2.1 0.56 0.42 0.42 1.68 1.68 2.24 1.12 1.12 0.56 14
2005 1.89 1.26 3.15 0.84 0.63 0.63 2.52 2.52 3.36 1.68 1.68 0.84 21
2006 2.61 1.74 4.53 1.16 0.87 0.87 3.48 3.48 4.64 2.32 2.32 1.16 29
2007 2.25 1.5 3.75 1 0.75 0.75 3 3 4 2 2 1 25
2008 1.35 0.9 2.25 0.6 0.45 0.45 1.8 1.8 2.4 1.2 1.2 0.6 15
2009 1.08 0.72 1.8 0.48 0.36 0.36 1.44 1.44 1.92 0.96 0.96 0.48 12
2010 0.81 0.54 1.35 0.36 0.27 0.27 1.08 1.08 1.44 0.72 0.72 0.36 9

123
A Study of Using Back-Propagation Neural Network for the Sales
Forecasting of the Thin Film Sputtering Process Material
Tian-Syung Lan, Pin-Chang Chen, Chun-Hsiung Lan, Kai-Chi Chuang

Figure 5. Data collection and data preprocess

4.2 Neural network design

After data collection and data preprocess, the learning rate is set at 0.1 and the convergence of
learning cycles can be accomplished within 20,000 cycles. The rest of the neural network parameters
are shown in Table 2 and the neural network design is shown in Figure 6.

Table 2. Neural network parameters


Parameter Setting
Input Layer 12 neurons
Hidden Layer 8 neurons
Output Layer 1 neurons
Learning Cycles 20000
Learning rate 0.1

124
A Study of Using Back-Propagation Neural Network for the Sales
Forecasting of the Thin Film Sputtering Process Material
Tian-Syung Lan, Pin-Chang Chen, Chun-Hsiung Lan, Kai-Chi Chuang

Figure 6. Neural network design

4.3 Training and testing

After conducting the training and testing of the input and output data which are shown in Figure 7, 8,
the forecasted output data and the comparison of actual values and forecast values is shown in Table 3.

Figure 7. Training

125
A Study of Using Back-Propagation Neural Network for the Sales
Forecasting of the Thin Film Sputtering Process Material
Tian-Syung Lan, Pin-Chang Chen, Chun-Hsiung Lan, Kai-Chi Chuang

Figure 8. Testing

Table 3. Output data


1 2 3 4 5 6 7 8 9 10 11 12 Actual Forecast
2004 1.26 0.84 2.1 0.56 0.42 0.42 1.68 1.68 2.24 1.12 1.12 0.56 14 13.947322
2005 1.89 1.26 3.15 0.84 0.63 0.63 2.52 2.52 3.36 1.68 1.68 0.84 21 20.931175
2006 2.61 1.74 4.53 1.16 0.87 0.87 3.48 3.48 4.64 2.32 2.32 1.16 29 28.828844
2007 2.25 1.5 3.75 1 0.75 0.75 3 3 4 2 2 1 25 24.984166
2008 1.35 0.9 2.25 0.6 0.45 0.45 1.8 1.8 2.4 1.2 1.2 0.6 15 14.962315
2009 1.08 0.72 1.8 0.48 0.36 0.36 1.44 1.44 1.92 0.96 0.96 0.48 12 12.488675
2010 0.81 0.54 1.35 0.36 0.27 0.27 1.08 1.08 1.44 0.72 0.72 0.36 9 11.503177

4.4 Evaluation

According to the results of this study, the correlation and R-squared values of the proposed model
are 99.3967% and 0.975987, as well as the absolute error (AE) indicates the actual values and forecast
values of the proposed model are significantly correlated, which is shown in Table 4.

Table 4. Forecast results


Actual Forecast AE
Mean 17.857143 18.235096 0.476861
Std. Dev. 6.770283 6.239076 0.841037
Min. 9 11.503177 0.015834
Max. 29 28.828844 2.503177
Correlation : 0.993967
R-squared : 0.975987

5. Conclusions

By comparing the actual values with the values produced through our proposed trained model, it is
found that the forecasting values provided by the proposed forecasting model are almost identical to

126
A Study of Using Back-Propagation Neural Network for the Sales
Forecasting of the Thin Film Sputtering Process Material
Tian-Syung Lan, Pin-Chang Chen, Chun-Hsiung Lan, Kai-Chi Chuang

the actual sales amounts. The correlation and the R-squared values are 99.3967% and 0.975987,
respectively. Additionally, the trend of the trained forecasted values change similarly as the actual
values, which the disparity between the values is minimal. Thus, a highly reliable and trustworthy
forecasting model using the back-propagation neural network is established for the target industry.
The model developed in this study can surely be used as a reference in the future. The advantages of
the back-propagation neural network can also be additionally applied to other industries for accurate
forecasting, that will enable the companies to develop coping strategies in advance and identify the
range that may be affected by observing the overall forecast data. The results of this study can also be
moreover used to provide managers with estimates of future production demand to minimize the
company losses and strengthen companys sustainable competitiveness.

6. Acknowledgement
The authors would like to thank Mr. York Tsai, Mrs. Lillian Lai, Mr. Benny Chu and Mr. Tommy
Huang at Polema S.A. Taiwan Branch who kindly provided the data and suggestions to improve this
work.

7. References
[1] C.H. Tam, S.C. Lee, S.H. Chang, T.P. Tang, H.H. Ho and H.Y. Bor, Effects of the temperature of
hot isostatic pressing treatment on Cr-Si targets, Ceramics International, vol. 35, no. 2, pp.565-570,
2009.
[2] S.Z.M. Hashim, M.O. Tokhi and I.Z.M. Darus, Nonlinear dynamic modelling of flexible beam
structure using neural networks, Proceedings of the IEEE International Conference on
Mechatronics, pp.171-175, 2004.
[3] C. Purna, Y.R. Nayak, R. Satyaji and K.P. Sudheer, Groundwater Forcasting in a Shallow Aquifer
Using Artificial Neural Network Approach, Water Resource Management , vol. 20, pp.77-90,
2006.
[4] T.Y. Pai, Y.P. Tsai, H.M. Lo, C.H. Tsai and C.Y. Lin, Grey and neural network prediction of
suspended solids and chemical oxygen demand in hospital wastewater treatment plant Effluent,
Computers & Chemical Engineering, vol. 31, no. 10, pp.1272-1281, 2007.
[5] T.Y. Pai, S.C. Wang, C.F. Chiang, H.C. Su, L.F. Yu, P.J. Sung, C.Y. Lin and H.C Hu, Improving
neural network prediction of effluent from biological wastewater treatment plant of industrial park
using fuzzy learning approach, Bioprocess and Biosystems Engineering, vol. 32, no. 6,
pp.781-790, 2009.
[6] T.Y. Pai, T.J. Wan, S.T. Hsu, T.C. Chang, Y.P. Tsai, C.Y. Lin, H.C. Su and L.F.Yu , Using fuzzy
inference system to improve neural network for predicting hospital wastewater treatment plant
effluent, Computers & Chemical Engineering, vol. 33, no. 7, pp.1272-1278, 2009.
[7] P.C. Chen, C.Y. Lo, H.T. Chang and Y. Lo Cho, A Study of Applying Artificial Neural Network
and Genetic Algorithm in Sales Forecasting Model, Journal of Convergence Information
Technology, vol. 6, no. 9, pp.352-362, 2011.
[8] B.H. Song, K.W. Park and T.Y. Kim, U-health Expert System with Statistical Neural Network,
AISS : Advances in Information Sciences and Service Sciences, vol. 3, no. 1, pp.54-61, 2011.
[9] D.E. Rumelhart, G.E. Hinton and R.J. Williams, Learning Internal Representation by Error
Propagation, Parallel Distributed Processing, vol. 1, pp.318-362, 1986.
[10] B. Kosko, A dynamical approach to machine intelligence: Neural networks and fuzzy systems,
Prentice-Hall, USA, 1992.

127

You might also like