Transcript Elman Network

PREDICTING YARN
TENSILE STRENGTH
USING ELMAN NETWORK
Josphat Igadwa Mwasiagi,
Huang XiuBao and Wang XinHou
Department of Textile Engineering, Donghua
University, Shanghai
1
Introduction
The
influence of fiber properties to the yarn
strength characteristics has been a subject of study
by many researchers, [Jackowski et al, 2002;
Ureyen et al, 2006, Ghosh et al, 2005(a); Ghosh et
al, 2005(b)]
The reported artificial Neural Networks (NN)
models use HVI characteristics together with yarn
fineness and TPI as inputs
The reported models have different models for
rotor and ringframe yarns
 This paper discusses the design of a single NN
model to predict the tensile strength of rotor and ring
spun yarns
2
Elman Network
-The Elman network is a type of a
recurrent feed forward neural network,
with a feedback connection from the
output of the hidden layer neurons to
the input of the network,
-The Elman network has tansig neurons
in its hidden (recurrent) layer, and
purelin neurons in its output layer
3
The architecture of Elman Network
4
Training Algorithm
 Training
involves adjusting the weights
and biases of the network so as to
minimize the network’s performance
function
 This can be done using a by using a
technique called backpropagation (BP),
which involves performing computations
backwards through the network [Ham et
al, 2003]
5
Fletcher-Reeves Update method
- Fletcher-Reeves Update is a modification
of the BP technique
-it is much faster than the original BP
technique
6
Materials
Cotton lint and yarn samples were collected from
four textile factories in Kenya. For every yarn
sample collected, a sample of the corresponding
cotton lint mixture used to spin the yarn was also
collected
 The details of the cotton and yarn samples
collected are given in table 1. A total of 410
samples were collected.
 The quality characteristics of the cotton lint and
yarn samples were measured under standards
laboratory conditions in Shanghai-China

7
Table 1: Cotton Lint and Yarn samples
Cotton
Lint
Mill
Code
Meru AR
Meru AR
Meru AR
Voi AR
Voi AR
WT AR
Kitui AR
Kitui AR
Kitui AR
D
D
D
B
B
A
A
A
C
Machine
Type
Yarn
Ne
Rotor
Rotor
Rotor
Ringframe
Ringframe
Ringframe
Ringframe
Ringframe
Ringframe
27
12.5
7.5
30
20
30
30
24
24
Spinning
Speed (rpm)
68,0000
68,0000
57,0000
11,0000
10, 000
12,000
12,0000
11,0000
8,000
8
Methods
A strength prediction algorithm was design as
shown below.
9
Methods




The NN model used Elman network with FletcherReeves Update as the BP network training algorithm and
gradient descent with momentum as the weight/bias
learning function was designed.
Several options for the number of neurons in the hidden
layer were tried, in order to arrive at an optimum design
During training, mean square error (mse), which is the
average squared error between the networks and the
targeted outputs, was used as the performance function
To investigate the performance of the network in more
details, a regression analysis between the network’s
response and the corresponding targets was performed.
10
Results and Discussions
0.2
mse
0.15
0.1
0.05
0
2
4
6
8 10 12 14 16 18 20 22
No. of Neurons
Fig. 2. The performance of the Network.
11
Results and Discussions
R-Value
1
0.95
0.9
0.85
2
4
6
8 10 12 14 16 18 20 22
No. of Neurons
Fig. 3. Prediction ability of the Network.
12
Results and Discussions

From figures 2 and 3 the network
stabilizes at 6 neurons. Increase of
the number of neurons above 6 does
not cause any significant change in
the performance or prediction ability
of the network.
13
Results and Discussions
The performance of the trained network for
the training, validation and test subsets
are given in figure 4. Figure 5 shows the
linear regression between the network
outputs and the corresponding targets.
14
Results and Discussions
15
Results and Discussions
The final mse value for the test data was
0.0156.
 The network outputs tracks targets
reasonable well, with a correlation
coefficient (R-value) of 0.974.

16
Conclusions
An Elman Network model was trained using
Fletcher-Reeves Update conjugate gradient
training algorithm.
The network predicted the tensile strength of
cotton yarn samples consisting of ring and
rotor spun yarns, giving an mse value of
0.0156. The correlation coefficient (R-Value)
between predicted and targeted values for
the network was 0.974.
17
Thank You for Your Kind Attention
Q and A
18