The Development of Cotton-yarn-quality Predicting System

2. Neural Network Model Development
Aiming at the above issue, the strength and evenness 
CV % of the yarns was predicted when the 3 layers and 4 
layers BP networks were developed with neural network 
978-1-4577-0860-2/11/$26.00 ©2011 IEEE

---

toolbox in Matlab environment and trained 
[2-3]
; viewed 
from the training effect, not only the training time of 
3-layer-network was shorter than the 4-layer-network, the 
prediction precision was also higher. Therefore, 3-layer 
network model network structure was chosen in this 
yarn-quality predicting system, which consisted of an 
input layer, a hidden layer and an output layer. 
A. Neurons Number in Input Layer
This system was mainly used to predict the yarn 
tenacity and its evenness CV% under any other conditions 
unchanged, such as processing, temperature and humidity. 
To reach the above objective, the yarns quality was 
predicted through inputting six indicators of fiber quality 
such as impurity rate, the principal length, the percentage 
of short fiber, the degree of maturity, fiber strength and 
Micronaire 
[4]
.Firstly, the 11 fiber quality indicators the 
spinning plants often measured were statistically analyzed 
and then the experts' opinions were gotten in this field, 
finally six neurons were taken in this layer to meet the 
requirement that the network neurons number should 
correspond to the fiber index number.
B. Neurons Number in Hidden Layer
There was no effective principle to follow to 
determine the neurons numbers in this layer, so the 
optional one was mainly determined by training and 
contrast on different neurons numbers. It was revealed that 
the network predicting accuracy was highest when the 
neurons number in this layer was 16 through training and 
contrast in the course of modeling the network. However, 
this was only for the training data used by the author, if 
the training sample changed, the optional number of 
neurons would have a corresponding change, therefore the 
neurons number in hidden layer could be set to a variables 
with a recommended value 16 in the quality predicting 
system, and the best suited hidden nodes number for 
samples could be determined by the system operator 
through training and contrast. 
C. Neurons Number in Output Layer
The yarn tenacity and evenness CV % were chosen for 
the predicting output items in this system. As a result, the 
corresponding output layer neurons number in the network 
was taken as 2.
D. Activation Function of Each Layer
The common hyperbolic tangent function 
)
1
(
)
1
(
)
(
2
2
n
n
e
e
n
f
−
−
+
−
=
in BP network was adopted 
as activation function in the hidden layer. This function 
could compress each input to a value between -1 and 1; 
while the linear function 
b
n
w
n
f
+
×
=
)
(
could be adopted 
as the activation function in the output layer 
[5]
. 
E. Network Training Algorithm
Levenberg-Marquardt algorithm was adopted to train 
the network model in the system to overcome the 
disadvantages of criteria BP algorithm, such as long 
training time and vulnerability to local optimum. It has 
been proved that the improved algorithm could markedly 
reduce the time for network training. Meanwhile, the 
network model structure was not too complicated and the 
problems were relatively simple in the quality predicting 
system, the actual total memory consumption was not too 
much when the network was trained by this method. 
The network model has been built completely after the 
above design, according to which the topology map could 
be drawn in Figure 2. The BP network could be divided 
into three layers: the input layer, the hidden layer, and the 
output layer. There were 6 nodes in the input layer, an 
adjustable number in the intermediate hidden layer, and 2 
in the output layer. 
Figure2. The BP Net Structure 
IV. 
 
S
YSTEM 
D
EVELOPMENTS