Você está na página 1de 3

Fibers and Polymers 2008, Vol.9, No.

6, 782-784

Comparison of the New Methodologies for Predicting the


CSP Strength of Rotor Yarn
Deogratias Nurwaha* and Xin Hou Wang
Department of Textile Engineering, Dong Hua University, Shanghai 200051, P.R. China
(Received January 31, 2007; Revised April 22, 2008; Accepted August 6, 2008)

Abstract: This paper provides preliminary results on the relative performance of the adaptive neuro-fuzzy system inference
(ANFIS) model versus linear multiple regression method, when applied to the use of cotton fiber properties to predict spun
yarn strength obtained from open-end rotor spinning. Fiber properties and yarn count are used as inputs to train the two mod-
els and the output (dependent variable) would be the count-strength-product (CSP) of the yarn. The predictive performances
of the two models are estimated and compared. We found that the ANFIS has a better average prediction successful in com-
parison with linear multiple regression model.
Keywords: ANFIS, Yarn strength, Rotor spun yarn, HVI

Introduction network presentation of a Takagi-Sugeno fuzzy inference


system [6], and the if-then rules in Takagi-Sugeno are
In textile engineering, strength has been considered to be comprised in the network structure.
one of the most important yarn properties. High strength
yarns generally produce high strength fabrics. Also, yarns of Experimental
high strength process more efficiently at spinning and at
subsequent manufacturing operations. The relationship between Data Collection
fiber properties and yarn properties has been the focus of Cotton crop study data of 1997 published by the International
research, and considerably success has been achieved [1-4]. Textile Center were used in this study. This publication [1]
The complexity of applying a mathematical treatment to contains all fiber and yarn data, along with detailed explana-
staple yarn structure often prevents some of the theoretical tions of equipment and procedures. Cotton fiber properties
models from being useful practical tools for prediction. measured by (HVI) and Uster AFIS are given in Table 1.
Some previous studies have shown the nonlinear relationship The spinning machine parameters: Nominal yarn number
between yarn strength and fiber properties [5]. Since the (30’s), rotor speed (55,000 rpm), opening roller speed (6700
relationship between yarn strength and fiber properties is rpm) and twist magnitude (454.43 turns/10 cm) are held
nonlinear, it is very difficult to use conventional techniques constant during the processing. The skein method was used
such regression to predict the yarn strength. One of the to test the yarn strength. Skein strength was measured for
difficulties in regression model is that the model usually a each sample yarn and multiplied by yarn count to express
linear model should be predefined. On the other hand, the breaking force as the count-strength-product (CSP) in
artificial neural networks and fuzzy logic, generally called (kN·m/kg). The CSP was used as the output (dependent
intelligent techniques, are capable of mapping very nonlinear variable). Since multiple yarn sizes were spun from each
and complex relations and they do not require knowledge or cotton bale sampled, the yarn count of each produced yarn
estimation of any mathematical model in advance [6]. The sample was also included as an input variable. From the crop
aim of this work is to describe these new approaches in study results we selected 34 samples of input-output data for
textile engineering as well as to use them for developing carded rotor spun yarns. We used 21 and 13 samples,
rotor spun yarn prediction system.
This study will involve two stages. In the first stage, we Table 1. Cotton fiber properties
use regression analysis to model rotor spun yarn strength; Variable Description of variable Unit
that is, we identify the relationship between yarn strength FS Fiber strength g/tex
and fiber properties through regression function. In the UHML Upper half mean length mm
second stage, we use Adaptive Neuro-Fuzzy System Inference UI Uniformity index %
(ANFIS) to predict the rotor spun yarn strength. The predictive M Micronaire μg/mm
performances of the two models will be estimated and compared. R Grayness Rd
There are already some examples of applications of ANFIS Y Yellowness +b
method in textile issues [7,8]. ANFIS can be considered as a FE Elongation %
SFC Short fiber content* % by weight
*Corresponding author: nurwahadeo@yahoo.fr *Measured by uster AFIS.
Comparison of the New Methodologies for Predicting the CSP Strength Fibers and Polymers 2008, Vol.9, No.6 783

Table 2. ANOVA table


Source SS df MS F-ratio p-value
Columns 144399.2 8 18049.9 2344.1 0.00
Error 1382.9 180 7.7
Total 145782.1 188

respectively for training and testing the two models.

Linear Multiple Regression


Regression models are frequently used to predict yarn
properties. Fiber properties along with the yarn count were
used as inputs (independent variables) whereas count-
strength-product was used as output (dependant variable).
The ANOVA results are shown in Table 2.
Regression equation:
CSP = 33.39FS + 553.13UHML + 34.88UI − 132.475M
+ 5.565R + 37.19Y − 63.669FE − 8.147SFC
+ 3.28YC − 2244.343, [R2 = 0.90] (1) Figure 1. RMSE curve within 50 epochs.
2
The coefficient of multiple determination R defines the
fraction of the variability in the dependent variable (CSP) the count- strength product acts as the output. The learning
explained by the regression model. Since the R2 is very high, algorithm teaches the ANFIS to map the fiber properties to
we can conclude that the regression model fits the data very the yarn strength through a process called training. At the
well. end of training, the trained ANFIS network would have
learned the input-output map and be ready to be deployed
Adaptive Neuro-fuzzy Inference System (ANFIS) into the larger control system solution.
ANFIS integrated neural network and fuzzy inference
system together, constructing data model with fuzzy inference Prediction of Rotor Spun Yarn Strength with the ANFIS
techniques and choosing model parameter via adaptive neural For implementation, the Fuzzy Toolbox in MATLAB
network method. The adaptive capability of ANFIS makes it software [10], which provides functions of constructing,
almost directly applicable to adaptive control and learning editing and training of ANFIS, was employed. Since the input
control [9]. Two important components of ANFIS are data set has a high dimension, we used clustering function
membership functions and fuzzy inference rules. Membership (genfis2) as a pre-processor to ANFIS for determining the
function (MF) is a curve that defines how each point in the initial rules. An important advantage of using a clustering
input space is mapped to a membership value (or degree of method to find rules is that the resultant rules are more tailored
membership) between 0 and 1 of the associated fuzzy sets. to the input data than they are in an FIS generated without
Fuzzy inference rules are some if-then rules to define output clustering. This reduces the problem of combinatorial
behavior from the inputs. explosion of rules when the input data has a high dimension.
If x is A and y is B then z=f(x, y) where A and B are fuzzy After the 50 epochs, the final results are expressed by the root
sets in the antecedent and f(x, y) is generally a polynomial in mean squared error (RMSE). At the end of training, RMSE is
the input variables, although it can be any function that can 0.0021 and is quite small but the final comment on overall
appropriately describe the output model. The first order prediction performance should made by analyzing the testing
Sugeno model type fuzzy inference system in the following results. As we can see the minimal error occurred at epoch 46.
form [10]: Usually, the RMSE decreases with an increasing number of
If x is A and y is B then z=p×x + q×y + r, where {p, q, r} learning epochs as shown in the Figure 1.
are consequent parameters.
ANFIS stands for Adaptive Neuro-Fuzzy Inference System. Validation of the Two Models
It is a hybrid neuro-fuzzy technique that brings learning After the training, all the two models have been subjected
capabilities of neural networks to fuzzy inference systems. to the unseen checking (testing) data. The root mean squared
The learning algorithm tunes the membership functions of a errors for both ANFIS and regression have been calculated.
Sugeno-type Fuzzy Inference System using the training The results are the following:
input-output data. In this case, the input-output data refers to For ANFIS: ANFISchkRMSE=98.5792
the fiber properties-count-strength product (yarn strength) For regression: RegchkRMSE=110.8004
dataset. The fiber properties act as input to the ANFIS and We see that the RMSE provided by ANFIS is better than
784 Fibers and Polymers 2008, Vol.9, No.6 Deogratias Nurwaha and Xin Hou Wang

Table 4. Comparison analysis of the prediction performance of the Conclusion


two models
Statistical parameter ANFIS model Regression model In this work, we have predicted the yarn strength of rotor
RMSE 98.5792 110.8004 spun yarns with multiple linear regression and ANFIS. We
MAE 79.80 92.43 found that ANFIS model has the better predictive power in
MaxMAE 163 231 comparison with the regression model. Since we know that
MinMAE 3.67 8.30 the relationship between fiber properties and yarn strength is
nonlinear, ANFIS should be a potentially better data analytic
method that needs to be explored more in depth to assess the
practical impact of fiber properties on the rotor spun yarn
strength and compared to more techniques that are already in
use in textile engineering.

References

1. N. Ucar and S. Ertugrul, Text. Res. J., 72, 361 (2002).


2. S. Ertugrul and N. Ucar, Text. Res. J., 70, 845 (2000).
3. P. K. Majumdar and A. Majumdar, Text. Res. J., 74, 652
(2004).
4. L. Cheng and D. L. Adams, Text. Res. J., 65(9), 495
(1995).
5. M. D. Ethridge, J. D. Towery, and J. F. Hembree, Text. Res.
Inst., 52, 35 (1982).
6. J. S.-R. Jang, C.-T. Sun, and E. Mizutani, “Neuro-Fuzzy
Figure 2. Curves of testing results for both two models and and Soft Computing”, Prentice Hall Inc., 1997.
experimental results. 7. T. Takagi and M. Sugeno, IEEE Trans. on Systems, Man,
and Cybernetics, 15, 116 (1985).
that provided by regression method. 8. Texas Cotton Quality Evaluation of Crop of 1997,
Inernational Textile Center, Lubbock, Texas, 1997.
Comparison of Prediction Performance of the two Models 9. J. S.-R. Jang, IEEE Trans. on Systems, Man, and
The goal of this part of research is comparing ANFIS Cybernetics, 23, 665 (1993).
model and linear multiple regression method model. Various 10. Fuzzy Logic Toolbox for Use with MATLAB, MathWorks
methods of comparison were used to judge the performance Inc., 2001.
of the two models: the mean and standard deviation of the 11. Y. Zhu, “ANFIS: Adaptive Neuro-Fuzzy Inference
predicted data, maximum and minimum of absolute errors, Systems”, 2000.
root mean squared error (RMSE) and the mean of absolute 12. D. Naga Jyothi, Map India Conference, 2004.
errors. All these methods of comparisons are defined as 13. C.-B. Cheng, Int. J. Appl. Sci. Eng., 2, 59 (2004).
following: 14. C. B. Cheng, C. J. Cheng, and E. S. Lee, Computers and
n Mathematics with Applications, 44, 1503 (2002).
1 2
RMSE= --- ∑ ( Oi – Op ) (2) 15. J. Galindo, A. Urrutia, and M. Piattini, “Fuzzy Databases:
n Modelling, Design, and Implementation”, Idea Group
n=1 n
1
Mean of absolute error: MSE= --- ∑ Oi – Op (3) Publishing, 2006.
n 16. O. Pribyl, “Faculty of Electrical Engineering”, Prague,
n=1
Maximum of absolute error: MaxMAE (4) 2000.
17. W. S. Sarle, SAS institute Inc., Cary, NC, USA, 1994.
Minimum of absolute error: MinMAE (5)
18. F.-J. Chang and Y.-T. Chang, Advances in Water
where n is the number of pairs; Oi and Op are i-th desired Resources, 29, 1 (2006).
output and calculated output, respectively. The results are 19. J. José Vieira, F. Morgado Dias, and A. Mota, Eng.
summarized in Table 4 and the comparative curves are Applications of Artificial Intelligence, 17, 265 (2004).
plotted in the Figure 2. 20. J. Kim and N. Kasabov, Neural Networks, 12, 1301 (1999).
21. C.-B. Cheng, Neural Network World, 10, 545 (2000).

Você também pode gostar