You are on page 1of 14

Estimation of Uster H Hairiness Values From Zweigle Hairiness Test Results

Ylmaz ERBL1 and Osman BABAARSLAN2


1,2

ukurova University, Textile Engineering Department, Adana/TURKEY erbily@cu.edu.tr

Abstract: Hairiness is an important physical property for yarns. There are a lot of test instruments for hairiness test. Uster and Zweigle are two of most using hairiness instruments. At some cases the yarn producers use Zweigle but the clients are Uster. Thus some problems occur like reporting yarn hairiness values to customers. Producers tell values at Zweigle instrument but customer controls these values at Uster and vice versa. This study is focused to transform Zweigle hairiness test results to Uster H Hairiness values. Two set of data are used for statistical analysis at SPSS 11.5 software and for creating neural networks at Neuro Solutions 5.0 software. One set of data has 140 test results for 14 different products. And the other set of data has 450 test results for 3 different products.

1. Introduction The hairiness tests are indispensable part of yarn quality control. This physical property is very effective on goods comfort properties. Also it is important for production processes. There are a lot of test instruments for hairiness testing. Most using of them are Uster and Zweigle Hairiness Test instruments. At some cases the producers have problems with their customers for reporting of production informations because of different test instruments at them and at their customers. The producer has Zweigle hairiness test instrument and customer has Uster and vice versa. This causes problems between producers and customers. At this study we focused on transformation of Zweigle hairiness test results to Uster H test results. We used two set of data and tried to create networks for this transformations with Neuro Solutions 5.0 software. 2. Material and Method We used two data sets for this study. First data set has 140 test results for 14 different types of yarns. These yarns are produced at Open-End Rotor Spinning Machine with 4 different raw materials and 4 different types of take-off nozzles. The variables are raw material, take-off nozzle type, Zweigle - Index values and Uster H Hairiness values. The second data set is taken from Ring Spinning Production variables. There are 450 test results for 3 different yarn type. The variables for second data set are color type, blending rate and Zweigle - S3 and Zweigle - ndex values and Uster H Hairiness values.. Data sets are analyzed with both statistical regression methods and neural network models. 2.1. Statistical Approach For statistical approach the SPSS 11.5 software was used. Curve estimation and linear regression methods were used at statistical approach. At these analyzes only Zweigle Index variables and Uster H Hairiness variables were used for regression. Each data set was analyzed with all curve types and the most suitable type was presented at this paper.

2.1.1 Statistical Approach for Data Set 1 Scatter plot graphic of all variables at data set 1 is shown at Figure 1.
7

USTER

0 200 400 600 800 1000 1200 1400 1600 1800

ZWEIGLE

Figure 1 Scatter Plot of Data Set 1 For Data Set 1 which has 140 test results the most suitable model was cubic curve model. Because the R value of this model was the biggest. It was %63. Analyze informations and results are shown at below.
MODEL: MOD_19. Dependent variable.. USTER Listwise Deletion of Missing Data Multiple R .62756 R Square .39383 Adjusted R Square .38046 Standard Error 1.52934 Method.. CUBIC

Analysis of Variance: DF Sum of Squares Mean Square Regression 3 206.66389 68.887964 Residuals 136 318.08741 2.338878 F = 29.45342 Signif F = .0000 -------------------- Variables in the Equation -------------------Variable B SE B Beta T Sig T ZWEIGLE ZWEIGLE**2 ZWEIGLE**3 (Constant) .002835 -1.19398410E-05 5.69516470E-09 5.809021 .009128 9.9594E-06 3.3396E-09 2.523266 .581056 -4.914545 3.919389 .311 -1.199 . 2.302 .7566 .2327 . .0228

The curve for this model is shown at Figure 2.

USTER
7 6

1 0 200 400 600 800 1000 1200 1400 1600 1800

Observed Cubic

ZWEIGLE

Figure 2 Curve Fit Graphic for Data Set 1

2.1.2 Statistical Approach for Data Set 2 Scatter plot graphic of variables is at Figure 3.
7

USTER_H

2 800 1000 1200 1400 1600 1800 2000 2200

INDEX

Figure 3 Scatter Plot of Data Set 2 For Data Set 2 which has 450 test results the most suitable model was quadratic curve model. Because the R value of this model was the biggest. It was %75. Analyze informations and results are shown at below.
MODEL: MOD_13. Dependent variable.. USTER_H Method.. QUADRATI

Listwise Multiple R Square Adjusted Standard

Deletion of Missing Data R .75294 .56692 R Square .56498 Error .77754

Analysis of Variance: DF Sum of Squares Mean Square Regression 2 353.75698 176.87849 Residuals 447 270.24167 .60457 F = 292.57029 Signif F = .0000 -------------------- Variables in the Equation -------------------Variable B SE B Beta T Sig T INDEX .006778 .001205 1.699637 5.626 .0000 INDEX**2 -3.16159540E-06 3.9422E-07 -2.422886 -8.020 .0000 (Constant) 1.929809 .897505 2.150 .0321

The curve for this model is shown at Figure 4.

USTER_H
7

2 Observed 1 800 1000 1200 1400 1600 1800 2000 2200 Quadratic

INDEX

Figure 4 Curve Fit Graphic for Data Set 2 2.1.3 Linear Regression Analysis for Data Sets In addition to curve fit analysis linear regression analysis were studied on both data sets. But no acceptable results were obtained at linear regression. For first data set % 56 success rates was obtained and for the second data set % 75 success rate was obtained. There is a decrease at first data sets success but the rates around %70-80 are not acceptable for textile estimations. Analyze report for first data set at linear regression is shown below.

Regression
Variables Enter ed/Re m ovebd Model 1 Variables Entered INDEX, S3a Variables Remov ed . Method Enter

a. All requested variables entered. b. Dependent Variable: USTER_H


b Model Sum m ary

Model 1

R .560 a

R Square .314

Adjusted R Square .304

Std. Error of the Estimate 1.62087

a. Predictors: (Constant), INDEX, S3 b. Dependent Variable: USTER_H

ANOVAb Model 1 Sum of Squares 164.821 359.930 524.751 df 2 137 139 Mean Square 82.411 2.627 F 31.368 Sig. .000 a

Regression Residual Total

a. Predictors: (Constant), INDEX, S3 b. Dependent Variable: USTER_H


a Coe fficients

Model 1

(Cons tant) S3 INDEX

Unstandardiz ed Coef f icients B Std. Error 5.722 .354 .002 .001 -.006 .001

Standardized Coef f icients Beta .839 -1.331

t 16.154 2.999 -4.755

Sig. .000 .003 .000

a. Dependent Variable: USTER_H


a Res iduals Statistics

Predicted Value Residual Std. Predicted Value Std. Residual

Minimum 1.0689 -3.0177 -2.118 -1.862

Max imum 4.9002 2.6070 1.401 1.608

Mean 3.3751 .0000 .000 .000

Std. Deviation 1.08893 1.60917 1.000 .993

N 140 140 140 140

a. Dependent Variable: USTER_H

Charts
Histogram Dependent Variable: USTER_H
30

20

Frequency

10 Std. Dev = .99 Mean = 0.00 0 N = 140.00

Regression Standardized Residual

1.0 0

.75

.50

.25

0.0 0 0.0 0 .25 .50 .75 1.0 0

Analyze report for second data set at linear regression is shown below.

Regression
Variables Enter ed/Re m ovebd Model 1 Variables Entered INDEX, S3a Variables Remov ed . Method Enter

a. All requested variables entered. b. Dependent Variable: USTER_H

8 .3 3 .1 3 -.1 8 -.3 3 -.6 8 -.8 3 .1 -1 8 .3 -1 3 .6 -1 8 .8 -1

Normal P-P Plot of Regression Standardized Residual Dependent Variable: USTER_H

Observ ed Cum Prob

63 1. 38 1. 13 1. 8 .8 3 .6

b Model Sum m ary

Model 1

R .758 a

R Square .575

Adjusted R Square .573

Std. Error of the Estimate .77019

a. Predictors: (Constant), INDEX, S3 b. Dependent Variable: USTER_H

ANOVAb Model 1 Sum of Squares 358.842 265.156 623.999 df 2 447 449 Mean Square 179.421 .593 F 302.468 Sig. .000 a

Regression Residual Total

a. Predictors: (Constant), INDEX, S3 b. Dependent Variable: USTER_H


a Coe fficients

Model 1

(Cons tant) S3 INDEX

Unstandardiz ed Coef f icients B Std. Error 8.571 .201 -.003 .000 .002 .001

Standardized Coef f icients Beta -1.317 .580

t 42.688 -8.610 3.790

Sig. .000 .000 .000

a. Dependent Variable: USTER_H


a Cas ew is e Diagnostics

Case Number 205 305

Std. Residual -3.148 -3.286

USTER_H 2.95 2.08

a. Dependent Variable: USTER_H


a Res iduals Statistics

Predicted Value Residual Std. Predicted Value Std. Residual

Minimum 2.5150 -2.5306 -2.239 -3.286

Max imum 6.3327 2.0807 2.032 2.702

Mean 4.5163 .0000 .000 .000

Std. Deviation .89398 .76847 1.000 .998

N 450 450 450 450

a. Dependent Variable: USTER_H

Charts
Histogram Dependent Variable: USTER_H
60 50 40 30 20 10 0 Std. Dev = 1.00 Mean = 0.00 N = 450.00

Frequency

Regression Standardized Residual

1.0 0

.75

.50

.25

0.0 0 0.0 0 .25 .50 .75 1.0 0

2.2. Neural Network Approach 2.2.1 Network 1 (Data Set 1) For first data set Raw material type, Take-Off Nozzle type and Zweigle index values were defined as inputs. Uster H Hairiness values were defined as desired/output. At the beginning of study the scatter plot of values was drawed for preprocess (Figure 5). This graphic showed us values are 6 different clusters. The study was directed with this information. At first stage for creating neural network, SOFM (Self Organized Feature Maps) type of network selected for data set. The network was designed for 6 clusters and at hidden layer we used softmax axon for transfer function and at hidden layer softmax axon was used for transfer function and at output layer sigmoid axon was used transfer function. Before creating network the rows were randomized automatically for best training and testing. The numbers of training and testing values were selected by percentage as % 65 for training and % 35 for testing. Thus, the network was created and trained with reserved values for maximum epoch number as 2000 (1000 for unsupervised learning and 1000 for supervised learning). The MSE versus Epoch graphic showed at Figure 6 and Training report showed at Table 1. The minimum and final MSE of network is 0,02. 8

5 -.7 5 .2 -1 5 .7 -1 5 .2 -2 5 .7 -2 5 .2 -3

75 2. 25 2. 75 1. 25 1.

Normal P-P Plot of Regression Standardized Residual Dependent Variable: USTER_H

Obs erv ed Cum Prob

5 .2 5 -.2

5 .7

Scatter Plot
7 6 5 Uster 4 3 2 1 0 0 200 400 600 800 1000 1200 1400 1600 1800 Zweigle

Figure 5 Scatter Plot of Data Set 1


MSE versus Epoch
0,07 0,06 0,05

MSE

0,04 Training MSE 0,03 0,02 0,01 0 1 100 199 298 397 496 595 694 793 892 991

Epoch

Figure 6 MSE versus Epoch Graphic for Network 1 Table 1 Training Report for Network 1
Best Network Epoch # Minimum MSE Final MSE Training 999 0,020699026 0,020699026

After training of network it was tested. The test results are showed at Figure 7 and Table 2.

Desired Output and Actual Network Output


7 6 5

Output

4 3 2 1 0 1 5 9 13 17 21 25 29 33 37 41 45 49

Uster Uster Output

Exemplar

Figure 7 Desired Output and Actual Network Output for Network 1 Table 2 Test Results for Network 1
Performance MSE NMSE MAE Min Abs Error Max Abs Error r Uster 1,697693373 0,415308176 1,127838541 0,100209528 2,383335686 0,810369577

As seen Table 2 the success rate of network is %81. This is a good success rate but as seen at Figure 7 at some points there are very big differences between network output and real values. These are not acceptable at textile productions. The properties of textile goods are very sensitive. 2.2.2 Network 2 (Data Set 1) At the second stage we decided to try another network type for this data set. The Radial Basis Function RBF type of network was used for the second essay. At creating RBF network 1 hidden layer was used for unknown effects to results. At hidden layer and output layer linear axon was used as transfer function. Number of cluster centers was setted as 6. Competitive rule was selected as ConscienceFull and metric (the distance type) was selected as Euclidean. The number of maximum epoch was selected as 1000 at both unsupervised and supervised learnings. Thus maximum epoch of network was reached to 2000 totally. After creating network it was trained. The train results are showed at Figure 8 and Table 3.

10

MSE versus Epoch


0.8 0.7 0.6 0.5

MSE

0.4 0.3 0.2 0.1 0 1 100 199 298 397 496 595 694 793 892 991

Training MSE

Epoch

Figure 8 MSE versus Epoch Graphic for Training of Network 2 Table 3 Training Report for Network 2
Best Network Epoch # Minimum MSE Final MSE Training 999 0,006697365 0,006697365

After training the network was tested. The test results are showed at Figure 9 and table 4. Desired Output and Actual Network Output
7 6 5

Output

4 3 2 1 0 1 5 9 13 17 21 25 29 33 37 41 45 49

Uster Uster Output

Exemplar

Figure 9 Desired Output and Actual Network Output for Network 2

11

Table 4 Test Results for Network 2


Performance MSE NMSE MAE Min Abs Error Max Abs Error r Uster 0,151083915 0,036959787 0,303178133 0,014516129 0,923944029 0,982985944

As seen from Table 4 the success rate of Network 2 is %98. And this shows that RBF type networks are most suitable for our data set. 2.2.3 Network 3 (Data Set 2) At third stage data set 2 was used with a network which has same construction with network 2. But the number of cluster centers was selected as 3 because at this data set we have 3 different products and as seen Figure 10 the variables can be in 3 different clusters. With same settings as network 2 the new network was created for 3 cluster centers and trained for 2000 maximum epochs totally. The train results are showed at Figure 11 and Table 5.
6.5 6 5.5 5 4.5 4 3.5 3 2.5 2 800 1000 1200
Ekru50/50

1400

1600

1800
Ekru80/20

2000

2200

Melanj50/50

Figure 10 Scatter Plot of Data Set 2 By Different Yarn Types

12

MSE versus Epoch


0.4 0.35 0.3 0.25

MSE

0.2 0.15 0.1 0.05 0 1 100 199 298 397 496 595 694 793 892 991

Training MSE

Epoch

Figure 11 MSE versus Epoch Graphic for Training of Network 3 Table 5 Training Report for Network 3
Best Network Epoch # Minimum MSE Final MSE Training 455 0,008949153 0,008949153

After training the metwork 3 was tested. The test results are showed at Figure 12 and Table 6.

Desired Output and Actual Network Output


7 6 5

Output

4 3 2 1 0 1 16 31 46 61 76 91 106 121 136 151

USTER H USTER H Output

Exemplar

Figure 12 Desired Output and Actual Network Output for Network 3

13

Table 6 Test Results for Network 3


Performance MSE NMSE MAE Min Abs Error Max Abs Error r USTER H 0,070112326 0,051184517 0,199285619 0,002170438 0,949624876 0,974117478

As seen Table 6 the success rate of network 3 is % 97. 3. RESULTS and DISCUSSION For first data set we reached to % 98 success rates and for the second data set we reached to % 97 success rates at neural network models. The statistical analyses were deficient for explaining the relationship between Zweigle Hairiness results and Uster H Hairiness results. But Neural Network models are very adequate and successful for estimation Uster H values from Zweigle Hairiness results and production variables. From neural network models, the RBF method is the successful model for both two data sets. But these estimations which obtained with neural network models are not suitable for all textile yarns hairiness test results. Because the production variables like raw material, number of yarn, type of yarn, type of spinning machine and even color of yarn are effective on hairiness properties. The obtained models at this study can be used for similar yarn types which are produced at similar production type. And the other type of yarns and different productions can be added to model and it can be modified for acquiring model large yarn types ands productions.

14

You might also like