7
Computers in Industry 17 (1991) 217-223 217 Elsevier IMS '91 Learning in IMS The prediction of cotton yarn properties using artificial intelligence Zoran Stjepanovi~ and Anton Jezernik University of Maribor, Faculty of Technical Sciences, Smetanova 17, 62000 Maribor, Slovenia, Yugoslavia The possibility of determining cotton yarn properties using artificial intelligence techniques is studied. The advantages of this approach are presented. On the basis of experimental results the usage of machine learning and expert systems in the domain of spinning textile technology is estimated and sugges- tions for further research work are given. Kevwords: Textile industry, Spinning technology, Cotton yarns. Properties prediction, Artificial intelligence, Ma- chine learning. 1, Introduction In a spinning mill, production and quality are of equal importance. Also a uniform, permanent and high quality of spun yarn is gaining impor- tance for the success of textile plant. The quality of spun yarn is influenced by the raw material (cotton fibers), the technological procedure, char- acteristics and adjustments of machinery and common manufacturing conditions (temperature, humidity, uncleanliness). Nowadays many enhanced procedures for estimation of mechanical and chemical properties of textile fibers are known. The off-line test meth- ods for yarn quality and the quality of woven and knitted fabrics have been also very improved. For estimation of the mixing proportions the price of the mixture components is also of great impor- tance. The relationship between quality parameters of cotton fibers and construction and other proper- ties of produced cotton yarn is determined by standard equations which contain many experi- mentally determined constants. The yarn strength is of great importance for estimation of the quality properties and characteristics of cotton yarn. By using longer, finer and stronger cotton fibers we can get yarns with greater strength. Combining the quality parameters of cotton fibers, we can predict the yarn strength much more exactly. The aim of the present work is the determina- tion of the connection between the quality param- eters of the cotton fiber mixture and the quality properties of the produced yarn. Not only well known experimentally determined formulas and regression equations are considered but also the expert knowledge, the experience of textile special- ists and results of testing of former spinning series. In the first phase of research and development work, The Assistant Professional Expert System Shell [1] has been used. As a final result of this work we expect to determine an efficient and accurate method of yarn quality prediction considering the cotton fiber quality parameters and the influence of mixture composition. The aim of this kind of work would also be the confirmation or improvement of old experimentally determined spinning equations. 2. Cotton fiber properties In today's textile climate, knitting and weaving machinery are operating at higher speeds. Fewer units are producing more fabric, so downtime 0166-3615/91/$03.50 ,~Zi 1991 - Elsevier Science Publishers B.V. All rights reserved

The prediction of cotton yarn properties using artificial intelligence

Embed Size (px)

Citation preview

Page 1: The prediction of cotton yarn properties using artificial intelligence

Computers in Industry 17 (1991) 217-223 217 Elsevier

IMS '91 Learning in IMS

The prediction of cotton yarn properties using artificial intelligence

Zoran Stjepanovi~ and Anton Jezernik

University of Maribor, Faculty of Technical Sciences, Smetanova 17, 62000 Maribor, Slovenia, Yugoslavia

The possibility of determining cotton yarn properties using artificial intelligence techniques is studied. The advantages of this approach are presented. On the basis of experimental results the usage of machine learning and expert systems in the domain of spinning textile technology is estimated and sugges- tions for further research work are given.

Kevwords: Textile industry, Spinning technology, Cotton yarns. Properties prediction, Artificial intelligence, Ma- chine learning.

1, Introduction

In a spinning mill, production and quality are of equal importance. Also a uniform, permanent and high quality of spun yarn is gaining impor- tance for the success of textile plant. The quality of spun yarn is influenced by the raw material (cotton fibers), the technological procedure, char- acteristics and adjustments of machinery and common manufacturing conditions (temperature, humidity, uncleanliness).

Nowadays many enhanced procedures for estimation of mechanical and chemical properties of textile fibers are known. The off-line test meth- ods for yarn quality and the quality of woven and knitted fabrics have been also very improved. For estimation of the mixing proportions the price of the mixture components is also of great impor- tance.

The relationship between quality parameters of cotton fibers and construction and other proper- ties of produced cotton yarn is determined by

standard equations which contain many experi- mentally determined constants. The yarn strength is of great importance for estimation of the quality properties and characteristics of cotton yarn. By using longer, finer and stronger cotton fibers we can get yarns with greater strength. Combining the quality parameters of cotton fibers, we can predict the yarn strength much more exactly.

The aim of the present work is the determina- tion of the connection between the quality param- eters of the cotton fiber mixture and the quality properties of the produced yarn. Not only well known experimentally determined formulas and regression equations are considered but also the expert knowledge, the experience of textile special- ists and results of testing of former spinning series. In the first phase of research and development work, The Assistant Professional Expert System Shell [1] has been used.

As a final result of this work we expect to determine an efficient and accurate method of yarn quality prediction considering the cotton fiber quality parameters and the influence of mixture composition. The aim of this kind of work would also be the confirmation or improvement of old experimentally determined spinning equations.

2. Cotton fiber properties

In today's textile climate, knitting and weaving machinery are operating at higher speeds. Fewer units are producing more fabric, so downtime

0166-3615/91/$03.50 ,~Zi 1991 - Elsevier Science Publishers B.V. All rights reserved

Page 2: The prediction of cotton yarn properties using artificial intelligence

218 IMS "91--Learning in 1MS ('omputers in lndust(~

becomes more costly. Because of this, the yarn manufacturing process has to deliver optimum-qu- ality yarn to the fabric forming processess [2].

Traditional methods of measuring cotton qual- ity are no longer adequate for modern informa- tion-based manufacturing systems. Zellweger Us- ter's Advanced Fiber Information System, for ex- ample, provides information on neps and fiber length [3]. Its modularity and basic data yield distributions of measured parameters, not merely averages. It measures fibers in an individualized state. The Microdust and Trash Monitor, on the other hand, provides information on the non-lint content of fiber, specifically trash, microdust and fiber fragments. Another fiber testing instrument comes from Motion Control, Inc. [4]. For the cotton merchant it is important to know the qual- ity of raw material. All bales can be identified by their fiber properties, especially strength, for proper distribution to mills and accurate fulfill- ment of contracts. HVI testing instruments enable testing results of the most important properties of the cotton fiber: micronaire (fineness), length, strength, elongation, uniformity, color and trash content. Other important fiber properties which can be used as input data for prediction of yarn properties are classer's length, fiber maturity and content of honeydew.

Zoran Stjepanovi~ finished his study at the Mechanical Engineering De- partment, Textile Section at the Fa- culty of Technical Sciences, University in Maribor. He received the MSc de- gree in 1989 at the University of Maribor. He is working as senior re- search engineer at the Institute of Textile and Apparel Processes. His re- search interests include the use of computer-aided techniques in textile industry, especially in cotton spinning plants, knowledge acquisition and ex- pert systems.

Anton Jezernik finished his studies at the Faculty of Mechanical Engineer- ing, University of Ljubljana in 1962. Later he completed the MSc degree in 1968 and PhD degree in 1971 at Im- perial College of Science and Technol- ogy, University of London. After nearly three years' stay at Electricity Generating Industry in UK, he has been working since 1973 at the Uni- versity of Maribor, Faculty of Techni- cal Sciences. At present he is professor

- - of Mechanical Engineering and head of the Institute for Structures and Engine Design. His research interests include FEM, CAD, CIM and AI methods.

3. C o t t o n yarn p r o p e r t i e s

Yarn is generally defined as a structure of circular form, built up either of staple fibers, mostly from 20 up to 200 mm long, or of continu- ous filaments, the consistent fibers in either case being bound to one another by twist. The yarn may be single, if it is onefold, or twisted, if more singles are folded and bound together by twist [5].

3.1. The diameter and density o! yarn

Consequently to twisting, the fibers are sub- jected to a complex of strains, because the yarn being spun is simultaneously twisted, bent and axially stretched. This fact makes an analysis of the phenomena not at all easy. Diameter and density of yarn are influenced by the fineness and rigidity of the fiber used in the spinning process. If yarns are spun to the same count and at the same twist factor, then the use of coarser fibers or fibers of greater rigidity wilt bring out a larger yarn diameter, although the count of the yarn remains unchanged. Coarser or more rigid fibers resist more to bending while being twisted into yarn, therefore the result is a larger diameter of yarn. Fiber length on the other hand, also in- fluences the diameter and density of yarn. The spinning of longer fibers into yarn involves a higher density of the yarn, hence the diameter of the yarn spun of long fibers is smaller than that of short fibers, although the count and twist are the same. The way the fibers have been prepared for spinning is also of considerable importance. The more parallel the arrangement of fibers, the more compact is the yarn. Finer yarns have always a greater density, because of a better arrangement of fibers.

3.2. The strength of the yarn

Mechanical properties of a yarn spun from staple fibers constitute one of the most difficult fields for an analytical research on yarn structure. The main assumption has been that the yarn was of an ideally regular fineness and composition of fiber lengths. Consequently to that it is under- stood that all the considerations refer to a hypo- thetical sliver in the cross section of which there are as many fibers as in the mean. furthermore

Page 3: The prediction of cotton yarn properties using artificial intelligence

Computers in Industry Z. Stjepanot:i?, A. Jezernik / Prediction of cotton yarn quality 219

their length distribution reflects the proportion of particular lengths in all the yarn [5].

It is a fact that the strength of the yarn depends on the strength of the assembled fibers. Therefore we can account for the strength by introducing a factor representing the effect of the number of fibers. The magnitude of this factor has been found experimentally. Such an approach to the problem is represented by some researchers [6,7].

3.3. The elasticity of the yarn

The elasticity of yarn spun from continuous filaments is troublesome enough if one tries to define it analytically, but the elasticity of staple fiber yarn is even more troublesome in that re- spect. Fibers of such a yarn play with their locked parts the same role as continuous filaments, so the elasticity of yarn spun from staple fibers must be lower than that of yarn spun from similar continu- ous filaments.

3.4• The irregularity of the yarn

Yarn unevenness is defined as the irregularity of the linear density along its length. Modern testing instruments measure linear density of the yarn both during production (on-line) and in the laboratory (off-line) over sufficiently long sample lengths. The outcomes of the data processing re- veal different kinds of unevenness [8]: - overall or general unevenness (U or CV); - periodic unevenness (correlation and spectral

functions); length-specific unevenness (variance-length function), and

- nodular unevenness (imperfections and faults).

3.5. Imperfections in cotton yarns

Imperfections in yarns are serious faults that greatly detract from the quality. They are classi- fied into three categories according to their linear density variations: thin/ thick places and neps [9].

3.6. Other important yarn properties

It is a fact that the strength (and its variation), elasticity (and its variation), irregularity (yarn number variations) and imperfections (thin/thick

places, neps) are the most important properties of cotton yarn. But there are several more yarn prop- erties that should be taken into consideration: yarn hairiness, yarn bulk, yarn twist (and its varia- tion) and visual egality of the cotton yarn.

4. The processing stages of cotton spinning

The aim of the spinning mill should be to produce yarn with as little outlay as possible, but also yarn which exhibits a suitable and constant quality [10]. The quality of a yarn is very much dependent on the quality of raw material and the quality of the material being produced at each stage of processing prior to spinning. Quality downgrades in the spinning preparation which are not determined by the last pass of drawing, at the latest, cannot be corrected anymore, so they can result in high costs.

The carded cotton spinning production process is composed of the following stages [11]: - cleaning, opening and mixing of material (open-

ing pressed bales into tufts, intensive cleaning and blending of fibers); machinery: automatic bale opener, cleaning, blending and feeding machines;

- c a r d i n g (intensive opening and cleaning of fibers, web and sliver formation): machinery: cards;

- d r a w i n g , two passes (equalizing the sliver, parallelizing the fibers, blending the raw material); machinery: draw frames;

- roving (roving formation by drawing and pre- twisting); machinery: roving frame;

- spinning (yarn formation by drawing and twist- ing); machinery: ring or OE-rotor spinning machines;

- cone winding (cleaning and winding of the yarn); machinery: cone winding and cleaning mac- hines.

Basically, each production position can be consid- ered as a potential source of disturbance. In a spinning plan we use the combination of on-line and off-line quality assurance, which ensures that the process faults are detected at an early stage, so that the amount of downgrade material can be reduced.

The plant, from which we have collected the testing data, has both on- and off-line quality control and well adjusted production machines, so

Page 4: The prediction of cotton yarn properties using artificial intelligence

220 1MS "91--Learning in I M S

we can assume that the influence of machinery the yarn quality could be neglected.

5. Prediction of yarn properties

5.1. Conventional method

o n

The most important is the prediction (calcula- tion) of yarn strength. The commercially produced yarn we have to do with, has in reality an irregular fineness as well as various structural deficiencies. As far as strength of cotton yarn is considered, Solovev's formula [6] is mostly used.

The general formula for the strength of yarn looks as follows:

&= Sfk~kvk,,

where Sv is the strength of the yarn (in grams per tex), Sf is the strength of fibers (in grams per tex). k~ is the coefficient of twist, k , is the factor of fiber and yarn fineness irregularity, k I is the fac- tor of fiber length irregularity.

Solovev's formula is a combination of theoreti- cal considerations with empirically established factors. Strength of cotton yarn is accordingly a function of fiber properties, as well as yarn parameters, having the following form:

Sy= Sf 1 - 0 . 0 3 7 5 H 0 - 2 . 6 5 ~ / ~ )zK,,v,

where Sy is the fiber strength (g tex 1), Tty is the linear density of yarn (tex), Ttf is the linear den- sity of fiber (tex), H 0 is an index of the techno- logical process ( H 0 = 3.5 to 4.0 for combed cotton systems, 4.5 to 5.0 for carded cotton systems, and 3.0 to 4.0 for man-made staple fibers), v is a correction factor for the quality of equipment ( I ,=0 .95 to 1.10; for normal machinery 1,= 1), K,~ is a correction factor for twist (ranging from 0.70 to 1.00), and z is a correction factor for fiber length, calculated from the formula:

42(1+ z . = l - -

lk -- 1 6 0 / B

where l k is the classer's length, and B is the fiber base.Similar equations are used to calculate the elasticity and irregularity of the spun yarn.

~i :)m:~uler~' in lnduxtrv

5.2. G~'ing artificial intelligence prmcipies

In the literature on artificmJ intelligence, we can find several definitions for an expert system. It is a program that behaves like an expert system in some, usually narrow, domain of application [12]. Not every knowledge-based system could be considered an expert system, because an expert system also has to have the ability to explain its behaviour and its decisions to the user [12].

A well established method of learning in artifi- cial intelligence is learning from examples, also called inductive learning [1121. tn this area many good results have been produced. Machine learn- ing techniques have been applied to many prob- lems dealing with the learning concepts from ex- amples. The most obvious kind of application is in association with knowldege acquisition for expert systems, thus helping lo alleviate the knowledge- acquisition bottleneck.

5.2, I. Assistant Professional Assistant has been developed in 1985 by {

Kononenko and I. Bratko at the Faculty of Elec- trical Engineering in Ljubl jana The software was running on DEC-10 computer under the TOPS-10 operating system. In 1987 an implementation for the personal computer under the MS-DOS operat- ing system was developed by B. Cestnik a', the Jozef Stefan Institute in Ljubljana I1].

Assistant Professional is intended to help solv- ing decision problems that are hard or impossible to solve by hand, It is an environment for auto- mated knowledge acquisition for expert systems, The induction of decision trees from examples i>, the main vehicle for automatic construction of ~ knowledge base. Assistant Professional has a facil.- ity to use generated rules for solving new decision problems, so it can be used as an automatic expert system generator [4]. As described in numerous papers, e.g. [13,14], Assistant Professional has been successfully applied to many decision problems, such as:

medical diagnosis (primary tumor location, breast cancer recurrence, thyroid diseases, diag- noses in rheumatology etc.): economics prediction (problems of forecasting the trends in a certain index about the success of the company, automatic generation of prog- nostic rules);

.... weather prediction,

Page 5: The prediction of cotton yarn properties using artificial intelligence

Computers m lndustrv

Table 1 Number of classes, attributes and examples of two spinning domains

Domain No. of No. of No. of classes attributes examples

Fiber quality (cotton carded processl 5 12 170

Yarn quality (cotton carded process, Tt = 30 tex} 5 12 242

- quality control (paper industry: prediction of paper quality with respect to the process parameters; steel industry: estimation of steel quality of steel after one stage in the process of steel production by generation of a decision tree from the data collected in the past).

Assistant belongs to the learning systems which use the TDIDT technique (Top Down Induction of Decision Trees). The basic algorithm of ID3 for constructing a decision tree is as follows [3]:

if all the learnig examples belong to a single class then terminate with a leaf labelled with that class

else begin

on the basis of the learning set choose the most informative attribute (using the entropy mea- sure) for the root of the tree and partition the

Z Stjepanooi{, A. Jezernik / Prediction of cotton yarn quali(v 221

learning set into subsets according to the values of the selected attribute; for each value do

recursively construct a subtree with the corresponding subset of examples

end.

The improvements of the ID3 basic algorithm in assistant Professional are described in [12-14].

6. The data

We took the data for training/learning and for testing of the knowledge based system from a big spinning plant in Slovenia (Tables 1 and 2). We decided to take one production line of carded cotton yarn. The production process and its stages are shortly described in Section 4.

The structure of data needed for the planning/ production modules of the spinning process is shown in Fig. 1 [15].

7. Discussion of results and conclusion

Artificial intelligence methods and especially expert systems can be used to anlyze any practical problem lying within defined parameters and to provide solutions that hitherto required the capa- bilities of human experts. To make it possible, the use of specified principles and appropriate data and knowledge bases is required.

Table 2 The structure of attributes of fibers and yarn

FI BRES YARN

Attribute Type[No] Attribute Type[No]

1. Strength (g tex 1) CONT[20] Strength (cN) CONT[20] 2. Mean staple length (mm) CONT[20] Strength variation (%) CONT[20] 3. ('lasser's length (mm) CONT[20] Count/fineness (text) CONT[20] 4. % of fiber under 10 mm (%1 CONT[10] Count variation (%) CONT[20] 5. % of fiber above mean length (%) CONT[10] Elongation (%) CONT[20] 6. Pressley index (lbs mg 1) CONT[20] Twist (t m 1) CONT[20] 7. Fineness (Micronaire) CONT[20] Twist variation (%) CONT[20] 8. Content of honeydew (%) CONT[20] Thin places CONT[20] 9. Color DIST[5] Thick places CONT[20]

10. Maturity (%) CONT[20] Neps CONT[20] 11. (;lass of fibres DIST[4] Irregularity (CV%) CONT[20] 12. Trash content CONT[10] Visual egality DIST[7]

Page 6: The prediction of cotton yarn properties using artificial intelligence

2 2 2 1MS "91--Learning in 1MS Computers in Industry

__[~]1 - l b r e length

- bre

- f i b r e t r e n t

- f i b r e m a t u r i t y

- w a s t e content

- c l s~fl-

- m o i s t u r e c o n t e n t

- e r c e p t - ~e o

~ o r t ~ r ~ l b e r

P R O P O R T I ON l]

A !! x i[

c × I I: D X ~i

. . . . . . . . 'I .~×TuREII ....

L K N F I I G Y - S P I N N I N G

, P R O C E S S

~ORK I . . • i

TECHNOL W A < ! ' F

DATA F R O M S T O R E ~iO ~

A B A U T F I l E OUAN'IIT v .L"

RAW MATERIAL

PRODUCED YARN

- y a r n .~ .I:,L

- m r f c t lor~- p l a c e a , n e D ~

~ s D e c t r o ~ r a ~

~d.agram

na.~ines~

~ b l e a k ~ n g e l o n ~ a t , , ~ ~

x a r n t w , s ~

-->-' - n Fault,- OUT Y~assim~tT

-enG breaK~

--vat . ~nc length ~ u : v ~ .

WOVEN CLOTH

-Drea~n< ,

vJ~u~

e g a l i : . .

- w e t f a u : t s a u s e a p ~ ~ m p e r £ e < _ . , ' - .

P R O D U C I O N

- y a r n

PLANNING MODULE properties prediction

- spinning schedule

- m a c h i n e r y p l a n

- p r o d u c t i o n s c h e d u l e

- m a c h i n e s p e c i f i k a t i o n s - e x p e n d i t u r e

F i g . 1. Model of the sp inning process and the c o n n e c t i o n of i n p u t / o u t p u t data for p l a n n i n g / p r o d u c t i o n modules .

The knowledge-based system for the prediction of yarn properties could be both a decision-mak- ing instrument and a reference tool for yarn pro- ducers.

In the early stage of research and development work, the system enables the user to predict the quality of carded cotton yarns by means of predic- tion of main quafity parameters: strength, elastic- ity, irregularity (yarn number variations), imper- fection indicator (the number of thick/thin places

and neps per 1000 m of yarn) and the visual egality of the yarn.

Next the testing values are being compared with the standards. On that base the quality of the yarn that will be produced is predicted. It is divided into 5 classes: excellent, very good, good, medium good and poor quality. 242 sets of testing values (12 attributes) were used for the knowl- edge-based system--about 70% of them for learn- ing and 30% for testing. The classification accu-

Page 7: The prediction of cotton yarn properties using artificial intelligence

Computers" in lndusto' Z. Stjepanovi?, A. Jezernik / Prediction of cotton .yarn quality 223

racy, c o m p a r e d to the real d a t a of f in ished pro-

duc t ion , was nea r 65%. Th e resul ts c o n f i r m tha t

we can get be t te r y a r n qua l i ty by us ing finer, longer and s t ronger c o t t o n fibers.

In order to i m p r o v e the accuracy a n d i n f o r m a - t ivi ty of the system, we i n t e n d to supp ly even

more examples (for t r a i n i n g / l e a r n i n g a n d for test- ing tasks). We bel ieve that by a d d i n g a larger

n u m b e r of va lues we can m a k e the system m u c h more accura te a n d reliable.

Acknowledgement

W e wish to express our t hanks to prof. dr. I van Bra tko for useful sugges t ions r ega rd ing the devel-

o p m e n t of the mode l of the k n o w l e d g e - b a s e d sys- tem for p red ic t ion of y a r n qual i ty .

References

[1] I. Kononenko, B. Cestnik and I. Bratko, Assistant Profes- sional User's" Guide, Jozef Stefan Institute, Ljubljana, Slovenia, 1988.

[2] W.N. Rozelle, "Technology updates keep mills on top of the latest", Text. World, November 1990.

[3] "Uster Polyguard Q-Pack', Uster News Bull., No. 37, October 1990, Zellweger Uster, Switzerland,

[4] Peyer, "Fiber fineness and high volume testing instru- ments from Motion Control, Inc.", Dallas, TX, 1987.

[5] W. Zurek, The Structure of Yarn, US Department of Commerce, National Technical Information Service, Springfield, VA, 1975.

[6] A.N. Solovev, The Prediction of Yarn Properties in Cotton Processing, Moskva, 1958.

[7] V.A. Usenko, The Use of Staple Fibers in Spinning, Moskva, 1958.

[8] M.I. Zeidman, M.W. Suh and S.K. Batra, "A new per- spective on yarn unevenness: Components and determi- nants of general unevenness", Text. Res. J., January 1990, pp. 1 6.

[9] A.R. Padmanabhan and A. Balasubramanian, "An ex- ploratory study of imperfections in cotton yarns", Text. Res. J., January 1990, pp. 17 22.

[10] H. Luechinger, "Quality in yarn production", Int. Text. Bull., January 1990 (in German).

[11] Documentation of Short-Staple Spinning, Rieter Machine Works Ltd., Winterthur, Switzerland.

[12] I. Bratko, Prolog Programming for Artificial Intelligence, Addison-Wesley, Reading, MA, 1990.

[13] I. Bratko and I. Kononenko, "Learning diagnostic rules from incomplete and noisy data", in: Bob Phelps (ed.), Interactions in Artificial Intelligence and Statistical Meth- ods, Unicorn Seminar, London, December 1986.

[14] B. Cesmik, 1. Kononenko and I. Bratko, '~ASS~STANT 86: A knowledge elictitation tool for sophisticated users", in: I. Bratko and N. Lavrac (eds.), Progress in Machine Learn- ing, Bled, May 1987, Sigma Press, Wilmslow, UK, distrib- uted by Wiley.

[15] Z. Stjepanovic, "The model of computer aided process planning and production control in a spinning mill", MSc Thesis, University of Maribor, Faculty of Technical Scien- ces, Maribor, Slovenia, 1989 (in Slovene).