235

Textile Research Journal

Article

Predicting Seam Appearance Quality Abstract

The appearance of a garment is affected by the quality of the fabrics used in its manufacture, as well as a number of factors determined by the technology of the garment manufacturing process. Since fabric quality, as the most important element of garment appearance, is determined by its mechanical properties, it is obvious that these properties directly impact fabric processing properties. It can be seen through various forms of fabric behavior under the loads that occur in sewing. Investigations of the correlations of the stress and fabric behavior are aimed at constructing a system to predict fabric behavior in garment manufacturing processes, as well as to predict the appearance of the garment to be manufactured. The investigation presented here deals with the impact of fabric mechanical properties on the quality of seam appearance, as defined by seam puckering and work-piece flotation. Machine learning methods included in the Orange software package were used to establish the importance of mechanical properties with respect to fabric behavior.

Daniela Zavec Pavlinic1 and Jelka Geršak Department of Textiles, Faculty of Mechanical Engineering, University of Maribor, Institute of Textiles, SI-2000 Maribor, Slovenia

Janez Demšar and Ivan Bratko Artificial Intelligence Laboratory, Faculty of Computer and Information Science, University of Ljubljana, SI-Ljubljana, Slovenia

Key words garment engineering, mechanical properties, garment manufacture, seam quality, prediction, machine learning methods

Contemporary garment manufacturing asks for the application of new technologies and usage of increasingly more demanding fabrics. The fabrics often present serious problems in manufacturing, primarily due to the complex properties dictated by ever-changing fashion trends. Fabrics are key components of the garment to be made, and differ by mass, raw material content and construction parameters. Fabrics exhibit even more diversity when their mechanical properties are investigated in relation to load–deformation behavior. Loads occurring in fabric processing result in complex mechanisms of non-linear behavior, determined by the inhomogeneity of the fabric structure. A high degree of deformation is often the result of even quite low loads. Fabrics are exposed to various levels of load in transforming them from a two-dimensional form into the three-dimensional (3-D) form of an article of clothing. It is extremely

Textile Research Journal Vol 76(3): 235–242 DOI: 10.1177/0040517506061533

important to be familiar with the loads the fabrics are exposed to and the deformations that are the response of the fabric mechanical properties to the loads. It is thus possible, knowing the fabric mechanical properties at a particular load level, to predict its reaction behavior, investigate interactions of the parameters and their correlations, as well as to predict potential problems in the garment manufacturing processes.1 The investigations described here present the continuation of the development of the system for predicting seam appearance quality. The impact of fabric mechanical property parameters on the seam appearance quality is investigated, as it is the basis of a prediction system. This system

1 Corresponding author: tel.: +38622207970; +38622207960; fax: +38622207990; e-mail: [email protected]

www.trj.sagepub.com © 2006 SAGE Publications

TRJ

236

Textile Research Journal 76(3)

is aimed at offering predictions for the behavior of a new, carefully selected fabric, for which mechanical properties have been previously determined. Having proper information on the behavior of the fabric selected in processing, potential problems can be eliminated in advance, which will also contribute to the final appearance of the garment manufactured [1]. To establish attributes that impact the quality of the seam made, as well as further predictions of seam puckering and flotation of the work-piece, we applied methods of machine learning from learning examples. We primarily used the methods of regression trees and k-nearest neighbors (k-NN), i.e. the distance weighted k-nearest neighbors [2]. Regression trees induced from data define mappings from the attributes’ values to possible predictions. On the other hand, the method of distance-weighted k-NN respects the importance of the distance of the query point to the nearest examples. This means that nearer examples are more important for the prediction of the class value than those faraway in the attribute space. The k-NN method has proved to be more adequate for our prediction task. Predictions are thus made by the distance weighted k-NN. We used an implementation of k-NN included in the Orange software package, developed by the Artificial Intelligence Laboratory, Faculty of Computer and Information Science, University of Ljubljana, Slovenia [3]. During the study a high degree of correlation was found between the important attributes selected by the above machine learning methods, and the attributes suggested for this prediction task by an expert in the field of garment engineering.

Theoretical Basis Impact of Fabric Mechanical Properties on Seam Quality An article of clothing is an unmistakable part of the wholeness of an individual, complementing his/her personality. In addition to fulfilling the functions of use, it should be of an appropriate visual appeal, which is in itself a result of numerous factors occurring in the garment manufacturing process, including the mechanical properties of the fabrics used. The quality of a garment as a whole can be described as a harmony of the fabrics incorporated and the properties of these fabrics, divided into basic, end-use and processing properties, as well as the techniques and mechanisms employed in processing. These can have a serious impact, together with the fabric mechanical properties, on the fullness of form of the work-piece, proper drape and the fitting of the garment to the body contours [4]. One of the key factors impacting the quality of garment appearance is the visual appearance of the seams. This primarily depends upon fabric behavior in sewing and upon

the technique used in re-shaping the fabric into an article of clothing, as well as upon sewing processing parameters. The fabric to be sewn is affected by the technology-determined forces of the sewing machine on one hand, and on the forces depending on the process of processing the fabric. Fabric behavior in sewing, i.e. its resistance to the loads at the point of sewing, sewing needle penetration force and the force originating at the toothed feed element, is primarily affected by fabric elongation, surface and longitudinal compression properties of the fabric processed [5]. On the other hand, the work-piece is exposed, in processing the fabric into a 3-D shape, to more or less intensive tension, bending and shear loads, as well as to compression. Fabric behavior in processing depends upon its formability, bending rigidity, fabric elasticity and relaxation ability, shrinkability, shear rigidity, shear hysteresis and geometrical roughness. Formability is a specific fabric property, defined as the ability of the fabric to be re-shaped from a plane into the 3-D form of an article of clothing. It is based on the maximum longitudinal compression that the fabric can stand before buckling occurs [6]. Tension properties at low loads were used to describe longitudinal compression [4, 7]. They define fabric formability (F in mm2) as a ratio of fabric elasticity (EL = EMT/LT in %), bending rigidity by a length unit (B in cN cm2/cm), shear rigidity (G in cN/cm °) and shear hysteresis (2HG5 in cN/cm) at the shear angle ± 5° (equation (1)): EMT G F = ------------------- ⋅ B ⋅ ---------------F m ⋅ LT 2HG5

(1)

where EMT is extension in % achieved at maximum load; Fm = 490.35 cN/cm; and LT is linearity [8]. Seam appearance quality is also affected by the ratio of fabric extension weft wise and warp wise, denoted as α (alpha = EMT2/ EMT1). For ladies’ outerwear this value approaches 1. Seam puckering can occur with fabrics that exhibit values of the ratio described above that are either too high or too low.

Seam Puckering Investigations of the quality of seam visual appearance from the point of view of fabric mechanical properties, have resulted in the awareness that seam puckering occurs in most cases because yarns in the fabric weave are pushed aside, as the sewing needle with the thread inserted pushes warp and weft yarns aside at each puncturing, which results in certain degree of tension due to the fact that they have to take up new position in the fabric. The degree depends on the structure, construction and fineness of the fabric, its mechanical properties, sewing needle gauge and stitch length. When the sewing needle leaves the work-piece in the course of stitch formation, total or partial relaxation occurs of the warp and weft yarns in the area of the punc-

Predicting Seam Appearance Quality D. Z. Pavlinic et al.

237

Figure 1 Seam puckering as a detrimental factor of garment appearance.

turing aperture. It depends upon the elastic properties of the warp and weft yarns, since, due to abrasion forces in the interlocking area of the needle and catcher thread, the work-piece compacts in the area of the puncturing aperture at the moment of tightening the stitch. The deformation created in this way, which is reduced in the direction of the stitch center, can cause additional tension in warp or weft yarns. During repeated puncturing, the sewing needle must make its way between individual warp and weft yarns, and the yarns are again pushed aside. As in sewing needle puncturing the puncturing point is changed at each stitch formed, in principle by the stitch length, namely 2 to 5 mm, the same system of the warp and weft threads is established at the distance of the technologically determined puncturing point, e.g. stitch length. Due to repetitious pushing of the warp and weft yarns, construction deformation occurs in the textile surface. If the tension caused is higher than the margin of elasticity they are seen as plastic deformation of the fabric [9]. The deformation is reflected as seam puckering and has a detrimental impact on the quality of the seam, as well as on the appearance of the article of clothing as a whole (Figure 1). The fabrics in plain weave are more prone to seam puckering, due to higher density of warp and weft yarns, which cause higher values of shear properties, primarily shear hysteresis (2HG) at the shear angle ± 0.5°.

ric extension in the area of the seam, which is reflected as a gain in longitudinal length. Seam flotation is a deformation of the seam that occurs because of the interaction of the shear load (affecting the fabric through the sewing thread) and fabric extension in the area of the seam [10, 11].

Statistical methods Fabrics respond in different manners to the loads occurring in garment manufacturing processes. It is extremely important to have an understanding of the interactions of the fabric mechanical properties, to be able to study and understand properly the key parameters of these properties. Fabric behavior in sewing is, from the point of view of the aesthetics of seam appearance, reflected in its smoothness. The seam that is not smooth is either puckered or floated [11]. The degree of correlation between seam puckering and flotation can be measured by the Pearson’s correlation coefficient, as shown in equation (2) [12]:

∑  X – X  Y – Y ---------------------------------------------------------n

r =

i=1

i

( n – 1 )S x S y

i

(2)

where Xi and Yi are the values of variables X and Y in experiment i, X and Y are their average values, Sx and Sy are their standard deviations and n is number of samples.

Seam Flotation Investigations of the seam flotation, from the point of view of fabric mechanical properties, are affected by the values of bending and shear properties, the formability of the fabric in question, as well as by extension. Low bending rigidity is detrimental to the seam appearance because of the resistance the fabric offers to the bending. However, too low formability prevents the fabric from adapting to the forces occurring within it. The seam, as a joining element, causes fabric shearing deformation in the area of the seam, due to the mass of the thread interlocked in the stitch. The shear deformation of the fabric caused in this way results in fab-

Machine Learning Methods Machine learning is an area of artificial intelligence that includes procedures that, from a number of learning examples from a problem domain, discover new knowledge on the topic. This knowledge, represented by a formal model, is obtained from a kind of analysis and generalization of the learning examples. The knowledge can be used to process new, previously unseen, examples. Often such discovered knowledge itself is of interest to domain experts. Procedures of machine learning can be used to discover

TRJ

TRJ

238

Textile Research Journal 76(3)

previously unknown regularities within a set of given data. This is often referred to as data mining, or KDD – knowledge discovery in data bases. In this way the rules, previously unknown to the expert in the field, but quite meaningful, can be discovered and statistically verified for validity. A model thus obtained can also be tested on some other examples that differ from those used for inducing the model. The validity of induced models is typically measured by their classification accuracy, sensitivity, etc. Other qualities of a model are also of interest: its interpretability from the expert’s point of view, its brevity, etc. In attribute-value machine learning, each learning example is described in terms of values of attributes and the class it belongs to. Class can be categorical (in classification problems) or numerical (in regression problems). In the case presented here, the learning examples are the individual fabrics; the attributes are the measured parameters of mechanical properties, while the associated classes are the various responses of the fabrics behavior. The task of machine learning is to find a model, i.e. function, which could be used to predict the response of new fabrics. Two methods of machine learning have been used for the purpose in our study: regression tree learning, and learning by the k-nearest neighbors method (k-NN). The classification trees method was also used in preliminary experiments, as well as naive Bayesian classifier and locally weighted regression, which all proved to yield significantly less accurate results than regression trees and k-NN. Modeling by regression trees is based on recursive partitioning of the example set into ever-smaller subsets, until the variance in the class values obtained in the subsets is below a specified threshold [13]. The division is based on the values of attributes, chosen by heuristic that is supposed to lead to smaller and (according to the holy grail of machine learning, the Ockham’s principle) more accurate trees. The induced regression trees can be shown to a domain expert, who can then decide whether they agree with the existing theoretical knowledge. K-nearest neighbors method (k-NN) predicts the outcome for an example by computing the average outcome over k learning examples most similar to the query example, the so-called “nearest neighbors” [2]. Neighbors’ influence can also be weighted with regard to their distance from the query example. Unfortunately, k-NN models cannot be visualized, which makes their interpretation very difficult. In order to gain some insight into such k-NN models, we conducted a series of tests in which k-NN selected the most appropriate attribute subset for predicting a specific response. Using such a randomized procedure, we were able to statistically assess the importance of individual attributes by counting the number of trials in which they were selected. To evaluate the compliance of the model with the expectations made by experts, we compared the attribute selections by k-NN with those made by a domain expert. Statistical comparison of the attribute rankings by k-NN and a human

expert was conducted using the Spearman’s rank-correlation coefficient [12]: N

∑

2

6 Di i=1 ------------------------ρ = 1– N  N 2 – 1  

(3)

where Di is the difference between the two ranks of the corresponding values of two variables and N is the number of pairs of values. To assess the predictive accuracy of the induced regression trees and the k-NN classifiers, standard ten-fold cross validation was used. Examples were randomly divided into ten folds and each method was tried for ten times, where the examples from nine folds were used for learning and the remaining fold was used for testing. The accuracy of the models induced was measured by the root of mean squared error (RMSE), a standard measure for the evaluation of regression models (equation (4)):

N

RMSE =

2

 x i – p i   i=1 -----------------------------N

∑

(4)

where the pi is a predicted value; the xi is the corresponding measured value; and N is the number of examples. All the experiments were conducted using the component-based machine learning system Orange [3]. It includes a range of preprocessing algorithms (feature subset selection, categorization, and feature utility estimation for predictive tasks), modeling techniques (classification and regression trees, naive Bayesian classifier, k-NN, majority classifier, support vector machines, linear and logistic regression, ensemble methods), validation procedures (cross-validation, random sampling, leave-one-out) and visualization methods.

Methodology The research described here was aimed at investigating the impact of parameters of fabric mechanical properties on seam puckering and flotation of work-piece components in the area of the seam. The aim was to construct predictive models that can be used to predict the quality of the seams for a new randomly selected fabric, provided its mechanical properties are known.

Predicting Seam Appearance Quality D. Z. Pavlinic et al.

239

Figure 2 Development of prediction system.

Learning Examples

Scheme of Research

In the process of machine learning described, the attributes were the parameters of mechanical properties for the 489 fabrics, intended for the manufacture of ladies’ outerwear, that were analyzed. The fabrics were selected in the course of the garment manufacturing process. They differed in raw material content, mass, weave and construction parameters. The parameters of fabric mechanical properties were determined in the area of low loads, using the KES-FB measuring system at standard measuring conditions [8]. The classes in the learning subset were the fabric processing properties, i.e. the fabric’s responses to sewing, which depend on the fabric’s mechanical properties. The fabric’s responses to sewing that impact directly on the quality of garment appearance are seam puckering (s1) and flotation of work-piece components (s7). A control system, based on determining the most important fabric behavior elements for seam quality, was constructed for the purpose of determining fabric response behavior, from the point of view of seam quality. Each element was analyzed regarding its impact on appearance quality, e.g. smoothness of the seams, as well as regarding the degree of roughness (non-smoothness), e.g. seam puckering. This was performed by following the regulations of the AATCC standard, as was the case for the degree of seam flotation [14]. The grades were divided into five classes [11]: 5, high quality seam appearance; 4, good appearance, insignificant seam puckering or flotation; 3, acceptable appearance, noticeable puckering or flotation; 2, below average appearance, significant puckering or flotation; 1, poor appearance, unacceptable puckering or flotation of the seam.

The development of the model for predicting seam quality was based on learning examples. The model should identify the attributes that affect individual fabric behavior response in sewing, and predict seam puckering and flotation of work-piece components in the area of the seam. This is one of the criteria of assessing the quality of garment appearance. The overall picture of the development a system for predicting seam quality can be seen in Figure 2.

Results and Discussion Development of the system of predicting seam quality, namely seam puckering (s1) and flotation of work-piece components in the area of the seam (s7) included the analysis of correlation of the response of behavior described. The correlation between the degree of seam puckering and flotation of work-piece components in the area of the seam, as behavior response in sewing was measured by the Pearson correlation coefficient. A Pearson coefficient of 0.533 indicates correlation that is significant at the level p < 0.01. Behavior responses of the fabric were graded from 1 to 5 (no sample was graded as 1). The situation is presented in Figure 3. Based on the investigations of interaction of fabric mechanical properties and their impact on seam puckering and flotation of work-piece components, experienced experts proposed particularly important attributes, affecting particular responses (Table 1). The attributes are given separately for warp (mark 1) and weft wise (mark 2). Learning sets were formed which include important attributes for both classes as selected by the experts. The

TRJ

TRJ

240

Textile Research Journal 76(3)

Table 2 Root mean squared errors of the estimates. Response s1 s7

Figure 3 Correlation of seam puckering (s1) and flotation of work-piece components (s7).

Table 1 Important attributes affecting seam puckering (s1) and work-piece flotation in the area of the seam (s7). Response

Attributes selected for a particular response

s1 s7

EMT-1, RT-1*, alpha, 2HG-2, 2HG5-2, EL-1 EMT-2, RT-2*, alpha, G-2, 2HG-2, 2HG5-2, B-2

* tensile resilience in % [8]

root mean squared error (RMSE; equation (4)) of the models constructed are given in Table 2. The models compared in this table are: regression tree model, k-NN predictor, and the trivial model which always predicts the average class values.

Prediction by average

Regression tree

k-nearest neighbors

0.726 0.784

0.693 0.897

0.561 0.569

The results obtained by the experiments for response s1, show that regression trees are marginally better than prediction by the average value (Wilcoxon test for paired observations shows difference with the significance < 5%). For both responses, s1 and s7, the best learning method for our problem seemed to be k-NN, which by far outperformed the other two methods. Table 3 shows the results of the experiments in which k-NN was run on 30 randomly selected subsets of the examples described by the initial, complete set of attributes. To confirm the correlation between the results obtained using the k-NN method and the attributes given by the experts, these attributes were given rank following their importance for a particular response. The rank of the attributes was thus obtained. In the experts-proposed list of attributes, the most important attribute for a particular response was ranked as the first, and so on. When sorting the attributes obtained using the k-NN method, the most important one for a particular response was the one that was used most frequently in the course of the 30 experiments. The results thus obtained indicate that the parameter α was the most important one for seam puckering, as it offered the ratio of fabric elongation weftwise and warpwise. Properties exhibited weftwise, particularly bending rigidity B, shear hysteresis 2HG and tensile resilience RT, are the properties that have the highest impact on work-piece flotation in the area of the seam. The Spearman’s rank-correlation coefficient was then established between the two ranks obtained in the above manner [equation (3) (significance level p < 0.05); Table 4]. The results for the Spearman’s rank-correlation coefficient show a high degree of correlation among the attributes selected as important by the experts and those selected by the k-NN method. This indicates that the k-NN method is

Table 3 Number of trials (out of 30) in which the attributes were selected. The attributes that were used in less than four times are omitted for the sake of claritya. Responses s1 s7 a

EMT-1 19

EMT-2

RT-1

RT-2

a

G-2

2HG-2

2HG5-2

28

30 18

11

28

9 14

29

EL-1

EL-2

B-1

B-2 28

Justification for this threshold is that if the probability of the attribute to be chosen at random is 0.5, the probability that an attribute is chosen in four or more trials is less than p = 0.05. Therefore, attributes occurring in less than four trials can be considered statistically insignificant.

Predicting Seam Appearance Quality D. Z. Pavlinic et al.

Table 4 Spearman’s rank-correlation coefficient. Responses s1 s7

Spearman’s rank-correlation coefficient 0.943 0.815

appropriate and can be used to find the attributes that affect fabric behavior in sewing processes.

Conclusion The results obtained in constructing a model for predicting seam quality using machine learning, indicate that in solving these problems the k-NN algorithm is more appropriate for the purpose than regression trees. Despite the fact that regression trees offer, in principle, the formation of more complex models for prediction, and that the interactions among the attributes can be more clearly seen, the trees yield poor results. The k-NN method uses learning examples by memorizing them, and each new example which is given its class, makes its own contribution in addition to the previous learning examples. Spearman’s rankcorrelation coefficient showed relatively high correlation between the ranks of the attributes selected by the experts and the k-NN. The results also confirm the conjectured impact of the mechanical property parameters on seam puckering and flotation of the work-piece components in the area of the seam. It was confirmed that fabric elasticity had the most prominent impact on seam puckering. Seam puckering was most noticeable with inelastic fabrics, where the tension in warp and weft threads were higher than with elastic fabrics, due to the fact that they were pushed aside as each sewing needle penetration. Work-piece component flotation in the area of the seam occurred because of inadequate shear and bending properties, expressed as shear and bending rigidity and the level of associated hysteresis. In making predictions of work-piece component flotation in the area of the seam, using the k-NN method (k = 30), the above parameters of mechanical properties, namely shear hysteresis 2HG and bending rigidity B weftwise, were the most frequently selected ones in the course of 30 repetitions. The parameter of tensile resilience, RT, was also frequently selected in learning (28 times out of 30 repetitions). It denotes the ratio between the reversible and the energy needed for the deformation, and has a serious impact on the seam appearance. Elastic fabrics exhibit lower relaxation values, meaning that work-piece components made of such fabrics can more readily compensate for the loads in the area of the seam, which makes these fabrics more flexible in the end-use products, namely articles of clothing. Investigation of the mechanical property parameters of the fabrics analyzed and their impact on seam quality, rep-

241

resenting one of the key criteria for assessing the quality of garment appearance, is a proper basis for constructing the system for predicting garment appearance quality. Using kNN, the software package Orange enables successful determination of the mechanical property parameters that impact seam puckering and work-piece component flotation in the area of the seam. The selection of important attributes is specific for a particular response of fabric behavior, and can be used to predict the degree of seam puckering and work-piece component flotation, so that the proper fabric can be selected just by knowing its mechanical property parameters. Such an approach to solving the problem of seam puckering and work-piece component flotation, even before the actual process of garment manufacture has started, is of considerable importance in the process of designing high-quality garments. It offers, apart from savings in the amount of the fabric to be used, clear criteria for the required fabric quality parameters.

Literature Cited 1.

Zavec, P., D., and Geršak, J., Evaluation of Garment Appearance Quality, Tekstil 53(10), 497–509 (2004). 2. Hastie, T., Tibshirani, R., and Friedman, J., “The Elements of Statistical Learning,” Springer Verlag, New York, 2001. 3. Demšar, J., and Zupan, B., “Orange,” http://magix.fri.uni-lj.si/ orange/, 2004. 4. Geršak, J., Investigations of the Impact of Fabric Mechanical Properties on Garment Appearance, Tekstil 52(8), 368–379 (2003). 5. Zavec, D., and Geršak, J., The Influence of Mechanical and Physical Properties of Fabrics on Their Behaviour in Garment Manufacturing Processes, in “International Conference Innovation and Modeling of Garment Manufacturing Processes, IMCEP 2000,” University of Maribor, Faculty of Mechanical Engineering, Maribor, pp. 249–257, 2000. 6. Lindberg, J., Westerberg, L., and Svennson, R., Wool Fabrics as Garment Construction Materials, J. Textile Inst. 51(2), 1475–1493 (1960). 7. Morooka, H., and Niwa, M., Physical Properties of Fabrics Relating to Making-up and Good Appearance, J. Text. Machin. Soc. Japan 24(4), 105–114 (1978). 8. Kawabata, S., “The Standardization and Analysis of Hand Evaluation, The Hand Evaluation and Standardization Committee,” 2nd Edn, The Textile Machinery Society of Japan, Osaka, 1980. 9. Geršak, J., The Impact of Technical-Technological Parameters of Sewing on Seam Quality, in “Proceedings Clothing Engineering 1992,” Technical Faculty, Department of Mechanical Engineering, Institute for Textile and Garment Manufacture Processes, Maribor, Slovenia, pp. 49–63, 1992, (in Slovene). 10. Geršak, J., “Expertise with the Analysis of Fabric Mechanical and Physical Properties:[for] Mura European Fashion Design,” Murska Sobota, University of Maribor, Faculty of Mechanical Engineering, Maribor, Slovenia, 1998, (in Slovene).

TRJ

TRJ

242

Textile Research Journal 76(3)

11. Geršak, J., Development of the System for Qualitative Prediction of Garment Appearance Quality, Int. J. Clothing Sci. Technol. 14(3/4), 169–180 (2002). 12. Zar, J. H., “Biostatistical Analysis,” Prentice Hall, Englewood Cliffs, New Jersey, 1998.

13. Breiman, L., Friedman, J. H., Olshen, R. A., and Stone, C. J., “Classification and Regression Trees,” Wadsworth Int. Group, Belmont, CA, 1984. 14. AATCC, “Test Method 88B: Smoothness of Seams in Fabrics after Repeated Home Laundering,” AATCC, Research Triangle Park, NC, 1996.