A Neural Network System for Designing New Stretch Fabrics

Hamza ALIBI, Faten FAYALA, Abdelmajid JEMNI Xianyi ZENG GEMTEX research laboratory Laboratory of Study of the Thermal and Energy Systems National School Of Arts And Textiles Industries (ENSAIT) (LESTE) National School Engineers of Monastir, University of Roubaix, University North Lille of France, Monastir Avenue de l’Hermitage, Roubaix Cedex, France Avenue IBN ELJAZZAR, 5019 Monastir, TUNISIA [email protected]

Abstract— In this paper, an artificial neural network (ANN) methods used consisted of pattern construction and how aided system for designing knit stretch materials based on the produce a reduced pattern and this has been applied to a virtual leave one out approach is presented. This system aims garment produced from knitted stretch fabric of known at modeling the relation between functional properties extension and recovery properties [5, 6]. (outputs) and structural parameters (inputs) of knitted fabrics made from pure yarn cotton (cellulose) and viscose In the literature, some studies aimed to conceive new (regenerated cellulose) fibers and plated knitted with elasthane plating devices or to design new plated fabrics [7-11]. (Lycra) fibers. Knitted fabric structure type, yarn count, yarn Cuden, Srdjak, and Pelko [12] focused on the measurement composition, gauge, elasthane fiber proportion (%), elasthane of plated fabric properties such as elasticity and shrinkage yarn linear density, fabric thickness and fabric areal density, before and after laundering, but did not investigate the effect were used as inputs to ANN model. These models have been of elastane ratio on these characteristics. Some researches validated by a testing data. The developed neural model allows [13,14] concerned only weaved fabrics made with Lycra® designers to optimize the structure of knit stretch materials core-spun weft yarns. according to the functional properties. Other woks [15-21] have studied the effect of some Keywords: Comfort-stretch, Knit fabrics, Elongation, Elastic single factors on comfort stretch properties. These properties recovery, Artificial neural network, Virtual leave one out. are influenced by the finishing treatment, the amount of elastane, yarn count, yarn twist, yarn composition, fabric’s I. INTRODUCTION structure and the operating parameters (gauge, loop length, Elastic fabrics are an important route to achieve comfort ….). by freedom of movement for body fitted with sports and They haven’t considered the combinational effects of outdoor wear. Elastic garments used in athletics and sports several factors. Without considering the complex interactions may improve the athlete’s performance in cycling, of the various factors at the different processing stages, the swimming and so on. They are also important for inner wear. weight of each factor and their synergistic effect on This type of fabric enables freedom of body movement by extension and recovery properties cannot be fully reducing the fabric resistance to body stretch. A simple body understood. movement may extend the body skin by about 50% and the fabric must easily accompany the stretch and recover on Numerical techniques consider the combinational effects relaxation. Strenuous movements involved in active sports of several factors without considering the complex may require even greater garment stretch. Drastic differences interactions of the various factors at the different processing between skin and fabric movements result in restrictions of stages. movement to the wearer. Elastic fibre, yarn and fabric In this work, we contribute to develop an ANN-based provide the necessary elasticity to a garment [1]. model to predict elongation, growth and recovery test results Fabrics containing elastane stretch fiber have a wide of knits stretch fabric for course and wale direction. Based application value, especially because of their increased on these models, designers, industrial and researchers can extensibility, elasticity, high degree of recovery, good optimize the comfort stretch of knit product according to the dimensional stability, and simple care [2-4]. specifications, also permit to reduce machines adjustment duration. Many researches interest for dealing with the problems of designing for today’s stretch fabrics. They tried to establish a flexible and economical system for designing well-fitting body contouring apparel with knitted elastomeric. The

U.S. Government work not protected by U.S. copyright II. MATERIAL AND METHOD Table 1: Statistical values of input parameters of training set fabrics.

We produced a series of 340 knitted fabrics commonly Standard Maximum Minimum Mean value used in the clothing industry by using different industrial Inputs deviation value value circular knitting machines (single jersey, double jersey, parameters interlock; tubular and large-diameter; Diameter = 16, 34 Train Test Train Test Train Test Train Test inch, gauge = 18 to 28). Ground yarn was a 100% combed Knitted ------9 9 1 1 cotton (1) and 100% viscose yarn (2) (Nm=28 to 80) and Structure’s Yarn plating yarn was a Lycra® monofilament (22, 33 and 44 ------2 2 1 1 dtex) plated at half feeder. The fabric samples were Composition comprised of nine different knitted structures, single jersey Yarn Count 49 50 10 11 80 80 20 20 (1), single lacoste (2), double lacoste (3), polo pique (4), 1/1 Gauge 24 24 4 4 28 28 16 14 rib (5), 2/2 rib (6), interlock (7), visible molleton (8) and Lycra Proportion invisible molleton (9). 1 1 2 2 10 8 0 0 (%) Lycra Yarn First, we note that the fabrics samples were conditioned 8 6 15 13 44 44 0 0 in the testing laboratory under standard atmospheric Count (dtex) Weight per Unit conditions of 20 ± 2°C and 65 ± 2% relative humidity after a 2 221 209 64 62 548 422 120 119 minimum period of 24 hours conditioning in an NF Area (g/m ) Thickness (m x 78 76 21 22 153 160 44 44 ISO17025 certified laboratory. In this study, the tests carried 10-3) out were concerning the determination of these parameters according to the French national organization for standardization (AFNOR). Table 2: Statistical values of outputs parameters of training and test set fabrics. The evaluations of elastic properties (outputs) are executed according the AFNOR standard (NF EN 14704-1, Standard Maximum Minimum Mean value 2005) [22] with a cyclic loading standard test by using a deviation value value Outputs constant speed gradient dynamometer LRX 2.5 K (LLOYD, parameters England) (500 mm.min-1). The fabric sample (50x300 mm) Train Test Train Test Train Test Train Test is grabbed with two roll clamps. Specimen loading starts by Elongation in gradually increasing the load to 15N and then decreasing the wale direction 31 26 25 17 139 107 13 10 load until load zero. (%) Elastic The 340 measurements were randomly divided into a Recovery in 77 76 8 8 96 92 56 53 training database of 244 values for training and model wale direction selection, and a test database of 96 values for the final (%) assessment of the generalization performance of the model. Residual Extension in Table 1 and Table 2 present the statistical values of inputs 6 6 2 2 17 15 1.5 2 wale direction and output parameters of training and test set fabrics. (%) Elongation in course 92 82 58 35 564 167 13 28 direction (%) Elastic Recovery in 68 66 14 14 92 92 31 29 course direction (%) Residual Extension in 31 31 24 22 159 109 2 3 course direction (%)

A. Model design and network training III. RESULTS AND DISCUSSION The data in neural networks are categorized into two sets: Models of increasing number of hidden neurons (HN) training or learning sets, and test or overfitting test sets. The (i.e. Increasing complexity) was trained started from zero learning set is used to determine the adjusted weights and HN (linear model), and the virtual leave-one-out score E p biases of a network. of each model was computed. The root mean square error The test set is used for calibration, which prevents 2 ( RMSEtr ) and coefficient of correlation ( Rtr ) on the overtraining networks. training datasets were also calculated; results are shown in The overfitting test set should consist of a representative data Table 2. set. It should be approximately 10–40% of the size of the training set of data. The sigmoid function is used [23]. As expected, E p decreases as the number of HN increases and starts increasing when the number of The overall function represented by the network type is: parameters is big enough for over-fitting to arise. In the other HN n == ⎛ − ⎞ −θ hand, RMSEtr on the training dataset decreases as the xfy )( ∑∑⎜ jijij ⎟vVwxSigm j j ==11⎝ i ⎠ (2) number of HN increases (Table 3 and Table 4). Furthermore, it should be noted that when the number of HN exceeds the where x is a n-dimensional input vector, w is the weight eight, the learning task improved but the generalization vector connecting the input units with the single output ability decreased. In fact, the generalization error degraded neuron, and θ is the output neuron’s bias value; HN is the and the over-fitting phenomenon start to occur. number of hidden neurons, v is the weight vector In our case, the generalization error does not increase connecting the hidden layer with the output neuron, and V considerably when the number of HN exceeds the eight. is the hidden neurons’ bias values. Sigm(x) is the common Therefore, with the purpose to reduce the number of model’s sigmoid transfer function: parameters, eight HN were chosen. The optimized ANN architectures are shown in Fig. 1, corresponding to 81 1 xSigm )( = parameters. []−− x )exp(1 (3) The Levenberg-Maquardt fast learning procedure, based on Table 3: Optimization of the number of hidden neurons for the a second order error back propagation algorithm, is then neural networks (course direction) used for determining the parameters of the neural network from the learning data sets. B. Selecting the optimal model architecture Number Elongation in Elastic Recovery in Residual Extension of course direction course direction in course direction Model selection was performed essentially by estimating hidden the generalization ability of the models trained as described, neurons using the score RMSEtr q Ep RMSEtr q Ep RMSEtr q Ep 0 (linear 2 46.589 01 68.609 9.96 01 10.63 53.435 01 33.450 n model) = []RE −k)( p ∑ k 1 36.266 11 52.124 8.23 11 8.55 47.121 11 27.299 k=1 (4) 2 28.674 21 42.478 7.274 21 8.138 10.963 21 14.318 −k )( where Rk is the prediction error on the example k when 3 24.018 31 41.99 6.729 31 7.959 10.183 31 13.828 the latter has been withdrawn from the training set and the model has been trained with all other examples. The leave- 4 19.281 41 36.23 6.343 41 8.443 8.33 41 15.457 −k)( one-out errors R were computed by the “virtual leave- k 5 16.778 51 32.781 5.541 51 8.757 8.907 51 13.625 one-out” method, described in [23]. In this application, the model is based on p samples of 6 15.126 61 25.639 5.465 61 8.242 8.284 61 15.028 knits fabrics. Training uses the leave one out technique. 7 13.926 71 28.52 5.153 71 19.064 7.249 71 16.06 After training, the optimal model architecture was chosen by using a selection methodology [24-26]. 8 (optimal 5.405 81 10.79 4.725 81 8.752 6.445 81 12.526 model) 9 4.342 91 23.076 3.765 91 40.57 5.105 91 14.611

10 3.861 101 33.082 3.452 10121.637 4.067 10119.464 Table 4: Optimization of the number of hidden neurons for the neural networks (wale direction)

Number Elongation in wale Elastic Recovery in Residual Extension of direction wale direction in wale direction hidden neurons RMSEtr q Ep RMSEtr q Ep RMSEtr q Ep 0 (linear 41.92 01 31.256 7.876 01 10.36 2.83 01 2.74 model) 1 34.85 11 28.883 6.017 11 8.231 2.194 11 2.137

2 26.215 21 30.652 5.656 21 6.217 2.134 21 2.056

3 7.48 31 12.246 5.235 31 6.094 1.966 31 1.996

4 6.891 41 17.519 4.873 41 6.034 1.968 41 1.901

5 7.237 51 12.075 4.466 51 6.704 2.037 51 2.102

6 5.969 61 12.616 4.502 61 7.19 1.095 61 1.909

7 5.056 71 11.503 3.84 71 6.674 1.088 71 1.926

8 (optimal 3.966 81 11.801 3.689 81 6.796 1.014 81 1.707 model)

9 2.625 91 12.906 2.629 91 5.957 0.88 91 2.55 10 1.561 101 13.071 2.877 101 6.931 0.556 101 2.603 Fig. 1: The optimal ANN architecture

Fig. 2: Experimental vs. Predicted values for studied properties: Dataset validation. According to these results, choose the ‘virtual leave one understand, evaluate and predict comfort stretch by knitting out’ approach can be argued in view of the fact that it could parameters and fabric features. takes over fitting into account based on the leverages of the learning samples, i.e. on the influence that each sample has on the parameters of the ANN model. Therefore it permits to REFERENCES resolve during the learning of ANN the over-fitting phenomenon. [1] J. Voyce, P. Dafniotis, and S. Towlson, Textiles in Sport , 1st ed. To test the generalization performance of the optimal Cambridge: Wood Head Publication, 2005. trained network, validating processes was applied using the test database (Table 1 and Table 2). The main quality [2] F. Ceken, “Some Investigations of the Dimensional Properties of Knitted Fabrics Containing Different Materials”, doctoral thesis, Ege indicator of a neural network is its generalization ability, its University, Izmir, Turkey, 1996. ability to predict accurately the output of unseen data. The [3] A. Mukhopadhyay, I. C. Sharma, and A. Mohanty, “Impact of Lycra experimental versus predicted values of test dataset is shown Filament on Extension and Recovery Characteristics of Cotton in Fig. 2, as it can be observed, the predictability of ANN fits Knitted Fabric”, Indian Journal of Fibre & Textile Research, vol. 28, very well ( >0.85). pp. 423–430, 2003. [4] M. Tasmac, “Effects of Spandex Yarn on Single Jersey Fabrics”, The mean levels of error (6%) are satisfactory and Tekstil Konfeksiyon, vol. 6, pp. 422–426, 2000. smaller than errors that normally arise due to experimental [5] D. T. I-Chin, C. Cassidy, T. Cassidy, and J. Shen, “The influence of variation and instrumentation accuracy as shown in Table 5. woven stretch fabric properties on garment design and pattern construction”, Transactions of the Institute of Measurement and Table 5: Summary result of training and testing neural network Control, vol. 24(1), pp. 3–14, 2002. model. [6] B. Ziegert, and G. Keil, “Stretch Fabric Interaction with Action Wearables: Defining a Body Contouring Pattern System”, Clothing and Textiles Research Journal, vol. 6, pp. 54-64, 1988. Training model Testing model [7] K. Baozhu, and Z. Weiyuan, “The optimal design of three-layer plated fabrics”, Fibres & Textiles in Eastern Europe, vol. 15, pp. 59- Output Avg Avg 2 2 61, 2007. Error SD (%) Error SD (%) Rtr RT (%) (%) [8] S. M. Bruer, M. Powell, and G. Smith, “Three dimensionally knit Elongation in spacer fabrics: a review of production techniques”, Journal of Textile wale direction 5% 6% 0.935 5% 7% 0.903 and Apparel, Technology and Management, vol. 4, pp. 1-31, 2005. (%) [9] P.J. Doyle, “ Fundamental aspects of the design of knitted fabrics”, Elastic Recovery Journal of The Textile Institute, vol. 44, pp. 561-578, 1953. in wale direction 3% 4% 0.88 4% 8% 0.857 [10] D. L. Munden, “The geometry and dimensional properties of plain- (%) Knit fabrics”, Journal of The Textile Institute, vol. 50, pp. 448-471, Residual 1959. Extension in wale 9.5% 11% 0.9 8% 4% 0.864 [11] T. Pusch, I. Wünsch, and P. Offermann, “Dynamics of yarn tension direction (%) on knitting machines”, AUTEX Research Journal, vol. 1, pp. 55-63, Elongation in 2000. course direction 8% 6% 0.911 7% 8% 0.879 [12] A. P. Cuden, , M. Srdjak, and H. Pelko, “Optimization of the (%) cotton/Lycra plain knitted Fabric parameters”, International Journal Elastic Recovery of Polymeric Materials, vol. 47, pp. 633-648, 2000. in course 5% 5% 0.87 6% 5% 0.91 [13] S. Gorjanc, and V. Bukosek, “The behaviour of fabric with elastane direction (%) yarn during stretching”, Fibres and Textiles in Eastern Europe, vol. Residual 16(3), pp. 63-68, 2008. Extension in 7% 8% 0.89 8% 10% 0.9 [14] N. Özdil, “Stretch and Bagging Properties of Denim Fabrics course direction Containing Different Rates of Elastane”, Fibres and Textiles in (%) Eastern Europe, vol. 16(1), pp. 63-67, 2008.

[15] S. Ben Abdessalem, Y. Ben Abdelkader, S. Mokhtar, and S. Elmarzougui, “Influence of Elastane Consumption on Plated Plain IV. CONCLUSIONS Knitted Fabric Characteristics”, Journal of Engineered Fibers and In this paper, a support system based on artificial neural Fabrics, vol. 4 (4), pp. 30-35, 2009. networks is proposed for optimizing the design of knit [16] B. Chemani, and R. Halfaoui, “Influence of twist, diameter and account of the weft yarn on fabric elasticity: application to pancake stretch fabrics made from pure yarn cotton (cellulose) and band”, In Proceedings of the International Conference of Applied viscose (regenerated cellulose) fibers and plated knitted with Research in Textile, Monastir-Tunisia, MAT 30, 2006. elasthane (Lycra) fibers using an ANN technique and virtual [17] A. Marmarali, “Dimensional and Physical Properties of Cotton leave one out approach dealing with over fitting /Spandex Single Jersey Fabrics”, Textile Research Journal, vol. 73(1), phenomenon and allowing the selection of the optimal pp. 11-14, 2003. neural network architecture. The prediction accuracy was [18] A. Marmarali, and N. Özdil, “Stretching Properties of Cotton/Elastane Single Jersey Fabrics”, In Proceedings of the very good for the training as well as the testing dataset. The International Conference of Applied Research in Textile. Monastir- mean absolute error of prediction was lower than 6% and Tunisia, PDT 47, 2006. the correlation coefficient was higher than 0.85 for both the [19] R. A. Miller, J. D. Atkins, and D. R. Rummel, “Cotton jersey fabric datasets. Thus, it is believed that neural network models construction having improved stretch characteristics”, US Patent could efficiently be applied to the knit industry to 7040124, 2006. [20] T. Pfangen, “Finishing of woven and knitted fabrics with elastane fibers”, Asian Textile Business, vol. 585, pp. 46-51, 2003. [21] S. A. P. Singh, “The Effect of Fabric Structure on the Properties of Two-Way Stretch Fabrics Made from Elastic Core-Spun Yarns of Cotton and Wool Blend”, Textile Research Journal, vol. 44, pp. 506- 512, 1974. [22] NF EN 14704-1, “Determination of the elasticity of fabrics- Part 1: Strip tests”, Paris: French Association of Normalization (AFNOR), 2005. [23] Y. Oussar, G. Monari, and G. Dreyfus, “Repley to the comments on “Local overfitting control via leverages” in “ Jacobian Conditioning Analysis for Model Validation” by I.Rivals and L.Personnaz”, Neural computation, vol. 16, pp. 419-443, 2004. [24] H. Alibi, F. Fayala, A. Jemni, and X. Zeng, “A neural network system for prediction of thermal resistance of knit fabrics”. Special Topic in Porous Media, vol. 3(1), pp. 35-53, 2012. [25] G. Monari, and G. Dreyfus, “Local overfitting control via leverages”, Neural Computation, vol. 14, pp. 1481–1506, 2002. [26] V. Vapnik, The nature of statistical learning theory. New York: Springer, 2000.