Modeling the Properties of Core-Compact Spun Yarn Using Artificial Neural Network

Transcription

1 JOURNAL OF TEXTILES AND POLYMERS, VOL. 4, NO. 2, JUNE Modelng the Propertes of Core-Compact Spun Yarn Usng Artfcal Neural Network Parvaneh Kherkhah Barzok, Morteza Vadood, and Majd Safar Johar 1 Abstract In ths research, the compact-core spun yarns have been produced usng RoCoS roller and the effects of flament pre-tenson, yarn count and type of sheath fbers were nvestgated on the physcal and mechancal propertes of produced yarns such as strength, elongaton percentage, harness, and abrason. After statstcally analyss on the obtaned results, for modelng the core-compact yarn propertes, the regresson and artfcal neural network (ANN) were used to predct the physcal and mechancal propertes. Tral and error method was consdered for determnng the best of ANN topology. For ths am, 1110 topologes of ANN (wth dfferent hdden layers and neurons n each hdden layer) were nvestgated for each property. Moreover, to evaluate the accuracy of the created ANN three ndexes were used, namely mean absolute percentage error (MAPE), mean square error (MSE), and correlaton coeffcent (R-value). It was observed that the most accurate results were obtaned based on MAPE and the best topology for predctng all propertes s a two-hdden layer ANN (maxmum MAPE < 0.10) except for the abrason whch s a three-hdden layer ANN (MAPE < 0.17). Keyword: artfcal neural network, compact-core yarn, modelng, physcal and mechancal propertes, RoCoS roller I. INTRODUCTION In the past decades, core-spun spnnng has been developed to acheve a better yarn qualty and mechancal propertes as well as hgher producton per spnnng unt. The specal structure of core-spun yarns, n whch a flament core s covered by staple fbers, permts to deally combne the advantages of flaments lke hgh strength wth those of the staple fbers lke appearance or absorbency propertes. Core-spun yarns are used n a wde spectrum of varous applcatons such as mltary, ndustral, techncal textles and sport clothng. Rng and Sro spnnng systems are the most conventonal systems for producton of core-spun yarns. Some researchers [1,2] employed a novel method usng rng spnnng frame to produce core-spun yarns. Also, a modfed rng spnnng system has been ntroduced for producng core-spun yarns [3]. Ths system utlzes an ar jet for better formng of the sheath fber around the core. Jou and East [4] desgned a flament chargng devce, whch was based on the prncple of a two electrode system to separate a mult flament yarn. Embeddable and Locatable spnnng (ELS) have P. Barzok and M. S. Johar are wth the Department of Textle Engneerng, Amrkabr Unversty of Technology, Tehran, Iran. M. Vadood s wth the Department of Textle Engneerng, Yazd Unversty, Yazd, Iran. Correspondence should be addressed to M. Vadood (e-mal: mortezavadood@yazd.ac.r). been ntroduced n another work [5], n whch locatng technology s employed to locate flaments and staple fbers so that each staple strand could be renforced by the flaments, and the staple fber could be well embedded nto the stem of the yarn. Pourahmad and Johar [6] nvestgated the physcal and mechancal propertes of Rng, Solo and Sro core spun yarns at dfferent controllable parameters. Compact spnnng s another method for producng core-spun yarns. Ths system has two dfferent types whch are called Elcore and Elcore Twst. Brunk [7] reported that core-spun yarns produced by these systems have better evenness and abrason resstant n comparson wth Rng core-spun yarns. The artfcal neural network (ANN) s one of the ntellgent technques for data processng whch has been employed extensvely n varous textle felds. Ths technque s useful when there are nonlnear relatonshps between parameters. There are many publshed work, n whch ANN has been employed to predct the propertes of dfferent yarns and fabrcs and many other characterstcs of textle materals [8-15]. It seems there s a lack of research focused on the predctng propertes of corecompact yarns based on the spnnng parameters, therefore ths paper presents the applcaton of ANN models to predct the propertes of core-compact yarns based on the statstcally sgnfcant controllable factors such as flament pre-tenson, yarn count and knd of sheath fbers. II. NEURAL NETWORK ANN s a structure nspred from the human bran. ANN s very useful for modelng nonlnear problems and complex functons. ANN conssts of three layers ncludng nput, hdden, and output layers. Neurons n each layer are connected by assocated weghts to other neurons n the next layer. The nput data s receved n nput layer and the output s obtaned n the output layer by a mathematcal functon through hdden layers [16]. In ANN there are three operatons ncludng tranng, valdaton and testng sets. Tranng s used to tran the ANN. Valdaton s useful when the network begns to overft the data, and testng group s used to control the error durng the tranng process [17]. In ths study for predctng mechancal and physcal propertes of compact-core yarns, a feed forward multlayer ANN model was used. III. MATERIALS AND METHODS In ths study, 56 dfferent types of yarn samples were produced on a compact-core spnnng system. A blended vscose/polyester and cotton fbers were used as the sheath fber and mult flament nylon yarn wth a count of 100

2 102 JOURNAL OF TEXTILES AND POLYMERS, VOL. 4, NO. 2, JUNE 2016 dener was used as the core flament. The sheath fber propertes are shown n Table I. Type of sheath fber Vscose Polyester Cotton Length (mm) TABLE I THE PROPERTIES OF SHEATH FIBERS Dener Percentage of fber blendng (%) Tenacty (gf/tex) Elongaton (%) The cotton and vscose/polyester rovng count were 0.72 and 1.09 Ne, respectvely. To produce the compact-core yarns the RoCoS system was nstalled on SKF lab spnner nstead of delvery top roller (Fg. 1), and each rovng was fed to the draftng system of the compact-core spnnng frame In order to produce dfferent types of compact-core yarns, the core flament should be pre-drawn before enterng the front rollers (RoCoS roller). The flament also should be fed to compactor groove of RoCoS roller. For ths am, gude rod and pre tensoner were used. Fg. 2. shows the process of core-compact yarn producton and Table II shows the machne settngs for producng compact-core spun yarns. TABLE II MACHINE SETTINGS Settng parameters Value Twst per meter 900 Spndel speed (rpm) Rng dameter (mm) 36 Fg. 2. Producton of core-compact yarns. Fg. 1. Rotorcraft compact spnnng roller (RoCoS). In ths study, the tenacty, elongaton percentage, harness and abrason were consdered as the physcal and mechancal propertes of compact-core yarns TABLE III LIST OF YARN SAMPLES AND CONTROLLABLE FACTORS No. Flament pre-tenson Yarn count Type of fber No. Flament pre-tenson Yarn count Type of fber Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Cotton Cotton Cotton Cotton Cotton Cotton Cotton Cotton Cotton Cotton Cotton Cotton Cotton Cotton Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Vscose/polyester Cotton Cotton Cotton Cotton Cotton Cotton Cotton Cotton Cotton Cotton Cotton Cotton Cotton Cotton

3 BARZOKI et al.: MODELING THE PROPERTIES OF CORE-COMPACT SPUN YARN 103 and to determne the effect of controllable factors on these propertes, 7 levels of flament pre-tenson (25, 50, 75, 100, 125, 160, and 180 g), 4 levels of yarn count (41.5, 43.5, 48 and 59 tex) and two knd of sheath fbers (cotton and polyester/vscose) were chosen. The lst of yarn samples and consdered controllable factors are shown n Table III. Instron testng machne (Model M ) was used to measure the tenacty and elongaton at breakage of yarns wth a gauge length of 25 cm. For measurng harness (number of hars, longer than or equal to 3 mm), Shrley Harness Tester (Model SDL096/8) was used. The measurement was carred out on 20 m of each yarn sample at the speed of 60 m/mn. was determned by Shrley Tester (Model Y027). All experments were conducted under the condtons of 22ºC and 65 RH%. In order to determne the tenacty and elongaton, each test was repeated 10 tmes and for abrason and harness the tests were repeated 5 tmes, and the average values were consdered for each measured property. TABLE IV THE STATISTICAL RESULTS OF MANOVA TEST WITH A %95 CONFIDENCE INTERVAL Characterstc source SSE df MSE F P- value Yarn count Flament pretenson Harness Type of sheath fber Error total Yarn count Flament pretenson Tenacty Type of sheath fber Error total Yarn count Flament pretenson Type of sheath fber Error total Yarn count Flament pretenson Elongaton Type of sheath fber Error total IV. RESULTS AND DISCUSSION To evaluate the effectveness of consdered parameters on the yarn propertes, multvarate analyss of varance (MANOVA) was conducted on the obtaned expermental data. The results of MANOVA test wth a 95% confdence nterval, for each property are shown n Table IV. Ths test compares the varance explaned by factors to the left over varance that cannot be explaned. If the calculated P-value s lower than 5%, t means that the effect of the correspondng factor s sgnfcant on the nvestgated property. The statstcally analyss showed that the controllable factors have sgnfcant effects on all nvestgated propertes. Moreover, n ths study, a mult compare test wth a 95% confdence nterval on the measured propertes was conducted to determne whether the collected data are all the same, aganst the general alternatve that they are not all the same. The obtaned results are presented n Table V. TABLE V THE RESULTS OF MULTI COMPARE TEST WITH A %95 CONFIDENCE INTERVAL FOR EACH GROUP Group 1 Group 2 Lower boundary for the true mean Mean of group 1 mnus the mean of group 2 Upper boundary for the true mean Harness Tenacty Harness Harness Elongaton Tenacty Tenacty Elongaton Elongaton If the confdence nterval contans 0, the dfference would not be sgnfcant. As can be observed n Table V, for none of the pars of nvestgated propertes the confdence nterval s 0, therefore the dfference s sgnfcant. As a result, separate models were used to predct each property. A. Regresson Model Lnear multple regresson analyss was used to establsh a relatonshp between the core-compact yarn propertes and the nvestgated controllable factors. To ths am, all data were dvded randomly nto two groups; namely regtran and test group. Reg-tran group (44 data sets) was used to determne the regresson coeffcents and test group (12 data sets) was used to evaluate the accuracy of obtaned regresson equaton n predctng measured propertes. In ths paper, three ndexes were consdered for measurng accuracy; namely mean absolute percentage error (MAPE, Eq. (1)), mean square error (MSE, Eq. (2)) and correlaton coeffcent (R-value). 1 n y x MAPE = 100 (1) n = 1 x 1 n 2 MSE = (y x ) 1 n = where x s the actual value and y s the correspondng predcted value. The range of R-value s between -1 to +1 and n predcton a hgher R-value means hgher accuracy. But for the other ndexes such as MAPE and MSE, hgher (2)

4 104 JOURNAL OF TEXTILES AND POLYMERS, VOL. 4, NO. 2, JUNE 2016 accuracy n modelng obtans when they are 0 or very close to 0. Eqs. (3) to (6) and Table VI present the obtaned results for regresson analyss. Tenacty = N T F (3) Elongaton = N T F (4) Harness = N T F (5) = N T 2.14 F (6) where N, T and F are the yarn count, the flament pretenson and the type of sheath fbers, respectvely. Here, cotton and vscose/polyester types were consdered 1 and 0, respectvely. As can be seen n Table VI, the accuracy of regresson model evaluated by R-value s hgh, but MSE and MAPE are not close to 0 whch means the lnear regresson s not approprate enough to model the measured propertes. It should be mentoned that although usng a hgher order regresson lke quadratc leads to better results n predcton as shown n Table VI, ths type of regresson due to exstence of numerous terms and calculaton complexty s not easy to use. Hence, a more powerful model such as ANN s requred for modelng. TABLE VI R-VALUE, MSE AND MAPE BETWEEN REGRESSION PREDICTION AND CORRESPONDING ACTUAL DATA (TEST GROUP) Model type Property MAPE MSE R-value Lnear Quadratc Tenacty Elongaton Harness Tenacty Elongaton Harness B. ANN Model ANN ncludes varous parameters whch nfluence drectly the predcton accuracy, but the most effectve ones are the number of hdden layers and the number of neurons n each hdden layer. In ths study to fnd the best set of ANN parameters for each nvestgated property, the tral and error method was appled. Regardng the lterature revew, the number of hdden layers and neurons n each hdden layer were consdered between 1 to 3 and 1 to 10, respectvely. The actvaton functons for all the hdden and output layers were consdered Tangent hyperbolc shown n Eq. (5), and lnear functons, respectvely. x e Tanh = x e x e x + e ANNs were traned wth the error back propagaton algorthm usng Tranlm functon. To tran ANN, the data (5) sets n reg-tran group were dvded randomly nto two groups; namely ANN-tran (34 data sets) and valdaton (10 data sets) groups. As ntal weghts n ANN were selected randomly, each ANN topology was consdered fve tmes and the best result for that topology was consdered. To evaluate the accuracy of ANN models, the same test group for regresson analyss was used and all three mentoned ndexes were calculated. The best topology of ANNs for each nvestgated property correspondng to dfferent ndexes are shown n Table VII. TABLE VII THE BEST TOPOLOGY OF ANN FOR EACH INVESTIGATED PROPERTIES BASED ON MAPE, MSE AND R-VALUE Consdered ndex to select the best ANN MAPE MSE R-value Propertes Hdden layer (best topology of ANN) Accuracy ndexes between ANN output wth best topology and correspondng actual values for testng group MAPE MSE R-value Tenacty {8 3} Elongaton {9 9} Harness {6 6} {2 4 4} Tenacty {8 4} Elongaton {7 9} Harness {4 6} {2 5 5} Tenacty {8 6} Elongaton {9 7} Harness {6 8} {2 5 5} In Table VII, for example {2 4 4} n the hdden layer column means that ANN contans three hdden layers wth 2, 4 and 4 neurons at frst, second and thrd hdden layers, respectvely, and ths ANN can predct the abrason wth the hghest accuracy accordng to MAPE ndex. But n usng MSE as the accuracy ndex, the best topology for predctng abrason s {2 5 5}. Besdes the crtera ndex to select the best ANN topology, the other two ndexes were calculated as shown n Table VII. Accordng to Table VII, the perfect predcton ablty of ANN model s revealed, but a closer look ndcates that consderng the MAPE ndex for choosng the best ANN topology leads to hgher accuracy n predcton. Fg. 3 llustrates the ANN outputs based on MAPE ndex to select ANN topology along wth correspondng actual values for dfferent propertes. Regardng the hgh accuracy of ANN, the physcal and mechancal propertes of compact-core spun yarns can be predcted nstead of arduous task of expermental analyss. So the yarn parameters can be adjusted to produce compact-core yarn wth desred physcal and mechancal propertes.

5 BARZOKI et al.: MODELING THE PROPERTIES OF CORE-COMPACT SPUN YARN 105 Fg. 3. The ANN output (selected based on the MAPE ndex shown n Table VI) along wth the correspondng actual values for dfferent propertes (testng group). V. CONCLUSION Core-spun yarns are used n a wde spectrum of enduses such as mltary textles and ndustral textles. The yarn parameters such as flament pre-tenson, yarn count and the type of sheath fber have a sgnfcant nfluence on physcal and mechancal propertes of compact-core spun yarns such as tenacty, elongaton, harness and abrason. Therefore, modelng these parameters can gve n-depth nformaton about yarn propertes. In the frst step, the sgnfcant effect of yarn parameters on the measured propertes were nvestgated statstcally. Regardng the results of mult compare test, to predct yarn propertes separately, regresson and ANN models were consdered based on the yarn parameters. To acheve the best result for modelng, three ndexes namely MAPE, MSE and R-value were evaluated and fnally t was found that consderng MAPE as a crteron for selectng the best ANN topology leads to the hghest accuracy n predcton. Moreover, the results showed that the best topology for predctng tenacty, elongaton and harness s a twohdden layer ANN (maxmum MAPE < 0.10) wth {8 3}, {9 9} and {6 6} formats, respectvely, whle for the abrason the best one s {2 4 4} (MAPE < 0.17). REFERENCES [1] A. Sawhney, K. Robert, G. Ruppencker, and L. Kmmel, Improved method of producng a cotton covered/polyester staple-core yarn on a rng spnnng frame, Text. Res. J., vol. 62, no. 1, pp , [2] A. Sawhney, G. Ruppencker, L. Kmmel, and K. Robert, Comparson of flament-core spun yarns produced by new and conventonal methods, Text. Res. J., vol. 62, no. 2, pp , [3] G. L. Lous, H. Salaum, and L. B. Kmmel, Rng spun staple core warp yarn- a progress report, J. Text. Inst., vol. 59, no. 4, pp , [4] G. Jou, G. East, C. Lawrence, and W. Oxenham, The physcal propertes of composte yarns produced by an electrostatc flament-chargng method, J. Text. Inst., vol. 87, no. 1, pp , [5] W. Xu, Z. Xa, X. Wang, J. Chen, W. Cu, W. Ye, C. Dng, and X. Wang, Embeddable and locatable spnnng, Text. Res. J., vol. 81, no. 3, pp , [6] A. Pourahmad and M. S. Johar, Comparson of the propertes of Rng, Solo, and Sro core spun yarns, J. Text. Inst., vol. 102, no. 6, pp , [7] N. Brunk, ElCore and ElCoreTwst producton of compact core yarns, Spnnovaton, vol. 21, pp. 4-9, [8] O. Balc, S. N. Oğulata, C. Şahn, and R. T. Oğulata, An artfcal neural network approach to predcton of the colormetrc values of the strpped cotton fabrcs, Fber. Polym., vol. 9, no. 5, pp , [9] B. Behera and R. Guruprasad, Predctng bendng rgdty of woven fabrcs usng artfcal neural networks, Fber. Polym., vol. 11, no. 8, pp , [10] A. A. Gharehaghaj, M. Shanbeh, and M. Palhang, Analyss of two modelng methodologes for predctng the tensle propertes of cotton-covered nylon core yarns, Text. Res. J., vol. 77, no. 8, pp , [11] A. K. Soe, M. Takahash, M. Nakajma, T. Matsuo, and T. Matsumoto, Structure and propertes of MVS yarns n comparson wth rng yarns and open-end rotor spun yarns, Text. Res. J., vol. 74, no. 9, pp , [12] V. K. Mdha, V. Kothar, R. Chattopadhyay, and A. Mukhopadhyay, A neural network model for predcton of strength loss n threads durng hgh speed ndustral sewng, Fber. Polym., vol. 11, no. 4, pp , [13] M. Nasr, M. Shanbeh, and H. Tavana, Comparson of statstcal regresson, fuzzy regresson and artfcal neural network modelng methodologes n polyester dyeng, In: Internatonal Conference on Computatonal Intellgence for Modellng, Control and Automaton, Venna, [14] P. Soltan, M. Vadood, and M. S. Johar, Modelng spun yarns mgratory propertes usng artfcal neural network, Fber. Polym., vol. 13, no. 9, pp , [15] E. Naghashzargar, D. Semnan, S. Karbas, and H. Nekoee, Applcaton of ntellgent neural network method for predcton of mechancal behavor of wre-rope scaffold n tssue engneerng, J. Text. Inst., vol. 105, no. 3, pp , [16] D. Semnan and M. Vadood, Improvement of ntellgent methods for evaluatng the apparent qualty of kntted fabrcs, Eng. Appl. Artf. Intell., vol. 23, no. 2, pp , [17] M. Vadood, D. Semnan, and M. Morshed, Optmzaton of acrylc dry spnnng producton lne by usng artfcal neural network and genetc algorthm, J. Appl. Polym. Sc., vol. 120, no. 2, pp , 2011.