168 Natalja Fjodorova1, Marjana Novič1, Tamara Diankova2 and Olga Baskova2 'National Institute of Chemistry, Laboratory of Chemometrics, Hajdrihova 19, SI-1000 Ljubljana 2St. Petersburg State University of Technology and Design, Department of Chemical Engineering and Textile Design, Bolshaya Morskaya Str. 18, RU-191186, St. Petersburg, Russia New Approach for Optimising the Impregnations of Individual Batches of Aramid Fabrics Nov pristop pri optimizaciji različnih serij aramidnih tkanin Original Scientific Article/Izvirni znanstveni članek Received/Prispelo 04-2015 • Accepted/Sprejeto 07-2015 Abstract Innovate textile materials for special clothing are intended for providing trauma protection for the wearer. Fabrics made from high performance aramid fibres are widely used nowadays for manufacturing athletic sportswear for extreme sports due to their high specific tensile modulus and strength. The aim of our study was to illustrate a new approach when searching for optimal settings for impregnating individual batches of textile materials on the basis of para aramid fibres. We demonstrate a feed-forward bottleneck (FFBN) neural network mapping technique that makes it possible to see all optima (optimal settings for best quality) in the studied process. The selections of optimal settings are based on making decisions allowing us to choose optimal settings for processes in relation to the best quality and smallest (minimal) expense. This new approach can be applied for searching optimal settings regarding different chemical treatments. If a standard statistical regression model (in the cases of non-linear relationships) experiences lack of fit, it can be successfully substituted with the FFBN neural network mapping technique. This method can also be recommended as a double check of a studied process when we use other approaches. Keywords: aramid fabrics, optimisation, impregnation, feed-forward bottleneck neural network, design of experiment Izvleček Inovativni tekstilni materiali za specialna oblačila so namenjeni za zaščito telesa pred poškodbami. Tkanine, narejene iz visokozmogljivih aramidnih vlaken, se dandanes na široko uporabljajo za izdelavo atletskih športnih oblačil za ekstremne športe, ker imajo visoko natezno trdnost in elastični modul. Cilj naše študije je bil prikazati nov pristop pri iskanju optimalnih nastavitev za impregnacijo posameznih serij tekstilnih materialov na osnovi paraaramidnih vlaken. V raziskavah je predstavljena metoda 2D (2-dimenzionalnega) nevronskega ma-piranja s tako imenovanim pristopom »feed toward bottleneck neural network (FFBN NN)«, ki omogoča vizualno ugotovitev optimalnih rešitev (optimalnih nastavitev procesa, pri katerih dobimo najboljšo kakovost). Optimalne nastavitve procesa izberemo s pomočjo odločitev, pri katerih lahko izberemo najboljšo kakovost pri najnižjih stroških. Takšen pristop je uporaben za ugotovitev optimalnih nastavitev pri različnih kemijskih obdelavah. V primerih, ko pri standardnem regresijskem modelu (zaradi nelinearnih zvez) pride do nezadostnega ali pomanjkljivega ujemanja, ga lahko uspešno nadomestimo z modelom v obliki nevronske mreže FFBN. Metodo mapiranja FFBN NN lahko uporabljamo sočasno s standardnimi statističnimi metodami. Tako omogočimo dvojno kontrolo sistema. Ključne besede: aramidna vlakna, optimizacija, impregnacija, feed-forward bottleneck neural network, načrtovanje poskusa Corresponding author/ Korespondenčna avtorica: Tekstilec 2015,58(3), 168-176 D.Sc. Natalja Fjodorova DOl: 10.14502/Tekstilec2015.58.168-176 Telephone: +386 1 4760441 E-mail: natalja.fjodorova@ki.si New Approach for Optimising the impregnations of individual 169 Batches of Aramid Fabrics 1 Introduction Aramid fibres have been produced in Russia since 1970. The history of the first Russian para aramid fibres under the name SVM (super high-modulus fibre) is given on the web-page of the Alchemie Group [1]. This group was founded in 1999 to commercialise Russian materials' technologies focusing on unique materials' processes and materials that enable the Alchemie Group to supply ultra-efficient, ultra-light, ultra-high strength materials, composites, components, solutions, systems and products for the 21st century. Aramid fibres are currently in production as AuTx-WE fibre based on Rusar (Russian Aramid) technology with a tenacity of over 250 cN/tex. The process of fibre forming is principally different from that used while forming Kevlar and Twaron aramid fibres. The SVM fibre has high strength (190-220 cN/tex), high modulus (75-100 GPa), and elongation at break (3,0-4,0%). This fibre is used for the manufacturing high strength lightweight composites and as a fabric was applied in the creation of the first flexible Russian bullet proof vests. The current generation of fibres that have become the base for AuTx materials appeared in 1997 under the tradename RUSAR (an abreviation for Russian Aramid). This fibre has very high strength (230-270 cN/tex for AuTx WE and >300 cN/tex for AuTx DWE), high modulus (100-140 GPa), elongation at break at (2,6-3,0%) and is extremly environmentally resilient. The technology of AuTx has a large potential for development [1]. Nowadays the high interest for goods produced from aramide fibres and fabrics has been proved by numerous research [2-8]. In this paper we examined the impregnation processes of high performance fabrics made from ara-mid fibres designed for the manufacturing of athletic sportswear for extreme sports using the neural network mapping technique when searching for optimal conditions regarding the studied process. Traditional statistical methods based on the designing of experiments (for example, surface response method or others) are often used for improving the properties of textile materials [9-11]. In some cases the artificial neural network methods have been employed for predicting the properties of fabrics [12-13]. The application of the feed-forward bottleneck neural network (FFBN NN) mapping technique for optimisation is a relatively new method that is easy to use and non-time consuming [14-19]. The projection of multidimensional data into a 2D map enables obtaining of the input and output parameters within the same coordinates. Thus, a contour plot of output parameters (responses) overlapping with locations corresponding to the combinations of input parameters (setting points) enables visualisation of the optimal setting parameters of the technological processes in the 2D map. Implementation of the FFBN NN mapping technique enables the finding of several optimal solutions during the development of the new products, as well as to improve the qualities of industrial products. Application of the FFBN neural network mapping technique for pigment dyeing of aramid and arimid fibres was published in a paper [20]. It was of interest for considering optimisation of the impregnation process of aramid fabrics using the FFBN NN method. The FFBN NN approach in combination with the criteria functions for plotting coded diagrams is innovative within the field of enhancing textile properties via the adjusting of impregnation parameters. 2 Materials and methods 2.1 Materials Aramid fabrics (Russian name SVM) with mass per unit areas of 131 g/m2 (fabric 1) and 216 g/m2 (fabric 2) produced from polyamide benzimida-zole (PABI) filaments were used during the impregnation treatment. The PABI fibres have extremely high modulus and strength, and are heat-resistant. They are widely Figure 1: A pattern of a sport jacket and trousers with the pieces of aramid fabrics (darker coloured) Tekstilec 2015,58(3), 168-176 170 New Approach for Optimising the impregnations of individual Batches of Aramid Fabrics used for the production of protective clothing (i.e. bulletproof vests) [21] as well as for the recreational industry in a variety of applications ranging from boating to skiing. In this paper we studied those ar-amid fabrics used for the manufacturing of athletic sport wear for extreme sports, particularly sport jackets and trousers (see Figure 1). The quality of lighter fabric 1 (131 g/m2) was evaluated using output parameters (responses) Y1-Y4, where Y1 represents weight gain I (%), Y2 stiffness (mm), Y3 tensile strength (daN) and Y4 elongation (%). The quality offabric 2 (216 g/m2 ) was evaluated using output parameters Y5-Y7 where Y5 designates weight gain II (%), Y6 wear-resistance (cycles) and Y7 tear force (daN). 2.2 Plan of experimental design for the impregnation process The diagram of the impregnation process of aramid fabrics considered in this study is represented in Figure 2. The fabric passes through a bath (2) and subsequently through squeeze rollers 3, then is directed towards drying in the air and thermo-fixation at 150°C. Figure 2: A diagram of the impregnation process Statistical designs of experiments (DOEs) is commonly used in many industries (chemical, polymer, car manufacturing, biotech, food and dairy, pulp and paper, steel and mining, plastic and paints, electronic, telecom, etc.). DOEs can be used for the developments of new products and processes, enhancements of existing products and processes, optimising the qualities and performances of products, screening important factors, minimization-of products' costs and pollution, robustness testing of products and processes, and so on. The goal of our study was improvement of the impregnation process. The following five independent variables (which affect the qualities of impregnated fabrics) were chosen for the study: the concentration of latex (AH7) (xp g/l), concentration of gelatin (x2, g/l), concentration of binder (Carbamol) (x3, g/l), concentration of catalyst, MgCl2 x 6 H20 (x4, g/l), and heat fixation time, (x5, s). Each of the 5 independent variables were explored on 5 levels: -2, -1, 0, +1 and +2. The contents and concentrations of components in the impregnation bath used during this study as coded and non-coded values are represented in Table 1. The parameters were varied according to the specification requirements of the technological impregnation process. A design matrix including the 32 runs (number of experiments) was composed and is represented in Table 2, where Y1-Y4 are response variables related to fabric 1 while Y5-Y7 relate to fabric 2. The input parameters X1-X5 are represented in the coded units (as code levels). The relationships were examined between the output properties of fabrics (Y1-Y7) depending on the concentrations of the ingredients in the impregnating bath and conditions (times) of the heat settings of film-forming compositions. The goal of optimisation was to discover the maximal or minimal values of the Table 1: Coded and non-coded values of the independent input variables X1-X5 at 5 levels (-2, -1, 0, +1, +2) for the impregnating of aramid fabrics Input Factors/ Coded levels -2 -1 0 +1 +2 X1 Latex (AH7) (g/l) 100 80 60 40 20 X2 Gelatin 4%-solution (g/l) 0 20 40 60 80 X3 Binder 50%-emulsion (g/l) 20 40 60 80 100 X4 Catalyst MgCl2 x 6 H20 (g/l) 0 2 4 6 8 X5 Heat fixation time (s) 150 180 210 240 270 Tekstilec 2015,58(3), 168-176 New Approach for Optimising the impregnations of individual 34 Batches of Aramid Fabrics Table 2 Experimental layout of the design matrix employed for five independent variables (X1-X5) and output responses (Y1-Y7) for the impregnation process of aramid fabrics N Factors Fabric 1 (131 g/m2) Fabric 2 (216 g/m2) X1 X2 X3 X4 X5 Weight gain I (%) Y1 Stiffness (mm) Y2 Tensile strength (daN) Y3 Elongation (%) Y4 Weight gain II (%) Y5 Wear resistance (cycles) Y6 Tear force (daN) Y7 1 + 1 +1 + 1 +1 + 1 5.7 82 228 6.2 4.3 270 40 2 +1 + 1 +1 -1 10.8 90 244 6.5 5.8 280 42 3 + 1 + 1 +1 -1 7.5 96 204 6.5 6.5 212 40 4 + 1 +1 + 1 8.1 93 250 6.3 10.2 216 40 5 + 1 +1 +1 -1 8.5 73 216 6.0 3.4 285 25 6 +1 +1 + 1 7.2 70 242 5.7 6.7 216 36 7 + 1 +1 + 1 6.2 64 214 5.3 4.2 220 40 8 +1 -1 8.6 70 252 5.7 6.1 225 49 9 + 1 +1 + 1 -1 8.1 67 211 5.8 6.1 177 54 10 +1 + 1 + 1 11.6 54 243 4.7 6.5 277 57 11 + 1 + 1 + 1 10.2 49 228 6.4 7.8 192 58 12 + 1 -1 11.0 56 238 6.2 8.2 289 56 13 + 1 +1 + 1 9.6 52 228 5.9 4.2 221 50 14 +1 -1 9.6 51 229 6.5 5.6 270 67 15 + 1 -1 5.0 51 233 6.4 4.3 199 60 16 + 1 7.4 46 220 6.0 4.7 216 64 17 -2 0 0 0 0 10.8 57 224 5.0 7.9 325 63 18 +2 0 0 0 0 6.0 54 228 6.5 4.3 258 55 19 0 -2 0 0 0 13.3 52 198 4.5 6.5 234 56 20 0 +2 0 0 0 7.1 53 238 6.5 5.7 206 53 21 0 0 -2 0 0 8.3 52 220 5.8 4.3 227 52 22 0 0 +2 0 0 11.7 52 222 5.9 7.0 203 61 23 0 0 0 -2 0 4.8 65 198 6.0 3.5 188 51 24 0 0 0 +2 0 16. 47 236 6.2 8.8 367 62 25 0 0 0 0 -2 9.2 53 220 6.5 6.0 242 50 26 0 0 0 0 +2 6.2 95 232 5.5 3.2 140 60 27 0 0 0 0 0 8.8 97 252 6.5 4.6 281 60 28 0 0 0 0 0 9.0 98 248 6.5 4.6 304 59 29 0 0 0 0 0 9.2 96 248 6.4 4.6 288 65 30 0 0 0 0 0 8.7 96 247 6.4 4.6 279 62 31 0 0 0 0 0 8.9 97 252 5.7 4.5 291 60 32 0 0 0 0 0 8.8 93 251 6.0 4.5 295 63 Target min min max max min max max Tekstilec 2015,58(3), 168-176 72 New Approach for Optimising the impregnations of individual Batches of Aramid Fabrics objective functions in order to reach the best qualities for the products. In this case the generalised response should be determined. Normalisation of the data was performed in order to obtain a generalised response. In order to obtain the normalized values of the response variables (yn), each y value was divided by its maximal value (ynl = yj/yjmax' yn2 = y2/y2max> .... ynk = yk/ykmax). As a result values no greater than one were obtained for each of the response and could be compared. The goal of our optimisation was the maximisations of Y3, Y4, Y6 and Y7 and the minimisations of Yl, Y2 and Y5. Therefore, the generalised values of the response variables was obtained using Equations 1 and 2 for fabric 1 and fabric 2, correspondingly. (Yn_gen1) = y^i^ yn1*yn2 (1) (Yn_gen2) = y^*^ yn5 (2) It should be noted that the normalised values of "Y" were used in the neural network method described below. 2.3 Feed-forward bottleneck (FFBN) neural network The general concept of artificial neural networks (ANN) is based on simplified imitation of the human nervous system. A simple network has a feed-forward structure: signals flow from inputs forward through any hidden units, eventually reaching the output units. Input data are organised as vectors (based on linear algebra). In other words the input layer serves for introducing the values of the input variables. (In Figure 3 one can see fragments of vectors from the design matrix). Neural networks are typically organised in layers. The hidden and output layer neurons are connected to all the units in the preceding layer. In the study we applied the FFBN NN-containing input layer, hidden layer and de-mapping of input (output layer) (see Figure 4). Layers are made up of a number of interconnected "nodes" which contain an "activation function". The FFBN neural network applied in the study refers to an auto associative neural network. The feed-forward nets here are trained to produce an approximation of the identity mapping between network inputs and outputs using back propagation or similar learning procedures [14-19]. This neural network can deal with linear and nonlinear correlation amongst variables. Multidimensional data sets are difficult to interpret and visualise. The FFBN neural network was used for compression and visualisation of the data in 2D maps. The FFBN neural network is formed by means of mapping and demapping the hidden layer. The signals in the two hidden nodes are taken as two coordinates for each input object, enabling a 2D projection of experimental objects onto a 2D map. The architecture of FFBN NN is represented in Figure 3. The DOE matrics containing »m« factors (X1-Xm) with »k« numbers of runs is shown at the top of Figure 3. The principle of bottleneck layer mapping: the two neurons within the hidden layer produce, for each input object x, a corresponding pair of coordinates (h; = {hi 1, hi2}). The run number »i« marked as a vector xi = {xi1 x^ xi3...xjm}. In the FFBN each i-th object is projected onto a two dimensional map with coordinate hi1/hi2 (see blue arrow). In our experiment we have got 32 projection points (k = 32) corresponding to 32 experimental settings. Figure 3: The architecture of feed-forward bottleneck (FFBN) neural network The 2D map with distribution of 32 experimental settings (as was determined in the plan of the experiment) is shown at the right side of Figure 4 in coordinate H1/H2. For each of the 32 experimental settings the corresponding value of response Y was determined during the course of the experiment. The projection of Y onto H1/H2 coordinates gave the contour plots of response Y. Overlapping the projections of 32 experimental objects (obtained from the FFBN neural network 2D map) with responses' contour plots in the same coordinates (H1/H2) enables visualisation and the determining of optimal settings (the dark colour in the 2D map corresponds to the highest values of response Y). Tekstilec 2015,58(3), 168-176 New Approach for Optimising the impregnations of individual 173 Batches of Aramid Fabrics 3 Results and discussions 3.1 Searching optima using FFBN neural network mapping The architecture of the FFBN neural network with input matrix (X1-X5) applied in the study and projection of experimantal setting points (1-32) is shown in Figure 4. It complemented the projection of generalised response Ygenl (for fabric 1) onto a 2D map with the same coordinate H1/H2. The right (bottom) side of Figure 4 illustrates the contour plot of Ygenl overlapped with 32 input setting points. "tl » l ■■ h- m ---—- » f 4 1 it j <1- 11(1 ii *4 » Figure 4: The architecture of FFBN neural network with input matrix (X1-X5) and projection of experimental setting (points 1-32), as well as values of generalised response Ygenl (for fabric 1) onto 2D map with coordinate H1/H2 Input data are represented as vectors X1-X5 corresponding to 5 columns in the design matrix with 32 rows related to the 32 setting points (see Table 2). A special architecture of error back-propagation neural network was used (5, 2, 5), in which the data are fed into the 5-nodes input layer and then transferred through the 2-nodes hidden layer (compared to a bottleneck) to the 5-nodes output layer. The number of nodes in the input and output layers correspond to the number of independent variables which is equal to 5 (number of factors in DOE). During the training process we are able to get within the output nodes, the values more similar to the input variables of the samples, after passing the bottleneck of the two-node hidden layer. The signals in the two hidden nodes are then taken as two coordinates for each input object acting as a 2D projection of samples into a map. For each of the 32 experimental settings the corresponding value of Y (Y1-Y7) was determined in the course of the experiment. The projection of 32 combinations of experimental conditions and values of response within the same coordinates H1/H2 enables determination of optima in the impregnation process. Figure 5 illustrates the experimental data for fabric 1 with surface density 131 g/m2. It represents the 34 setting points overlapped with the contour plots of responses: Y1- weight gain, Y2- stiffness, Y3- tensile strength, Y4- elongation and generalised values Ygen1 = Y3*Y4/ Y1*Y2. The dark green (in colour picture)/grey (in the grey scale picture) area corresponds to maximal values. For fabric 1 it corresponds to combination 15. Thus, for fabric 1 optimal condition settings correspond to levels of X1-X5 (+,-,-,-,-) with parameter X1 at level +1 and parameters X2-X5 at the level -1. Thus, the optimal settings for fabric 1 enable obtaining of the following quality characteristics: Y1 = 5.0%, Y2 = 51mm, Y3 = 233daN; Y4 = 6.4%. Figure 6 represents the experimental data for fabric 2 with surface density 216 g/m2. The thirty two (32) experimental setting points are overlapped with the contour plots of responses: Y5- weight gain, Y6-wear resistance, Y7- tear force; and generalised values Ygen2 = Y6*Y7/ Y5. For fabric 2 optima for generalised response correspond to combination 14, 18 and central values (zero level (27-32)) (marked with red circles in Figure 6). Non-coded values of Y5-Y7 in that points are seen in Table 2. We can also select the optimal settings by taking into account individual responses Y6, Y7 and Y5 (not generalised). If the weight gain (Y5) is insignificant (or the values for Y5 at point 24 satisfy our requrements) and we would like to get the highest wear resistance (Y6) and tear force (Y7) we obviously would select point 24 (see Figure 6). This point 24 corresponds to the highest values of Y7 (Y7 = 62daN) and Y6 (Y6 = 367cycles) keeping the Y5 at the middle level (Y5 = 8.8%). Thus, three (multimum) optima were recommended for fabric 2. The first optimum corresponds to the setting point 14 (-,+,-,-,-); the second optimum belongs to point 18 (2,0,0,0,0); and the third Tekstilec 2015,58(3), 168-176 174 New Approach for Optimising the impregnations of individual Batches of Aramid Fabrics Figure 5: Experimental data for fabric 1 with surface density 131 g/m2. Thirty two (32) setting points overlapped with the contour plots of responses: Y1- weight gain; Y2- stiffness; Y3- tensile strength; Y4- elongation, and generalised values Ygen1 = Y3*Y4/ Y1*Y2 optimum is located at the centre point 27-32 (0,0,0,0,0). To obtain the highest wear resistance and tear force keeping the weight gain at the middle level we recommend setting point number 24 (0,0,0,+2,0). Visualisation of the process in the 2D map provides significant information for making decisions and selecting a suitable condition for the experiment. An optimal solution was chosen on the basis of a compromise decision. Tekstilec 2015,58(3), 168-176 New Approach for Optimising the impregnations of individual 175 Batches of Aramid Fabrics Optimum for Y6 and Y7 in point 24 Optimum for Ygen2 in points 14, 18 and in center point — (27-32) Figure 6: Experimental data for fabric 2 with surface density 216 g/m2. Thirty two (32) setting points overlapped with the contour plots of responses: Y5- weight gain; Y6- wear resistance; Y7- tear force, and generalised values Ygen2 = Y6*Y7/ Y5 4 Conclusion Implementation of the feed-forward bottleneck neural network technique enabled the finding-out of one optimum for fabric 1 that corresponded to setting parameters in setting point 15 (+,-,-,-,-) with parameter X1 at level +1 and parameters X2-X5 at the level -1. For fabric 2 three (multimum) optima were recommended: corresponding to the setting points 14 (-,+,-,-,-), 18 (2, 0,0,0,0) and centre point 27-32 (0,0,0,0,0). To get the highest wear resistance and tear force keeping the weight gain at the middle level and recommended setting point number 24 (0,0,0,+2,0). The FFBN mapping technique enables the finding-out multiple optima that could be selected upon compromise decisions by taking into account the desired quality of material, as well as the technical and economic viewpoint and safety of the process. Translation of the multidimensional process input parameter's space into the 2D coordinate system and application of the criteria function as a combination of the process output parameters can be very informative and opening a whole new area of finding optimal conditions in the field of the textile impregnation industry. References 1. Alchemie Group home page [online], [accessed 22. 4. 2015]. Available on Word Wide Web: . 2. KIM, Hyun Ah and KIM, Seung Jin. Physical properties of para-aramid/nylon hybrid air textured yarns for protective clothing. Fibers and Polymers, 2014, 15(11), 2428-2436, doi: 10.1007/s12221-014-2428-5. 3. ZHENG, Huanda and ZHENG, Laijiu. Dyeing of meta-aramid fibers with disperse dyes in supercritical carbon dioxide. Fibers and Polymers, 2014, 15(8), 1627-1634, doi: 10.1007/s12221-014-1627-4. Tekstilec 2015,58(3), 168-176 176 New Approach for Optimising the impregnations of individual Batches of Aramid Fabrics 4. BILISIK, Kadir. Experimental determination of yarn pull-out properties of para-aramid (Kev-lar®) woven fabric. Journal of Industrial Textiles January, 2012 41(3), 201-221, doi: 10.1177/ 1528083711413411. 5. UPPAL, Rohit, RAMASWAMY, Gita N. and LOUGHIN, Thomas. A novel method to assess degree of crystallinity of aramid filament yarns. Journal of Industrial Textiles, 2013, 43(1), 3-19, doi: 10.1177/1528083712444648. 6. LIN, Lantian, SHEN, Yiping and ZHANG, Qiuping. Analysis of environmental impact on mechanical properties of aramid filaments. Journal of Industrial Textiles, 2013, 42(4), 489-500, doi: 10.1177/1528083712446383. 7. PARK, Jong Lyoul, YOON, Byung Il, PAIK, Jong Gyu and KANG, Tae Jin. Ballistic performance of p-aramid fabrics impregnated with shear thickening fluid. Part I - Effect of laminating sequence. Textile Research Journal, 2011, 82(6), 527-541, doi: 10.1177/0040517511420753. 8. ZIELINSKA, Dorota, DELCZYK-OLEJNICZAK, Bogumila, WIERZBICKI, Lukasz, WILBIK-HAEGAS, Bozena, STRUSZCZYK, Marcin Henryk and LEONOWICZ, Marcin. Investigation of the effect of para-aramid fabric impregnation with shear thickening fluid on quasi-static stab resistance. Textile Research Journal, 2014, 84 (15), 1569-1577, doi:10.1177/0040517514525881. 9. LI, Ting-Ting, WANG, Rui, LOU, Ching Wen and LIN, Jia-Horng. Evaluation of high-modulus, puncture-resistance composite nonwoven fabrics by response surface methodology. Journal of Industrial Textiles, 2013, 43(2), 247-263, doi: 10.1177/1528083712452900. 10. ZHIYING, Cui, YANMIN, Wan and WEIY-UAN, Zhang. Thermal protective performance and moisture transmission of firefighter protective clothing based on orthogonal design. Journal of Industrial Textiles, 2010, 39(4), 347-356, doi: 10.1177/1528083709347126. 11. LEE, Kyulin and CHO, Gilsoo. The optimum coating condition by response surface methodology for maximizing vapor-permeable water resistance and minimizing frictional sound of combat uniform fabric. Textile Research Journal, 2014, 84(7), 684-693, doi: 10.1177/0040517513509870. 12. SARAVANA, kumar T. and SAMPATH, V. R. Prediction of dimensional properties of weft knitted cardigan fabric by artificial neural network system. Journal of Industrial Textiles, 2013, 42(4), 446-458, doi: 10.1177/1528083712444296. 13. BEHERA, B. K. and GOYAL, Y. Artificial neural network system for the design of airbag fabrics. Journal of Industrial Textiles, 2009, 39(1), 45-55, doi: 10.1177/1528083708093335. 14. KRAMER, Mark A. Nonlinear principal component analysis using autoassociative neural networks. AIChE Journal, 1991, 37(2), 233-243, doi: 10.1002/aic.690370209. 15. NOVIC, Marjana and GROSELJ, Neva. Bottleneck type of neural network as a mapping device towards food specifications. Analytica Chimica Acta, 2009, 649(1), 68-74, doi: 10.1016/ j.aca.2009.07.018. 16. DASZYKOWSKI, M., WALCZAK, B. and MASSART, D.L. A journey into low-dimensional spaces with autoassociative neural networks. Talanta, 2003, 59(6), 1095-1105, doi: 10.1016/ S0039-9140(03)00018-3. 17. DASZYKOWSKI, M., WALCZAK, B. and MASSART, D.L. Projection methods in chemistry. Chemometrics and Intelligent Laboratory Systems, 2003, 65(1), 97-112, doi: 10.1016/S0169-7439(02)00107-7. 18. LIVINGSTONE, D. J., HESKETH, G. and CLAYWORTH, D. Novel method for the display of multivariate data using neural networks. Journal of Molecular Graphics, 1991, 9(2), 115118, doi: 10.1016/0263-7855(91)85008-M. 19. KOCJANCIC, Robert and ZUPAN, Jure. Application of a feed-forward artificial neural network as a mapping device. Journal of Chemical Information and Modeling, 1997, 37 (6), 985989, doi: 10.1021/ci970223h. 20. FJODOROVA, Natalja, NOVIC, Marjana and DI-ANKOVA, Tamara. Optimization of pigment dyeing process of high performance fibers using feed-forward bottleneck neural networks mapping technique. Analytica Chimica Acta, 2011, 705(1-2), 148-154, doi: 10.1016/j.aca.2011.04.041. 21. High-performance fibres. Edited by J. W. S. Hear-le. Cambridge: Woodhead Publishing, 2001, pp. 329, doi: 10.1533/9781855737549. Acknowledgments Authors thank for the European Commission for the financial support under project CAESAR (SSPI- 022674) and the Slovenian Ministry of Higher Education, Science and Technology (grant P1-017). Tekstilec 2015,58(3), 168-176