Intelligent Model to Investigate the Sound of Woven Fabric by Using Signal Processing and Artificial Neural Network

Transcription

1 Columbia International Publishing Journal of Signal Processing Theory and Applications (2014) Vol. 3 No. 1 pp doi: /jspta Research Article Intelligent Model to Investigate the Sound of Woven Fabric by Using Signal Processing and Artificial Neural Network Maryam Gholamzadeh 1, Dariush Semnani 1*, Mohammad Zarrebini 1, and Rassoul Amirfattahi 2 Received 3 June 2014; Published online 11 October 2014 The author(s) Published with open access at Abstract Fabric comfort comprises fabric sound which is the sound generated by fabric during wear by users. In this work sound generated by samples of polyester and cotton-polyester woven fabric was investigated. The sample specifications in warp wise direction were identical. In order to stimulate sound generated by the samples, an apparatus capable of sound induction was designed and developed. The recorded sound signals were analyzed, using Discrete Fourier Transform, together with Discrete Wavelet Transform. Sub-bands energy of FFT and energy coefficients of wavelet transform were calculated. The effect of Physical and mechanical properties on sound features of samples was predicted using Multilayer Feed forward network with back propagation learning algorithm. The network generated results showed that, fabric surface roughness and drape coefficient are the most and the least effective parameters affecting sound volume respectively. Keywords: Comfort; Sound; Signal Processing; Wavelet; Woven Fabric; Neural Network 1. Introduction Fabrics are regarded as flexible flat sheets. Sound generated by fabrics can be indicative of other properties such as fabric hand. Thus, fabric sound can be a means for manipulation of fabric properties (Sukigara and Ishibashi, 2001). Fabric sound is generated when various parts of a garment worn by a user rub against each other or when the fabric hand is judged by feeling the fabric between thumb and forefinger (Yi and Cho, 2000). Fabric sound can affect psychological stability of the garment wearer and others in the vicinity of the garment user (Cho et al., 2009). Fabric sound can be either pleasant or irksome, the extent of which is dependent on circumstances (Yi and Cho, 2000; Yi et al., 2002; Yi, 2005). *Corresponding d_semnani@cc.iut.ac.ir 1* Department of Textile Engineering, Isfahan University of Technology, Isfahan , Iran 2 Department of Electrical Engineering, Isfahan University of Technology, Isfahan , Iran 20

2 Suiting fabrics are expected to not to be obnoxious as far as fabric sound is concerned. Fabric sound amplitude and frequency are considered to be affected by factors such fabric friction, roughness, shear, and bending stiffness (Yi and Cho, 2000). Previous researches in the field of fabric sound have investigated the fabric sound in relation to mechanical properties of the fabric and have attempted to characterize frictional generated fabric sound (Yi et al., 2002; Cho and Yi, 2000). Fabric sound has been objectively analyzed in terms of the effect of constituent fiber cross-section shape, fiber type, and weaves (Kim et al., 2003; Kim and Cho, 2003; Kim et al., 2002). Additionally subjective sensation and the physiological responses to fabric sounds have been reported (Cho and Cho, 2007; Cho et al., 2005; Cho et al., 2005). Cho et al. (2005) have conducted extensive research in the field of fabric sound. Pioneering research in stimulation of silk fabric sound by polyester fabric was under taken by Japanese researcher Fukuhara (1993). The aim of this research is prediction of woven fabric sound by using digital signal processing technique (DSP) and artificial neural network (ANN). To achieve this goal, we designed a apparatus with ability of rubbing sound measurement. After that various fabric samples which are commonly used in apparel industry have been experienced. The fabric sound features extracted by DSP- have been trained in to a back propagation neural network. Also, effect of fabrics mechanical properties on their sound features have been investigated through ANN. 2. Signal Processing Signal processing refers to those techniques which are applied to signals in order to extract some informaton from them. Signals could be repesented in time-domain but in order to extract useful information, it's better to map them into other domains. Mathematical transforms such as Fourier transform has been used to map time-dependent signals into frequency domain. Short-time Fourier transform and wavelet transform represent frequency content of signals in a short time duration. In other words, those transforms could describe signals in a time-frequency or time-scale domain (Chui, 1992; Graps, 1995; Haigh et al., 2001). Fourier transform is the most applicable for time-series analysis. The frequency content of signal could be determined through decomposin gto a series of harmonic functions of different frequencies. Equation 1 shows the Fourier transform relation for time-domain signal f(t). F(w) = 1 2 f ( t) e i t where i 1 and is the angular frequency in radians per second. dt (1) Spectral analysis, the analysis of time-domain signals through Fourier transform, is a standard technique for information recovery from frequency content of periodic signals with infinite alternation. However, all signals are not necessarily continuous time but they are resulting from sampling process and are discrete in time. To analysis of such signals, Discter Fourier Transform (DFT) or in practice, a faster version called Fast Fourier Transform (FFT) has been applied. 21

3 Therefore, Fourier transform could be used for determination of frequency content of time domain signals. The main fault of Fourier transform is its complete seperation from time domain. It may be causes to loss of time domain information of signals specially for those signals with rapid changes in time characteristics. Short-Time Fourier Transform (STFT) can overcome this problem by taking Fourier transform from a windowed frame of signal. STFT devides signal to short time frames through passing it from a window function. Each frame then could be thought as stationary signal. Widow shape and duration are very critical in STFT. According to Heisenberg uncertainty principle, there is a trade-off between time and frequency resolution in STFT analysis. In the other words, long duration windows redound to high frequency resolution while the time resolution remain very low and vise versa. Wavelet transform tries to overcome this problem through variable window length. Windows become longer whenever the signal variations become slow (Chui, 1992; Graps, 1995; Haigh et al., 2001). Discrete Wavlet Transform (DWT) is a realizable form of wavelet transform by using digital filters which are implanted on computer software. In the first decomposition order, a frame of signal is passed through a set of low-pass and high-pass quadrature digital filter to make the first order approximation and detail wavelet coefiecients respectively. Approximation coeficients of each order are passed through a similar set of digital filters after downsampling by 2 to make the detail and approximation coeficients of subsequent order (Polikar, 1995; Strang and Nguyen, 1995; Titchmarsh, 1948). 3. Artificial Neural Networks ANN is a simplified model of how the ordinary neural systems work. They have found abundant applications in various field of engineering and sciences (Rojas, 1996). Moreover, ANN's are an incomplete imitation from Biological neural system and brain structure of a human body. The main idea in ANN's is based on the training and communication of a lot of processing units, called neuron, in decision making. Neurons are connected together in a network structure in such a way that each connection has an individual weight in final decision. ANN's have trained through altering in connection weights according to observed input data. After the training process using a large training data set, ANN's could predict the behavior of a system (Fausett, 1994). Although the designed artificial models are very simpler than Biological structures, they have so much capabilities in pattern recognition and classification. In other words, ANN's introduce a particular organization to solve serious problems which could not been solved through regular methods. The essential featurs of ANN's can be summerized as ability of learning, generalizing and mapping, robustness aginst the noise and interferences and parallel processing (Patterson, 1996). Learning is predicated to the process of changing the connected weights in order to attain to the best match between training and output patterns (Gurney, 1997). There are different models of ANN's according to their architecture and learning process. In this paper, we used of feedforward Multi- Layer Perceptron (MLP) neural network with error backpropagation learning (BP) algorithm in order to predict the sound of warp-and-woof textile. The BP algorithm is belong to supervised categories against to non-supervised ones. BP was proposed firstly by Romelhart et., al. in 1986 to learning MLP networks with arbitrary hidden layers. This algorithm consists of two main part. Feed-forwrad that input data are fed to the network and are propagated to the output through hidden layers and Feed-back taht difference between determined and desired output is computed as error. Then new connection weights are computed according to the bacpropagated error. First of 22

4 all, the output layer connection weights are adjusted based on the difference value of determined and desired output and according to delta rule. The hidden layers connection weights are adjusted subsequently. Sicne the desired output does not defined for the hidden layers, weight adjusting is more difficult for their neurons (Patterson, 1996; Gurney, 1997; Krose and Smagt, 1996). 4. Materials and Methods Twenty-four samples of woven fabric were woven using polyester and cotton-polyester yarns. The characteristics of the experimental samples are shown in Table 1. Samples have been woven in different warp/weft kinds, warp/weft densities and weaving design. Table 1 Characteristics of cotton-polyester samples sample Yarn Type Warp Weft Warp linear density Weft linear density Warp /Weft Density (w/cm) Weave pattern 1 spun 40/2 Ne 40/2 Ne 26/24 Twill 2 spun 40/2 Ne 40/2 Ne 26/21 Twill 3 spun 40/2 Ne 40/2 Ne 26/17 Twill 4 spun 40/2 Ne 30/2 Ne 26/24 Twill 5 spun 40/2 Ne 30/2 Ne 26/21 Twill 6 spun 40/2 Ne 30/2 Ne 26/17 Twill 7 spun (80/20) 40/2 Ne 20/2 Ne 26/24 Twill 8 spun (80/20) 40/2 Ne 20/2 Ne 26/21 Twill 9 spun (80/20) 40/2 Ne 20/2 Ne 26/17 Twill 10 spun polyester/ cotton polyester 2/40 Ne 2/30 Ne 26/24 Twill 11 spun polyester/ cotton polyester 2/40 Ne 2/30 Ne 26/21 Twill 12 spun polyester/ cotton polyester 2/40 Ne 2/30 Ne 26/17 Twill 13 Filament polyester polyester 150 den 150 den 33/14 Plain 14 Filament polyester polyester 150 den 150 den 33/17 Plain 15 Filament polyester polyester 150 den 150 den 33/20 Plain 16 Filament polyester polyester 150 den 150 den 33/26 Plain 23

5 17 Filament polyester polyester 150 den 150 den 60/21 Satin 18 Filament polyester polyester 150 den 150 den 60/24 Satin 19 Filament polyester polyester 150 den 150 den 60/27 Satin 20 Filament polyester polyester 150 den 150 den 60/30 Satin 21 Filament polyester polyester 150 den 150 den 24/15 Plain 22 Filament polyester polyester 150 den 150 den 24/19.5 Plain 23 Filament polyester polyester 150 den 150 den 24/23 Plain 24 Filament polyester polyester 150 den 150 den 24/27 Plain In order to creation and production sound of fabric, was developed requirement apparatus as shown in Figure 1. Sound is generated by sliding of the contact lever covered by a harsh fabric against the surface of the test sample. The generated sound was picked up by a microphone. Using a mixer and a sound card the generated sound is fed to a computer and recorded. The sound inducing apparatus consists of the following: A Sample holding part comprising of a bottom fixed and top movable mounting pinned jaws. The sample is mounted flat on the pins between the jaws (Figure 2(A)). B Sound inducing sliding belt driven lever with 1 2 cm rectangular head. The head is covered with a harsh fabric (Figure 2(B)). C - The sound registration and recording system consists of a microphone, mixer, sound card and computer. Induced sound is picked up by a 20 Hz to 20 khz JTS-JS-1T type microphone with connection a mixer. Mixer output is transmitted to sound card input through an interface cable. Sound card output is fed to the computer where it is recorded monophonically at frequency Hz of using Cool Record Edit Deluxe sound software (Figure 3). Fig. 1. Fabric sound induction apparatus 24

6 A Fig. 2. Main parts of the sound induction apparatus; (A) sample holding part, and (B) sound inducing part B A Fig. 3. Sound registration and recording system; (A) mixer, sound card and computer, and (B) the microphone Rectangular cm samples cut along weft direction were prepared. The samples were mounted flat on the pins while elongated by 2%. Experiments were conducted by moving the sliding lever head against the back of the sample in bottom to top direction at a speed of 0.2 m/s. Microphone was positioned half way between the jaws at a distance of 8 cm away from the sample. In order to eliminate the interfering noises experiments were conducted in an insulated space. Recording ends upon return of sliding lever to bottom position again. Sound is recorded under the sampling frequency of Hz and as mono. In order to extract sound characteristics from sound signal, sampling frequency reduction of signals or pre-processing step must be performed. Thus, all signals were reduced to 12,000 Hz sampling B 25

7 rate. Sound frequency spectra of each sample were obtained using FFT by separation of 100 milliseconds from each signal. The sound spectra were transformed to six frequency sub-bands from 0 to 6000 Hz in intervals of 1000 Hz. The energy spectra sub- bands were calculated as a characteristic of the sound signal. Energy in the sub bands was calculated using equation 2. P = subband where, X ( f ) is the discrete Fourier transform of sound signal. 2 X ( f ) (2) Additionally discrete wavelet transform was also used. Energy coefficients of the sub-bands of the wavelet were extracted as additional characteristics of sound signals. In order to obtain highest level of decomposition and wavelets signals were evaluated at different levels and wavelets. Thus, four levels were selected for decomposition and wavelets sym7 was selected among various type wavelets. Apart from symmetric part of the wavelets, wavelets families' sym are almost symmetrical. All other wavelets characteristics are similar to those of Daubechies families. There are a various types of sym wavelets from 2 to 8, as shown in Figure 4. Fig. 4. Various types of sym mother wavelets Characteristics of p1 to p6, is related to the spectrum of FFT. Energy values of approximation coefficients and detail of discrete wavelet transform is shown at different levels of decomposition respectively with Ea4, Ed4, Ed3, Ed2, and Ed1. Tables 2 and 3 show the amount of energy in different frequency sub-bands in the spectrum of FFT and Wavelet Transform. In this study, Physical and mechanical properties of the samples such as tensile strength, shear strength, bending strength, friction, drape coefficient, thickness and cover factor as shown in Tables 4 and 5 were measured. 5. Results and Discussion Average FFT spectrum for polyester fabrics and cotton-polyester fabrics are shown in Figure 5. It can be observed that amplitude varies between -40 to 40 db at all frequency spectrum range. Additionally peak amplitude for all samples occurs at frequency sub-band 4000 to 5000 Hz. due to similarities of spectral diagrams of the samples, In order to extract information from the frequency spectrum, sub-bands energy of FFT and energy coefficients of wavelet transform were calculated. 26

8 A B Fig. 5. FFT spectrum; (A) polyester samples, and (B) cotton-polyester samples Table 2 FFT spectrum of samples Sample p1 p2 p3 p4 p5 p

9 Table 3 Wavelet Transform of samples sample Ea4 Ed4 Ed3 Ed2 Ed Characteristics of FFT spectrum of polyester and cotton-polyester samples are shown in Figure 6(A) and 6(B) respectively. Characteristics of wavelet transform of polyester and cotton-polyester samples are shown in Figure 7(A) and 7(B) respectively. Considering the results shown in Figures 6 and 7, it can be stated that fabrics can be separated based on the different spectral energies in FFT and wavelet coefficients sub-band spectrum. Energy in the sub-band 0 to 1000 Hz in FFT spectrum is at peak value for all samples. This represents the generated low frequency nature of sound by the fabrics. The highest amount of energy in the wavelet transform coefficients represents energy in approximation coefficients, which are illustrated energy in sub-band Hz. This also represents the generated low frequency nature of sounds by the fabrics. As mentioned earlier, a feed-forward MLP neural network with error BP algorithm has been designed to predict relationship between physical and mechanical properties of samples based on 28

10 their sound characteristics. In this network, 14 physical and mechanical features of the samples were selected as network input. These included properties such as tensile, shear, surface, bending in direction of weft, drape coefficient, weight, thickness and weft fractional cover of the samples. As far as the network output layer was concerned, a total of 11 characteristics were selected. Six energy characteristics of sub-band frequency in the spectrum FFT, and 5 energy characteristics in sub-bands frequency wavelet transform. A B Fig. 6. Characteristics of FFT spectrum; (A) polyester samples, and (B) cotton-polyester samples 29

11 Table 4 Physical characteristics of samples Sample Breaking force (N) Elongation at break (mm) Max. shear force (N) Max. shear elongation Shear modulus (N/mm 2 ) Bending length (cm) Flexural rigidity (μnm) Bending modulus (N/m 2 ) drape coefficient (mm) / The network contained 15 neurons in hidden layer. Network results, are reported based on the minimum amount of error percent. In general, the neural network does not emphasize the importance of input variables over output. In order to understand the effect of each input variables on output variables, weight percent of inputs variables can be used. Percent weight can identify the input variables that can cause the greatest change in network output. These variables are considered to have higher importance in comparison to other variables. In this study in order to determine the contribution of individual fabric physical and mechanical parameters on the amount of energy spectral, weight percent of input variables was calculated using equation 3. The calculated results are shown in Figure 8. n i 1 wij w = 100 n m wi i 1 i 1 j (3) 30

12 A B Fig. 7. Wavelet Transform coefficients; (A) polyester samples, and (B) cotton-polyester samples In equation 3, wij represents input or output weight in layers. As shown in Figure 8, fabric surface roughness (SMD) with a contribution of 14.04% has the largest effect on the energy spectral. This is followed by fabric shear modulus with a contribution of 12.40% and coefficient of variation of friction force (MMD) with a contribution of 8.16%. Other fabric parameters that can influence fabric sound are fabric maximum shear force, flexural rigidity, elongation at break, bending modulus, thickness, coefficient of friction, shear force, weight surface, weft fractional cover, breaking force. Fabric drape with a contribution of 3.87% is the least affecting parameters as far as fabric sound is concerned. The calculated weight percent of the output layer used for determination of the most 31

13 effective output variables are shown in Figure 9. As shown in Figure 9, it can be said that Ed3 and P4 have the most and least contribution. Fig. 8. Contribution of input variables on sound energy Fig. 9. Contribution of sound output variables 32

14 Table 5 Mechanical characteristics of samples Sample SMD MIU MMD Weight (gr/m2) thickness (mm) partial coverage of fabric (%) Conclusions Fabric comfort comprises fabric sound which is the sound generated by fabric during wear by users. The amount of sound generated by fabric can be considered as an index of apparel comfort. This index can determine the suitability of fabrics for their intended end-use. In this work sound generated by samples of polyester and cotton-polyester woven fabrics was investigated. Since warp wise features of the samples were identical weft wise sound characteristics was investigated. The recorded sound signals were analyzed, using Discrete Fourier Transform, together with Discrete Wavelet Transform. In order to extract information from the frequency spectrum, sub-bands energy of FFT and energy coefficients of wavelet transform were calculated. This was due to similarities in spectral diagrams of the samples. It was found that the amount of energy in subbands of FFT spectrum and sub-band of wavelet transform coefficients are different. This was considered to be due to differences in the characteristics of the experimental samples. All samples showed the highest amount of energy in sub-band 0 to 1000 Hz in FFT spectrum. This was found to 33

15 correspond to the low frequency nature of sounds that can be produced by the fabrics. The highest amount of energy in the wavelet transform coefficients approximated the amount of energy coefficients in the sub-band 0 to 375 Hz. This corresponded to the low frequency nature of sounds produced by the fabrics. The use of a neural network feed forward with back propagation learning algorithm with one hidden layer reveled that among fabric physical and mechanical parameters surface roughness and drape coefficient were the most and the least effective parameter affecting fabric sound respectively. Additionally as far as ANN is concerned it was found that output variables Ed3 and P4 were the most and the least effective parameter respectively. References Cho, G., Cho, J., Kim, C. & Ha, J. (2005). Physiological and subjective evaluation of the rustling sounds of polyester warp knitted fabrics. Textile Research Journal, 75(4), Cho, G., Kim, Ch., Cho, J., & Ha, J. (2005), "physiological signal analyses of frictional sound by structural parameters of warp knitted fabrics. Fibers and Polymers, 6(1), Cho, G., & Yi, E. (2000). Fabric sound parameters and their relationship with mechanical properties. Textile Research Journal, 70, Cho, J., & Cho, G. (2007). Determining the psychoacoustic parameters that affect subjective sensation of fabric sounds at given sound pressures. Textile Research Journal, 77(1), Cho, S., Cho, G., & Kim, Ch. (2009). Fabric sound depends on fiber and stitch types in weft knitted fabrics. Textile Research Journal, 79(8), Chui, C. K. (1992). An introduction to wavelets. Academic Press Inc., San Diego, California. Fausett, L. (1994). Fundamentals of Neural Network. Prentice Hall, New Jersey. Fukuhara, M. (1993). Innovation in polyester fibers: from silk- like to new polyester. Textile Research Journal, 63, Graps, A. (1995). An introduction to wavelets. IEEE Computational Science and Engineering, (Vol. 2, No. 2, Summer 1995) Gurney, K. (1997). An Introduction to neural networks. 1st edition, UCL Press, London. Haigh, S. K., Teymur, B., Madabhushi, S. P. G., & Newland, D. E. (2001). Applications of wavelet analysis to the investigation of the dynamic behavior of geotechnical structures. Soil Dynamic and Earthquake Engineering, 22, Kim, Ch., & Cho, G. (2003). Characteristics of rustling sounds created by the structure of polyester warp knitted fabrics. Textile Research Journal, 73(8), Kim, Ch., Cho, G., Hong, K., & Shim, H. (2003). Sound characteristics according to cross-sectional shapes of fibers. Fibers and Polymers, 4, Kim, C., Cho, G., & Na, Y. (2002). Effects of the basic weave differences in silk fabric and the yarn type variations in satin weave on the sound parameters. Textile Research Journal, 72,

16 Krose, B., & Smagt, P. (1996). An Introduction to Neural Networks. 8th edition, The University of Amsterdam, Amesterdam, Netherland. Patterson, D.W. (1996). Artificial Neural Networks. 1st edition, Prentice Hall, Singapore. Polikar, R. (1995). The Wavelet tutorial, Rown university, Department of Electrical and Computer Engineering, Technical Report, 2nd Edition. Rojas, R. (1996). Neural Network: A Systematic Introduction. Springer, Berlin. Strang, G., & Nguyen, T. (1995). Wavelets and Filter Banks. Wellesley-Cambridge Press, Wellesley, MA. Sukigara, S., & Ishibashi, T. (2001). Subjective evaluation of the rustling sound of fabrics. SEN'I GAKKAISHI, 57(9), Titchmarsh, E. C. (1948). Introduction to the Theory of Fourier Integrals. Oxford: Oxford University Press. Yi, E., & Cho, G. (2000). Fabric sound classification by autoregressive parameters. The Journal of the Textile Institute, 91, Yi, E., Cho, G., Na, Y., & Casali, J. G. (2002). A fabric sound evaluation system for auditory-sensible textiles. Textile Research Journal, 72, Yi, E. (2005). Cross-cultural comparison of sound sensation and its prediction models for korean traditional silk fabrics. Fibers and Polymers, 6(3),