VIEWS: 294 PAGES: 10 CATEGORY: Emerging Technologies POSTED ON: 12/4/2010
(IJCSIS) International Journal of Computer Science and Information Security, Vol.8, No. 8, November 2010 Evaluation of Vision based Surface Roughness using Wavelet Transforms with Neural Network Approach *T.K.Thivakaran **Dr.RM.Chandrasekaran *Research Scholar, **Professor, Department of CSE, MS University, Annamalai University, Tirunelveli – 627012.INDIA Chidambaram – 620 024.INDIA Abstract---Machine vision for industry has generated a great result in inconsistent estimation of roughness of deal of interest in the technical community over the past components using machine vision primarily due to the fact several years. Extensive research has been performed on that illumination, shadow on the images is likely to be machine vision applications in manufacturing, because it has different. the advantage of being non-contact and as well faster than the contact methods. Using Machine Vision, it is possible to In this work, the machined surfaces are captured evaluate and analyze the area of the surface, in which machine using a Machine Vision system. Following the image vision extracted the information with the help of array of enhancement, the features are extracted and then the sensors to enable the user to make intelligent decision based on roughness parameters are estimated and analyzed. Here the applications. In this work, Estimation of surface roughness wavelet is used to extract the features of the enhanced has been done and analyzed using digital images of machined image, and an artificial neural network (ANN) is developed surface obtained by Machine vision system. Features are to predict the surface roughness. The results are compared extracted from the enhanced images in spatial frequency with that obtained using the standard stylus method. domain using a two dimensional Fourier Transform and Wavelet Transform. An artificial neural network (ANN) is trained using feature extracted values as input obtained from II. ROUGHNESS PARAMETERS wavelet Transform and tested to get Rt as output. The estimated roughness parameter (Rt) results based on ANN is compared with the Rt values obtained from Stylus method The machined surfaces are generally characterized by three and the best correlation between both the values are kinds of errors (i) form errors, (ii) waviness, and (iii) determined. surface roughness. The concept of roughness is often described with terms such as ‘uneven’,’ irregular’, ‘coarse Keywords--- Surface roughness, Machine vision, Milling, in texture’, broken by prominences’, and other similar ones Grinding, Wavelet Transform, Neural Network. (Thomas,1999). Similar to some surface properties such as hardness, the value of surface roughness depends on the I. INTRODUCTION scale of measurement. In addition, the concept roughness has statistical implications as it considers factors such as The quality of components produced is of main concern to sample size and sampling interval. The one parameter that the manufacturing industry, which normally refers to is standardized all over the world and is specified and dimensional accuracy, form and surface finish. Therefore, measured far more frequently than any other is the the inspection of surface roughness of the work piece is arithmetic average roughness height, or Roughness very important to assess the quality of a component, which Average. Universally called Ra, it was formerly known as is normally performed using stylus type devices, which AA (Arithmetic Average) in the United States and CLA correlate the vertical displacement of a diamond-tipped (Center Line Average) in the United Kingdom. It is defined stylus to the roughness of the surface under investigation. as the arithmetic mean of the departures of the profile from But, the limitations of stylus techniques have already been the mean line. reported in detail in [6, 5, 4]. Machine Vision typically employs a camera, a frame grabber, a digitizer and a Rq (or also known as RMS) is the root mean-square processor for inspection tasks where precision, repetition average of the departures of the roughness profile from the and/or high speed are needed. The histograms of the surface mean line. Rq has statistical significance because it image have been utilized to characterize surface roughness represents the standard deviation of the profile heights and and quality. Fourier transform (FT) of the digitized surface it is used in the more complex computation of skewness, image in which the magnitude and frequency information the measure of the symmetry of a profile about the mean obtained from the FT are used as measurement parameters line. of the surface finish. These methods use the basic assumption that the surface of the specimen is completely flat and there is no inclination when the images are captured. Even a small inclination of the specimen may … (1) 243 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol.8, No. 8, November 2010 Fig. 1(a), (c) The profiles obtained for a turned component A. Profiles for Turned Machined Components with a stylus instrument. (b), (d) the gap profiles obtained for the same turned components by diffraction method. Figure 1(a) and figure 1(c) shows the profiles obtained for a turned component with a stylus instrument. Similarly, B. Profile for Ground Machined Components figure 1(b) and 1(d) shows the gap profiles obtained for the Figure 2(a) shows the profiles obtained for a ground same turned components by diffraction method. In both component with a stylus instrument. Similarly, figure 3(b) graphs, ‘z’ is the deviation of the points on the profile from shows the gap profiles obtained for the same ground the mean-line. It can be observed that appreciable components by diffraction method. In both graphs, ‘z’ is the differences in the diffraction pattern are seen for large deviation of the points on the profile from the mean-line. variations in the gap and therefore good comparison of results is guaranteed in both only for turned components of medium roughness. For very rough surfaces scattering is observed. A limitation in the usage of the different methods is that the smoothness of the edge plays a crucial role in the evaluation of the finish of the components. Fig. 2(a) The profile obtained from a ground component with a stylus instrument, (b) The gap profile obtained for the same ground component by diffraction method. III. SPECTRUM TECHNIQUES FOR FEATURE EXTRACTION A. Fourier spectrum The Fourier spectrum is the frequency domain counterpart of the autocorrelation function. The FT of the correlation is used, which corresponds to the power spectral density function and describe how the power in a signal is distributed over frequency. The power spectrum can reveal the presence of offset, or periodic structures in a data set. B. Wavelet Transform (WT) The wavelet is a tool in surface texture analysis and can decompose a surface into multi-scale representation in a very efficient way. The wavelet transform (WT) is a mapping of the signal to the time-scale joint representation. By WT, the decomposition of a signal with a real 244 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol.8, No. 8, November 2010 orthonormal bases Ψmn(x) obtained through translation and where,ψ H ( x, y ) , ψ V ( x, y ) and ψ D ( x, y ) are called the dilation of a kernel function Ψ(x) known as mother wavelet horizontal, vertical and diagonal wavelets. Thus, DWT is as given in eqn. [2], well localized and allows decomposition in three directions, namely, horizontal, vertical and diagonal respectively. … (2) D. Features of Wavelets Where, m,n are integers. To construct the mother wavelet In this application, the features are extracted using a Ψ(x), it is required to determine a scaling function φ(x) wavelet which belongs to a family of orthogonal wavelets. given in eqn. [3], The mother wavelet (DB4), its corresponding scaling and wavelet functions and the decomposition filters are shown in Figure 3 and Figure 4 respectively. … (3) Then, the mother wavelet Ψ(x) is related to the scaling function as in eqn. [4], … (4) where, The coefficients h(k) have to meet several conditions for the set of basis wavelet functions to be unique, be orthonormal Fig. 3 wavelet extraction and also have a certain degree of regularity. C. Wavelet Transform for Signals In two dimensional cases, the one dimensional wavelet transforms are applied along both the horizontal and vertical directions φ ( x) is a one dimensional real, sequence integral scaling function defined as in [5] j φ j , k ( x ) = 2 2 φ (2 j x − k ) … (5) Fig. 4 Decomposition of low-pass filter h φ(-n) and high- Translation k determines the position of this one pass filter h ψ(-m) dimensional function along the x- axis, scale j determine its j The DB4 scaling function is given by width along x axis and 2 2 controls its height and ai = h0 s2i + h1s2i +1 + h2 s2i + 2 + h3 s2i +3 …(10) amplitude. This one dimensional scaling function satisfies these conditions: a [i ] = h0 s [ 2i ] + h1s [ 2i + 1] + h2 s [ 2i + 2] + h3s [ 2i + 3] … (11) φ j ,k is orthogonal to its integer translates. The Daubechies DB4 wavelet function is given by The set of functions that can be represented as a series expansion of φ j ,k at low scale is contained ci = g0 s2i + g1s2i +1 + g 2 s2i + 2 + g3 s2i +3 … (12) c [i] = g0 s [ 2i] + g1s [ 2i +1] + g2 s [ 2i + 2] + g3s [ 2i + 3] … (13) within those at higher scale. So, the difference between any two sets of φ j ,k is IV. NEURAL NETWORKS FOR SURFACE ROUGHNESS represented by a companion wavelet function ψ j , k defined ASSESSMENT The roughness features extracted from the machined j images, are fed as input to an ANN to predict the roughness in eqn. [6], ψ j , k ( x) = 2 2 ψ (2 j x − k ) value Rt. ANN consists of a number of elementary units … (6) called neurons. A neuron is a simple processor, which can Then, the 2 dimensional DWT functions are the take multiple inputs and produce an output. Each input into linear products of scaling and wavelet functions φ ( x) the neuron has an associated weight that determines the ‘‘intensity’’ of the input. The processes that a neuron and ψ ( x) yielding the eqn. [7] through eqn. [9]. performs are: multiplication of each of the inputs by its ψ ( x, y ) = ψ ( x).φ ( y ) H … (7) respective weight, adding up the resulting numbers for all the inputs and determination of the output according to the ψ V ( x, y ) = φ ( x).ψ ( y ) … (8) result of this summation and an activation function. Data is ψ D ( x, y) = ψ ( x).ψ ( y ) … (9) fed into the network through an input layer, it is processed through one or more intermediate hidden layers and finally 245 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol.8, No. 8, November 2010 fed out of the network through an output layer as shown in Horizontal, Vertical and Diagonal Figure 5. components). Calculate the weighted standard deviations of three detailed images. …(14) where, = Standard deviation of the M detail image at ith Fig. 5 Typical ANN network Level M=H(Horizontal)/V(Vertical)/D(Diagonal) V. PROPOSED SYSTEM FOR SURFACE ROUGHNESS component EVALUATION The standard deviation of each sub image at level i is weighted by the factor (1/2i-1), The methodology and block diagram of proposed Machine vision system is shown in Figure 6(a) and Figure 6(b). (iv) Repeat steps 1-4 four times for original image and images at orientation 90º, 180º, and 270 º (achieved by rotating original image). The final feature set consists of 4*(3L) features. Fig.6 (a) Block diagram of proposed system B. Wavelet based Feature Extraction Since, the wavelet coefficient are orthogonal, the original profile can be re-obtained after wavelet decomposition by simply adding the sub-scales signals as shown in Figure 7. Furthermore, using this simple summation technique the concepts of roughness, waviness and form can be preserved. This is reflected in Figure 8 where, an arbitrary decomposition of a surface texture is obtained by casting into three frequency components, representing the form, waviness and roughness, using Daubechies wavelet of order 20. A dimensional step can now be cleared. Indeed, the same kind of decomposition process can be performed using images instead of profiles, because surface roughness can be measured precisely using for instance optical surface measurement systems. The arbitrary decomposition into form waviness and roughness of surface textures obtained by casting, grinding and vertical milling respectively, using wavelet of order 20 is shown. The roughness average of each component (i.e. Fig.6 (b) Methodology in the proposed computer vision form, waviness and roughness) is also shows in order to system for measuring surface roughness illustrate the roughness scale. The measured area is of a few millimeters square. Hence, the wavelet tool allows the A. Algorithm for feature extraction decomposition of surfaces into form, waviness and roughness components and can successfully replace (i) Carry out image enhancement of machined standard filters that are commonly used in surface texture image characterization and hence, give a solid theoretical base for (ii) Subject the enhanced image to a L-level the standardization of these filters. discrete wavelet decomposition. (iii) At each level (i=1, 2, … L), there are four sub-images. One approximation image and three components/images (LH, HL and HH or 246 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol.8, No. 8, November 2010 Fig. 7 Multiscale decomposition of a surface texture profile Fig. 9 Comparison of multiscale decomposition of a surface obtained by casting under seven different scales using the texture profile obtained by casting under seven different wavelet of order 20 scales using the wavelet of order 20 and its frequency normalized equivalent. C. ANN based surface roughness estimation ANN with a variation of the classic back-propagation algorithm is employed to predict surface roughness. Compared with more conventional approaches, ANN demonstrates certain advantages that make them much more attractive. They have the ability to recognize patterns that are similar, but not identical, it can store information and generalize it. There is no need for explicit statement of Fig. 8 Multiscale decomposition of a surface texture profile the problem or for a problem-solving algorithm. Due to obtained by casting under three different components (form their massive parallelism, ANNs display increased waviness and roughness) using the wavelet of order 20. computational power that can be used to deal with complex problems. Back-propagation neural network used for In Figure 9, the FNWT maxima indicate at each scale the estimating the surface roughness of the machined surfaces location of a frequency component. Those features can also with is a four layer network with six nodes in the input be quantified according to both the shape of the layer, six nodes in the first hidden layer, five nodes in the corresponding peak and its height. For an image, when second hidden layer and one single node in the output layer. using a multiresolution scheme for a dyadic standard Each layer is fully connected to the succeeding decomposition of a function into sub-bands a filter bank layer. The outputs of nodes of one layer are transmitted to with a power of two number of filters should be used. When nodes in another layer through links. The structure of an using orthogonal wavelets like ones, one can easily simplify ANN is shown in Figure 10 where the Energy maps are fed the problem by gathering the channels by scale in both as inputs into the trained neural network and the surface directions. This process applied to a discrete wavelet is roughness parameter (Rt) is estimated . In the training called the scaled DWT. The frequency normalization can phase, the desired value of the node in the output layer is then be performed based on these filters. the actual roughness value, Rt obtained by stylus method. The ANN adjusts the weights in all connecting links such that the mean square error, i.e. the averaged squared error 247 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol.8, No. 8, November 2010 between the network output and the desired output is minimized. Training of ANN is stopped as soon as the specified number of epochs has reached and the values of weights corresponding to the minimum error are restored. … (14) Once trained, the ANN is then tested for different sets of input data. In the testing phase of the neural network, the predicted roughness, Rt is the value of the node in the output layer. … (15) … (16) … (17) Fig. 10 The System architecture of ANN used for predicting In Figure 12, f(x,y) is the highest resolution representation Ra for surfaces of the image being transformed. It serves as the input for the first iteration and for the succeeding iterations; the VI. RESULTS AND DISCUSSION approximation coefficients Wφ (j, m, n) are given as input to the filter bank, to obtain the next set of wavelet coefficients. Case 1: Feature Extraction using Scaled DWT The blocks contain time reversed scaling and wavelet vectors. The hφ (-n) and hΨ (-m) are low pass and high pass decomposition filters. Blocks are containing a down arrow and represent down sampling extracting every other point from a sequence of points. Each pass through the filter bank in Figure 11 decomposes the input signal into four lower resolutions (or lower scale) components. The Wφ Fig. 12 Sub-band image decomposition for wavelet based coefficients are created by two low pass (hφ based) filters feature extraction and are thus called the approximation coefficients and {Wφ i for i = H, V, D} are the horizontal, vertical and diagonal detail coefficients. Thus the energy for each subband is calculated up to 4 levels of decomposition and the image features Et, Eh, Ev and Ed are obtained from the energy map which is determined using tree-structured wavelet transform for each image. Few Sample enhanced machine images [Figure 13(a) to 18(a)] are applied with DWT and the respective transform outcomes are shown in [Figure 13(b) to 18(b)] along with the energy details in Table 1. Fig. 11 2D DWT filter bank Mathematically, the series of filtering and down sampling operations are used to compute the DWT coefficients Wφ (j,m,n) and {Wφi (j,m,n) for i = H,V,D} at scale j. Figure 13(a) 248 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol.8, No. 8, November 2010 Figure 16(a) Figure 13(b) Transform absolute coefficient Figure 16(b) Transform absolute coefficient Figure 14(a) Figure 17(a). Figure 14(b) Transform absolute coefficient Figure 15(a). Figure 17(b) Transform absolute coefficient Figure 18(a) Figure 15(b) Transform absolute coefficient 249 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol.8, No. 8, November 2010 Table 2 ANN estimated Rt for milling parameters (Image enhancement done and image features extracted using FT) Figure 18(b) Transform absolute coefficient Table 1 Energy maps obtained from DWT Fig. Ea Edetail Values Name Values Et Eh Ev Ed 13(a) 99.0286 0.0396 0.1714 0.4482 0.3122 14(a) 99.0088 0.0204 0.1183 0.3862 0.4663 15(a) 97.7414 0.0272 0.2686 0.6214 1.3414 16(a) 98.5546 0.0393 0.3785 0.7579 0.2697 17(a) 98.5984 0.0606 0.2933 0.5217 0.5260 Table 3 ANN estimated Rt for milling parameters (Image enhancement done and image features extracted using WT) 18(a) 96.8104 0.0282 0.1311 0.5938 2.4364 Where Et is Energy total, Eh is Energy horizontal, Ev is Energy Vertical and Ed is Energy diagonal. Ea is Energy Approximation. Case 2: Estimation of Rt using ANN (a) For Milled surfaces Two types of feature extraction and surface roughness estimation using ANN is performed in this work. The first one extracts the features using FT and the second uses the WT. In FT approach the key input features collected for training the network consist of (i) average grey scale value (Ga) (ii) major peak frequency (F1) and (iii) Principal component magnitude squared value (F2). The WT based feature extraction is already discussed in case (i) of section VI. In the training phase (for both FT and WT) the desired value of the node in the output layer is the surface roughness Rt obtained using the stylus method. The surface roughness Rt from ANN along with the stylus The results obtained are validated by plotting the measurement values for the milled samples after image correlation graph between stylus measured (conventional enhancement with FT (WT) extracted features is given in method) Rr and vision measured (proposed) Rt for both the Table 2 (Table 3). FT and WT techniques for milled components is shown in Figure 15. 250 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol.8, No. 8, November 2010 Table 5 ANN estimated Rt for grinding parameters (Image enhancement done and image features extracted using WT) Figure 15 Comparison between predicted roughness values using vision approach and stylus approach for FT features and WT features (milling) (b) For grinding operations The Rt value predicted using the trained ANN and that measured using the stylus approach for the grinding process after image enhancement with features extracted using FT (WT) is given in Table 4 (Table 5). The results obtained are validated by plotting the correlation graph between stylus measured (conventional method) Rr and vision measured (proposed) Rt for both the Table 4 ANN estimated Rt for grinding parameters (Image FT and WT techniques for grinding components is shown enhancement done and image features extracted using FT) in Figure 16. Figure 16 Comparison between predicted roughness values using vision approach and stylus approach for FT features and WT features (grinding). VII. CONCLUSION AND FUTURE ENHANCEMENT. The developed model is tested online on images of specimens grabbed by computer vision systems with linearly decreasing intensity. The features of the grabbed enhanced image (to remove noise present in the captured image) are extracted using two different schemes, one using Fourier transform (FT) and the other using wavelet 251 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol.8, No. 8, November 2010 decomposition. The FT method is used to extract the [10] Daubechies, The wavelet transform, time-frequency features of image texture, namely, the major peak frequency localization and signal analysis, IEEE Trans. Inform. F1, and the principal component magnitude squared value Theory 36 (1990) 961–1005. F2. Using the wavelet (Db4) multi resolution decomposition [11] Xiaodong Gu, Daoheng Yu, Liming Zhang, Image algorithm, the energy details of the sub band images, shadow removal using pulse coupled neural network, namely, energy total (Et), energy horizontal (Eh), energy IEEE Transactions on Neural Networks 16 (3) (2005) vertical (Ev) and energy diagonal (Ed) are extracted. These 692–698. extracted features of the enhanced image are given as input [12] X.Q. Jiang, L. Blunt, K.J. Stout, Three-dimensional to a trained neural network (back propagation network) and surface characterization for orthopaedic joint the surface roughness parameter Rt is estimated. From the prostheses, proceedings of institution of mechanical obtained results, it is concluded that the wavelet based engineers. Part H, J. Eng. Med. 213 (1) (1999) 49–68. image feature extraction of the enhanced images gives [13] Grzesik W., Rech J., Wanat T.: Comparative study of better correlation between vision Rt and the stylus Rt both the surface roughness produced in various hard for milled and grinding surfaces. machining processes. 3rd International Congress of Precision Machining, Vienna, Austria, 2005, pp. 119- Future direction of research shall focus on 124. implementing the proposed algorithms using high speed [14] Josso B., Burton D., Lalor M.: Frequency normalized hardware units thus making the present work ideally for wavelet transform for surface roughness analysis and high speed real-time machine vision applications. characterization. Wear, Vol. 252, 2002, pp. 491-500. [15] S.S. Liu, M.E. Jernigan, Texture analysis and discrimination in additive noise, Computer Vision, VIII REFERENCES Graphics and Image Processing 49 (1) (1990) 52–67. [16] Zawada-Tomkiewicz A., Storch B.: Introduction of the [1] G.A. Al-Kindi, R.M. Baul, K.F. Gill, An application of wavelet analysis of a machined surface profile. machine vision in the automated inspection of Advances in Manufacturing Science and Technology, engineering surfaces, International Journal of Vol. 28, No. 2, 2004, pp. 91-100. Production Research 30 (2) (1992) 241–253 [17] S.-H. Lee, H. Zahouani, R. Caterini, T.G. Mathia, [2] M.B. Kiran, B. Ramamoorthy, B. Radhakrishnan, Morphological characterization of engineered surfaces Evaluation of surface roughness by vision system, by wavelet transform, in: Proceedings of the 7th International Journal of Machine Tools & Manufacture International Conference on Metrology and Properties 38 (5–6) (1998) 685–690. of Engineering Surfaces, Götebarg, Sweden, 1997, pp. [3] Du-Ming Tsai, Jeng-Jong Chen, Jeng-Fung Chen, A 182–190. vision system for surface roughness assessment using IX AUTHORS PROFILE neural networks, International Journal of Advanced Manufacturing Technology 14 (6) (1998) 412–422. [4] M.Y. Rafiq, G. Bugmann, D.J. Easterbrook, Neural 1. Mr. T.K.Thivakaran is presently a research scholar in network design for engineering applications, MS university, Thirunelveli in the faculty of Computer Computers and Structures 79 (17) (2001) 1541–1552. Science and Engineering. He is working as Assistant [5] K. Venkata Ramana, B. Ramamoorthy, Statistical Professor in the faculty of Information Technology, Sri methods to compare the texture features of machine Venkateswara college of Engineering, Chennai. His area of surfaces, Pattern Recognition 29 (9) (1996) 1447– research includes Image Processing, Cryptography and 1459. Network Security. [6] P.G. Benardos, G.C. Vosniakos, Prediction of surface roughness in CNC face milling using neural networks 2. Dr.RM.Chandrasekaran is currently working as a and Taguchi’s design of experiments, Robotics and Professor at the Department of Computer Science and Computer Integrated Manufacturing 18 (5–6) (2002) Engineering, Annamalai University, Annamalai Nagar, 343–354. Tamilnadu, India. From 1999 to 2001 he worked as a [7] Shengyu Fu, B. Muralikrishnan, J. Raja, “Engineering software consultant in Etiam, Inc, California, USA. He Surface Analysis with different wavelet bases”, Trans. received his Ph.D degree in 2006 from Annamalai of ASME, vol 125, Nov. 2003, pp. 844-852. University, Chidambaram. He has conducted workshops [8] H.T. Hingle and J.H. Rakels, The practical application and conferences in the area of Multimedia, Business of diffraction techniques to assess surface finish of Intelligence, Analysis of Algorithms and Data Mining. He diamond turned parts, Ann. CARP, 32(1)(1983)499- has presented and published more than 32 papers in 501. conferences and journals and is the co-author of the book [9] B. Josso, D.R. Burton, M.J. Lalor, Wavelet strategy for Numerical Methods with C++ Program( PHI,2005). His surface roughness analysis and characterisation, research interests include Data Mining, Algorithms, Image Comput. Methods Applications. Mech. Eng. 191 (8– processing and Mobile Computing. He is life member of the 10) (2001) 829–842. Computer Society of India, Indian Society for Technical Education, Institute of Engineers etc. 252 http://sites.google.com/site/ijcsis/ ISSN 1947-5500