Advances in Production Engineering & Management Volume 11 | Number 2 | June 2016 | pp 93-104 http://dx.doi.Org/10.14743/apem2016.2.212 ISSN 1854-6250 Journal home: apem-journal.org Original scientific paper Surface roughness assessing based on digital image features Simunovic, G.a*, Svalina, I.a, Simunovic, K.a, Saric, T.a, Havrlisan, S.a, Vukelic, D.b aMechanical Engineering Faculty in Slavonski Brod, University of Osijek, Croatia bFaculty of Technical Sciences, University of Novi Sad, Serbia A B S T R A C T A R T I C L E I N F O The paper gives an account of the machined surface roughness investigation based on the features of a digital image taken subsequent to the technological operation of milling of aluminium alloy A16060. The data used for investigation were obtained by mixed-level factorial design with two replicates. Input variables (factors) are represented by the face milling basic machining parameters: spindle speed (at five levels: 2000; 3500; 5000; 6500; 8000 rev/min, respectively), feed per tooth (at six levels: 0.025; 0.1; 0.175; 0.25; 0.325; 0.4 mm/tooth, respectively) and depth of cut (at two levels: 1; 2 mm, respectively). Output variable or response is the most frequently used surface roughness parameter - arithmetic average of the roughness profile, Ra. Digital image of the machined surface is provided for every test sample. Based on experimental design and obtained results of roughness measuring, a base has been created of input data (features) extracted from digital images of the samples' machined surfaces. This base was later used for generating the fuzzy inference system for prediction of the surface roughness using the adaptive neuro-fuzzy inference system (ANFIS). Assessing error, i.e. comparison of the assessed value Ra provided by the system with real Ra values, is expressed with the normalized root mean square error (NRMSE) and it is 0.0698 (6.98 %). © 2016 PEI, University of Maribor. All rights reserved. Keywords: Surface roughness Face milling Digital image Adaptive neuro-fuzzy inference system *Corresponding author: goran.simunovic@sfsb.hr (Simunovic, G.) Article history: Received 7 March 2016 Revised 15 May 2016 Accepted 16 May 2016 1. Introduction Surface roughness is an important technological parameter and indicator of the machined surface quality. Requirements for lower values of surface roughness simultaneously affect the prolongation of machining time and increase of production costs. Surface roughness is conditioned by a larger number of controlled and uncontrolled process parameters (including cutting speed, depth of cut and feed rate, raw material properties, cutting conditions, tool properties, tool machine vibrations, tool wear etc.). By regular monitoring the results of a machining process and expanding the knowledge base about the monitored parameters of observed processes, it is possible to continuously improve a product characteristic and production results. There is a great number of scientific investigations aimed at prediction and control of surface roughness [1-4]. The models defined in these investigations can be divided into regression (statistic), analytic (mathematic) and those based on the application of artificial intelligence (AI) [5-8]. It is often the case that the digital image features of the machined surfaces are used in controlling or assessing the machined surface roughness. The image features are used as input variables for the assessing model [9-13], and they are mostly represented by statistic values such as arithmetic mean and standard deviation [14], different kinds of standards such as the Euclidean and the Hamming norm [15], wave transformations such as the Haar wavelet transform [16] and 93 Simunovic, Svalina, Simunovic, Saric, Havrlisan, Vukelic the two-dimensional Fourier transform [17] etc. Adaptive neuro-fuzzy inference system (ANFIS), artificial neural networks (ANN), regression analysis and others are the methods mostly used for assessing. Lee et al. [14] propose a method using an adaptive neuro-fuzzy inference system (ANFIS) to establish the relationship between actual surface roughness and texture features of the surface image. The input parameters of a training model are spatial frequency, arithmetic mean value, and standard deviation of grey levels from the surface image. In paper [18] the ANFIS is also used in assessing the surface roughness using cutting parameters (cutting speed, feed rate, and depth of cut) and grey level of the surface image. The assessing model error is less than 4.6 %. In papers [19] and [20] machine vision system is also used integrated with ANFIS. Paper [19] assesses Ra, tool wear ratio and metal removal rate in micro-turning process. The assessing error is less than 3.5 %. Investigation in paper [20] is directed to assessment of surface roughness of end milled parts. Using a two dimensional Fourier transform (2D FT) features of image texture are extracted, such as peak frequency, principal component magnitude squared value and the average grey level. The ANFIS and the neural networks methods used in assessing roughness are compared and the assessing errors are very close and less than 2.5 %. In paper [21] the Euclidean and Hamming distances of the surface images are used for surface recognition. Machined surface images with different values of surface roughness were collected. The base is formed of referent images with known values of surface roughness. The Euclidean and Hamming distances between the tested surface and the referent surface image were used in the base to predict the surface roughness of the unknown surface. Excellent concluding results were obtained and the system is suitable for online surface characterization of machined surfaces. The paper [16] presents methodology based on the extraction of texture features from part surface images in the frequency domain using wavelet transform. One-level Haar wavelet transform is applied to the original surface images. Surface evaluation was accomplished by means of the analysis of grey levels in the vertical detail sub-image of surface images. Experimental results show that the proposed approach achieves error rates between 2.59% and 4.17%. Paper [22] is also based on the application of the wavelet transform. Authors apply vision system for acquiring digital images of machined surfaces, analyse the image parameters and connect them with the roughness of the surface machined by turning. The machined surface digital images are described using the one-dimensional digital wavelet transform. The neural networks based system is used for assessing roughness. The testing phase error is a bit more than 5 %. Papers [23, 24] also apply Machine Vision and ANN. In paper [23] they are used to assess Ra values using the input obtained from the digital images of inclined surfaces which include optical roughness parameters (major peak frequency, principal component magnitude squared, average power spectrum, central power spectrum percentage, ratio of major axis to minor axis). To improve the quality of the images shadow removal algorithm is used. The high value of the ANN model correlation coefficient (87 %) confirms its applicability. Through computer vision system authors in paper [24] collect features of image texture of machined surface (major peak frequency, principal component magnitude squared value and the standard deviation of grey level and by the application of abductive networks they assess surface roughness of turned parts. The assessment error is around 15 %. Authors in papers [25-27] analyse interconnection between the machined surface texture and the machining time, in other words condition and wear of tools. Authors [25] have investigated the relationship between texture features of the grey-level co-occurrence matrix and the machining time in turning operations. Results of investigation have shown that the error between the actual and the calculated machining time ranges from -4.65 % to 7.79 %. Authors in paper [26] used machine vision technique to detect the condition of tools on the basis of turned surface images using an accurate grey level co-occurrence matrix. In paper [27] authors investigated cutting tool nose wear area and surface roughness of turned parts using machine vision system. They developed an algorithm that uses Wiener filtering and simple thresholding on backlit images in order to reduce the impact of ambient factors (ambient lighting) and vibrations. The developed system roughness assessing error was 10 %. In paper [28] authors investigated connection between surface roughness of AA 6061 alloy end milling and image texture features of milled surface. They used grey level co-occurrence matrix to ex- 94 Advances in Production Engineering & Management 11(2) 2016 Surface roughness assessing based on digital image features tract the image texture features (contrast, homogeneity, correlation and energy). For establishing the relation between surface roughness and image texture the regression analysis is applied. The paper [29] demonstrates that some roughness parameters (Ra, Rq, Rv, Rt and Rz) can be estimated using only image-extracted features and models, without the knowledge of machining parameters. Authors observe three image texture features of turned surface: gradient factor of surface, average cycle of texture and average grey level. There is a very high correlation between surface roughness and the given features of digital image. Authors in [9] investigated surface lay in the surface roughness evaluation using machine vision. Numerous parameters of digital image are considered such as grey level average, grey level co-occurrence matrix based image quantification parameters (contrast, correlation, energy or uniformity, maximum probability and differential box courting based fractal dimension) of machined surface while changing the angle of taking images. Therefore, it can be concluded that investigations were directed towards building a system (machine vision system) for a quicker and cheaper control, i.e. assessment of the roughness of machined surfaces in real time. The actual paper investigations are an additional contribution for assessing roughness of machined surface based on the features of digital image using the adaptive neuro-fuzzy inference system (ANFIS). 2. Experimental and methodology 2.1 Experimental Investigation is conducted on the material of samples Al6060 T66 (in accordance with the European norms EN AW-6060 T66 [AlMgSi]). Chemical composition of the alloy Al6060 or EN AW-6060 according to EN 573-3 is given in Table 1. Mass fraction of other elements can be to the maximum of 0.15 %, and individually 0.05 %. Mechanical and physical properties (at 20 °C) of alloy Al6060 or EN AW-6060 according to EN 755-2 are given in Table 2. Dimensions of samples are 100 x 60 x 10 mm. Samples (Fig. 1) are machined from flat bars of transverse section 60 x 10 mm. For face milling of 100 x 60 mm surface vertical CNC milling machine was used produced by HASS type VF-2 and face milling cutter of diameter 40 mm produced by Walter with four cutting inserts (holder mark: F 4042.B.040.Z04.15 and inserts mark: ADMT160608R-F56 WKP35S). The machining was carried out based on CNC programme that repeated the same path of the tool. The following machining parameters were being changed: spindle speed, feed per tooth, and depth of cut in an order defined by the selected mixed-level factorial design. For clamping of samples a hydraulic machine vice Alfa NCO-A was used. As the applied fuzzy inference system will have three inputs, the suggested factorial design has three factors. By a detailed analysis, considering the total number if input/output experimental data for the training phase and the checking phase of inference system, a mixed-level factorial design was selected. Five spindle speeds (2000; 3500; 5000; 6500; 8000 rev/min, respectively), six feeds per tooth (0.025; 0.1; 0.175; 0.25; 0.325; 0.4 mm/tooth, respectively) and two depths of cut (1; 2 mm, respectively) are used, and two replicates of a mixed-level factorial design are run. Table 1 Chemical composition of A16060 according to EN 573-3 (wt%) Si Fe Cu Mn Mg Cr Zn Ti 0.3-0.6 0.1-0.3 max 0.1 max 0.1 0.35-0.6 max 0.1 max 0.1 max 0.1 Table 2 Mechanical and physical properties (at 20 °C) of material Al6060 according to EN 755-2 Yield tensile strength, MPa 150 Density, kg/m3 2700 Ultimate tensile strength, MPa 195 Melting point, °C 585-650 Elongation at break, % 8 Electrical conductivity, mS/m 28-34 Hardness, HB 65 Thermal conductivity, W/mK 200-220 Modulus of elasticity, GPa 70 Coefficient of thermal expansion, 10-6/K 23.4 Advances in Production Engineering & Management 11(2) 2016 95 Simunovic, Svalina, Simunovic, Saric, Havrlisan, Vukelic Fig. 1 Preparation of samples All other features specific for end milling: tool stepover between neighbour paths, number of passes, total length of paths, are held-constant factors, the same as the Maxol cooling/lubricating fluid produced by Forol d.d., tool (milling cutter) and material of the sample. 2.2 Methodology Adaptive neuro-fuzzy inference system (ANFIS) method for generating fuzzy inference system requires a set of input/output experimental data. The fuzzy inference system (FIS) has three input variables. In generating FIS the ANFIS method with three membership functions per each input was used. Therefore, 27 different fuzzy rules are used to form the base of fuzzy rules. For the FIS first row the base of fuzzy rules can be written as: Rule 1: If x is A1 and y is B1 and w is C1 then z is f1 (x, y, w) Rule 2: If x is A2 and y is B2 and w is C2 then z is f2 (x, y, w) Rule 27: If x is A3 and y is B3 and w is C3 then z is f27(x,y, w) where x, y and w present ANFIS inputs, Aj,Bj and Cj fuzzy sets, and (x, y, w) is the polynomial first row and represents the output of the first row of Sugeno FIS. The system has adaptive nodes (the sets' parameters that are changeable-adaptive) and fixed nodes (the sets' parameters that are fixed-unchangeable). By arrangement, the nodes outputs are marked as Qi t where I represents the layer and i the number of nodes. Five layers are usually used to explain the ANFIS architecture. Layer 1 contains adaptive nodes. Layer 1 node functions are described as: (j = 1fori = 1.....9 Qi,i=HAjW\j = 2fori = 10.....18 [j = 3 for i = 19,...,27 (j = 1 for i = 1,2,3,10,11,12,19,20,21 Qi,i =HBj(y) \j = 2 for i = 4,5,6,13,14,15,22,23,24 (; = 3 for i = 7,8,9,16,17,18,25,26,27 j7 = 1 for ¿ = 1,4,7.....25 Qi,i =^cj(w)lj = 2 for i = 2,5,8,...,26 (y = 3fori = 3,6,9, ...,27 (1) 96 Advances in Production Engineering & Management 11(2) 2016 Surface roughness assessing based on digital image features where x, y and w present node inputs, Aj,Bj and Cj linguistic labels and ju^.,^. and membership functions. Membership functions determine the degree in which some variable satisfies a specific rule premise. There are various membership functions. In this paper a bell-shape membership function is applied whose general form can be written as: 111 Kx) =-r ; J"(y) =-r =-r m "Fir1? "Pi?? i + F?1) (2) where a^bi and ct are parameters of fuzzy sets. Layer 2 contains fixed nodes n. This layer fixed nodes represent multiplication of input signals whose product is the output for each node. Q2i -nCj, for i = 1, .„,27 (3) Output is called the firing strength of a fuzzy rule. Layer 3 contains fixed nodes N. This layer node functions calculate the ratio of the i-th firing strength of a rule and the firing strength of all rules. Q3,i =<¡»1= fj-, for i = 1.....27 (4) Output <¿>1 is called normalized firing strength of a fuzzy rule. Layer 4 contains adaptive nodes. This layer node functions are expressed as: Q*,i -fi, for i = 1, .,27 (5) where ft represents conclusions of fuzzy rules for which is valid: fi = PiX + qty + riW + Si, for i = 1,... ,27 (6) where pi,qi,ri and Sj are called consequent parameters. Layer 5 contains only one fixed node. Function of this node is to calculate the overall output using: 27 Qs = font = -fi = iX)pi + (coiy)qi + (coiw)ri + (¿¿¿^ (7) ¿=i The arithmetic average of the roughness profile Ra can be expressed as: n n Ra = ^coi (¿¡(ki o + kil-n + ki2 - fz + ki3 -ap) (8) ¿=i ¿=i where k = [k10,k11,k12,k13,k2o,k2i,k22,k23,..-,kn0,knl,kn2,kn3] represents the consequent parameters vector, n spindle speed, fz feed per tooth, ap depth of cut, normalized firing strength and as output the arithmetic average of the roughness profile Ra. Output of every fuzzy rule is connected with the output function defined by three different consequent parameters. It can be concluded from the foregoing that in training of the system 81 parameters are being adapted which then requires minimally 81 sets of input/output experimental data for the training of the FIS for assessing surface roughness. In addition to the training phase, the input/output experimental data are necessary for the checking phase too. For the checking phase 10 % of input/output experimental data are to be provided. Advances in Production Engineering & Management 11(2) 2016 97 Simunovic, Svalina, Simunovic, Saric, Havrlisan, Vukelic 3. Results and discussion Fig. 2 shows a machined sample. The arithmetic average of the roughness profile Ra is measured according to the standard ISO 4288 by means of a portable surface roughness tester produced by Taylor & Hobson model Surtronic S128. The arithmetic average of the roughness profile Ra is measured on mid part of samples (between two white horizontal lines) as shown in Fig. 2 for each run separately. The upper and lower lines are 40 mm apart from the ends of the samples so that the central part width is 20 mm. The arithmetic average of the roughness profile Ra is measured vertically to the tangents of tool traces on the line where the tool traces are most apart, this line being at 10 mm distance from the left and the right edge of the sample and is parallel with them. For the measuring data processing the Talyprofile software produced by Taylor & Hobson is applied, designed to be used with the Surtronic series S-100 instruments. After the experiment the acquisition of machined surfaces digital images of all samples was carried out using table scanner Scanjet 3100. The scanner optical resolution of 1200 points per inch was used to obtain greyscale image, i.e. image of the grey colour shades. For the greyscale image 8 bits per pixel were used while the grey colour shades values were represented in 256 levels. After the acquisition all digital images are registered in matrix form from which input variables are quantified: greyscale mean value of all digital image matrix members, greyscale standard deviation of all digital image matrix members and the digital image matrix greyscale entropy that are, along with the measured arithmetic average of the roughness profile Ra, used as an input/output data base in creating the FIS for surface roughness assessing. The greyscale mean value of digital image matrix is: N N Mean=^XX/(x,y) (9) x=ly=1 where N presents the number of columns and number of rows of digital image, and f(x,y) is a greyscale intensity value of the digital image matrix member defined by x andy. Fig. 2 Machined sample and roughness measuring points 98 Advances in Production Engineering & Management 11(2) 2016 Surface roughness assessing based on digital image features The greyscale standard deviation of all digital image matrix members (N = 250) can be described as: 1 Std = — N N N ¿¿CfOcoO-/) (10) x=ly=1 where N presents the number of columns and number of rows of the digital image matrix, f(x, y) is a greyscale intensity value of the digital image matrix defined by x andy while f is the digital image matrix mean greyscale value. Entropy is a statistical measure of randomness that can be used to characterize the texture of the input image. The digital image matrix greyscale entropy is described as: 256 £ = ^(PiXlog2Pi) (11) ¿=i where E is a scalar value representing the entropy of greyscale image I, and p is a vector which contains the histogram counts. Resolution of all digital images used in this investigation was 250 x 250 pixels. The used resolution represents the size of the machined surface digital images denoting the place where the arithmetic average of the roughness profile Ra was measured along with the surrounding surface. The surface shown in the used digital images is between the white horizontal lines of the samples displayed in Fig. 2 for each run separately. To serve the needs of the current paper the digital images are used of those runs in which a higher value of the arithmetic average of the roughness profile Ra was measured. The digital image matrix consists of 250 rows and 250 columns. Table 3 displays the extracted values of roughness (higher values Ramax are displayed of two repeated measurements), and the earlier described input variables for creating the fuzzy inference system (mean greyscale value of all digital image matrix members, greyscale standard deviation of all digital image matrix members and the digital image greyscale matrix entropy). The given data were used for generating FIS for assessing surface roughness using ANFIS. This system assesses the Ra values on the basis of the machined surfaces digital images and their features. The error of assessing i.e. of comparisons provided by the system with real values Ra, is expressed by the average normalized root mean square error (NRMSE). The assessing error of the fuzzy inference system created in this investigation is 0.0698 or 6.98 %. The figures that follow demonstrate the arithmetic average of the roughness profile Ra dependence on greyscale standard deviation of all digital image matrix members and greyscale mean value of all digital image matrix members (Fig. 3), on entropy of digital image greyscale matrix and greyscale standard deviation of all digital image matrix members (Fig. 4), on entropy of digital image greyscale matrix and on the greyscale mean value of all digital image matrix members (Fig. 5). The arithmetic average of the roughness profile Ra ranging from 0.194 |j.m to 1.68 |j.m has been measured in the experimental investigations. The measured outputs are ranged in four classes of surface roughness: N3 (from 0.1 |j.m to 0.2 |j.m), N4 (from 0.2 |j.m to 0.4 |j.m), N5 (from 0.4 |j.m to 0.8 |j.m), N6 (from 0.8 |j.m to 1.6 |j.m) and N7 (from 1.6 |j.m to 3.2 |j.m). The applicable parts of response surface in this particular case are given separately in Fig. 6. It can be seen from Fig. 6 that slight changes in input variables have a considerable effect on output variable, the arithmetic average of the roughness profile Ra. Advances in Production Engineering & Management 11(2) 2016 99 Simunovic, Svalina, Simunovic, Saric, Havrlisan, Vukelic Table 3 Extracted values of roughness depending on technological parameters of machining, and the digital image values of variables for creating the fuzzy inference system Feed per Bench Spindle Depth RQmax, ^m Greyscale Greyscale Greyscale STD RUN tooth, feed, speed, of cut, mean standard matrix mm/tooth mm/min rev/min mm value deviation entropy 1 99 0.175 2450 3500 2 0.236 201.3327 12.9087 5.7077 2 32 0.1 800 2000 2 0.194 215.8912 14.4114 5.8688 3 40 0.25 3500 3500 2 0.340 216.4442 10.9891 5.4273 4 22 0.25 6500 6500 1 0.325 211.9918 14.9249 5.8883 5 34 0.25 2000 2000 2 0.291 211.6711 18.7212 6.2131 6 92 0.1 800 2000 2 0.259 207.0648 11.9313 5.6197 7 91 0.025 200 2000 2 0.295 205.5052 14.2197 5.8618 8 119 0.325 10400 8000 2 0.284 202.4441 21.4677 6.3535 9 112 0.25 6500 6500 2 0.892 172.1966 23.6160 6.5007 10 35 0.325 2600 2000 2 1.270 203.2657 26.0274 6,6797 11 6 0.4 3200 2000 1 1.050 181.5731 33.5879 7.0212 12 55 0.025 800 8000 2 0.549 200.8293 9.9470 5.3400 13 3 0.175 1400 2000 1 0.510 138.9439 30.9279 6.8194 14 96 0.4 3200 2000 2 1.300 192.3719 27.7303 6.7241 15 68 0.1 1400 3500 1 0.405 138.9501 18.3823 6.2209 16 16 0.25 5000 5000 1 0.845 146.9739 29.7534 6.8952 17 69 0.175 2450 3500 1 0.504 134.2705 27.7919 6.7490 18 115 0.025 800 8000 2 0.333 197.0068 8.3465 5.0556 19 11 0.325 4550 3500 1 0.755 124.9713 27.7997 6.6697 20 101 0.325 4550 3500 2 1.340 169.0206 24.9581 6.6209 21 54 0.4 10400 6500 2 1.560 175.4106 26.5330 6.5932 22 105 0.175 3500 5000 2 0.631 141.6713 18.3205 6.0764 23 30 0.4 12800 8000 1 1.030 149.9594 29.6730 6.8415 24 45 0.175 3500 5000 2 0.462 139.8943 19.5561 6.1107 25 77 0.325 6500 5000 1 0.684 142.7332 23.7919 6.5355 26 60 0.4 12800 8000 2 1.680 171.5458 25.1323 6.5677 27 113 0.325 8450 6500 2 1.660 150.6172 23.1638 6.5385 28 74 0.1 2000 5000 1 0.413 142.7963 19.5778 6.2314 29 78 0.4 8000 5000 1 1.040 149.5937 21.4949 6.4195 30 72 0.4 5600 3500 1 1.310 149.2047 21.1949 6.4186 31 62 0.1 800 2000 1 0.408 146.4589 17.7637 6.0372 32 70 0.25 3500 3500 1 0.742 159.4245 25.2728 6.5954 33 107 0.325 6500 5000 2 1.570 157.0338 26.5375 6.7189 34 51 0.175 4550 6500 2 0.359 142.6573 22.1164 6.4169 35 33 0.175 1400 2000 2 0.417 155.2297 20.8527 6.3764 36 7 0.025 350 3500 1 0.459 197.1356 8.5898 5.0740 37 86 0.1 3200 8000 1 0.376 150.2156 20.6194 6.3526 38 103 0.025 500 5000 2 0.342 216.3073 10.3385 5.3718 39 38 0.1 1400 3500 2 0.526 166.4716 21.4820 6.4444 40 100 0.25 3500 3500 2 1.140 164.7204 33.5821 7.0123 41 58 0.25 8000 8000 2 1.270 177.2273 29.6023 6.8859 42 76 0.25 5000 5000 1 0.802 157.8756 26.7976 6.6881 43 81 0.175 4550 6500 1 0.521 134.4788 22.0851 6.3085 44 89 0.325 10400 8000 1 0.761 152.5526 20.9315 6.3611 45 42 0.4 5600 3500 2 1.210 167.3245 21.3705 6.4070 46 28 0.25 8000 8000 1 0.855 145.4409 25.8796 6.5799 47 17 0.325 6500 5000 1 0.812 147.7410 20.9663 6.4149 48 41 0.325 4550 3500 2 1.540 171.2488 20.8178 6.3214 49 120 0.4 12800 8000 2 1.470 156.4950 21.1037 6.4185 50 47 0.325 6500 5000 2 1.090 176.9611 27.9537 6.8013 51 111 0.175 4550 6500 2 0.470 140.2953 20.8329 6.3067 52 98 0.1 1400 3500 2 0.374 145.4403 24.1626 6.4329 53 80 0.1 2600 6500 1 0.399 162.8991 21.2284 6.3343 54 14 0.1 2000 5000 1 0.366 154.3155 19.6337 6.2506 55 46 0.25 5000 5000 2 1.040 176.2519 32.7465 6.9622 56 56 0.1 3200 8000 2 0.364 155.8688 17.3140 6.0190 57 63 0.175 1400 2000 1 0.597 153.3269 29.7856 6.8004 58 104 0.1 2000 5000 2 0.386 159.6362 20.5342 6.2324 59 93 0.175 1400 2000 2 0.422 165.1706 23.2286 6.4068 60 117 0.175 5600 8000 2 0.508 149.4792 19.4560 6.0821 100 Advances in Production Engineering & Management 11(2) 2016 Surface roughness assessing based on digital image features Table 3 Extracted values of roughness depending on technological parameters of machining, and the digital image values of variables for creating the fuzzy inference system (continuation) Feed per Bench Spindle Depth Rßmax, ^m Greyscale Greyscale Greyscale STD RUN tooth, feed, speed, of cut, mean standard matrix mm/tooth mm/min rev/min mm value deviation entropy 6l 8 0.1 1400 3500 1 0.402 156.5143 28.9983 6.7787 62 67 0.025 350 3500 1 0.463 224.1073 11.3163 5.4469 63 84 0.4 10400 6500 1 0.977 151.3250 27.0345 6.6960 64 90 0.4 12800 8000 1 0.945 155.1779 25.9010 6.6127 65 97 0.025 350 3500 2 0.202 218.8573 11.8558 5.5605 66 83 0.325 8450 6500 1 0.666 154.6065 27.3491 6.7255 67 59 0.325 10400 8000 2 1.300 152.1983 24.2644 6.5952 68 106 0.25 5000 5000 2 1.150 157.4970 30.3303 6.7487 69 79 0.025 650 6500 1 0.336 174.5179 16.0944 5.9967 70 5 0.325 2600 2000 1 0.778 152.4728 23.3246 6.5248 7l 48 0.4 8000 5000 2 1.600 163.3946 25.3079 6.6660 72 108 0.4 8000 5000 2 1.560 158.2985 19.6642 6.2761 73 53 0.325 8450 6500 2 1.460 176.6186 27.1825 6.7422 74 29 0.325 10400 8000 1 0.715 173.9498 23.3387 6.5550 75 21 0.175 4550 6500 1 0.513 157.2126 25.0313 6.5654 76 25 0.025 800 8000 1 0.257 194.3272 17.8550 6.0991 77 52 0.25 6500 6500 2 1.500 176.9590 34.1330 7.0064 78 109 0.025 650 6500 2 0.510 215.7320 7.6283 4.9336 79 37 0.025 350 3500 2 0.264 184.7601 12.6105 5.6820 80 64 0.25 2000 2000 1 0.932 164.4880 23.3275 6.4361 8l 31 0.025 200 2000 2 0.326 196.8387 9.8417 5.3327 82 49 0.025 650 6500 2 0.277 196.4967 18.2400 6.1901 83 27 0.175 5600 8000 1 0.418 180.7377 30.4711 6.7616 84 110 0.1 2600 6500 2 0.319 194.5997 16.2075 6.0494 85 61 0.025 200 2000 1 0.449 202.4482 12.0212 5.6169 86 88 0.25 8000 8000 1 0.345 208.5057 18.1894 6.2054 87 50 0.1 2600 6500 2 0.362 216.7653 12.7031 5.6936 88 87 0.175 5600 8000 1 0.315 203.7227 16.5828 6.0807 89 26 0.1 3200 8000 1 0.377 204.2278 17.9229 6.1923 90 43 0.025 500 5000 2 0.322 225.7510 4.8023 4.3056 9l 94 0.25 2000 2000 2 0.391 185.6580 17.7899 6.1080 92 19 0.025 650 6500 1 0.412 226.0397 4.4847 4.2012 93 102 0.4 5600 3500 2 0.657 193.8495 17.3608 6.1385 94 44 0.1 2000 5000 2 0.302 205.5300 15.0375 5.9108 95 15 0.175 3500 5000 1 0.317 198.4703 17.8107 6.1613 96 73 0.025 500 5000 1 0.361 217.4270 5.6124 4.5311 97 1 0.025 200 2000 1 0.404 210.2537 11.3657 5.5270 98 36 0.4 3200 2000 2 0.652 183.9130 18.6225 6.1525 99 116 0.1 3200 8000 2 0.371 210.2263 13.8650 5.8338 l00 82 0.25 6500 6500 1 0.359 194.8133 13.8527 5.8269 l0l 71 0.325 4550 3500 1 0.454 179.7367 21.4364 6.3853 l02 23 0.325 8450 6500 1 0.413 200.2917 16.8814 6.0965 l03 65 0.325 2600 2000 1 0.550 183.4905 18.8896 6.2459 l04 2 0.1 800 2000 1 0.371 220.3295 13.6538 5.8036 l05 114 0.4 10400 6500 2 0.561 185.6083 18.8118 6.2636 l06 4 0.25 2000 2000 1 0.369 201.7676 15.8028 6.0238 l07 18 0.4 8000 5000 1 0.544 186.0058 22.0295 6.4740 l08 85 0.025 800 8000 1 0.408 219.5023 5.5210 4.4854 l09 75 0.175 3500 5000 1 0.347 187.1892 15.5448 5.9636 ll0 24 0.4 10400 6500 1 0.529 205.4299 15.7066 6.0073 lll 95 0.325 2600 2000 2 0.557 176.9512 21.9477 6.4261 ll2 39 0.175 2450 3500 2 0.336 190.3445 19.2155 6.2940 ll3 13 0.025 500 5000 1 0.415 225.9157 5.7998 4.5690 114 9 0.175 2450 3500 1 0.323 188.4408 17.3426 6.1525 ll5 66 0.4 3200 2000 1 0.566 182.2851 19.7799 6.3321 116 20 0.1 2600 6500 1 0.378 218.3453 12.7629 5.6885 117 57 0.175 5600 8000 2 0.332 186.9872 18.1653 6.2152 118 10 0.25 3500 3500 1 0.356 189.2162 16.7248 6.1067 119 12 0.4 5600 3500 1 0.495 184.4284 19.5615 6.3077 120 118 0.25 8000 8000 2 0.321 192.7420 17.5391 6.1643 Advances in Production Engineering & Management 11(2) 2016 101 Simunovic, Svalina, Simunovic, Saric, Havrlisan, Vukelic Fig. 3 Dependence of Ra on standard deviation and mean Fig. 4 Dependence of Ra on entropy and standard devia-value of digital image matrix members tion of digital image matrix members It can be seen from the papers that deal with assessing the roughness of machined surfaces on the basis of the features of digital image that the range of measured roughness has a great impact on the level of error assessing. The wider the range of measured roughness, with the uniform distribution by the roughness classes, the lower the error of assessment. The error of assessment in this study (6.98 %) was significantly influenced by outlier values. Specifically, almost 97 % of the measured values of roughness belong to the roughness classes N4, N5 and N6. The remaining values and part of roughness values in class N6 are outliers. Without outliers the error of assessment of the machined surface roughness is expected to be significantly lower. 4. Conclusion The conducted investigation is part of a project whose ultimate objective is to build an online system for machined surface roughness monitoring i.e. roughness monitoring in real time. The system should faster carry out the activities of required control of machined surfaces, testing would be cheaper, and monitoring during machining would help to timely react to possible deviations and to reduce subsequent costs. The investigations in this paper are focused on assessing the machined surface roughness based on the features of digital image with the use of adaptive neuro-fuzzy inference system (ANFIS). A controlled parameter of surface roughness is the arithmetic average of the roughness profile Ra. The following features of digital image are studied in the paper: mean greyscale values of all digital image matrix members, standard greyscale deviation of all digital image matrix members and entropy of digital image greyscale matrix. Comparison of real values Ra and the values provided by the built system is shown by the nor- 102 Advances in Production Engineering & Management 11(2) 2016 Surface roughness assessing based on digital image features malized root mean square error (NRMSE), or assessing error. The conducted investigation enters the area of high speed machining. Therefore the machined surfaces are of high quality and the measured roughness is very small. Thus the features of digital images become quite similar and a higher assessing error is expected. The fuzzy inference system obtained in the present investigation has an assessing error of 6.98 %. However, even with such an error, the technical requirements set on the workpiece as regards quality of machining, should not be diminished. The plan is to expand the research on existing material, but also conduct research on other materials. This would be a way to expand the base of digital photos and their features and to accumulate sufficient knowledge to influence the reduction of assessing errors. Acknowledgement This research is accomplished within the projects Nos. IZIP-2014-95 and INGI-2015-28 financed by the Josip Juraj Strossmayer University of Osijek. References [1] Stankovic, I., Perinic, M., Jurkovic, Z., Mandic, V., Maricic, S. (2012). Usage of neural network for the prediction of surface roughness after the roller burnishing, Metalurgija, Vol. 51, No. 2, 207-210. [2] Simunovic, K., Simunovic, G., Saric, T. (2015). Single and multiple goal optimization of structural steel face milling process considering different methods of cooling/lubricating, Journal of Cleaner Production, Vol. 94, 321-329, doi: 10.1016/j.jclepro.2015.02.015. [3] Zuperl, U., Cus, F. (2015). Simulation and visual control of chip size for constant surface roughness, International Journal of Simulation Modelling, Vol. 14, No. 3, 392-403, doi: 10.2507/IJSIMM14(3)2.282. [4] Vukelic, D., Tadic, B., Miljanic, D., Budak, I., Todorovic, P.M., Randjelovic, S., Jeremic, B.M. (2012). Novel workpiece clamping method for increased machining performance, Tehnički vjesnik - Technical Gazette, Vol. 19, No. 4, 837846. [5] Brezocnik, M., Kovacic, M., Ficko, M. (2004). Prediction of surface roughness with genetic programming, Journal of Materials Processing Technology, Vol. 157-158, 28-36, doi: 10.1016/j.jmatprotec.2004.09.004. [6] Pare, V., Agnihotri, G., Krishna, C. (2015). Selection of optimum process parameters in high speed CNC end-milling of composite materials using meta heuristic techniques - A comparative study, Strojniški vestnik - Journal of Mechanical Engineering, Vol. 61, No. 2, 176-186, doi: 10.5545/sv-jme.2014.1914. [7] Simunovic, G., Simunovic, K., Saric, T. (2013). Modelling and simulation of surface roughness in face milling, International Journal of Simulation Modelling, Vol. 12, No. 3, 141-153, doi: 10.2507/IISIMM12(3)1.219. [8] £olak, o. (2014). Optimization of machining performance in high-pressure assisted turning of Ti6Al4V alloy, Strojniški vestnik - Journal of Mechanical Engineering, Vol. 60, No. 10, 675-681, doi: 10.5545/sv-jme.2014.1914. [9] Nammi, S., Ramamoorthy, B. (2014). Effect of surface lay in the surface roughness evaluation using machine vision, Optik - International Journal for Light and Electron Optics, Vol. 125, No. 15, 3954-3960, doi: 10.1016/ j.ijleo.2014.01.152. [10] Klancnik, S., Ficko, M., Balic J., Pahole, I. (2015). Computer vision-based approach to end mill tool monitoring, International Journal of Simulation Modelling, Vol. 14, No. 4, 571-583, doi: 10.2507/IJSIMM14(4)1.301. [11] Krolczyk, G., Raos, P., Legutko, S. (2014). Experimental analysis of surface roughness and surface texture of machined and fused deposition modelled parts, Tehnički vjesnik - Technical Gazette, Vol. 21, No. 1, 217-221. [12] Samta§, G. (2014). Measurement and evaluation of surface roughness based on optic system using image processing and artificial neural network, The International Journal of Advanced Manufacturing Technology, Vol. 73, No. 1, 353-364, doi: 10.1007/s00170-014-5828-1. [13] St^pien, K., Makiela, W., Stoic, A., Samardžic, I. (2015). Defining the criteria to select the wavelet type for the assessment of surface quality, Tehnički vjesnik - Technical Gazette, Vol. 22, No. 3, 781-784, doi: 10.17559/TV-20140124110406. [14] Lee, K.C., Ho, S.J., Ho, S.Y. (2005). Accurate estimation of surface roughness from texture features of the surface image using an adaptive neuro-fuzzy inference system, Precision Engineering, Vol. 29, No. 1, 95-100, doi: 10.1016/j.precisioneng.2004. 05.002. [15] Jeyapoovan, T., Murugan, M. (2013). Surface roughness classification using image processing, Measurement, Vol. 46, No. 7, 2065-2072, doi: 10.1016/j.measurement.2013.03.014. [16] Morala-Arguello, P., Barreiro, J., Alegre, E. (2012). A evaluation of surface roughness classes by computer vision using wavelet transform in the frequency domain, The International Journal of Advanced Manufacturing Technology, Vol. 59, No. 1, 213-220, doi: 10.1007/s00170-011-3480-6. [17] Palani, S., Natarajan, U. (2011). Prediction of surface roughness in CNC end milling by machine vision system using artificial neural network based on 2D Fourier transform, The International Journal of Advanced Manufacturing Technology, Vol. 54, No. 9, 1033-1042, doi: 10.1007/s00170-010-3018-3. Advances in Production Engineering & Management 11(2) 2016 103 Simunovic, Svalina, Simunovic, Saric, Havrlisan, Vukelic [18] Ho, S.Y., Lee, K.C., Chen, S.S., Ho, S.J. (2002). Accurate modeling and prediction of surface roughness by computer vision in turning operations using an adaptive neuro-fuzzy inference system, International Journal of Machine Tools and Manufacture, Vol. 42, No. 13, 1441-1446, doi: 10.1016/S0890-6955(02)00078-0. [19] Palani, S., Natarajan, U., Chellamalai, M. (2013). On-line prediction of micro-turning multi-response variables by machine vision system using adaptive neuro-fuzzy inference system (ANFIS), Machine Vision and Applications, Vol. 24, No. 1, 19-32, doi: 10.1007/s00138-011-0378-0. [20] Natarajan, U., Palani, S., Anandampilai, B. (2012). Prediction of surface roughness in milling by machine vision using ANFIS, Computer-Aided Design & Applications, Vol. 9, No. 3, 269-288, doi: 10.3722/cadaps.2012.269-288. [21] Jeyapoovan, T., Murugan, M. (2013). Surface roughness classification using image processing, Measurement, Vol. 46, No. 7, 2065-2072, doi: 10.1016/j.measurement.2013.03.014. [22] Zawada-Tomkiewicz, A. (2010). Estimation of surface roughness parameter based on machined surface image, Metrology and Measurement Systems, Vol. 17, No. 3, 493-504. [23] Priya, P., Ramamoorthy, B. (2007). The influence of component inclination on surface finish evaluation using digital image processing, International Journal of Machine Tools and Manufacture, Vol. 47, No. 3-4, 570-579, doi: 10.1016/j.ijmachtools.2006.05.005. [24] Lee, B.Y., Yu, S.F., Juan, H. (2004). The model of surface roughness inspection by vision system in turning, Mecha-tronics, Vol. 14, No. 1, 129-141, doi: 10.1016/S0957-4158(02)00096-X. [25] Gadelmawla, E.S., Al-Mufadi, F.A., Al-Aboodi, A.S. (2014). Calculation of the machining time of cutting tools from captured images of machined parts using image texture features, Proceedings of the Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture, Vol. 228, No. 2, 203-214, doi: 10.1177/0954405413481291. [26] Dutta, S., Datta, A., Das Chakladar, N., Pal, S.K., Mukhopadhyay, S., Sen, R. (2012). Detection of tool condition from the turned surface images using an accurate grey level co-occurrence technique, Precision Engineering, Vol. 36, No. 3, 458-466, doi: 10.1016/j.precisioneng.2012.02.004. [27] Shahabi, H.H., Ratnam, M.M. (2009). In-cycle monitoring of tool nose wear and surface roughness of turned parts using machine vision, The International Journal of Advanced Manufacturing Technology, Vol. 40, No. 11, 11481157, doi: 10.1007/s00170-008-1430-8. [28] Nathan, D., Thanigaiyarasu, G., Vani, K. (2014). Study on the relationship between surface roughness of AA6061 alloy end milling and image texture features of milled surface, Procedia Engineering, Vol. 97, 150-157, doi: 10.1016/j.proeng.2014.12.236. [29] Kamguem, R., Tahan, S.A., Songmene, V. (2013). Evaluation of machined part surface roughness using image texture gradient factor, International Journal of Precision Engineering and Manufacturing, Vol. 14, No. 2, 183-190, doi: 10.1007/s12541-013-0026-x. 104 Advances in Production Engineering & Management 11(2) 2016