eduzhai > Physical Sciences > Materials Sciences >

Prediction of manufacturing process and surface roughness by artificial neural network

  • sky
  • (0) Download
  • 20211030
  • Save
https://www.eduzhai.net International Journal of M aterials Engineering 2013, 3(3): 47-58 DOI: 10.5923/j.ijme.20130303.04 Prediction of Manufacturing Processes and Surface Roughness Using ANN Ossama Abouelatta Production Engineering and M echanical Design Dept., Faculty of Engineering, M ansoura University, 35516, M ansoura, Egypt Abstract Surfaceroughness isconsidered as one of the most specified customer requirements in machin ing processes. For efficient use of machine tools, selection of machin ing process and determination of optimal cutting parameters (speed, feed and depth of cut) are required. Therefore, it is necessary to find a suitable way to select and to find optimal machining process and cutting parameters for a specified surface roughness values. In this work, mach ining process was carried out on AISI 1040 steel in dry cutting condition in a lathe, milling and grinding machines and surface roughness was measured. Forty five experiments have been conducted using varying speed, feed, and depth of cut in order to find the surface roughness parameters. This data has been divided into two sets on a random basis; 36 training data set and 9 testing data set. The training data set has been used to train different art ificia l neura l network (ANN) models in order to predict machining processes and surface roughness parameter valuesthrough back propagation network. Experimental data co llected fro m tests were used as input parameters ofa neural network to identify the sensitivity among machining operations, cutting parameters and surface roughness. A selected indexes were used to design a suitable algorithm for the prediction of machin ing processes. A software wasdeveloped and imp lemented to predict the machin ing processes and surface roughness values. The results showed that the proposed models are capable of pred icting machin ing operations, cutting parameters and surface roughness. Keywords Manufacturing Processes, Surface Roughness,Neural Network, Pred iction 1. Introduction Increased demands for higher product quality, reliability, and manufacturing efficiency levels have imposed stringent requirements on automated product measurement and evaluation. The progress in the development of predictive models, based on cutting theory, has not yet met the objective of maintain ing high quality production at low cost.The most essential cutting performance measures, such as, tool life, cutting force, roughness of the machined surface, energy consumption, … etc., should be defined using experimental studies. Therefore, further improvement and optimization for the technological and economic performance of machining operations depend on a well-based experimental methodology. Hayajneh et al.[1], aimed to develop a better understanding of the effects of cutting parameters on the surface roughness and to build a mu ltip le regression model. Suhail et al.[2], proposed a method for cutting parameters identificat ion using multi adaptive network based fuzzy inference system (MANFIS). Their results showed that the MANFIS model can be used successfully for machinability data selection. * Corresponding author: abouelatta@m ans.edu.eg (OssamaAbouelatta) Published online at https://www.eduzhai.net Copyright © 2013 Scientific & Academic Publishing. All Rights Reserved To deal with a nonlinear optimization problem wh ich aims to minimize the unit production cost in mu lti-pass turnings, Xie and Guo[3] proposed a new approach, which comb ines genetic algorith ms with a pass enumerating method. Co mputer simu lation results showed that the optimization approach can find the better results than other algorithms proposed previously to significantly reduce the unit production cost. Grzesik and Wanat[4] provided a comprehensive analysis of part surface finish in continuous dry turning of hardened construction steel when using mixed alu mina cutting tools. Their results showed that hard turning with mixed alu mina tools produces specific surface profiles and microstructures, although the Ra roughness values of about 0.25 µm can be co mparable to those produced by fin ish grinding. Surface roughness has been modeled by Patrikar[5], based on power series and Fast Fourier Transform methods. A method based on neural networks has been used to model these surfaces to map the process parameters to roughness parameters. In recent years, the trends were towards modeling of mach ining using artificial intelligence. Artificial neural networks are considered one of the important methods of artificial intelligence in the modeling of nonlinear problems like machining processes. ANN show good capability in prediction and optimizat ion of machining processes compared with tradit ional methods. In turn ing operations, Luong and Spedding[6] described the application of 48 Ossama Abouelatta: Prediction of M anufacturing Processes and Surface Roughness Using ANN neural-network technology to the selection of mach ining parameters and to the prediction of machin ing performance in metal cutting via an on-line imp lementation of the trained network using the C progra mming language. A neural-network-based methodology was proposed by Kohli and Dixit [7], for predicting the surface roughness in a turning process by taking the acceleration of the radial vibration of the tool holder as feedback. They stated that their methodology was able to make accurate predict ion of surface roughness by utilizing small sized training and testing datasets. ANN and mu ltiple regression approaches were used by Asiltürk and Çunkaş[8], to model the surface roughness of AISI 1040 steel. The A NN model was used to estimate the surface roughness with high accuracy compared to the multip le regression model. Models for p redicting surface roughness of AISI 1040 steel material using artific ial neural networks and mult iple regression were developed by Asiltürk[9]. Mean squared error of 0.00292% ach ieved using the developed ANN outperforms error rates reported in earlier studies and could also be considered admissible for real-t ime deploy ment of the developed ANN algorith m for robust prediction of the surface roughness in industrial s ettin gs . In milling operations, multip le regression and ANN techniques were applied by Murthy and Rajendran[10], to predict the surface roughness. The results of the prediction models were quite close with experiment values. They reported that the feed rate was the most dominant factor in influencing surface roughness. The results also showed that the highest cutting speed, mediu m feed rate and mediu m depth of cut produces lowest surface roughness. An emp irical approach using a statistical analysis was employed by Huang and Chen[11], to discover the proper cutting force in order to represent the uncontrollable factors in end milling operations using an in-process neural network-based surface roughness prediction system. Response surface model and an art ificial neural network were developed by Erzuru mlu and Oktem[12], to predict surface roughness values error on mo ld surfaces. The response surface model and an artific ia l neural network we re compared with manufacturing problems such as computational cost, cutting forces, tool life, d imensional accuracy, etc. A new technique was developed by Wibowo and Desa[13], using hybridizat ion of kernel principal component analysis based on nonlinear regression and GAs to estimatethe optimu m values of surface roughness. Their results showed that the presented technique gave more accurate prediction model than the ordinary linear regression’s approach. The need for precision components and parts in manufacturing industries has bought an increase in the need for finishing operations that can satisfy this demand. Grinding has the potential to meet these critical needs for accurate and economic means of finishing parts, and generate the required surface topography. An approach suggesting the combination of design of experiment method and ANN was developed by Fredj and Amamou[14]. The built ANNs showed low deviation fro m the training data, low deviation fro m the testing data and high sensibility to the inputs levels. The high prediction accuracy of the developed ANNs was confirmed by the good agreement with the results of empirical models developed by previous investigations. Mukherjee and Kumar Ray[15] attempted to provide a systematic methodology to develop a mult ivariate linear regression model, hypothesis testing for the influence of nonlinear terms to linear model, and accordingly selection of a suitable artificial neural network-based inferential model with imp roved prediction accuracy and control o f grinding behavior. The development of neural model-based control strategies for the optimization o f an industrial alu miniu m substrate disk grinding process was described by Govindhasamy et al.[16], using a nonlinear autoregressive exogenous based neural network model. Preliminary plant investigations show that thickness defects can be reduced by 50% o r more, co mpared to other schemesemp loyed. Ku mar and Choudhury[17] focused on the prediction of wheel wear and surface roughness using two techniques, namely design of experiments and neural network. Effect of p rocess parameters, such as pulse current, duty ratio, wheel speed and grain size on output responses, namely, wheel wear and surface roughness of high speed steel were investigated e xperimentally. This paper aims to reduce the huge number of two dimensional surface roughness parameters to a reasonable number of groups. It aims to propose some indexes may be suitable for the prediction of machining processes and mach ining parameters based on surface roughness analysis. This work was done experimentally and with the help of neural network. 2. Experimental Setup and Procedure 2.1. Design of Experiment Experiments have been performed in order to investigate the effects of mach ining parameters (speed, feed and depth of cut) on the surface fin ish of the machined surface. Forty five specimens made of AISI 1040 steel, 45 mm diameter with 10 mm in height, were used for experimentation using a turning (facing), milling and grinding machines. Chemical composition and mechanical properties of AISI 1040 steel are listed in Table 1. The specimens were divided into three groups according to the machining operation; each has 15 specimens (12 for training and 3 for evaluation). Befo re emp loying machining process, each specimen was polished using sandpapers grade 1200 to eliminate the effect of previous cuts. All experiments have been done under dry mach ining environ ment. Several variables were put under close control including the machining condition (the same mach ine was used for all experimental work for each group), and the operator (the same operator machined all specimens for each machine). The first group was produced by lathe turning machine International Journal of M aterials Engineering 2013, 3(3): 47-58 49 (Facing) under different cutting parameters. A cemented carbide cutting tool equipped with throwaway inserts was used in the experiments. The insert type was KPGN 160412 and the shank type was MSKNR/L 2525M16 for approach angle χ=35°. The second group was produced by a plain milling machine. A 3-inch (≈ 80 mm outer diameter), 14-flute high-speed steel cutters and a 80 mm width was used in the experiments. The third group was machined by a surface grinder machine with electro magnetic chuck. The used cutting tool for grinding operations is alu miniu m o xide grinding wheels (38A60KK5VBE). 2.2. Surface Roughness Measurements There is an industrial need for the measurement and classification of the topography of engineering surfaces.Two-dimensional (2D) analysis is fast, but limited in the usefulness of the results obtainable whilst the three-dimensional (3D) approach offers greater scope. Unfortunately, the data analysis step for 3D data characterizat ion can be time-consuming, and often requires a skilled metro logist[18]. In this investigation, a set of 2D surface roughness parameters, as well as profile and surface characteristics, such as the amplitude distribution functions, bearing area curves and symmetrical curves of geometrical contact obtained for the machined surface, were determined and analyzed. The surface roughness data were collected for each of the 45 samp les for the mach ining processes and conditions as defined in Table 2. A surface roughness measuring system consists of two major parts, hardware and software was used.The hardware consists of two main items as shown in Fig. 1. Mitutoyo SurfTest-SJ201 is a surface measuring instrument, used in contact method surface assessment. Two special programs written in MatlabTM, Prediction of Machining Process and Cutting Parameters (PMPCP) and Surface Roughness Calculation Program (SRCP) were used to give a fu ll assessment of surface roughness parameters fro m the resulted surface profile, Fig. 2. The PMPCP program is used to measure, import and export surface profile and surface roughness data from M itutoyo SurfTest-SJ201. By clicking the "Run SRCP" button, the SRCP program can be invoked to calculate more than 66 standard and non-standard surface roughness parameters as shown in Table 3[19]. Three measurements for workp iece surface roughness were made and averaged for each test. Figure 3 shows surface texture and surface profile for some manufacturing processes. Table 1. Chemical composition and mechanical properties of AISI 1040 steel Chemical composition (Wt. %) Mechanical properties Carbon Iron Manganese Phosphorus Sulphur Others σyield (MPa) σultimate (MPa) Vickers Hardness Modulus of Elasticity (GPa) 0.37 98.8 0.7 0.035 0.045 0.05 450 515 155 200 Table 2 (a). Cutting and measured roughness parameters at different machining processes (Training data) T urning operation No. f s dc Ra S mm/rev rpm mm µm mm Milling operation Pc f s dc Ra S Count mm/min rpm mm µm mm Pc Vw Count m/min Grinding operation Vc ae Ra S m/s µm µm mm Pc Count 1 0.16 140 0.3 7.05 112.2 72 28 157 0.2 3.41 120.7 67 6 35 20 0.57 53.8 164 2 0.16 224 0.3 4.69 79.3 101 55 157 0.2 3.73 127.3 64 6 35 30 0.83 49.3 166 3 0.16 560 0.3 6.77 115.4 70 157 157 0.2 4.15 135.6 63 6 35 40 0.76 46.1 176 4 0.16 900 0.3 3.25 105.4 83 110 157 0.2 5.23 165.4 53 6 35 50 0.84 51.5 157 5 0.16 560 0.1 8.01 116.0 71 140 157 0.2 3.58 126.0 67 6 20 20 0.90 59.7 136 6 0.16 560 0.75 5.00 121.0 76 110 80 0.2 3.09 149.8 57 6 20 30 1.00 53.1 155 7 0.16 560 1.5 2.94 75.9 107 41 157 0.45 3.66 179.3 47 6 20 40 0.70 49.9 163 8 0.08 560 0.3 6.00 111.9 74 110 80 0.2 3.07 142.6 59 6 20 50 0.69 50.5 168 9 0.23 560 0.3 4.41 89.0 94 110 80 0.1 2.99 121.2 67 8.4 20 20 0.64 46.7 178 10 0.32 560 0.3 6.59 113.6 71 110 157 0.35 5.11 165.5 50 7.2 35 20 0.66 47.8 169 11 0.46 560 0.3 6.38 95.1 82 110 157 0.25 5.06 147.1 62 4.8 35 20 0.73 42.7 192 12 0.59 560 0.3 9.59 106.8 75 110 157 0.3 3.71 150.2 54 3.6 20 20 0.65 48.3 167 Table 2 (b). Cutting and measured roughness parameters at different machining processes (Verification data) T urning operation No. f s dc Ra S mm/rev rpm mm µm mm Milling operation Pc f s dc Ra S Count mm/min rpm mm µm mm Pc Vw Count m/min Grinding operation Vc ae Ra S m/s µm µm mm Pc Count 13 0.26 710 1.25 5.22 120.0 68 280 439 0.4 6.19 205.6 40 10 20 20 0.82 57.4 141 14 0.32 180 1.25 5.91 90.3 89 20 40 0.3 6.38 158.9 60 40 35 60 0.78 46.3 176 15 0.52 450 1.75 6.03 95.2 84 110 157 0.15 3.93 123.5 65 5 35 10 0.94 56.5 145 50 Ossama Abouelatta: Prediction of M anufacturing Processes and Surface Roughness Using ANN 1 2 3 456 7 1. Personal computer 2. Mitutoyo interface program 3. Specimen 4. Stylus 5. SurfTest-SJ201 driving unit 6. Holder 7. SurfTest-SJ201 control unit Figure 1. SurfTest-SJ201 and personal computer (PC) (a) PMPCP, Prediction of Manufacturing Process and Cutting P aramet ers P rogram. (b) SRCP, Surface Roughness Calculation Program. Figure 2. Programs to predict manufacturing processes and cutting parameters, and assess surface roughness parameters 3. Results and Discussion 3.1. Analysis of Surface Roughness Parameters In recent years there has been a pro liferat ion of parameters with which to specify surface texture. So me of these parameters are useful, but most are not, (Whitehouse, 1982)[20]. The result of this rash is confusion and expense. So, a measure of the strength of the linear relationship between roughness parameters were applied using MS Office Excel to reduce the number of surface roughness parameters used in surface analysis. Table 4 shows a part of correlation coefficient between surface roughness parameter. Then, the surface roughness parameters regrouped with that having a strong correlation coefficient (R≥0.90). The results of this analysis are grouped in six categories (I-VI) and listed in Table 5. For example, the first group includes Ra, Rq, Rp, Rv, Rt, NPeaks, ACF10% , ACF, WF roughness parameters. This means that any roughness parameter within group (I) is enough to study the effect of cutting parameter, for example, on the resulting surface roughness. International Journal of M aterials Engineering 2013, 3(3): 47-58 51 (a) Turned workpiece No. (1) (b) Milled workpiece No. (5) (c) Ground workpiece No. (6) Figure 3. Samples of surface texture and surface profile for some machining processes Table 3(a). Standard and non-standard surface roughness parameters Surface roughness parameter Arithmetic average roughness, Ra Geometric average roughness, Rq Skewness, Rsk Kurtosis, Rku Mean of Rpi values, Rpm Mean of Rvi values, Rvm Mean of Rti values, Rtm Maximum height, Rp Maximum depth, Rv Maximum peak to valley height, Rt Largest Rti value, Ry Ten-point height, RzISO Ten-point height, RzDIN Mean of R3y, R3z Fullness grade, Km Emptiness grade, Kp Median Standard Deviation, SD T urning (Facing) Milling Grinding Mean 5.86 7.15 0.10 2.82 14.77 12.32 27.09 18.41 17.19 35.60 33.90 27.09 13.54 7.42 0.48 0.52 -0.57 7.15 Min. 2.94 3.69 -0.42 2.30 8.16 7.19 15.52 10.31 9.37 19.68 18.40 15.52 7.76 2.09 0.35 0.44 -2.98 3.69 Max. 9.59 11.21 0.69 4.03 25.99 18.54 40.24 28.71 25.52 48.93 47.71 40.24 20.12 16.31 0.56 0.65 0.39 11.22 Mean 4.22 5.34 -0.53 3.56 8.56 10.53 19.09 11.91 16.27 28.18 26.34 19.09 9.54 3.14 0.58 0.42 0.52 5.34 Min. 2.99 3.96 -1.28 2.57 5.87 8.71 15.08 8.61 13.56 23.05 18.51 15.08 7.54 0.53 0.46 0.33 0.00 3.96 Max. 6.38 7.90 0.14 6.12 14.71 14.00 28.71 19.36 22.36 36.30 35.19 28.71 14.36 8.09 0.67 0.54 1.46 7.90 Mean 0.78 0.98 -0.27 3.15 1.95 2.58 4.53 2.52 3.45 5.98 5.51 4.53 2.27 0.32 0.58 0.42 0.05 0.98 Min. 0.57 0.71 -0.57 2.87 1.50 1.60 3.10 1.97 2.10 4.07 3.82 3.10 1.55 -0.37 0.51 0.38 0.00 0.71 Max. 1.00 1.25 0.06 3.58 2.43 3.47 5.91 3.28 4.48 7.75 6.93 5.91 2.95 1.17 0.62 0.49 0.11 1.25 Un it s µm µm --------µm µm µm µm µm µm µm µm µm µm --------µm µm 52 Ossama Abouelatta: Prediction of M anufacturing Processes and Surface Roughness Using ANN Table 3(b). St andard and non-st andard surface roughness paramet ers(Cont inued) Surface roughness parameter High spot count at 10%, HSC10% High spot count at 20%, HSC20% High spot count at 30%, HSC30% High spot count at 40%, HSC40% High spot count at 50%, HSC50% Peak count at 10%, Pc10% Peak count at 20%, Pc20% Peak count at 30%, Pc30% Peak count at 40%, Pc40% Peak count at 50%, Pc50% Mean spacing of adjacent local peaks, S Mean spacing between profile peaks, Sm10% Mean spacing between profile peaks, Sm20% Mean spacing between profile peaks, Sm30% Mean spacing between profile peaks, Sm40% Mean spacing between profile peaks, Sm50% No. of intersections of the profile at 10%, n(0)10% No. of intersections of the profile at 20%, n(0)20% No. of intersections of the profile at 30%, n(0)30% No. of intersections of the profile at 40%, n(0)40% No. of intersections of the profile at 50%, n(0)50% Number of inflection points, g Average slope, ∆a Root mean slope, ∆q Average wavelength, λa Root mean wavelength, λq Bearing ratio at 10%, tp10% Bearing ratio at 20%, tp20% Bearing ratio at 30%, tp30% Bearing ratio at 40%, tp40% Bearing ratio at 50%, tp50% Profile slope at line 10%, γ10% Profile slope at line 20%, γ20% Profile slope at line 30%, γ30% Profile slope at line 40%, γ40% Profile slope at line 50%, γ50% Relative length of the profile, lo Peak count, rp Am plitude Density Curve at 10%, ADC10% Am plitude Density Curve at 20%, ADC20% Am plitude Density Curve at 30%, ADC30% Am plitude Density Curve at 40%, ADC40% Am plitude Density Curve at 50%, ADC50% Autocorrelation function, ACF Correlation Length, CL Power spectrum density function, PSDF Roughness height uniformity, Hu Roughness height skewness, Hs Roughness pitch uniformity, Pu Roughness pitch skewness, Ps Steepness factor of the profile, SF Waviness factor of the profile, WF T urning (Facing) Mean Min. Max. 3.87 1.67 8.67 9.04 5.67 17.67 15.29 6.67 27.33 19.71 8.00 30.67 20.11 7.67 37.67 33.44 16.66 57.49 27.66 16.66 49.16 21.55 10.83 38.33 16.61 6.67 30.83 12.72 3.33 23.33 103.1 75.9 121.0 433.9 8.3 1353.2 287.2 13.8 778.2 156.4 0.9 510.8 33.6 0.3 224.3 1.2 0.2 10.8 7.8 3.3 17.7 18.4 11.7 35.7 31.0 14.0 54.7 40.3 16.7 62.3 41.3 16.3 76.0 4706.7 3977.0 4968.3 9.33 1.78 16.20 13.78 1.86 21.00 5.67 1.43 15.60 4.60 1.40 18.23 1.94 1.00 3.46 6.84 4.31 12.14 16.14 9.38 24.55 29.69 18.81 40.99 45.92 27.13 62.46 189.0 72.3 283.5 199.7 66.9 283.9 193.6 152.8 212.4 193.2 184.6 210.1 193.2 183.5 210.6 1.04 1.02 1.07 81 68 107 0.28 0.16 0.73 0.64 0.27 1.09 1.10 0.36 1.67 1.50 0.65 2.43 1.61 0.87 2.38 56.17 14.01 125.91 211.6 88.0 389.3 183.0 37.6 700.3 6.86 3.51 11.06 -0.54 -2.92 0.63 451.3 260.4 597.5 0.79 0.48 1.71 13.62 7.44 22.91 0.20 0.11 0.36 Milling Mean Min. Max. 5.33 1.67 12.33 11.67 6.33 22.67 19.40 5.67 32.67 21.56 11.33 32.67 17.04 7.67 27.67 28.94 10.83 37.50 21.05 9.17 26.66 15.61 6.67 20.83 11.67 5.83 16.66 8.17 2.50 15.00 147.9 120.7 205.6 601.7 17.7 1600.2 356.4 26.2 1181.8 132.1 1.1 517.2 24.0 0.3 275.2 0.6 0.3 1.6 10.9 3.3 25.0 23.7 13.3 45.7 39.5 11.7 66.7 44.2 23.3 65.7 35.4 17.0 56.0 4667.3 3642.0 4932.3 8.43 3.47 14.87 13.08 6.51 20.07 4.50 1.28 12.27 3.24 1.25 9.40 3.41 1.19 8.26 12.03 4.22 27.74 28.19 13.84 46.94 48.08 25.03 74.44 66.02 42.92 87.33 185.6 81.3 328.7 207.6 131.7 248.4 195.5 185.0 226.6 189.5 183.7 202.4 193.3 184.8 215.7 1.02 1.02 1.04 58 40 67 0.55 0.18 1.60 1.16 0.55 2.31 1.80 0.36 3.53 1.97 0.95 3.15 1.56 0.86 2.45 31.22 16.14 62.74 355.0 237.7 536.7 82.2 26.7 183.8 4.77 3.57 7.45 0.49 -0.25 1.70 450.3 207.0 657.0 0.95 0.73 1.27 9.45 3.55 17.06 0.27 0.16 0.36 Grinding Mean Min. Max. 10.94 6.00 16.50 36.99 19.25 59.00 72.91 42.50 107.75 93.75 63.00 123.50 84.42 64.50 116.50 125.26 76.87 186.85 71.97 40.00 116.24 42.51 26.25 71.24 26.91 16.25 49.37 17.31 10.00 28.75 51.1 42.7 68.1 589.0 16.4 1855.8 326.2 32.0 1116.1 168.9 33.5 482.1 171.5 36.2 504.8 144.9 28.5 454.6 21.9 12.0 33.0 74.3 38.5 118.5 146.3 85.5 216.3 188.3 126.5 247.0 170.0 130.5 234.0 3506.4 3169.0 3832.6 8.40 4.93 11.67 14.62 10.50 18.48 0.77 0.33 1.37 0.48 0.26 0.75 1.77 0.88 2.84 8.31 4.75 11.65 22.93 14.78 32.15 44.13 31.06 56.47 66.25 50.71 77.70 204.1 148.3 249.2 189.1 181.4 208.4 182.8 180.7 187.3 181.2 180.3 183.1 181.5 180.5 182.9 1.01 1.01 1.02 162 123 192 0.33 0.13 0.52 0.98 0.48 1.68 1.81 1.23 2.56 2.30 1.54 2.79 1.99 1.61 2.48 0.99 0.55 1.57 334.0 185.5 490.0 65.6 31.8 128.7 0.92 0.67 1.17 0.03 -0.04 0.12 502.2 332.2 673.3 21.52 9.73 38.01 0.58 0.00 2.67 1.37 1.03 1.90 Un it s P eak P eak P eak P eak P eak P eak /cm P eak /cm P eak /cm P eak /cm P eak /cm mm mm mm mm mm mm Count Count Count Count Count Count degree degree µm µm % % % % % degree degree degree degree degree µm Count % % % % % ----µm ----------------------------- International Journal of M aterials Engineering 2013, 3(3): 47-58 53 Table 4. Correlat ion coefficient bet ween surface roughness paramet ers (Correlat ion coefficient s greater than 0.9 are highlight ed wit h gray) Parameters Rq Rsk Rku Rpm Rvm Rtm Rp Rv Rt Ry RzISO Ra R3y R3z Km Kp Med SD RzDIN 0.750 0.421 0.484 0.942 0.874 1.000 0.780 0.533 0.715 0.915 1.000 0.750 0.385 0.463 0.436 0.436 0.451 0.750 Rq 0.398 0.456 0.652 0.735 0.750 0.948 0.808 0.952 0.772 0.769 0.998 0.865 0.266 0.314 0.314 0.633 1.000 Rsk 0.670 0.595 0.081 0.421 0.556 0.086 0.269 0.265 0.424 0.435 0.478 0.737 0.935 0.935 0.770 0.398 Rku 0.510 0.345 0.484 0.434 0.092 0.292 0.303 0.489 0.496 0.335 0.387 0.570 0.570 0.455 0.456 Rpm 0.660 0.942 0.713 0.286 0.550 0.762 0.939 0.665 0.348 0.732 0.657 0.657 0.621 0.652 Rvm 0.874 0.712 0.778 0.803 0.943 0.878 0.714 0.358 0.023 0.025 0.025 0.110 0.735 Rtm 0.780 0.533 0.715 0.915 1.000 0.750 0.385 0.463 0.436 0.436 0.451 0.750 Rp 0.715 0.932 0.820 0.795 0.946 0.846 0.355 0.488 0.488 0.647 0.948 Rv 0.919 0.737 0.550 0.775 0.650 0.276 0.243 0.243 0.159 0.808 Rt 0.843 0.732 0.933 0.813 0.058 0.150 0.150 0.446 0.952 Ry 0.918 0.751 0.439 0.161 0.228 0.228 0.276 0.772 RzISO 0.768 0.409 0.458 0.436 0.436 0.462 0.769 Ra 0.868 0.303 0.354 0.354 0.664 0.998 R3y 0.227 0.361 0.361 0.697 0.865 R3z 0.861 0.861 0.779 0.266 Km 0.728 0.314 Kp 0.728 0.314 Med 0.633 Table 5. Grouping of a highly correlated surface roughness parameters Group T urning (Facing) Milling Grinding All Processes No. P aramet ers No. P aramet ers No. P aramet ers No. P aramet ers No. I Ra, Rq, Rp, Rv, Rt, NPeaks, ACF10%, ACF, WF 9 Ra, Rq, Rpm, Rtm, Rt, RzISO, RzDIN, ACF10%, ACF, Hu 10 Ra, Rq, Rpm, Rtm, Rv, Rt, Ry, RzISO, RzDIN, ACF10%, ACF, 12 Hu Ra, Rq, Rp, Rv, Rt, NPeaks, ACF10%, ACF, WF 9 II Rsk, Km, Kp, tp40%, tp50%, ADC40% 6 Rsk, tp40%, tp50% 3 ADC30%, tp30%, tp40%, tp50% 4 tp40%, tp50% 2 III RzDIN, Rpm, Rtm, Ry, RzISO, Hu, Rvm 7 RzDIN, Rpm, Rtm, Ry, RzISO, Hu, Rvm 7 RzDIN, Rpm, Rtm, Ry, RzISO, Hu, Rvm 7 RzDIN, Rpm, Rtm, Ry, RzISO, Hu, Rvm 7 IV Pc10%, Pc20%, Pc30%, Pc40%, Pc50% 5 Pc10%, Pc20%, Pc30%, Pc40%, Pc50% 5 Pc10%, Pc20%, Pc30%, Pc40%, Pc50% 5 Pc10%, Pc20%, Pc30%, Pc40%, Pc50% 5 HSC10%, NIntr10%, HSC20%, HSC10%, NIntr10%, HSC20%, HSC10%, NIntr10%, HSC20%, HSC10%, NIntr10%, HSC20%, V NIntr20%, HSC30%, NIntr30%, HSC40%, NIntr40%, HSC50%, 10 NIntr20%, HSC30%, NIntr30%, HSC40%, NIntr40%, HSC50%, 10 NIntr20%, HSC30%, NIntr30%, HSC40%, NIntr40%, HSC50%, 10 NIntr20%, HSC30%, NIntr30%, HSC40%, NIntr40%, HSC50%, 10 NIntr50% NIntr50% NIntr50% NIntr50% VI ∆a, ∆q, λa, λq 4 ∆a, ∆q, λa, λq 4 ∆a, ∆q, λa, λq 4 ∆a, ∆q, λa, λq 4 * Correlation coeffi cients are equ al or more than 0.90 (P=0.05) 3.2. Analysis of Machining Process Signature The values of amp litude, spacing and hybrid parameters of surfaces manufactured by different manufacturing processes are overlapped significantly. Even if one manufacturing process results in a similar surface pattern, it produces surfaces whose amplitude, spacing and hybrid parameters vary widely depending on the manufacturing conditions. Thus, it is d ifficult to identify surface signatureof different manufacturing processes by investigating the magnitudes of the parameters[21]. It is considered here that a good method of classifying machin ing processes is to useindexes rather than absolute physical quantities. This section provides evidence of the relationships between manufacturing processes and some derived indexes. Ho wever by examining these indexes, it is possible to identify different classes of manufacturing processes. First of all, a paired of surface roughness parameters were selected to study its relation with the manufacturing process. The selected paired of surface roughness parameters, which called index, is based on some trails except for the condition of their poor correlat ion coefficient with each other. Then, by 54 Ossama Abouelatta: Prediction of M anufacturing Processes and Surface Roughness Using ANN plotting this pair of surface roughness parameters, the manufacturing processes can be partially or comp letely separated. If an appro ximate zone of a manufacturing process can be got based on the proposed indexes, the efficiency of the identification can be assessed. It was found that Pc/Ra and S/Ra indexes satisfy the condition of determining an appro ximate zone. The manufacturing process zones which enclosethe values of theindexes obtained under different cutting parameters are evident from Figs . 4 and 5. It was found that turning (Facing), milling and ground surfaces are distinctively separated from other manufacturing processes.In general, turned surfaces have the largest Pc/Ra indexand S/Ra index, whereas grinding surfaces have the lowest Pc/Ra index and S/Ra index. M illing surfaces have the second largest Pc/Ra index and S/Ra index. Each manufacturing process can be deduced in an area of ellipse as shown in Figs. 4 and 5. An ellipse is a smooth closed curve which is symmetric about its horizontal and vertical axes. The distance between antipodal points on the ellipse, or pairs of points whose midpoint is at the center of the ellipse, is maximum along the major axis or transverse diameter, and a minimum along the perpendicular minor axis or conjugate diameter[22]. An ellipse in general position can be expressed parametrically as the path of a point(????????(????????) , ????????(????????)), where: ????????(????????) = ???????????????? + ???????? cos(????????) cos(????????) − ???????? sin(????????) sin(????????) (1) ????????(????????) = ???????????????? + ???????? ????????????????????????(????????) ????????????????????????(????????) + ???????? ????????????????????????(????????) ????????????????????????(????????) (2) as the parameter t varies fro m 0 to 2π. Here(???????????????? , ????????????????) is the center of the ellipse, and???????? is the angle between the ????????-axis and the major axis of the ellipse. Table 6 lists the ellipse parameters (Xc, Yc, a, band ???????? ) for d ifferent machin ing p ro cess es . The range of the Pc/Ra and S/Ra indexes for d ifferent mach ining processes are given in Figs. 6-8, respectively. Although the conventional parameters listed in Table 7 vary considerably for each kind of surface, it is not difficult to observefrom the figures that the indexes identify the different manufacturing processes clearly. In addition, so me general conclusions for the relationship between the indexes and manufacturing processes can be drawn here with the help of Figs. 6-8. The indexes may also vary in range, see Table 7. It can be observed that if surfaces have clearly different surface roughness or height distributions, the distribution zones of the indexes of the surfaces would be distinctively separated. Moreover, the variation of the indexes of one kind of surface indicates differences of surface features although the surfaces may have similar surface roughness or height distributions. 3.3. Analysis of Cutting Parameters The influences of machine parameters on machined parts are not always precisely known, and hence, it becomes difficult to recommend the optimu m machinability data for mach ine process. Cutting parameter identification in mach ining operations, which included cutting speed, feed rate and depth of cut plays a very important role in the efficient utilization of machine tools and directly influences the product quality. Thus, it significantly influences the overall manufacturing costs. In practice, the machinists select cutting parameters fro m their specified ranges in mach ining handbooks, main ly based on experience, in order to satisfy the required accuracy of the final product[2]. The analysis of cutting parameters will be e xpla ined in details in section 4.2. Relation between Ra and S 220 Turning 200 Grinding Milling 180 Mean spacing of adjacent local peaks (mm) 160 140 120 100 80 60 40 0 2 4 6 8 10 12 Average roughness height (µm) Figure 4. A diagram of average roughness height (Ra) versus mean spacing of adjacent local peaks (S) of typical manufacturing processes Relation between Ra and Pc 200 Turning 180 Grinding Milling 160 Peak count (Count) 140 120 100 80 60 40 20 0 1 2 3 4 5 6 7 8 9 10 Average roughness height (µm) Figure 5. A diagram of average roughness height (Ra) versus peak count (Pc) of typical manufacturing processes Pc/S (Count/mm) 4.5 Turning 4.0 Milling 3.5 Grinding 3.0 2.5 2.0 1.5 1.0 0.5 0.0 Turning Milling Grinding Machining processes Figure 6. Surface index (Pc/S) International Journal of M aterials Engineering 2013, 3(3): 47-58 55 Pc/Ra (Count/µm) 350 300 Turning Milling 250 Grinding 200 150 100 50 network and the desired output pattern[23]. The overall training process used in this paper is illustrated in Fig. 9. Firstly, the roughness parameters were used as an input to predict the output which is the machining process. Secondly, both of roughness parameters and the predicted machin ing process were used to predict cutting parameters wh ich are: feed, speed and depth of cut. Surface Roughness Parameters Neural Network Processing Cutting Parameters Applying Neural Network Prediction of Machining Process Applying Neural Network 0 Turning Milling Grinding Machining processes Figure 7. Surface index (Pc/Ra) Average roughness height, Ra Peak count, Pc Mean spacing of adjacent local peaks, S Feed Speed Depth of cut S/Ra (mm/µm) 100 90 Turning 80 Milling 70 Grinding 60 50 40 30 20 10 0 Turning Milling Grinding Machining processes Figure 8. Surface index (S/Ra) Table 6. Ellipse parameters (Xc, Yc, a, band ????????) P aramet er Xc Yc a b ???????? T urning (Facing) Ra-S Ra- Pc 0.8 0.8 55 158 0.28 0.35 15 37 78 -70 Milling Ra-S Ra- Pc 4.8 6.1 162 88 1.8 3.4 46 25 86 -77 Grinding Ra-S Ra- Pc 6.1 4.8 102 55 3.3 1.9 30 16 0 0 4. Neural Network Modeling Neural network is a highly fle xible mode ling tool with the ability to learn the mapping between inputand output parameters. An art ificial neural network(A NN) is capable of learning fro m an experimental data set to describe the nonlinear and interactioneffects more effectively. The network consists of aninput layer used to present data, output layer toproduce ANN’s response, and one or more h idden layers in between. The network is characterized bytheir topology, weight vectors, and activation function that are used in hidden and output layers ofthe network. Net works with b iases, a sig moid layer,and a linear output layer are capable of appro ximating any function with a finite number of discontinuities. The knowledge is presented by the interconnectionweight, which is adjusted during the learning stageusing the back propagation learning algorithm tomin imize the mean square between the actual output of the Figure 9. Application of neural network to predict manufacturing processes and cutting parameters 4.1. Prediction of Machining Process A feed forward neural network with error back propagation algorithm was adopted for the NN system. Here itis used to develop machin ing process predictionmodel through surface roughness measurements. Fro m 45 e xperiments were conducted, 36 e xperimental datasets were used totrain the network. Be fore applying the neural network for modeling the architecture of the network has to be decided; i.e. the nu mber o f h idden layers and the nu mber of neurons in each layer and transfer function for each layer. As there are three inputs (Ra, Pc and Sroughness parameters)to produce oneoutput (Machining process which are: Turning “Facing”, Milling or Grinding), the number of neurons in the input and outputlayer has to be set to 3 and 1respectively as shown in Fig. 10.Considering one hidden layer, the number of neuronsin the hidden layer is optimized. A procedure was emp loyed to optimize the hidden layer neurons and select the transfer function for which a program was generated in MATLAB software . Accordingly, an e xperimental approach was adopted, which involves verification of the neural network against another 9 sets of data fro m 45 experimental d atas et. The performance of the network was evaluated by mean squared error between the experimental and the predicted values for every output nodes in respect of training the network. When the training is co mplete, the network performance could be checked and determined to search for if any changes need to be made to the training process, the network architecture or the data sets. The first thing to do is to check the training record, Fig. 11. As indicated, the iteration at wh ich the validation performance reached a minimu m was 6. The training continued for 6 mo re iteration before the training stopped.The next step in validating the networkisto create a regression plot, which shows the relationship between the outputs of the network and the targets, Fig. 12. If the training were perfect, the network outputs and the targets would be exactly equal, but the 56 Ossama Abouelatta: Prediction of M anufacturing Processes and Surface Roughness Using ANN relationship is rarely perfect in practice. The four axes represent the training, validation, testing and all data. The dashed line in each axis represents the perfect result between outputs and targets. The solid line represents the best fit linear regression line between outputs and targets. The R value is an indication of the relationship between the outputs and targets. If R = 1, this indicates that there is an exact linear relationship between outputs and targets. If R is close to zero, then there is no linear relat ionship between outputs and targ ets . Table 8 shows the output of neural network process. The feedback fro m that processing is called the “average erro r” or “performance”. Once the average erro r is below the required goal or reaches the required goal, the neural network stops training and beco mes ready to be verified. This process was repeated one hundred times and the lowest MSE run was chosen as the best trained net. After the training process, the model is tested for validation. In this work, the network is validated in terms of agreement with an e xtra 9 e xpe rimental data listed in Table 7(b). Average roughness height Ra Peak count Pc Mean spacing of S adjacent local peaks T Manufacturing process (Turning, milling and grinding) Input layer Hidden layer Output layer Figure 10. Application of neural network to predict manufacturing processes Table 7(a). Manufacturing processes signature indexes (Training data) No. f mm/rev 1 0.16 2 0.16 3 0.16 4 0.16 5 0.16 6 0.16 7 0.16 8 0.08 9 0.23 10 0.31 11 0.46 12 0.59 T urning (Facing) operation s dc Pc/Ra rpm mm Count/µm 140 0.3 10.2 224 0.3 21.5 560 0.3 10.3 900 0.3 25.4 560 0.1 8.9 560 0.75 15.3 560 1.5 36.3 560 0.3 12.3 560 0.3 21.3 560 0.3 10.8 560 0.3 12.9 560 0.3 7.9 S/Ra mm/µm 15.9 16.9 17.1 32.4 14.5 24.2 25.8 18.7 20.2 17.2 14.9 11.1 f mm/m in 28 55 157 110 140 110 41 110 110 110 110 110 Milling operation s dc Pc/Ra S/Ra rpm mm Count/µm mm/µm 157 0.2 19.8 35.4 157 0.2 17.2 34.1 157 0.2 15.1 32.7 157 0.2 10.2 31.6 157 0.2 18.6 35.2 80 0.2 6.5 33.2 157 0.45 18.4 48.6 80 0.2 9.4 24.9 80 0.1 12.8 49.0 157 0.35 19.2 46.4 157 0.25 22.3 40.6 157 0.3 9.7 32.4 Vw m /min 25 25 25 25 25 25 25 25 35 30 20 15 Grinding operation Vc ae Pc/Ra S/Ra m/s µm Count/µm mm/µm 35 20 287.3 94.3 35 30 200.2 59.5 35 40 231.6 60.6 35 50 185.9 61.0 20 20 151.4 66.6 20 30 155.2 53.0 20 40 231.6 70.8 20 50 241.1 72.8 20 20 278.5 73.3 35 20 257.4 72.7 35 20 264.2 58.8 20 20 256.5 74.3 Table 7(b). Manufacturing processes signature indexes (Verification data) T urning (Facing) operation Milling operation No. f s dc Pc/Ra S/Ra f s mm/rev rpm mm Count/µm mm/µm mm/min rpm dc Pc/Ra S/Ra Vw mm Count/µm mm/µm m/min Grinding operation Vc ae Pc/Ra S/Ra m/s µm Count/µm mm/µm 13 0.26 710 1.25 13.0 14 0.32 180 1.25 15.0 15 0.52 450 1.75 13.9 23.0 280 439 0.4 12.2 29.1 10 20 20 173.0 70.3 15.3 20 40 0.3 14.5 40.5 40 35 60 225.4 59.4 15.8 110 157 0.15 16.5 31.5 5 35 10 153.2 59.9 Table 8. Neural network output P aramet er Best trail number Coefficient of determ ination (R) Maximum error Average error Mean absolute error (MAE) Mean squared error (MSE) Mean squared normalized error (MSNE) Mean squared normalized error with regularization (MANEREG) Mean squared error with regularization (MAEREG) Mean squared error with regularization and economization (MSEREGEC) Sum squared error (SSE) Number of Iteration (NI) T urning 35 0.9938 10.19294 2.43334 0.026787 0.00187 1.91122 1.924798 0.206385 0.104221 0.059828 1954 Milling 93 0.9092 92.3164 12.2651 0.07882 0.02957 0.88089 0.96344 0.19726 0.11489 0.79843 20464 Grinding 91 0.93107 68.3333 5.2975 0.043183 0.01111 4.1716 4.0315 0.28706 0.14964 0.26664 22958

... pages left unread,continue reading

Document pages: 12 pages

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×