USING SUPPORT VECTOR REGRESSION AND NON-DOMINATED SORTING GENETIC ALGORITHM IN MULTI-OBJECTIVE OPTIMIZATION OF MILLING OF S50C STEEL UNDER MQL CONDITION

Canh Nguyen V., Nghien Nguyen B., Dung Hoang T., Que Nguyen V., Truong Nguyen X., Truong Nguyen T. (2022) USING SUPPORT VECTOR REGRESSION AND NON-DOMINATED SORTING GENETIC ALGORITHM IN MULTI-OBJECTIVE OPTIMIZATION OF MILLING OF S50C STEEL UNDER MQL CONDITION, Journal of Applied Engineering Science, 20(1), 123 130, DOI:10.5937/ jaes0-31366 Ba Nghien Nguyen Faculty of Information Technology, Hanoi University of Industry, 298 Cau Dien Street, Hanoi, Vietnam Tien Dung Hoang Faculty of Mechanical Engineering, Hanoi University of Industry, 298 Cau Dien Street, Hanoi, Vietnam


INTRODUCTION
Metal cutting fluid and energy consumption are important factors of sustainable development because of the scarcity of resources and environmental disruption. The related energy usage results in about 90 percent of the environmental impacts of machining machinery, based on preliminary studies of machining processes such as milling and turning... Hence, It is necessary to reduce energy consumption and to minimize the usage of cutting fluid in the machining process to reach sustainable production and environmental protection. a. Minimizing the usage of cutting fluid The method of applying a very limited volume of high-quality lubricant directly to the cutting area instead of using conventional flood coolants is the minimum quantity lubricant (MQL). Hence, MQL or dry-machining is now considered an excellent alternative to reduce the minimizing usage of cutting fluids. Many researchers claimed [1]- [5] that Oil droplets by the MQL system reduce fric-tion between tool and chips were removed from the cutting zone by the high compressed air. P. V. Krishna [6] carried out an experimental investigating on nano cutting fluids effect in milling process under MQL environment. The experiment shows that MQL technique with vegetable oil-based executed well to reduce temperatures at cutting zone and cutting forces, and hence, decline tool wear and improve the surface quality of a product. Singh [7] shows the suitability of MQL and NMQL with vegetable oil in face milling of Inconel 625. The author claimed the positive effects of MQL on reducing tool wear and average surface roughness value. Weinert [8] carried out the research study in effects of dry machining confirm that MQL condition is suitable with most of the different machining types, and could be applied for the great number of materials like steel, cast iron, aluminium. b. Reducing cutting energy According to the high competition in this industry, the manufacturers always try to reduce the cost of energy, Van Canh Nguyen, et al. -Using support vector regression and non-dominated sorting genetic algorithm in multi-objective optimization of milling of S50C steel under MQL condition machining tools and value of surface roughness simultaneously. Hence, besides the research in optimizing product quality and maximum production rate, reducing energy consumption is a considerable approach. Zhang, Xuewei [9] carried out a research study on optimising energy consumption considering cutting parameters in the milling process. The research result confirmed that energy consumption could be declined by control of the cutting parameter. However, it is complicated to optimize three or more responses at the same time. Many methods were proposed in the recent decade to solve this problem. In this field, Taguchi-Based such as TOPSIS, MOORA, COPRAS… were often selected because of the simplicity in use. Recently, because of the development of computing, artificial intelligence algorithms were becoming more and more popular. Murthy et al. [10] used an Artificial neural network (ANN) approach to guess the cutting criteria like wear-vibration-cutting forces of tool and surface roughness in a boring of titanium alloy. The ANN predicted results are almost the same as experimental results. Limited studies have used support vector regression to generate the regression model but for one response only [11]- [15]. In this paperwork, the hybrid method, including SVR combined with NSGA2, was applied to solve multiple objective optimizations of milling process considering three different responses.

RESEARCH METHODOLOGY
a. Optimization Issues In this study work, combining of NSGA2 and SVR were performed to build the regression model then optimize three machining responses such as the arithmetical mean roughness (Ra-µm), specific cutting energy (Ec-J/ cm3), and material removal rate (MRR-cm3/min) simultaneously. The average roughness value is calculated as: The specific cutting energy is obtained by equation (2) [16] [17] = 1 + 2 + 3 + 4 + 5 5 = = The production rate is calculated as eq. (3)   to classify information into two partitioned locales so that the removal between the closest point and the hyperplane is at most extreme. Figure 1 presents a hyperplane and a margin. Suppose that the equation for a hyperplane is w.x+b=0 (4). To maximize the margin, the goal of the SVM algorithm is to find w and b in this equation.
(1) The SVM algorithm applies to solving classification issues and finds solutions to regression subjects named Support Vector Regression (SVR). The SVR calculation depends on a Los function was proposed by Vapnik [14], [15], Which is error-tolerant for points within a small epsilon that is far from the true value. This implies that this function gives zero error for all points in the training set in the epsilon range. Fig. 1 and Fig. 2 present linear/nonlinear regression model under the epsilon range [13] For SVR, the input x is mapped into m dimension feature space by a nonlinear mapping function first, and then the linear model is built, which is based on this dimension feature space by Eq (4):  The accuracy of the estimate is evaluated by loss function L(y,f(x,w)). Support Vector Regression uses a loss function which proposed by Vapnik [14], [15]:

Figure 4: SVM model
Thus, SVM is performed linear regression in multiple dimension feature space using function L and minimizing |(|W|)|^2for diminishing complexity of the model. This problem can be solved by introducing slug variables ζ i and ζ i * with i = 1, 2, … n to measure the deviation of the training samples outside the epsilon range. Subsequently, SVR is minimized by the equation underneath: with constraints: The final function f(x) could be found by applying the duality hypothesis for minimizing issues: c. Multi-objective optimization Typically, multiple objective optimizations have a few criteria regarding disparity and balance imperatives to optimize [12]. The objective is to discover a set of arrangements that don't have any limitation infringement and are as great as conceivable with respect to all its objectives values. Consequently, the ultimate outcome of such a multiple criteria optimization issue may be a set of trade-off arrangements among diverse destinations. These trade-off focuses are named as Pareto ideal arrangements, which are not overwhelmed by any other arrangement and cannot be progressed within the case of at slightest one other objective without declining. The set of all these doable nondominated arrangements is named as Pareto ideal arrangement set, and the comparing objective values are named the Pareto front [11]. Figure 5 outlines the Pareto front for the optimization of two objectives, namely Obj-1 and Obj-2.

EXPERIMENTAL PROCEDURE
a. Workpiece In this present study, S50C grade carbon steel is applied as the workpiece material to research. The chemical composition of S50C steel has appeared in Table 2. This carbon steel is broadly utilized in apparatus fabricating since of the great mechanical properties of this steel. S50C carbon steel is, for the most part, utilized to form mechanical parts, like spring, gear, pressure bar, the roller. The machining workpiece was prepared with the dimension of 250x350x15mm.  to reduce the vibration in the milling process. c. Measurement Equipment The surface roughness of each experimental run was measured by Mitutoyo Surftest JS-210, the cutting force was obtained by Kisler dynamometer, specific cutting energy was determined by equation (2). The results were fill in Table 3.

RESULTS AND DISCUSSION
In this work, the Minitab was chosen to analyze the effect of cutting and lubrication variants on the average value of surface roughness (Ra,µm), specific cutting energy (Ec) and production rate (MRR, cm3/min). First, the regression model was performed by the Support Vector Regression method, and then the NSGA2 was applied to find the optimum set of cutting and lubricant parameters. а. Effect of cutting parameters and lubricant condition on surface roughness. Fig. 8, 9 show that cut's depth exhibited a considerable effect on surface roughness is the most, followed by feed rate then cutting speed. By contrast, the influence of the lubricant flow rate and air pressure is insignificant. In comparing cutting and lubricant parameters, cutting parameters greatly influence material removal rate MRR, while the flow rate of lubricant Q and air pressure P does not affect MRR. This is in accordance with formula (3). Graphs 10 and 11 clearly illustrate the effect of cutting and lubricant parameters on material removal rate MRR. Where the depth of cut (a p ) affects significant on the MRR, followed by cutting speed (V c ) and feed rate (fz). The MRR value is proportional to cutting speed V c ), depth of cut (a p ) and feed rate (f z ). c. Influence of cutting parameters and lubricant condition on specific cutting energy Ec. According to Fig. 12 and 13, the feed rate has the most significant impact on specific cutting energy (E c ), followed by the depth of cut (ap) and cutting speed (V c ). In this case, the effects of lubricant (Q) flow rate and air pressure (P) are significant. These research results are inconsistent with previous publications. The MQL condition particles the lubrication film between the cutting tool and workpiece, leading to reduced friction, thereby declining the cutting force, leading to a decrease in specific cutting energy E c . d. Support vector regression Support Vector Regression (SVR) is applied to build regression models for prediction surface roughness (Ra), specific cutting energy (Ec), Material Removal Rate (MRR) with the C = 10e5 and the ε = 0.01. Table 3 presents the performance of SVR models. Figures 14, 15, and 16 show SVR predicted output vs  Table 3 show that the errors of predicted output with expected output for Ra, Ec, MRR are nearly 100%, at 99,96%, 99,99% and 100%, respectively.
= 3 ( , , , , ) Next, the Non-dominated Sorting Genetic Algorithm (NSGA2) was applied to perform multi-object optimization. In this case, The approach of this work is to minimize f1, f2 and maximize f3. The Sklearn and Pymoo framework in Python was used to implement SVR and multi-object optimization. Pymoo method considers minimization problems only. Therefore, f3 should be multiplied by -1 and be minimized. NSGA2 was performed with a population size is 50, and the number of generations is 1000, and the optimal result is shown in table 4.

CONCLUSION
Though a lot of previous techniques for multiple response optimization in cutting processes are reported, a new hybrid model was proposed in this paperwork. A multiple response optimization methodologies, by using genetic algorithm based Non-dominated Sorting Genetic Algorithm (NSGA2) combined with Support Vector Regression (SVR) was also proposed to build the regression model and optimize the cutting and lubrication parameters under MQL condition. The following conclusions are made: The optimum lubrication and cutting parameters for minimum specific cutting energy and surface roughness, and maximum material removal rate are depth of cut of 0.899 mm, cutting speed of 298.735 m/min, a feed rate of 0.094 mm/tooth, and lubricant flow rate, air pressure of 149.122 ml/h, 2.0077 MPa, respectively. Among the lubricant parameters: flow rate of lubrication (Q), air pressure (P) and milling parameters: cutting speed (Vc), feed rate (fz) and depth of cut (ap), cutting parameters effects significant on all of the response. By contrast, the lubrication parameters have a significant effect on specific cutting energy (Ec) only, while the influence on surface roughness is insignificant and don't have an effect on production rate. SVR Model was applied successfully to build the regression model of responses in the cutting process due to the high accuracy. The errors between predicted and expected results for surface roughness Ra, specific cutting energy Ec and material removal rate MRR are 99.96%, 99.99%, 100%, respectively.
The hybrid method combining the SVR model and the Non-dominated Sorting Genetic Algorithm (NSGA2) could be used to find the optimum parameters in the milling process.