Prediction models for energy consumption and surface quality in stainless steel milling

Stainless steel is a kind of difficult-to-machine material, and the work hardening in milling easily leads to high energy consumption and poor surface quality. Thus, the influence of machined surface hardness on energy consumption and surface quality cannot be ignored. To solve this problem, the prediction models for machine tool specific energy consumption and surface roughness are developed with tool wear and machined surface hardness considered firstly. Then, the validity of the models is verified through AISI 304 stainless steel milling experiments. The results show that the prediction accuracy of the machine tool specific energy consumption model can reach 98.7%, and the roughness model can reach 96.8%. Later, according to the developed prediction models, the influence of milling parameters, surface hardness, and tool wear on the machine specific energy consumption and surface roughness is studied. Results show that in stainless steel milling, the most significant parameter for surface roughness is the machined surface hardness, while that for energy consumption is the feed per tooth. The machine specific energy consumption increases linearly with the increase of the tool wear and the machined surface hardness gradually. The proposed models are helpful to optimize the process parameters for energy-saving and high-quality machining of stainless steel.


Introduction
The manufacturing industry plays an important role in economic globalization and sustainable development, which has become the main cause of global warming due to the excessive consumption of energy and a large amount of greenhouse gas emissions in processing [1]. In addition, CNC machining technology has been widely used in the manufacturing industry with low processing efficiency, energy consumption, and other problems exposed, making green and efficient manufacturing become the goal pursued by modern enterprises. Therefore, the research on energy consumption of machine tools and surface quality of parts is more important. For the manufacturing industry, in the case of existing machining equipment, reducing energy consumption while improving the surface quality of machining to achieve green energyefficient manufacturing is a critical issue [2].
Specific energy consumption (SEC) is a crucial evaluation standard for machine efficiency and energy consumption. At present, both foreign and domestic scholars have made indepth studies on the energy consumption of machine tools. The study on energy consumption prediction models is composed of two main aspects: direct models and indirect models. For direct models, cutting parameters, material removal rate (MRR), and tool wear are mainly used as inputs, and cutting energy or specific energy as outputs to develop a mathematical model to quantitate the cutting energy consumption of machines. For example, Kara and Li [3] developed a relationship model between SEC and MRR, which accurately assessed energy consumption in actual production. Li, Yan, and Xing [4] presented an SEC model in relation to MRR and spindle speed, which was tested by medium-carbon steel milling experiments. Zhang et al. [5] developed an exponential model of SEC and cutting parameters according to the cutting force empirical formula and analyzed the influence on parameters in detail. Li et al. [6] optimized the relationship model between specific energy and process parameters to derive the optimal combination of cutting parameters for energyefficient milling. Liu et al. [7] developed a mathematical model at the machine tool, spindle, and process aspects to study the influence of cutting parameters on SEC at different stages in milling.
Though these models can accurately predict energy consumption, they fail to consider the influence of factors other than cutting parameters on energy consumption, such as tool parameters or tool wear. Based on this consideration, Zhao et al. [8] analyzed the effect of tool wear on machine tool energy consumption in 45# steel semi-finishing milling, and developed a machine power model with MRR and tool wear as the inputs. The study showed that cutting energy consumption can be reduced by monitoring and controlling tool wear. In a study by Sujan et al. [9], Taguchi was used to optimize cutting parameters to reduce the tool wear. It also proved that tool wear can influence energy consumption indirectly. Mativenga and Rajemi [10] developed a carbon emission prediction model, and the model took the tool parameters and cutting parameters as the inputs. Then, the cutting parameters were optimized to achieve a balance between minimum energy consumption and minimum processing time, which concluded that the increase of spindle speed and feed rate can improve machining efficiency and reduce energy consumption and rate of tool wear.
For indirect models, the available studies generally use gray correlation theory, neural networks, response surface methods (RSM), and so on to predict machine energy consumption by training and finding the optimal parameters. For instance, Al-Hazza et al. [11] studied the energy consumption variation regularity under different cutting speeds, feed rates, and amount of cutting depth through AISI 4340 steel cutting experiments; BP neural network was used to train the experimental data, and RSM was used to establish a prediction model of machine tool energy consumption and cutting power. Quintana, Ciurana, and Ribatallada [12] analyzed the electrical energy consumption of AISI H13 steel in high-speed milling, used artificial neural network to build a prediction model for electrical energy consumption, and investigated the influence of process parameters on electrical energy consumption. Carmita [13] used the RSM algorithm to develop and optimize prediction models for energy consumption and surface quality with cutting parameters and MRR as inputs, which was verified through the AISI 6061 T6 aluminum alloy turning experiment, and then they compared the results with empirical parameters experiment results, which showed that energy consumption was significantly reduced after optimization. Hanafi, Khamlichi, and Cabrera [14] studied the influence of cutting parameters on energy consumption and surface quality through PEEK-30 dry-turning experiments, optimized the multi-objective model by gray correlation theory, and concluded that cutting depth and cutting speed have significant effects on energy consumption and surface roughness.
Surface quality is a comprehensive standard for evaluating the working performance of parts. As one of the important indicators of surface quality, surface roughness has a significant impact on service life and reliability of mechanical products [15]. In recent years, the factors affecting surface roughness in turning and milling have been widely studied. Tool wear and cutting forces are gradually taken into account in surface roughness prediction models. Asit and Kalipada [16] used RSM to develop a surface roughness model for machined surfaces, with cutting parameters and ambient temperature as the inputs, and ultimately found a combination of cutting parameters which can reduce tool wear to a certain extent and obtain the best surface quality. Liu et al. [17] predicted surface roughness in slot milling through three models, exponential model, linear model, and power function model, and the results showed that the power function model could more accurately predict the relationship between surface roughness and cutting parameters. Wang et al. [18] developed a milling surface roughness prediction model considering mechanical factors, tool parameters, cutting parameters, and microhardness, and obtained the optimum milling parameters by analyzing the influence of parameters on surface roughness through Ti6AL4V milling experiment.
Based on roughness prediction models, some scholars have made multi-objective optimization of surface quality and energy consumption in relation to energy consumption. For instance, Kumar [19] selected C360 copper alloy material to perform micro-turning experiments, used genetic algorithm to optimize multi-objective model based on the best surface roughness and the maximum material removal rate, and ultimately obtained the best cutting parameters as spindle speed n =1686r/min, feed rate v f =10.62μm/r, and cutting depth a p =99.45μm. Kant and Sangwan [20] proposed a multiobjective prediction model based on minimum energy consumption and optimal surface roughness. The gray correlation and RSM were used to analyze and optimize the model, and the results showed that the feed rate is the most important parameter affecting the multi-objective model. Li et al. [21] developed a third-order polynomial prediction model of cutting force and surface roughness and optimized the multiobjective model of cutting force, surface roughness, and MRR based on RSM and ITLBO algorithm. The optimized cutting parameters can get better processing quality tested by 7050 aluminum alloy milling experiment. Although existing studies have considered the influence of the tool, workpiece, and cutting force on the surface roughness, they fail to consider the changes of the surface hardness during the machining process. Actually, in hard-to-cut materials processing, due to the high hardness of the material, a large amount of cutting heat is generated in cutting, which increases the temperature of the shear surface obviously and causes severe hardening of the material surface. The increase of the tool-chip contact area makes the friction between the machined surface and the cutter surface increase, which affected the machined surface quality and cutting energy consumption [22].
Presently, the problem of energy consumption and surface quality prediction in hard-to-cut material processing has not been solved yet. To address this problem, the article conducted the following three studies: (1) to develop an MSEC prediction model with surface hardness and tool wear considered based on machine energy consumption characteristics; (2) to develop a surface roughness prediction model based on cutting parameters and hardness; (3) to analyze the influence of cutting parameters, tool wear, and surface hardness on MSEC and surface roughness prediction model.

Energy consumption analysis in CNC milling
The CNC milling process has multiple energy consumption parts, complex regularities, and enormous energy consumption. Figure 1 shows the energy conversion process of machining a blank into a product with specific appearance characteristics.
Existing energy consumption models for CNC machine tools usually reflects relationship between power and machining time. In terms of the composition of energy consumption, the energy consumption of CNC systems can be divided into two categories. One is only related to the characteristics of the specific machine tool itself, mainly including standby energy consumption E standby , main drive system, and feed system no-load energy consumption E no-load . The other is load-related energy consumption, which is related to the process parameters, workpiece, and tool parameters of the machining process, including cutting energy E cutting , additional load energy E add-load , and processing-related energy consumption of auxiliary systems E auxiliary . Since the measurements of E add-load and E auxiliary are very small, the article does not take them into account. To establish an energy consumption model to calculate the consumption of the CNC milling quantitatively, it is necessary to analyze the period characteristics of the processing in conjunction with the consumption unit of the CNC system to determine the consumption state of each period, as shown in Fig. 2.
The article analyzes the energy consumption of VMC850 CNC machining center in combination with characteristics of the period of CNC milling as follows:

Standby stage energy consumption
Machine Standby stage refers to the period after the machine is started and before the spindle runs. Energy consumption components mainly include the machine control system, light system, and cooling system. The power of the machine at this period is the standby power P standby , the value of which is considered a constant [4]. According to the preliminary test, the different systems power of machine tools during the standby period is calculated by controlling the opening and closing of the cooling system and lighting system: the actual power of the machine measured with the cooling system and lighting system on is 710.15W, and the actual power of the machine measured with cooling system off and lighting system on is 402.89W, and the cooling system power P cooling is 307.26W by calculating the difference between the two power. Similarly, the control system power P control and the light system power P light are calculated, as shown in Table 1. Then, the machine tool energy consumption in standby stage E standby is: where t standby is the machine standby stage time in seconds.

No-load energy consumption
The period when the spindle is working but the tool is not touching the workpiece is the air-cutting stage, and the power consumed by machine tool at this stage is main drive system and feed system no-load power P no-load . According to the previous test, the spindle speed (n) is set in the range of 200 2500r/min, and the power measurement result is collected once for every 100r/min increase. The results are shown in Table 2.
Based on the measurement results, the relationship between no-load power and spindle speed is approximately within the segmented function, as shown in Fig. 3. Then, the machine tool no-load energy consumption E no-load is: where t air-cutting is the time of machine air-cutting stage in seconds and n is the spindle speed in r/min.

Cutting consumption
The cutting stage is the period from the time when the tool touches the workpiece until it leaves after cutting, and the power consumed by the machine during this period is cutting power P cutting , including P spindle , P feeding , P standby , and P no-load . Since P feeding is small and can be neglected in calculation, and the power of the cutting stage P cutting can be directly measured. Then cutting consumption E cutting is: where t cutting is the cutting stage time in seconds.

A machine tool specific energy consumption (MSEC) model based on hardness and tool wear
Use the specific energy consumption SEC (J/mm 3 ) to evaluate the energy efficiency of the machine tool, which is calculated as the ratio of the total energy consumption E total (J) to the material removal volume MRV (mm 3 ) in milling by Eq. (4). Select the milling depth a p (mm), milling width a e (mm), feed per tooth f z (mm/z), milling speed v c (m/min), tool flank wear VB (mm), and machined surface hardness H (HL) as the inputs to develop the machine specific energy consumption model (MSEC). Based on the previous analysis, the MSEC model is developed in two parts: the fixed energy consumption (FEC) and variable energy consumption (VEC). Similar to the exponential relationship between the turning parameters and the specific energy consumption, an exponential model is used to establish the TSEC as shown in Eq. (6), Eq. (7), and Eq. (8).   where v f is federate in mm/min, and the relationship between v f and f z is shown in Eq. (5).
where Z is the milling tool teeth number, Z=2, and d is the milling tool shank diameter, d=25mm.
where A, K, b, c, d, e, m, and n are coefficients to be determined.

Orthogonal experimental design
The orthogonal experiment was designed by Taguchi method, and the four parameters of milling (a p , a e , v c , and f z ) were chosen as controllable factors. The range of parameters is a p = (0.2~0.14) mm, a e = (2~15) mm, v c = (66~150) m/min, and f z = (0.1~0.25) mm/z, which is determined by the cutting capacity principle of the carbide milling tool in stainless steel processing. According to the parameter range, the orthogonal experiments with 25 groups of 4 factors and 5 levels were designed as shown in Table 3.

Experimental equipment and data collections
The hard-to-cut material AISI 304 stainless steel was used to perform plane wet milling experiments, and the workpiece dimensions are 50mm (length) × 50mm (width) × 30mm (height). Considering the difficulty of hard-to-cut processing, down milling was selected in processing in experiment as shown in Fig. 4 which can slow down the tool wear and improve the surface quality to some extent. Before the experiment, pre-milling of the upper surface of the workpiece of 1 mm depth was carried out to remove the rusted and hardened The processing equipment is VMC850E CNC machining center (SHENYANG Machine Tool, SMTCL). The cutting tool is a 25-mm-diameter right angle and rotatable-position mill cutter (TAP400R-2525-160) with two KYOCERA carbide inserts (APMT1604PDER-KZ) and specific parameters as shown in Table 4.
The measurement instruments required for the experiment are shown in Fig. 5. The power and energy signals during the milling process of CNC machine center are collected by a power analyzer (Yokogawa WT500), and the data are processed by WTVIEWEREFREE software, which can calculate the MSEC indirectly through Eq. (9).
where WP is the sum of positive and negative active power for each data update cycle collected by the power analyzer in W·h, N is the sample count of the integration time, u(n) and i(n) are the nth voltage and current measurements, and T is the total time of the sampling process in hours. A roughness tester (RTP120) was used to measure the machined surface roughness (Ra) of the workpiece, and 5 points evenly distributed on the workpiece surface were selected for measurement to average, which can obtain statistically significant Ra values. A microscope (ZEISS Axio Lab.A1 Mat) was used to measure VB values, and the average of the two measurements before and after the experiment was taken as the values of VB for this group experiments, and replace the inserts in the event of tool wear up to 0.2mm. Since the frequent removing and resetting of workpieces in continuous milling would cause discontinuous data on MSEC, VB, and Ra, and could not reflect the effect of changing cutting parameters on energy consumption and surface quality, thus, a Leeb hardness tester was chosen to measure the hardness values (H) of each group of experiments before processing without damaging the surface quality of the workpiece. The orthogonal experimental tables L25(5 4 ) and data collection results are shown in Table 5. 4 Machine tool specific energy consumption (MSEC) and surface roughness (Ra) prediction models

MSEC model fitting and parametric analysis
The MSEC prediction model can be obtained with least square method through nonlinear curve fitting with the 25 sets of data in Table 5 as Eq. (10). The determination coefficient R 2 is widely used to evaluate the regression effect in the model, and the results show that the R 2 of MSEC model was 99.3%, and R 2 (adjusted) reached 98.7%. The model is compared with the specific energy consumption model developed by Zhao [23] which only considers the tool wear and cutting parameters as shown in Eq. (13). The results show that the R 2 of the contrast   Fig. 6 and Table 6.
where k 0 , k 1 , and k 2 are coefficients to be determined. With the identification of setting machining parameters, the pattern of the effects of cutting parameters, tool wear, and hardness of the machined surface on MSEC is shown in Fig. 7. In hard-to-cut material milling, a p is the most significant impact followed by a e and f z on the machine tool specific energy, making the MRR increase and reducing MSEC. The effect of v c on MSEC is not significant in the range of cutting dosage of hard-to-cut materials (90-150m/min), and MSEC slowly reduces with the increase of v c . MSEC increases linearly within a certain range with the increase of tool wear and machined surface hardness. When cutting hard-to-cut materials, some workpiece materials and tool materials react at high temperature and separate from the tool to accelerate tool wear. With the increase of tool wear, the area of the tool and workpiece contact section increases so that the tip of the tool and the workpiece friction generates more heat in cutting. In addition, due to the low thermal conductivity of most hard-to-cut materials, cutting heat is difficult to diffuse, and the cutting edge is significantly affected by heat, which causes high temperature of cutting edge and leads to increase in surface hardening of the parts, resulting in MSEC increasing.

Development and parameter analysis of surface roughness model
Surface roughness is an important indicator to measure surface integrity. There are numerous factors affecting the surface roughness. On the one hand, it is influenced by cutting parameters, tool wear, residual stresses, and surface work hardening during the milling process. On the other hand, the geometry and mechanical properties of the workpiece are also affected the surface roughness, as shown in Fig. 8.
Since the relationship between surface roughness and process parameters is not simply linear, thus, the RSM was used to predict the surface roughness and establish a quadratic response surface model for surface roughness, as shown in Equation (12).
where y is the response, which indicates the surface roughness in Ra model, x i is the independent variable, β i is the coefficient of the regression equation, and ɛ is the error between the fitted and experimental values. Before response surface analysis, the model variables are linearly transformed to resolve the effects of different ranges and dimensions of the independent variables, and the results are shown in Table 7.
The milling depth a p , milling width a e , cutting speed v c , feed per tooth f z , and the machined surface hardness H were selected as the inputs, and the surface roughness was used as the response. To verify the effect of machined surface hardness on the model, the regression model Ra 2 considering hardness was compared with the model Ra 1 which only considered four cutting parameters. The fitting results are shown in Eq. (13), Eq. (14), and Fig. 9.
Significance level was chosen as α=0.05, which meant that the confidence level of the results was above 95%. F value is the ratio of regression error to mean square error, which was used to measure the significant effect of model terms on response, and the ANOVA results are shown in Table 8 and Table 9. The determination coefficient R 2 was used to assess the fitting of the model, and the results show that the R 2 of the Ra 1 model is 68.5% and the R 2 (adjusted) is 58.7%, and the R 2 of the Ra 2 model is 96.7% and the R 2 (adjusted) is 95.8%, which indicated that Ra 2 model considering hardness can predict surface roughness more accurately.
By determining a set of processing parameters and hardness values, a p = 0.6 mm, a e = 6 mm, v c = 100 m/min, f z = 0.    Fig. 10. According to ANOVA of Ra 2 model, a p , a e , and H are parameters that have more significant effects on surface roughness, followed by f z and v c . In actual machining, the increase of a p and a e will make the cutting resistance increase, which will cause severe deformation on workpiece surface and then lead to the increase of surface roughness. Therefore, it is possible to obtain a lower surface roughness by choosing smaller ap and ae. In addition, the surface hardness also has a significant influence on surface roughness especially in hard-to-cut materials cutting, as shown in Fig. 10. The hard-to-cut material has high surface hardness; it is easy to appear serrated chips and shear surface slip during cutting so that the cutting force acts intermittently on the cutting edge, causing vibration of the tool, which leads to the increase of surface roughness. However, during the increase of surface hardness, the rise of cutting temperature will make the surface of the material soften, making cutting easier, which will reduce the surface roughness to some extent (Fig. 11).

Validation of MSEC model and Ra model
Three new combinations of cutting parameters were selected to verify the prediction accuracy of MSEC model and Ra model, and the results are shown in Table 10 and Table 11. The accuracy of machine tool specific energy prediction results for all three groups of experiments is above 95%, which indicates that the MSEC model considering hardness and tool wear has high accuracy. For surface roughness prediction, the accuracy results of Ra 1 model are only about 50%, while that of Ra 2 model is above 90%, which shows that the Ra model considering the material hardness can predict the surface roughness more accurately in hard-to-cut materials processing (Fig. 12).

Multi-objective optimization for MSEC and Ra 2 model
The variables milling depth a p , milling width a e , cutting speed v c , and feed per tooth f z are represented by the independent variables x 1 , x 2 , x 3 , and x 4 , respectively; the machine tool specific energy prediction model MSEC is represented by the function f(x); the surface roughness prediction model Ra 2 is represented by the function g(x); VB=0 and H=405 are taken. f(x) and g(x) are optimized by genetic algorithm to obtain the Pareto front surface as shown in Fig. 13. The Pareto solution set is divided into three sections to analyze: the change of function g(x) in section ab is obvious while the change of f(x) is not obvious, which means that the change of cutting parameters in this range has a more significant impact on surface quality. The function f(x) in section cd increases significantly and the change of function g(x) is not significant, which means that the impact of cutting parameters on energy consumption is more obvious in this stage. In section bc, the small change of parameters is obvious for both energy consumption and surface quality with lower levels. Therefore, the value of the variables in this section is the cutting parameter closet to the lowest energy consumption and the best surface roughness, as shown in Table 12.

Conclusion and future work
Firstly, prediction models for MSEC and surface roughness in stainless steel milling were developed respectively. Then, the models were verified the reliability through AISI 304 stainless steel milling experiments. At last, the influence of cutting parameters, surface hardness, and tool wear on MSEC model and Ra model was studied. The main conclusions are as follows: (1) The prediction accuracy of MSEC model considering surface hardness and tool wear can reach 98.7%. The model can accurately predict the cutting energy consumption in hard-to-cut materials milling. The research has revealed the linear relationship between tool wear and energy consumption, which shows that it is helpful to reduce the energy consumption through reasonable control of tool wear in actual machining. (2) The prediction accuracy of Ra 2 model considering hardness can reach 96.8%, which is much greater than that of Ra 1 model considering only cutting parameters, indicating that surface hardness is a non-negligible factor affecting the machined surface roughness. (3) The hardness of the machined surface has a significant effect on energy consumption and surface quality; thus, the machine tool specific energy and surface roughness prediction models developed have the highest accuracy, which is of great importance in selecting energy-saving and high-quality processing schemes.
(4) The machine specific energy consumption increases linearly with the increase of machined surface hardness, which also leads to an increase in surface roughness. Therefore, in hard-to-cut materials processing, the cutting parameters can be reasonably selected to slow down the hardening, reduce tool wear, decrease cutting energy consumption, and improve surface quality.
The future work is as follows: (1) In order to verify the influence of hardness on surface roughness, tool wear is not considered in the Ra model. To develop a more comprehensive surface roughness    Fig. 12 Comparison of MSEC and Ra model prediction accuracy prediction model, combining hardness and tool wear is a problem that needs further study. (2) The multi-objective optimization of energy consumption and surface quality in hard-to-cut materials processing will be the focus of future study.