Materials and measurements

All the experimental tests are carried out at the workshops at the Higher Institute of Technological Studies of Gabes (ISET Gabes). The machine used is a Realmeca C300H 4-axis machining center. The center generates power between 5.5 kW and 7.5 kW with a maximum rotation frequency of 6500 rpm. This machine is powered by 380 V electrical energy and 6 bar pneumatic energy. The Catia V5 software used in this work makes it possible to create the tool paths and then generate the G-code program compatible with the Num director of the center. The tool used is an uncoated high-speed steel tool with a diameter of 8 mm and two teeth (Figure 4).

Figure 4. Measurement plan.

The material of the machined part is an aluminum alloy (2017A: AlCu4MgSi). This material is widely used in the aerospace, automotive, and so forth industries. This material has high mechanical characteristics after treatment and also good machinability, polishability, and good heat resistance between 100 and 250 °C. Table 3 shows the chemical composition of the material used.

Table 3. Chemical composition of the material (wt%)

A Chauvin-Arnoux CA8332 power analyzer was used in order to measure the power consumed by the machine at each instant during a machining operation. Data processing is performed using the Power Analyzer’s Dataview interface. Figure 5 illustrates the shape of the power consumed during the machining of the S₆ sequence. This graph consists of two main parts. The first part is the machining of the features F₂ with the IPC strategy, which in turn consists of three practically identical parts which correspond to the three layers of the same depth of cut (a_p). Also, note that each layer consists of a pace for a cutting width a_e = tool diameter and a second for a width a_e = 3 mm. This is why the power for the first width is greater than the second. Now for the second part, the tool, after completing the first feature F₂, will move on to execute F₁. During the machining of F₁, the tool, a large part of its trajectory passes empty (without cutting) because a large part of the material has been removed by the feature F₂. For this, the machining phase of this part is practically nonvisual in the shape of the power.

Figure 5. Graph of the power consumed for the sequence S₆.

According to Figure 5, the power consumed can be written in the following forms:

(1) $$ {P}_{pi}={P}_0+{P}_{c\_ pi}\hskip2.28em i=\mathrm{the}\;\mathrm{th}\;\mathrm{layer}\hskip1.68em 1\le i\leqslant 3 $$

(2) $$ {P}_{pi}^{\ast }={P}_0+{P}_{c\_ pi}^{\ast } $$

(3) $$ {P}_g={P}_0+{P}_{c\_g} $$

The arithmetic Ra is measured using a KR100 roughness meter. The stroke measured is 6 mm for a caliber of 2.5 mm. Two studies will be carried out in order to reduce energy consumption, minimize the cost of machining, and increase the surface quality of the machined part. The working method for the first case study consists of carrying out a set of machining sequences with different strategies while keeping the cutting parameters constant, namely cutting speed, feed rate, and depth of cut. This is in order to see the influence of sequences and machining strategies on energy consumption, cost, and surface quality. Following and after the realization of the first case and the prediction of the optimal sequence by the use of gray relational analysis (GRA), an experimental plan with a variation of the cutting parameters on this sequence will be carried out in the second case. The objective, in this case, is to predict the experimental results by the use of artificial intelligence and to determine the effects of the cutting parameters on the output responses. The work plan is shown in Figure 6.

Figure 6. Work plan diagram.

One of the big problems in companies is determining the time and cost of manufacturing a product [Reference Kutschenreiter16]. In the case of CNC machining, it is also interesting to determine the time and cost of each operation in order to choose the optimal machining sequence which takes into account QCE in its execution. The cost of a machining operation can be estimated by the equation:

(4) $$ {C}_{tot}={C}_{Machine}+{C}_{Tool}+{C}_{Energy} $$

(5) $$ {C}_{Machine}={\mu}_1+{t}_{cycle}.{\mu}_2 $$

(6) $$ {C}_{Tool}=\varepsilon\;\frac{t_c}{T} $$

(7) $$ {C}_{Energy}=\delta\;{E}_{tot} $$

First case

The first case study focuses on the variation of machining strategies and sequences by setting the cutting parameters for all the tests. The choice of technological machining parameters was made taking into account the catalog of the tool manufacturer and the capacity of the machining center. Table 4 presents the cutting parameter values used in this first part. This study takes into consideration the machining sequences and strategies (tool paths) in order to obtain the optimal sequence. Table 5 illustrates the values of energy consumed, cost, and roughness. Table 6 presents the energy consumed for feature F1 for the different strategies (IPC, zigzag, and zig). Figure 7 illustrates the effects of each strategy for this feature F₁ on the energy consumed (E_tot). The figure shows that the IPC strategy admits minimum energy consumption compared to the other two. The maximum energy consumption for F₁ is achieved with the zig strategy. This increase in energy for this strategy means that the cut is done in one direction and that the tool for a significant part of its cycle moves in the air. This conclusion converges with the results of Edem et al. [Reference Edem Isuamfon, Balogun and Mativenga17] and Altıntaş et al. [Reference Altıntaş Resul, Müge and Hakkı Özgür18]. Figure 8 illustrates graphs of energy consumed and machining cost. The figure shows that the two graphs follow the same variation, which proves that the energy consumed has a significant influence on the cost. As a result, good energy policy management in industries allows for cost reductions in manufacturing. On the other hand, the sequence S₈ = F₁-F₅-F₆ has the minimum cost and energy. This sequence is characterized by the fact that the tool along its trajectory is in full material (cutting phase). In other words, the tool along its trajectory does not cross empty areas (without material). This condition saves a significant amount of cutting time. On the other hand, the S₈ sequence is realized by the IPC strategy for all the features F₁-F₅-F₆, which proves that the choice of the tool path and the machining sequence has a significant influence on the cost and energy.

Table 4. Cutting parameters (first case)

Table 5. Experimental values (first case)

Table 6. The energy consumed for machining the feature (F₁) with the different strategies (first case)

Figure 7. Histogram of the energy consumed for feature (F₁) with the different strategies.

Figure 8. Energy and cost graphs for all sequences.

Industry and researchers face a big challenge when they try to study the surface quality of machined parts. Most researchers who have worked on surface quality as a function of the tool path have used a single machining feature in their studies. Examples of this claim are the work of Zaleski et al. [Reference Zaleski, Jakub and Andrzej19] and Ali Raneen et al. [Reference Ali Raneen20]. Indeed, the use of interacting features with a set of tool paths can lead to logic of the choice of surface quality. On the other hand, an overlay of tool paths when machining Fi features can modify the surface texture of the machined part.

Figure 9 shows the Ra of successive machining of F₁-F₂ entities for sequences S₁, S₅, and S₇ with IPC-IPC, zigzag-IPC, and zig-IPC strategies, respectively. The figure shows that machining F₁ and then F₂ (F₁-F₂) with zig-IPC strategies has the lowest roughness. According to the references by Aramcharoen et al. [Reference Aramcharoen and Paul21] and Ali Raneen et al. [Reference Ali Raneen20], the zig strategy for machining a single feature has the maximum roughness compared to the IPC strategy, which has the lowest roughness. In this study, the problem is a little different since the tool for the first feature F₁ creates several passes with the same strategy in order to complete the depth h of the part. Then, the tool will switch to executing F₂ with another strategy. The problem occurs in the last pass for feature F₂. Indeed, the tool causes the last pass with this new strategy, which passes over the texture left by the first strategy of the feature F₁. This phenomenon can cause a superfinish in the F₁ area. This is possible when the first strategy has a poor surface quality compared to the second strategy. The conclusion of Aramcharoen et al. [Reference Aramcharoen and Paul21] that there is a synergy between the roughness and the energy consumed is not valid in this study because the problem is more complicated given that the tool makes trajectories in the air (without cutting), which can disturb this conclusion, which is not the case in the studies by Aramcharoen et al. [Reference Aramcharoen and Paul21] and Ali Raneen et al. [Reference Ali Raneen20]. The next step is to determine the optimal sequence S_i, which gives minimum energy consumption, machining cost, and Ra.

Figure 9. Surface roughness histogram of F₁-F₂ features made with IPC-IPC, zigzag-IPC, and zig-IPC strategies.

Multi-objective optimization by GRA

After carrying out the experimental tests for this first case, it is now necessary to determine the optimal machining sequence. In this study, there are three output responses, namely energy consumed, cost, and roughness. In this case, it is necessary to have a tool capable of solving this multi-criteria optimization problem. GRA helps to solve this type of uncertainty problem given its ability to understand uncertain systems with partially known information [Reference Kant and Kuldip Singh22]. This theory was established by Deng in the late 1990s. The procedure required by this theory is presented as follows [Reference Ali Raneen20]: To obtain better surface quality, minimum cost, and low power consumption, the condition “the lower the better” will be chosen.

(8) $$ {x}_i=\frac{\mathit{\max}{y}_i(k)-{y}_i(k)}{\mathit{\max}{y}_i(k)-\mathit{\min}{y}_i(k)} $$

where $ {x}_i $ is the output response after normalization and $ {y}_i(k) $ is the experimental values of the output responses. The largest and smallest values are $ \max \left({y}_i(k)\right) $ and $ \min \left({y}_i(k)\right) $ , respectively.

After normalization, the next step is devoted to the calculation of the gray relational coefficient (GRC). The following equation presents this coefficient:

(9) $$ {\psi}_i=\frac{\Delta_{min}+\phi\;{\Delta}_{max}}{\Delta_{0i}(k)+\phi\;{\Delta}_{max}} $$

where $ {\Delta}_{0i}(k) $ represents the absolute value of the deviation between the test reference $ {y}_0(k) $ and $ {y}_i(k) $ , $ {\Delta}_{0i}(k)=\left\Vert {y}_0(k)-{y}_i(k)\right\Vert $ . $ {\Delta}_{min} $ and $ {\Delta}_{max} $ respectively represent the minimum and maximum value of $ {\Delta}_{0i}(k) $ . $ \phi $ represents the distinctive coefficient which includes the interval [0,1]. In this work, the coefficient $ \phi $ is 0.5.

The next step is to figure out the gray relational grade (GRG) after figuring out the GRC.

(10) $$ {\gamma}_i=\sum \limits_{k=1}^n{\omega}_k{\psi}_i $$

where $ {\omega}_k $ denotes the normalized weight of the kth experimental response.

In the methodology of GRA, the maximum value of $ {\gamma}_i $ indicates that the input parameters associated with this value are close to the optimum [Reference Kumar, Paramjit Singh and Sehijpal23]. In this work and according to Table 7, the maximum value of $ {\gamma}_i $ corresponds to the sequence S₈. This sequence shows the best way to combine the input parameters, taking into account their different weights. In this first study, the cutting parameters are constant for all the tests, so the combination F₁-F₅-F₆ with the same IPC machining strategy is the optimal combination.

Table 7. Gray relational grade (GRG) (first case)

In what follows, the study relates to this sequence S₈ with the combination F₁-F₅-F₆.

Second case

In this second part, the objective is to determine the effects of the cutting parameters on the output variables, E_tot, C_tot, and Ra. Predicting experiment results has been done by using neural networks (ANNs) and adaptive neuro-fuzzy inference systems (ANFISs). The last phase of this second part is a comparison between the two artificial models in order to choose the model that gives more refined results. The cutting parameters associated with this part are shown in Table 8.

Table 8. Cutting parameters (second case)

The depth of cut (a_p) is chosen to ensure that for each level, the tool removes the same depth. The total cutting height of the part is h = 1.5 mm. Table 9 presents the experimental values of the 15 trials of E_tot, C_tot, and surface quality for the machining of the S₈ sequence with a different combination of cutting parameters.

Table 9. Experimental values (second case)

Prediction with ANN

ANN models are programming techniques that simulate the human brain [Reference Çay24]. These models have been implemented in several manufacturing, scheduling, and other industry applications [Reference Vaishna, Ankit and Desai25]. An ANN is a digital adoption composed of manipulation (processing) elements called neurons. Neurons form a complex random system that generates a relationship between input parameters and output variables. Nonlinear regression analysis of network output responses [Reference Najafi26] is used to figure out how well neural network-based predictions work. In general, NN consists of three layers, input layers, hidden layers, and output layers.

Two proposals for neural architecture were presented in this study in order to determine the structural performance of a neural network. The first architecture consists of three inputs (V_C, f, and a_p) and three outputs (E_tot, C_tot, and Ra) at the same time (Figure 10a). The second consists of three inputs and a single output presented by the networks ANN1, ANN2, and ANN3 (Figure 10b).

Figure 10. Proposed neural architectures.

The objective of this study is to determine the optimal structure at the output level, number of hidden neurons, and level of the learning algorithm used. The performance of all structures is measured using the MSE during the algorithm learning process. The MSE is given by the following equation:

(11) $$ MSE=\frac{\sum \limits_{i=1}^n{\left({x}_i-{y}_i\right)}^2}{n} $$

The second performance criterion is the correlation coefficient (R²). The R² varies between −1 and + 1. R² close to +1 indicates a strong positive linear relationship between input parameters and output variables [Reference Sayin27]. The R² coefficient is calculated from this equation:

(12) $$ {R}^2=1-\left(\frac{\sum \limits_{i=1}^n{\left({x}_i-{y}_i\right)}^2}{\sum \limits_{i=1}^n{y_i}^2}\right) $$

Several learning algorithms use backpropagation in their processing systems to determine the weights and biases of neural structures. Figure 11 presents the general NNarchitecture with the different parameters.

Figure 11. General neural network architecture.

For structural NN modeling, the Levenberg-Marquardt (LM) algorithm, the scaled conjugate gradient (SCG) algorithm, and the BR algorithm were used. The present work uses the sigmoid function as an activation function in order to normalize the output response. The activation function is represented by the following equation:

(13) $$ f\left({p}_i\right)=\frac{2}{1+{e}^{-2{p}_i}}-1 $$

where p is the input conversion function.

For all neural architectures, the data are divided into three groups: learning, testing, and validation. Seventy percentage of the data is devoted to the learning phase, 15% to the testing phase, and finally 15% to validation. Table 10 shows how to figure out the best learning algorithm, the number of hidden layers, and the number of best outputs. After several tests, the optimal architecture in this study for a single output is {3-10-1} for the output variable E_tot and {3-14-1} for the variables C_tot and Ra. The same LM learning algorithm is used to realize all optimal structures. The optimal architecture with three outputs (E_tot, C_tot, and Ra) is now presented by the {3-10-3} architecture realized with the BR algorithm. The backpropagation algorithm was very good at predicting the amount of energy used, the cost of machining, and the quality of the surface. Figure 12 illustrates the predicted values of E_tot, C_tot, and Ra compared to the experimental values for the single output structure. The R² for Etot, Ctot, and Ra are 0.9984, 0.9963 and 0.8741, respectively. These values show a good correlation between the predicted values and the experimental values. On the other hand, the MSE for the output responses is low, namely 8.25 10⁻⁶ (kWh), 1.42 TD (0.44 $), and 1.17 10⁻⁴ μm, respectively, for E_tot, C_tot, and Ra. Figure 13 shows the predicted values of Etot, Ctot, and Ra versus the experimental values for the three-lead structure. The R² for Etot, Ctot, and Ra are 0.9992, 1, and 0.9117 respectively. So, from Table 10 and Figures 12 and 13, the conclusion is drawn that the LM algorithm has good learning with a single output response. On the other hand, the BR algorithm has good learning with a multi-criteria response (three for this work). By comparison between these architectures, the {3-10-3} architecture with the BR algorithm has a good correlation compared to the other architectures. The MSE global MSE for this architecture is 2.74 10⁻³. Therefore, the use of artificial intelligence by neural networks seems to be a useful methodology to simulate multi-criteria responses in mechanical manufacturing. In what follows, the {3-10-3} architecture will be retained.

Table 10. Different neural architectures (second case)

Figure 12. Experimental values versus predicted values (single output: {3-10-1} and {3-14-1}).

Figure 13. Experimental values versus predicted values (three outputs: {3-10-3}).

Prediction with ANFIS model

The ANFIS model is highlighted to predict experimental results in different fields of engineering [Reference Toghroli28, Reference Sedghi29 and Reference Zhou30]. ANFIS is a hybrid fuzzy inference approach from Sugeno [Reference Mandal31] which simultaneously combines a fuzzy inference system (FIS) with a NN. This combination aims to make the system efficient in order to solve complex problems [Reference Saw Lip32]. In the general case, the ANFIS model includes five consecutive treatment lines. The lines (layers) are connected to each other by different nodes. Each entry node is drawn from the previous line. Figure 14 presents the general architecture of the ANFIS model. The numerical interpretation of this architecture can be presented as follows:

Rule 1 = if x is A₁ and y is B₁, then f₁ = p₁ x + q₁ y + r₁

Figure 14. General architecture of the ANFIS model.

Rule 2 = if x is A₂ and y is B₂, then f₂ = p₂ x + q₂ y + r₂

where x and y are the input parameters of the experimental study, and p_i and q_i represent the variables of the fuzzy set. The output function of the ANFIS model is presented by

(14) $$ f=\overline{w_1}{f}_1+\overline{w_2}{f}_2 $$

where $ \overline{w_1}=\frac{w_1}{w_1+{w}_2} $ ; $ \overline{w_2}=\frac{w_2}{w_1+{w}_{2.}} $

where w_i is the trigger strength for each rule.

In this present work, in order to find the optimal architecture of the ANFIS model, a combination of numbers and types of membership functions (MFs) (Gaussian-shaped MFs, gaussmf and gauss2mf, and triangular-shaped MF, trimf) was carried out. The MFs used in this work are the trimf, gaussmf, and gauss2mf. The numbers of MF used are {2 2 2}, {2 2 3}, and {2 3 3}. The MF of constant and linear types is used for the output variables. The ANFIS modeling process runs with 75% of the data for training and 25% for testing. In this process, the iteration number for the mapping is equal to 100. Table 11 presents the different combinations of ANFIS architecture with the different MFs. According to this table, structure {2 2 2} is the optimal structure for the three output responses (E_tot, C_tot, and Ra). According to the results, the function (trimf) has the lowest test error for the energy consumed (E_tot) with a linear output. On the other hand, the (gaussmf) function has the lowest error for C_tot and Ra output responses with constant and linear output functions, respectively. The MSE_All values for the E_tot, C_tot, and Ra output responses are 0.01913, 3.399 and 4.516 E-3, respectively. Figure 15 illustrates the final {2 2 2} architecture for the three output variables. Figure 16 illustrates the comparison between predicted and experimental values for output responses. The R² values for the variables E_tot, C_tot, and Ra are 0.95, 0.965 and 0.968, respectively. These values present a good correlation but are less weak than those of the neural network, except for the surface quality for the ANFIS model, which is greater than that of the ANN (0.968 > 0.9117). Table 12 shows a comparison between the two ANN and ANFIS models’ global MSEs (training and testing) for the different output variables. The result shows that the global MSE of the {3-10-3} neural structure is lower compared to that of the ANFIS model. Indeed, Figures 12 and 13 (ANN), Figures 15 and 16 (ANFIS), and comparison Table 12 show that the use of the BR algorithm with a multi-criteria output response can give good results when compared with ANFIS.

Table 11. Different combinations of ANFIS architecture (second case)

Figure 15. Final {2 2 2} architecture.

Figure 16. Experimental values versus predicted values (ANFIS model).

Table 12. The MSE_ALL values of the ANN and ANFIS models (second case)

A 3D surface plot study was carried out in order to describe the effects of the cutting parameters on the output responses. Figure 17 shows the variation of energy consumed (E_tot), cost (C_tot), and surface quality as a function of the cutting parameters. Figure 17a and 17b shows that for a decrease in the feed rate (f), there is an increase in the energy (E_tot) and the cost (C_tot). This conclusion is because the machining time decreases when the feed rate increases, which positively influences the energy consumption and the cost. This conclusion is similar to the studies of references [Reference Bagaber Salem and Ahmed Razlan33, Reference Congbo34 and Reference Khan Aqib35]. Figure 17e shows that as the feed rate increases, the surface quality decreases. In other words, the roughness (Ra) increases. Indeed, the feed rate (f) has a remarkable influence on the surface quality compared with the cutting speed. In this present work, the depth of cut (a_p) is chosen so that all levels have the same depth of cut for a height of cut h. Figures 17(b) and (d) clearly show that as the depth of cut increases, so does the energy (E_tot) and the cost (C_tot). The decrease in the number of cutting levels, caused by an increase in the depth of cut, leads to a reduction in total machining time. Similarly, for the surface quality, the increase in the depth of cut has a negative influence on the roughness (Ra) (Figure 17(f)). This is because the tool tends to vibrate more as the depth of cut gets deeper, especially if the diameter of the tool is small. These conclusions converge with the studies of Khan Aqib et al. [Reference Khan Aqib35], Eser et al. [Reference Eser36] and Devarajaiah et al. [Reference Devarajaiah and Muthumari37].

Figure 17. 3D surface plot of energy (E_tot), cost (C_tot), and roughness (Ra).

Development of an intelligent simulator based on prediction models ANN and ANFIS

An interactive interface has been developed to help machinists predict the energy consumed, surface quality, and cost of a 2017A alloy part containing an interacting pocket and groove. The simulator was implemented using the “Guide” interface of the MATLAB 2016 software. Figure 18 shows this user interface named “QCE_Intelligent_simulator,” which is simple and extremely easy to learn. As shown in the figure, the user must first input the values of the cutting parameters (Inputs), that is, the cutting speed (V_c), the feed rate (f), and the depth of cut (a_p). Subsequently, the user has the choice to predict the results according to the ANN model or the ANFIS model by pressing the “Predict” button.

Figure 18. QCE_Intelligent_simulator.

The Catia-machine interaction is achieved through the production of a G-code program according to the ISO 6983 standard compatible with the CNC director Num of the machining center. The power is measured in real time during the execution of the machining operation. The actual results are compared to what was expected after the data and results have been analyzed [Reference Bousnina, Hamza and Yahia38, Reference Bousnina, Hamza and Ben Yahia39]. When another operation is executed, the machining parameters are updated.