1. Introduction
In the turning process, surface roughness performs a vital role in the creation of products, and also exerts great influence on machining cost because it is considered an index of product quality [
1]. However, surface roughness defines such mechanical properties as corrosion, wear, lubrication, electrical conductivity and fatigue behavior [
2]. Moreover, the surface roughness of any machining process has become prominent because of the heightened quality demands. The production of a desired surface finish on a piece of work is mainly affected by machining parameters, such as cutting speed, feed rate, depth of cut, tool geometry, workpiece material and other factors such as tool wear, vibrations, machine dynamics and temperature. Meanwhile, heat is generated during the turning process and the uses of cutting fluid provide lubrication and cooling, which affects and progresses the final quality of the workpiece. Cutting fluids improve the efficiency of machining in terms of improved surface finish, improved dimensional accuracy, reduced tool wear and reduced cutting temperature. Sen et al. [
3] presented the advance in capabilities of the ecofriendly minimum quantity lubrication (MQL) technique. The authors discussed the advantages of MQL and illustrated a review of literature of MQL assisted machining operations. Rapeti et al. [
4] use the application of vegetable oil based nano cutting fluids (coconut oil, sesame oil and canola oil) during the turning of AISI 1040 steel. Economic analysis for the application of nano cutting fluids is done to assess the viability of these fluids in the industry. Kanth et al. [
5] investigated the use of a mixture of nano crystalline graphite and sunflower oil as an alternative for cutting fluids for an improved surface roughness finish in a turning process. The study revealed that sunflower oil results in a better surface roughness finish when compared to other vegetable oils.
In the past, machining parameters used to be selected by the trial and error that was time consuming and costly, based on process planners’ experience and machining handbooks [
6]. A human process planner chooses proper machining process parameters that depend on his own experience or his machining tables. In most cases, the selected parameters are conventional and far from optimal. However, in machining it is significant to choose the proper parameters. If the machining parameters are not appropriate, excessive cutting tool wear is noticed and the choice may result in surface damage.
Surface roughness refers to the shape of the surface to be machined and combined with surface quality. The appearance of the surface roughness mechanism is very complex and mostly depends on highly analytical equations. The surface finish can be characterized by two main parameters, average roughness (
) and maximum peak to valley height (
). Theoretical models have been used to calculate these parameters [
7]. A basic theoretical model for surface roughness is given by Equation (1)
where
is the feed rate and
is the tool nose radius. Based on this model, one need only reduce the feed rate or increase the tool nose radius to produce the desired surface roughness. This model to some extent presumes a large nose radius and a slow feed. For a zero nose radius and a somewhat larger feed, the following model is suggested by Boothroyd and Knight [
8]
where
and
are the major and end cutting edge angles respectively, and
is the cotangent function. Fang and Safi-Jahanshahi [
9] present linear and exponential empirical models for surface roughness as functions of cutting speed (
), feed (
) and depth of cut (
):
where
is constant and
and
are the exponents.
In the present paper, empirical models are established with conventional methods such as a factorial design, statistical regression and response surface methodology. Artificial intelligence-based models are introduced using nonconventional approaches such as the artificial neural network (ANN), fuzzy logic (FL), support vector regression (SVR) and a genetic algorithm (GA) [
10]. Using conventional methods may not be enough to define the nonlinear complex relationship between machining parameters and machining performance. Lately a good deal of attention has been devoted to establishing predictive and optimization models in order to consider the effect of machining parameters on machining functioning, using artificial intelligence methods as an alternative to conventional methods. Trung-Thanh Nguyen [
11] applies a microgenetic algorithm (AMGA) for dry milling in order to resolve the trade-off analysis between the material removal rate, specific cutting energy and surface roughness. Camposeco-Negrete [
12] uses a robust design technique to control the results and contributions of four machining parameters on the above-mentioned response variables in wire-cut EDM. Soepangkat et al. [
13] propose a grey fuzzy analysis and BPNN-based GA to control and predict the optimal parameters in the drilling KFRP. Venkata and Murthy [
14] combine predictive models such as response surface methodology, artificial neural networks and support vector machine to predict the surface roughness and root mean square of work piece vibration in the boring process. Prasath et al. [
15] developed a mathematical model for prediction response employing Taguchi and response surface methodology (RSM). The model is confirmed and predicted the surface roughness and MRR with less than 6% of error. Matras et al. [
16] introduced new optimized method that involves the prediction of the curvilinear surface roughness in turning titanium alloy. The created model also results in a short machining time and low manufacturing cost. The machining time was significantly reduced in comparison to the non-optimized cutting process. Mia and Dhar [
17] presented a prediction model development of surface roughness in hard turning when the experimental runs were conducted under both dry and high-pressure coolant (HPC) conditions. the prediction model was prepared by employing support vector regression (SVR) and response surface methodology (RSM) and the optimization model was constructed by embracing the composite desirability function (CDF) and genetic algorithm (GA). The predictive model by SVR and optimization model by GA provided the highest accuracy. Yadav [
18] applied a hybrid approach of the Taguchi methodology-response surface methodology (TM-RSM), which has been implemented for modeling and optimization for the duplex turning process. The optimum condition obtained from TM has been used as a central value in RSM for the modeling and optimization. The result shows the significant improvement in surface finish with the hybrid approach as compared to the Taguchi analysis. Chabbi et al. [
19] investigated the influence of cutting parameters on the finish of surface roughness during the cutting of the polyoxymethylene (POM C) by utilizing the response surface methodology (RMS) method. The results revealed that the surface roughness was strongly influenced by the feed rate with a large contribution, followed by the cutting depth, whereas, the cutting speed has no influence. A recent study on the use of dry, mono-jet and dual-jet of cryogenic conditions in the turning process was presented in [
20]. The Taguchi full factorial orthogonal array design was used to study the machining responses of Ti-6Al-4V alloy and grey relational analysis (GRA) method has been utilized to optimize the parameters. The results illustrated that ideal responses can be achieved using the dual-jet LN2 cryogenic condition.
Other soft computing machine learning approaches, such as ANFIS, have been proposed to predict workpiece surface roughness in the turning operation. Jain and Raj [
21] introduce monitoring systems that use ANFIS to predict the surface roughness. This model shows the ability to estimate tool life for an unmanned manufacturing system related to surface roughness. Elbaz et al. [
22] propose a model based on the fuzzy C-mean (FCM) clustering method that combines enhanced particle swarm optimization (PSO) with ANFIS. The computational model was used to predict the performance of an earth pressure balance (EPB) shield during tunneling. The prediction results indicate an accurate prediction of the EPB and good agreement between the actual measurements and the predicted values. Zhang et al. [
23] develop two computational models based on the random forest (RF) algorithm. A hybrid algorithm PSO-RF is proposed to optimize operational parameters in real time during the tunneling process so that tunneling-induced settlement can be controlled within the tolerated values. The results demonstrate that the predicted results are accurate when compared with actual settlements. Chen et al. [
24] apply three artificial neural network (ANN) methods: back-propagation (BP), a neural network (the radial basis function (RBF) neural network) and the general regression neural network (GRNN) to predict the maximum surface settlement caused by EPB shield tunneling. The results of analysis show that close correlations were established between the predicted and the measured settlements in the GRNN model with MAE = 1.10, and RMSE = 1.35, respectively. Shivakoti et al. [
25] present predictions about the machining of stainless steel 202, based on the adaptive network-based fuzzy inference system and parametric analysis of CNC lathe-process parameters. The experimental outcomes and ANFIS predicted results are compared, confirming the precise prediction of ANFIS outcomes in the course of turning stainless steel 202. Maheshwera et al. [
26] analyze the influence of machining parameters on surface roughness by establishing regression analysis (RA) and artificial neural network (ANN) models during the turning of hard work material, AISI 52,100 steel. The prediction performance of the ANN model is shown to be better than that of the RA model and is expected to be a practical way of reducing the required time and expense of experimental runs. Palanisamy and Senthil [
27] introduce an adaptive neuro fuzzy inference system (ANFIS) to define the relationship between the count input machining conditions and output measures such as the cutting force and surface roughness of the machined surface. The achieved results reveal the development of output quality combined with lower production cost, which is evident of the efficiency of the established ANFIS model. Arapoglu et al. [
28] suggested new variable selection method based on artificial neural networks (ANN) for the prediction of the surface roughness. A statistical hypothesis test is used as an elimination criterion. The selection of variables does not change the prediction accuracy of the model at the 1% significance level.
The reported literature suggests that machine learning approaches, such as ANN and ANFIS, have shown efficacy in predicting the machining parameters of various applications. When compared to ANFIS, a hybrid ANN approaches such as ANN-RBF and ANN-BPFN have a more complex structure and require high computation power. In addition, the hybridization of ANFIS with evolutionary algorithms such as GA and BFA would require more computational time due to the nested populations in the GA and BFA algorithms. In addition, ANFIS-PSO have shown it to be effective in predicting surface roughness but ANFIS-PSO may not converge to global optima and could get trapped in local optima [
29]. QPSO, however, has been found highly effective, outperforming PSO in several applications due to its simple implementation and fast global optimum convergence [
29].
In this research, we propose the use of ANFIS-QPSO for predicting the surface roughness in the turning process. To the best of our knowledge, there is a gap in literature in utilizing ANFIS-QPSO to predict surface roughness in the turning process. In addition, no previous study has investigated the use of the ANFIS-QPSO approach for predicting the surface roughness in dry and cryogenic turning processes. The present study set out to examine the accuracy of ANFIS-QPSO in predicting the experimental dataset of a dry and cryogenic turning process involving AISI 304 stainless steel. In the next section, the methodology of the proposed ANFIS-QPSO approach is presented with nomenclature presented in
Table 1. The simulation results present the predicted results and highlight the prediction accuracy of ANFIS-QPSO when compared with the classical ANFIS approach.
4. Performance Comparison of ANFIS-QPSO versus Relevant Machine Learning Algorithms
In this section, the predictive accuracy of the proposed ANFIS-QPSO was compared with state of art evolutionary-optimized ANFIS algorithms that are widely used in literature such as in [
43,
44]. The ANFIS-QPSO performance was assessed against ANFIS-GA and ANFIS-PSO algorithms. The assessment was carried out based on the RMSE, MAPE and R
2 values in dry and cryogenic turning processes presented in this work. The integration of ANFIS with GA and PSO falls beyond the scope of this study.
Figure 15 presents a comparison between the measured and predicted surface roughness values by ANFIS-QPSO, ANFIS-GA and ANFIS-PSO algorithms for the dry turning process with the associated performance indicators presented in
Table 11. All of the algorithms performed well, with slight performance measures, in predicting the surface roughness values of the dry turning process. However, it can be noted that ANFIS-QPSO had a considerably better predictive performance that outperformed ANFIS-GA and ANFIS-PSO in terms of the three performance indicators.
Similarly, for the cryogenic turning process, simulations were carried out to compare the performances of ANFIS-QPSO, ANFIS-GA and ANFIS-PSO to predict the surface roughness values.
Figure 16 illustrates a comparison between the measured and predicted surface roughness values by the three algorithms. In addition, the performance indicators of the ANFIS-QPSO, ANFIS-GA and ANFIS-PSO are presented in
Table 12. In contrast, ANFIS-GA had the least RMSE value in comparison with ANFIS-QPSO and ANFIS-PSO. ANFIS-QPSO outperformed ANFIS-GA and ANFIS-PSO in terms of the MAPE and R
2 values, which was capable of reaching 5.15% and 0.988 respectively. Therefore, ANFIS-QPSO exhibited the highest prediction performance.
The superior performance of ANFIS-QPSO can be interpreted by its capability in converging to global optima solutions and avoiding local optima. The comparison results suggest that ANFIS-QPSO has a superior predictability performance in dry and cryogenic turning processes when compared with ANFIS-GA and ANFIS-PSO.