An application of hybrid Taguchi-ANN to predict tool wear for turning EN24 material Open Access

In the metal-removing process, the reduction in tool cost minimizes the manufacturing cost. Wear found in tools increases this cost. Tool wear has a huge direct impact on surface roughness, power consumption, dimensional precision of the finished part, and chatter in the tool and machine tool. More tool wear means less surface finish, more power consumption, less dimensional accuracy, and more vibration in the tool and machine tool. This ultimately increases the production cost. Hence, it becomes important to minimize tool wear through optimization techniques. In this work, the optimization is carried out by the Taguchi method along with an artificial neural network (ANN) to find the optimal setting of process parameters. Furthermore, the additive model is used to predict the optimal output parameter. The ANN has a wide range of applications in the prediction of responses. The systematic optimization capabilities of the Taguchi method are merged with the predictive power of artificial neural networks. Once the hybrid Taguchi-ANN model is trained and validated, it is used to predict tool wear for different sets of cutting parameters. This can help in the optimization of machining processes to minimize tool wear and improve efficiency. Hence, an effective tool wear prediction system for turning EN24 material is developed, which can lead to significant improvements in machining efficiency and cost savings. Figure 1 shows the combined Taguchi optimization and ANN for the turning process.

In the turning process, just like in other metal-removing processes, product quality in terms of finished surface and production rate in terms of MRR (material removal rate) are the most significant concepts. If MRR is increased, roughness in the surface gets increased, and vice versa. If a tool with minimal or no wear is used, the surface finish will increase even though the MRR is higher. In manufacturing industries, metal-removing methods such as turning, milling, drilling, etc., have special importance. Hence, advancement in such processes is always needed. The prediction of responses by new techniques such as artificial neural networks will fulfill this need. The main objective of this research is to predict tool wear using the ANN and the Taguchi method for turning EN24 material. A lot of work is carried out with respect to the application of the ANN in machining processes. Some important literature reviews are summarized here. The ANN, along with response surface methodology, can be used to predict the surface roughness and MRR of turning Ti–6Al–4.¹ ANOVA revealed that the most significant parameter is speed for surface roughness and depth of cut for MRR. Also mentioned is that the ANN exhibited 5.04% and 10.66% errors for surface roughness and MRR, respectively. The percentage error is very low in the ANN with respect to response surface modeling (RSM). The Taguchi method is also used for 16MnCr5 steel material for optimization of Computer Numerical Controlled (CNC) turning process parameters such as surface roughness and MRR. The workpiece material used is 16MnCr5 steel, and the tool material is a TiN-coated cutting tool. Taguchi orthogonal is adopted. The result revealed that feed and depth of cut are the most significant factors for surface roughness and MRR, respectively.² Multi-optimization was successfully performed by a genetic algorithm for hard turning in previous studies with 42CrMo4 alloy workpiece material.³ A full factorial method was used for the Department of Energy (DOE). The analysis of variance (ANOVA) result specified that cutting speed and hardness are the significant factors for cutting temperature and surface roughness, respectively. In a previous study, the ANN was used for the prediction of MRR and OC. Five input parameters, viz., feedforward neural network (FNN), Cable News Network (CNN), recurrent neural network (RNN), General Regression Neural Network (GRNN) (general regression neural network), and long short-term memory (LSTM), are used as NN models. These NN- Neural Network (NN) models successfully predict the responses.⁴ LSTM is able to solve problems with disappearing and exploding gradients during training. Experiments with the NN approach to predict tool wear are used to train the data.⁵ The Levenberg–Marquardt algorithm, Bayesian regularization, and scaled conjugate gradient are used for training the data. The result revealed that the NN can be successfully implemented with a very minimal variation in experimental and predicted values.

Particle swarm optimization, as a multiple optimization method, can be used for frequency division multiplexing (FDM). The process parameters used are layer thickness, build orientation, infill density, and extrusion temperature, whereas compressive strength and build time are response parameters.⁶ In a previous study,⁷ RSM and Grey Relational Analysis (GRA) were used to optimize Wire Electrical Discharge Machining (WEDM) process parameters. Inconel 690 was used as the workpiece material. The Box–Behnken design of response surface methodology is implemented for data analysis. Experimental and numerical methods can be used to optimize the injection-molded polymer gears.⁸ Shrinkage and sink marks have been investigated experimentally in previous studies. The responses time and cost to create the appropriate shape are considered. The result concluded that the depth of sink marks was optimized to 0.07 mm. The ANN is utilized for the prediction of cutting force, surface finish, and material removal rate for turning the titanium alloy Ti6Al4V. A full factorial method is adopted for the data analysis. Cutting speed, feed, depth of cut, and rake angle are the process parameters. Comparative studies are made for experimental and predicted values. It is observed that prediction accuracies are very high, which indicates the model’s effectiveness.⁹ The NN tool is capable of predicting the surface roughness during the end milling process.¹⁰ The conventional models are compared with vision-based approaches for prediction. The experimentation is carried out on a Taguchi L27 orthogonal array. The input parameters are spindle speed, feed/tooth, and depth of cut. ANOVA revealed a p-value of less than 0.01 with a 99% confidence level. Neural networks are trained, and the extracted features are used to predict the surface roughness. The outcome showed a better result for the ANN architecture 5-3-1 with three hidden layers. The multi-objective optimization method of gray relational analysis, along with the ANN, is also used.¹¹ The material used is Al5059/SiC/MoS2. ANOVA suggested that SiC% and the feed rate are the most significant parameters. Particle size has a negligible effect on response. The ANN and Taguchi method are used to optimize surface finish and cutting force for milling AA7039/Al₂O₃. The L8 Taguchi OA is used for the DOE. ANOVA determined the effect of input parameters on response, such as cutting force and surface roughness. The ANN is used to predict the responses.¹² The ANN and Taguchi method are also used for laser welding optimization.¹³ Data analysis is carried out to find the optimum level of process parameters. Neural network training validated the optimum solution. The Taguchi method was also used in the dry-turning process to optimize the responses.¹⁴ Signal-to-noise and ANOVA are imposed to predict the level of optimization. A confirmation test is further performed, and it is found that the experimental and predicted outputs have only an 8% deviation.

The Taguchi method is used along with the ANN and fuzzy inference systems to optimize the responses.^15,16 The workpiece material used is AISI 4340. It is observed that the ANN and FIS can be implemented successfully. The ANN is used to predict Ra for the milling process.¹⁷ The Taguchi method is adopted to find the number of experiments and data analyses. ANN training is performed by the Marquardt algorithm. It can be concluded that the ANN is a powerful technique for predicting Ra. It is concluded that the feed rate is the most significant parameter. A predictive model of output parameters for the turning process was also presented in a previous study.¹⁸ The ANN is used with the regression model. The comparative study demonstrated that the ANN, or the neural network, model produced a better result than the regression model for tool wear. The reviews on tool wear monitoring by the ANN are also available in the literature.¹⁹ The different types of wear and their measurement techniques are elaborated on. The publications are classified by considering features such as time, frequency statistics, and time-frequency domain. The application of the ANN and Taguchi method for estimation and optimization of process parameters in hard turning is also presented.¹¹ The three levels of all input parameters are considered. A Taguchi OA is used for experimentation. ANOVA and the S/N ratio are used for analysis. The methods of regression and adaptation are adopted for prediction. The result stated that the NN model is better than the regression model. According to the aforementioned literature review, there is a research gap because few studies have used the ANN with Taguchi approach to predict tool wear, very few have experimented with spindle vibration as a noise factor, and no studies have used EN24 as a workspace material. The hybrid Taguchi ANN can be a possible solution for tool wear prediction.²⁰ 

Through the use of a hybrid Taguchi-artificial neural network technique, the purpose of this study is to make an attempt to estimate tool wear when turning EN24 material. The wear on the tools should be reduced as much as possible. The variables that are considered independent include the cutting environment, feed rate, depth of cut, nose radius, and tool type. The cutting environment is referred to as CE. The spinner CNC lathe is utilized so that performance may be improved.

The work station for experimentation is Minar Hydro Ltd., MIDC, Nagpur. Experimentation is performed on a Spinner 15 CNC lathe. A Taguchi L₂₇ OA (Orthogonal Array) is implemented. A total of 81 observations [27 × 3 (spindle vibration level)] are performed. The machine tool and experimental set up are shown in Fig. 2.

Five input parameters and three levels of each parameters are considered.¹⁶  Table I shows the independent parameters and their level. Table II shows the L₂₇ Taguchi design matrix with actual values.

The unconventional method, i.e., image processing method, is used to measure the tool wear. The specially designed canny algorithm is implemented in MATLAB. CANNY shows some special features such as low disturbances in image handling and processing. This method is simple and quick, which also has high precision. Figure 3 shows the flowchart of the edge detection method for tool wear measurement disturbances.

The main steps of tool wear measurement are image reading, image processing, and measuring wear. In the image reading step, a high definition camera is used for image capturing. This captured image shown in Fig. 4 is sent as an input to MATLAB software. Then in the image processing step, the raw image is converted into a gray scale image to achieve good quality. Figure 5 shows the gray scale image. The shadow is removed by the CANNY algorithm. The weak edges are detected and then further included in the output. They are conspired as the output if they have strong edges. Then image can be saved in “jpeg” format for next processing. Finally in the measuring wear step, flank wear conditions are adopted for wear measurement. Confirmation of the suggested image is done by the MATLAB environment. The wear area is read by the 2D image process by the algorithm. Measurement of the depth is the limitation of this method. Figure 6 shows tool wear measurement by the algorithm.

The tool wear is measured for each level of the noise factor (spindle vibration). Table III shows the observation table with measured values of tool wear for every level of spindle vibration and the average value of tool wear. It also shows the signal to noise ratio for tool wear.

The 95% confidence interval of the population is [μTW − CI] < μTW < [μTW + CI], i.e., 0.0731 < 0.0911 < 0.109.

The architecture of the neural network is shown in Fig. 9. It is constructed for five input parameters, one hidden layer, and one output parameter. The ANN is performed in MATLAB 2019a software. It is accomplished by data input, ANN training, and validation of the ANN model.

Implementation of the ANN is carried out by data input, ANN training, and validation of the ANN model. One input file containing five input parameters is created and is fed as data in MATLAB software. In addition, one target file containing one output (tool wear) is created. This is treated as a target file to software. Table VIII shows the input data and target data to the ANN.

In ANN training, input is trained by providing the number of iterations and hidden layer. This is a crucial step, in which the output depends on the proper training of the input.

Table IX shows the coefficient of correlation (R) values for three topologies of three input parameters with one output parameter. The ANN model is trained on the basis of three algorithms, viz., Levenberg–Marquard, Bayesian regularization, and scaled conjugate gradient. The model is trained, tested, and validated for the three above- mentioned algorithms. It is observed that the overall R value is highest for topology 3-5-1, which used Levenberg–Marquard algorithm, as compared to that of 3-3-1 and 3-4-1 topologies. The model is further checked for mean square error (MSE), and it is also observed that the Levenberg–Marquard algorithm has the lowest value of MSE with the 3-5-1 topology.

After training the ANN, regression can be drawn out. Figure 10 shows the regression graph for training, testing, validation, and overall data with an overall value of R = 0.984 03. Figure 10(a) shows the training input data. The dotted line reveals a smooth fit. The direct fit is indicated by a blue color line. The model of ANN training is structured by projected value = 0.83 ∗ target value + 0.019. This presents the best relation in experimental and projected values. In Fig. 10(b), the dotted line and red line are shown. The model of ANN testing is structured by the equation projected value = 1.1 ∗ target value + 0.052. After training, validation of the model is performed. Figure 10(c) shows the validation of the ANN. Figure 11 shows the performance plot for the best validation when the model is trained, tested, and validated. The best mean square error is 0.000 162 31.

The Taguchi method is implemented successfully. The results of ANOVA (Table V) indicate that depth of cut is the most significant factor to achieve the best value of tool wear followed by feed rate, nose radius, cutting environment, and tool type. The optimal setting of process parameters are the cutting environment at level 3 (A₃ = MQL), nose radius at level 3 (B₃ = 1.2 mm), feed rate at level 3 (C₃ = 0.35 mm/rev), depth of cut at level 1 (D₁ = 0.5 mm), and tool type at level 3, E₃ = coated (CVD) or A₃B₃C₃D₁E₃. The interaction effect of cutting environment on the nose radius, feed rate, depth of cut, and tool type for tool wear is shown in Figs. 12–14, and it is observed that there is the synergic relationship between the parameters. As the relationship is synergic, the obtained optimal setting will be acceptable. The predicted value at optimal setting is 0.0401 mm. The experimental value at optimal setting is 0.0422 mm. Figure 12(a) shows the interaction plot for the mean S/N ratio and nose radius. The plot shows a synergetic relationship between input and output values. As per the Taguchi method, this type of relationship is acceptable for the analysis. The same type of relationship is obtained for feed rate, depth of cut, and tool type, shown in Figs. 12(b)–12(d), respectively.

These outputs obtained by optimal setting are also confirmed in Fig. 13. This figure shows that the increase in NR and FR reduces the tool wear. Figure 13(b) shows that a higher value of FR and lower value of DOC reduce the tool wear. Figure 13(c) shows that a higher value of NR and lower value of DOC reduced the tool wear. The established ANN is trained for input and output factors. 1000 epochs are implemented. Network simulation is carried out for nondependent values. Neural network training is successfully completed for the input during test readings. The input and the target are compared with the real output. Finally, the predicted (TW) and experimental values (TW) are compared. A negligible variation is seen. The surface plots are important to show the corresponding changes in the output with respect to each level of the input.²⁰ The experimental vs predicted tool wear is shown in Table X and Fig. 14. The predicted output values obtained by the Taguchi-ANN method are very close to the experimental output values. Hence, the Taguchi-ANN method is best suited for the current configuration.¹⁸ 

The artificial neural network combined with Taguchi philosophy is successfully implemented for optimization and prediction of tool wear. The hybrid Taguchi-ANN method is used in the present study to obtain the output results precisely. The current methodology shows the best possible outcomes compared to other existing methods. The tool wear is minimized to 0.041 mm. The optimal setting obtained is A₃, B₃, C₃, D₁, and E₁. The most significant factor is depth of cut (49.08% contribution), and the least significant factor is tool type (3.58% contribution). A cutting environment of MQL, a nose radius of 1.2 mm, a feed rate of 0.35 mm/rev, and a depth of cut of 1.0 mm with the uncoated tool are observed to provide the lowest tool wear of tungsten carbide for turning EN24 material. ANOVA analysis revealed that the depth of cut and feed rate significantly affected the tool wear around 54% and 21%, respectively. The cutting environment and nose radius achieved moderately significant tool wear by nearly 10% and 8%, respectively. The tool type is the least significant to tool wear. Its contribution is only 3.45%. The analysis further exposed that tool wear is minimum when the nose radius and feed rate are at a high level and the depth of cut is at medium for the uncoated tungsten carbide tool under minimum quantity lubrication. The conformity test discovered that the experimental and predicted tool wear is within the range given by the confidence interval. Negligible variation is found between the experimental and predicted tool wear. The regression model obtained by the ANN for training, testing, and validation is close to the ideal value of 1.

The authors have no conflicts to disclose.

All authors contributed equally to the conceptualization, formal analysis, investigation, methodology, and writing and editing of the original draft. All authors have read and agreed to the published version of the manuscript.

Prashant D. Kamble: Writing – original draft (equal); Writing – review & editing (equal). Jayant Giri: Conceptualization (equal); Resources (equal). Emad Makki: Formal analysis (equal); Investigation (equal); Resources (equal). Neeraj Sunheriya: Methodology (equal); Writing – review & editing (equal). Shilpa B. Sahare: Formal analysis (equal); Investigation (equal); Validation (equal). Rajkumar Chadge: Data curation (equal); Formal analysis (equal); Software (equal). Chetan Mahatme: Investigation (equal); Validation (equal); Visualization (equal). Pallavi Giri: Methodology (equal); Resources (equal). Sathish T.: Formal analysis (equal); Supervision (equal). Hitesh Panchal: Project administration (equal); Resources (equal).

Data sharing is not applicable to this article as no new data were created or analyzed in this study.