Application of ANN to estimate surface roughness using cutting parameters, force, sound and vibration in turning of Inconel 718
- 207 Downloads
Abstract
In this paper, artificial neural network approach is used to predict surface roughness using cutting parameters, force, sound and vibration in turning of Inconel 718. Experiments were performed by using cryogenically treated and untreated inserts, and various responses were measured. Then, these measured responses were used as input to the artificial neural network to predict surface roughness. It is found that the models developed by artificial neural network are predicting surface roughness with more than 98% accuracy. Further, the predictions obtained by artificial neural network are compared with the results of regression-based prediction models earlier proposed by the authors. The modified regression models were estimating surface roughness with more than 90% accuracy. Based on correlation coefficient values, the prediction results of modified regression model are compared with those obtained by artificial neural network. Finally, it is concluded that artificial neural network models are better for estimating surface roughness than the regression models and such predictions are useful for real-time control of the process to acquire the desired surface roughness.
Keywords
Artificial neural network Surface roughness Inconel 718Abbreviations
- CNC
Computer numerical control
- RSM
Response surface methodology
- FOE_{M}
Modified first-order equation
- ANN
Artificial neural network
- BP
Backpropagation
- BR
Bayesian regularization
- LM
Levenberg–Marquardt
- AE
Absolute error (%)
- MAE
Mean absolute error (%)
- MSE
Mean square error (%)
List of symbols
- v
Cutting speed (m/min)
- f
Feed rate (mm/rev)
- d
Depth of cut (mm)
- F_{c}
Cutting force (N)
- S
Sound pressure level (Pa)
- V_{v}
Vibration velocity of workpiece (m/s)
- n
Number of experiments
- \(R_{\text{ai}}\)
Average of measured surface roughness in μm
- \(\hat{R}_{\text{ai}}\)
Estimated surface roughness
- h
Number of neurons in single hidden layer
- R^{2}
Correlation coefficient
1 Introduction
Surface roughness (R_{a}) is a main performance indicator in machining, and it is important for different applications [1]. The existence of R_{a} is very complicated and method oriented. Machining can be modelled and confirmed for real-time control of the process to acquire the desired R_{a} [2], where modelling is the process of fitting input and output data [3]. Therefore, numerical models offering relationship amongst R_{a} and cutting parameters such as cutting speed, feed and depth of cut are developed [4, 5, 6].
Nickel-based super alloy (Inconel 718) is most popular in aerospace, petroleum, medical industries, etc. though it is costly and difficult to machine [7, 8, 9, 10, 11, 12]. Researchers have used regression analysis and response surface methodology (RSM) [13, 14, 15, 16, 17, 18, 19, 20, 21, 22] for prediction of R_{a}. Deshpande et al. [2] have used regression-based modelling for estimating R_{a} in turning of Inconel 718. This paper deals with use of artificial neural network (ANN) for predicting R_{a}, and therefore, some works reporting use of ANN in machining of different alloys are discussed below.
References | Workpiece material | Operation | Tool used | Parameters considered | Model used | Prediction accuracy based on | Remark |
---|---|---|---|---|---|---|---|
Pontes et al. [15] | AISI 52100 steel | Turning | Ceramic (Al_{2}O_{3} + TiC) inserts | Speed, feed and depth of cut | ANN | Mean absolute percentage error and percentage of maximum error | The use of DOE method in the use of ANN architecture to be an effective tool for estimation of surface roughness |
Mia and Dhar [16] | EN 24T steel | Dry turning | Coated carbide (TiCN, WC, Co) inserts | Speed, feed, hardness, dry and high-pressure coolant (HPC) | ANN with different training methods | Root mean square error (RMSE) and correlation coefficient (R-value) | The performance of the surface roughness is investigated for dry and HPC conditions. The Bayesian regularization proved the excellent prediction accuracy |
Asiltürk and Çunkaş [18] | AISI 1040 steel | Dry turning | Coated carbide (TiCN, Al_{2}O_{3}) inserts | Speed, feed and depth of cut | Multiple regression and ANN with different training methods | Mean absolute percentage error and the determination coefficient (R^{2}) | R^{2} for multiple regression models is 98.9%. The proposed model of ANN utilized successfully to predict the surface roughness compared to multiple regression |
Karayel [17] | St. 50.2 steel | Turning | Carbide inserts | Speed, feed and depth of cut | ANN | Average absolute error | Surface roughness can be controlled by the designed control circuit to certain accuracy by altering the cutting parameters |
Furthermore, researchers have used cutting parameters along with cutting forces, tool vibrations, etc. as input for estimation of R_{a} [23, 24, 25]. They used multiple regression and ANN techniques in machining of different alloys. It is found that using ANN approach, R_{a} is predicted more accurately than regression analysis when responses are combined with cutting parameters. They also concluded that responses are useful for process monitoring.
As reported by Deshpande et al. [2], attaining good finish is very tough for Inconel 718. Hence, authors have reported very extensive statistical study for estimating regression-based R_{a} models. Due to enormous statistical calculation of used method [2], it was thought to apply some pattern classifier technique such as ANN. ANN is a soft computing method commonly used in several applications such as forecasting, data control and recognition of pattern. ANN is an effective tool used for estimation of response factors by considering its easiness, speed and learning ability [18]. Therefore, to confirm the results obtained by Deshpande et al. [2], it was decided to use ANN to the experimental data.
ANN consists of input, hidden and output layers. The input and output layers are defined as nodes, and the hidden layer provides an association amongst them. The accuracy and efficiency of ANN structure depend on network pattern, training algorithm, training data pattern, learning rate, processing function, testing data pattern and output data representation [16, 26]. The network gives the sets of patterns to be learned and the preferred system response for each pattern. It can be reconstructed to obtain minimum error [27]. Risbood et al. [24] and Upadhyay et al. [25] have used LM algorithm along with back propagation to train the single hidden layer network for prediction of R_{a}. The multi-layer perceptron (MLP) network requires less number of data points for training and testing of the network. Therefore, researchers have used MLP neural network in almost all machining processes. They found efficient performance of network model [24, 28, 29]. Risbood et al. [24] used neural network to predict R_{a} in turning using data of 26 experiments for training and testing. Tamang and Chandrasekaran [28] used data of 22 experiments for training and 5 for validation. Kohli and Dixit [29] have proposed an MLP neural network to predict R_{a} in turning process using data of 30 experiments for training and testing, and they predicted acceptable results. According to Hagan and Menhaj [30], the LM algorithm runs more rapidly when it trains with the feed forward neural network.
However, there is no study found on estimation of R_{a} using ANN approach in machining of Inconel 718 by combining input and response parameters. Hence, the goal of this work is to employ ANN for prediction of R_{a} in turning of Inconel 718 and compare the results with previously published results of regression analysis [2].
2 Experimentation
Inconel 718 bars, 120 mm long with 22 mm diameter, were machined for uncoated with untreated (UT) and cryogenically treated (CT) tungsten carbide (WC) inserts using minimum quantity lubrication (MQL) along with selected cutting conditions. Authors have already reported the details of experimentation in previous study [2].
Input design matrix and measured responses [2]
Run order | v (m/min) | f (mm/rev) | d (mm) | (Untreated/cryo-treated) | |||
---|---|---|---|---|---|---|---|
F_{c} (N) | S (Pa) | V_{v} × 10^{−4} (m/s) | R_{a} (μm) | ||||
1 | 90 | 0.18 | 1.07 | 200/190 | 0.270/0.260 | 13/12 | 1.78/1.79 |
2 | 9.5 | 0.115 | 0.785 | 617/602 | 0.794/0.780 | 37/36 | 2.30/2.40 |
3 | 30 | 0.18 | 0.5 | 463/456 | 0.531/0.633 | 33/23 | 2.20/2.20 |
4 | 60 | 0.01 | 0.785 | 231/225 | 0.210/0.209 | 16/15 | 0.99/1.00 |
5 | 60 | 0.115 | 0.785 | 455/445 | 0.472/0.462 | 26/25 | 0.93/0.99 |
6 | 60 | 0.115 | 1.264 | 404/390 | 0.501/0.487 | 23/24 | 1.49/1.80 |
7 | 110.45 | 0.115 | 0.785 | 163/160 | 0.190/0.189 | 13/13 | 0.62/0.60 |
8 | 30 | 0.18 | 1.07 | 590/580 | 0.791/0.781 | 36/35 | 2.20/2.30 |
9 | 60 | 0.115 | 0.785 | 459/551 | 0.660/0.652 | 27/25 | 0.93/0.98 |
10 | 60 | 0.115 | 0.785 | 461/555 | 0.691/0.686 | 29/26 | 0.99/0.95 |
11 | 90 | 0.18 | 0.5 | 279/270 | 0.130/0.123 | 19/17 | 1.25/1.21 |
12 | 90 | 0.05 | 1.07 | 190/185 | 0.198/0.192 | 12/11 | 0.52/0.75 |
13 | 60 | 0.224 | 0.785 | 444/440 | 0.509/0.491 | 32/30 | 2.40/2.50 |
14 | 60 | 0.115 | 0.785 | 464/461 | 0.521/0.511 | 28/27 | 0.99/1.10 |
15 | 30 | 0.05 | 1.07 | 447/442 | 0.422/0.412 | 29/28 | 1.60/1.60 |
16 | 30 | 0.05 | 0.5 | 436/432 | 0.421/0.419 | 27/26 | 1.90/1.36 |
17 | 60 | 0.115 | 0.785 | 471/476 | 0.567/0.559 | 28/26 | 0.92/1.10 |
18 | 60 | 0.115 | 0.305 | 300/295 | 0.321/0.312 | 26/24 | 1.30/1.35 |
19 | 60 | 0.115 | 0.785 | 451/445 | 0.563/0.558 | 30/29 | 0.93/1.17 |
20 | 90 | 0.05 | 0.5 | 150/147 | 0.161/0.158 | 12/11 | 0.72/0.76 |
3 Application of ANN
In this study, a multi-layer feed forward ANN structure trained with an error BP algorithm is employed to estimate R_{a}. The various ANN parameters have been selected by referring research papers because no standard procedure is found to define the architecture. Similarly, the number of neurons in hidden layer is obtained by trial-and-error approach by inspecting and comparing different architectures [15, 31]. ANN models are developed using function ‘nntool’ in MATLAB R2015a. 5-h-1 network pattern is used for untreated inserts with v, f, F_{c}, S, V_{v} as inputs and R_{a} as output. In the previous study [2], it was reported that Pearson correlation analysis (PCA) for depth of cut was found to be very close to zero; which indicated insignificant effect on R_{a}. So, depth of cut was removed in the case of untreated inserts. However, significant effect of depth of cut on R_{a} was found in the case of treated inserts. Therefore, 6-h-1 pattern is used which includes v, f, d, F_{c}, S, V_{v} as inputs and R_{a} as output for the cryo-treated inserts.
With the aim, to improve the estimation accuracy of R_{a} in ANN, a total data of 150 sets (75 sets for untreated and 75 sets for treated) is generated using cutting parameters other than used in main experiment but within the range (Table 2). The data are obtained by varying one input parameter and keeping mid-values of other two input parameters constant and vice versa. Therefore, for untreated inserts, the estimated data of force, sound and vibration are obtained by placing cutting parameters in corresponding response equations published in Deshpande et al. [2]. R_{a} is estimated by inserting values of cutting parameters besides force, sound and vibration in Eq. (1). Similarly, data are generated for treated inserts, and the R_{a} is estimated using Eq. (2).
ANN training parameters for untreated and cryo-treated inserts
Parameters | Untreated inserts | Cryo-treated inserts |
---|---|---|
Neurons in the single hidden layer (h) | 12 numbers | 12 numbers |
Epochs (frequency of progress displays) | 4 iterations | 8 iterations |
Selection of maximum epochs to train | 1000 numbers | 1000 numbers |
Maximum validation set | 100 numbers | 100 numbers |
Sum-squared error goal | 1 × 10^{−2} | 1 × 10^{−2} |
All R^{2} values | 98.76% | 98.91% |
Performance of ANN in prediction of surface roughness
v (m/min) | f (mm/rev) | d (mm) | (Untreated/cryo-treated) | ||||||
---|---|---|---|---|---|---|---|---|---|
F_{c} (N) | S (Pa) | V_{v} × 10^{−4} (m/s) | R_{a} (μm) | ANN | |||||
\(\hat{R}_{\text{a}}\) (μm) | AE (%) | SE (%) | |||||||
10 | 0.12 | 0.79 | 600/698 | 0.74/0.77 | 35/36 | 2.20/1.98 | 2.15/1.82 | 2.27/8.08 | 0.25/2.56 |
60 | 0.12 | 0.79 | 450/410 | 0.455/0.52 | 26/29 | 1.30/1.32 | 1.16/1.11 | 10.77/15.91 | 1.96/4.41 |
80 | 0.10 | 0.70 | 275/290 | 0.294/0.30 | 19/18 | 1.00/1.10 | 1.02/0.98 | 2/10.91 | 0.04/1.44 |
110 | 0.12 | 0.79 | 158/149 | 0.215/0.14 | 11/27 | 0.60/0.83 | 0.66/0.71 | − 10.00/14.46 | 0.36/1.44 |
60 | 0.22 | 0.79 | 445/436 | 0.51/0.25 | 32/20 | 2.10/1.77 | 1.92/1.83 | 8.57/− 3.39 | 3.24/0.36 |
45 | 0.11 | 0.78 | 400/500 | 0.59/0.60 | 28/25 | 1.80/1.50 | 1.72/1.63 | 4/8.67 | 0.64/1.69 |
60 | 0.12 | 1.26 | 398/377 | 0.49/0.27 | 29/28 | 1.30/1.56 | 1.47/1.60 | − 13.08/− 2.56 | 2.89/0.16 |
Mean | 7.30/9.14 | 1.34/1.72 |
4 Model comparisons
Comparison of the models
Inserts type | Modified multiple regression model (FOE_{M}) | ANN model | ||||||
---|---|---|---|---|---|---|---|---|
Main expt. data | Confirmation test | Training all R^{2} (%) | Confirmation test | |||||
R^{2} (%) | MAE (%) | MSE (%) | MAE (%) | MSE (%) | MAE (%) | MSE (%) | ||
Untreated | 91.5 | 9.82 | 2.88 | 8.37 | 1.69 | 98.76 | 7.30 | 1.34 |
Treated | 94.8 | 8.29 | 1.69 | 10.71 | 2.18 | 98.91 | 9.14 | 1.72 |
5 Conclusions
The paper has presented multiple regression and ANN-based models by considering untreated and cryogenically treated carbide inserts for turning of Inconel 718. These models revealed different degrees of fitness. Therefore, to recommend the best model for estimation of R_{a}, analysis of the estimated results and related error investigation has been performed. In the previous study, authors have established multiple regression analysis to estimate R_{a} with desirable accuracy. R^{2} values of the modified regression model were obtained as 91.5% and 94.8%, and it indicates that the models can describe 91.5% and 94.8% of whole deviations in R_{a} for the untreated and treated inserts, respectively. Therefore, it is reported in the earlier study that using combination of specific cutting and response parameters, a good association was established amongst estimated and experimental data.
ANN prediction tool is proposed to test established regression models, 150 data sets are used for two types of inserts. ANN was trained and tested using combination of cutting and response parameters. All R^{2} values are found as 98.76% and 98.91% for untreated and treated inserts, respectively, which indicate very fine association and good fit for both types of inserts. Improvement in R^{2} values for training of neural network model for untreated and treated inserts indicates that ANN models are better for prediction than regression models. Furthermore, the regression plots of ANN confirm the fine association of fit between measured and estimated R_{a}. ANN performance was validated with independent data, which showed percentages of MAE are 7.30% and 9.14% for untreated and treated inserts, respectively. Similarly, percentages of MSE are 1.34% and 1.72% for the two cases. Therefore, the developed ANN models can be successfully used for estimation of in-process R_{a}.
It is found that modified regression model estimated R_{a} with more than 90%, whereas ANN is estimating roughness more than 98%, based on R^{2} values. Hence, it can be concluded that ANN is a better prediction tool for in-process monitoring of R_{a} for both the types of inserts. Such predictions can be useful for real-time control of the process to acquire the desired R_{a}.
6 Summary
In present research work, authors have proposed ANN modelling technique in turning of Inconel 718 (nickel-based supperalloy) using cryogenically treated and untreated carbide inserts. In this work, an attempt has been made to model, examine and compare the results of previously reported study, published by same authors with proposed ANN application.
The results of regression modelling technique are compared and analyzed using ANN application. The results obtained by ANN modelling using cutting parameters, force, sound and vibration are discussed for estimation of surface quality for treated and untreated inserts. This is the novelty of the article. Moreover, these results can be helpful to the machinist to attain good surface texture in finish cut for precision machining in modern era.
Notes
Acknowledgements
This research work was carried out within the scheme of Technical Education Quality Improvement Program, Phase II (TEQIP-II), VNIT, Nagpur, and financially supported with the assistance of World Bank under Ministry of Human Resource Development (MHRD), Government of India, New Delhi, India.
Compliance with ethical standards
Conflict of interest
The authors declare that they have no conflict of interest.
References
- 1.Blau PJ (2008) Friction science and technology: from concepts to applications. CRC Press, Boca RatonCrossRefGoogle Scholar
- 2.Deshpande Y, Andhare A, Sahu N (2017) Estimation of surface roughness using cutting parameters, force, sound, and vibration in turning of Inconel 718. J Braz Soc Mech Sci Eng. https://doi.org/10.1007/s40430-017-0819-4 CrossRefGoogle Scholar
- 3.Markos S, Viharos ZJ, Monostori L (1998) Quality-oriented, comprehensive modelling of machining processes. In: Sixth ISMQC IMEKO symposium on metrology for quality control in production, pp 67–74Google Scholar
- 4.Fang XD, Safi-Jahanshahi H (1997) A new algorithm for developing a reference-based model for predicting surface roughness in finish machining of steels. Int J Prod Res 35(1):179–199zbMATHCrossRefGoogle Scholar
- 5.Wang H, Li D (2002) Surface roughness prediction model for ultraprecision turning aluminium alloy with a single crystal diamond tool. Chin J Mech Eng (Engl Ed) 15(2):153–156CrossRefGoogle Scholar
- 6.Whitehouse DJ (1994) Handbook of surface metrology. Taylor & Francis, LondonGoogle Scholar
- 7.Pollock TM, Tin S (2006) Nickel-based superalloys for advanced turbine engines: chemistry, microstructure and properties. J Propul Power 22(2):361–374CrossRefGoogle Scholar
- 8.Ulutan D, Ozel T (2011) Machining induced surface integrity in titanium and nickel alloys: a review. Int J Mach Tools Manuf 51(3):250–280CrossRefGoogle Scholar
- 9.Cantero JL, Díaz-Álvarez J, Miguélez MH, Marín NC (2013) Analysis of tool wear patterns in finishing turning of Inconel 718. Wear 297(1):885–894CrossRefGoogle Scholar
- 10.Lu X, Wang F, Wang X, Lu Y, Si L (2017) A surface roughness prediction model using response surface methodology in micro-milling Inconel 718. Int J Mach Mach Mater 19(3):230–245Google Scholar
- 11.Olovsjö S, Hammersberg P, Avdovic P, Ståhl J-E, Nyborg L (2012) Methodology for evaluating effects of material characteristics on machinability—theory and statistics-based modelling applied on Alloy 718. Int J Adv Manuf Technol 59(1–4):55–66CrossRefGoogle Scholar
- 12.Zhou J, Bushlya V, Avdovic P, Ståhl JE (2012) Study of surface quality in high speed turning of Inconel 718 with uncoated and coated CBN tools. Int J Adv Manuf Technol 58(1):141–151CrossRefGoogle Scholar
- 13.Bhardwaj B, Kumar R, Singh PK (2014) Prediction of surface roughness in turning of EN 353 using response surface methodology. Trans Indian Inst Met 67(3):305–313. https://doi.org/10.1007/s12666-013-0346-7 CrossRefGoogle Scholar
- 14.Davoodi B, Tazehkandi AH (2014) Cutting forces and surface roughness in wet machining of Inconel alloy 738 with coated carbide tool. Proc Inst Mech Eng Part B J Eng Manuf. https://doi.org/10.1177/0954405414542990 CrossRefGoogle Scholar
- 15.Pontes FJ, Silva MB, Ferreira JR, Paiva APd, Balestrassi PP, Schönhorst GB (2010) A DOE based approach for the design of RBF artificial neural networks applied to prediction of surface roughness in AISI 52100 hardened steel turning. J Braz Soc Mech Sci Eng 32:503–510CrossRefGoogle Scholar
- 16.Mia M, Dhar NR (2016) Prediction of surface roughness in hard turning under high pressure coolant using Artificial Neural Network. Measurement 92:464–474CrossRefGoogle Scholar
- 17.Karayel D (2009) Prediction and control of surface roughness in CNC lathe using artificial neural network. J Mater Process Technol 209(7):3125–3137CrossRefGoogle Scholar
- 18.Asiltürk I, Çunkaş M (2011) Modeling and prediction of surface roughness in turning operations using artificial neural network and multiple regression method. Expert Syst Appl 38(5):5826–5832CrossRefGoogle Scholar
- 19.Bobbili R, Madhu V, Gogia AK (2015) Modelling and analysis of material removal rate and surface roughness in wire-cut EDM of armour materials. Eng Sci Technol Int J 18(4):664–668CrossRefGoogle Scholar
- 20.Ezilarasan C, Kumar VSS, Velayudham A, Palanikumar K (2011) Modeling and analysis of surface roughness on machining of Nimonic C-263 alloy by PVD coated carbide insert. Trans Nonferrous Met Soc China 21(9):1986–1994CrossRefGoogle Scholar
- 21.Kumar S, Singh R, Batish A, Singh TP (2015) Modeling the tool wear rate in powder mixed electro-discharge machining of titanium alloys using dimensional analysis of cryogenically treated electrodes and workpiece. Proc Inst Mech Eng Part E J Process Mech Eng. https://doi.org/10.1177/0954408915593875 CrossRefGoogle Scholar
- 22.Tsai K-M, Wang P-J (2001) Semi-empirical model of surface finish on electrical discharge machining. Int J Mach Tools Manuf 41(10):1455–1477CrossRefGoogle Scholar
- 23.Özel T, Karpat Y (2005) Predictive modeling of surface roughness and tool wear in hard turning using regression and neural networks. Int J Mach Tools Manuf 45(4):467–479CrossRefGoogle Scholar
- 24.Risbood KA, Dixit US, Sahasrabudhe AD (2003) Prediction of surface roughness and dimensional deviation by measuring cutting forces and vibrations in turning process. J Mater Process Technol 132(1):203–214CrossRefGoogle Scholar
- 25.Upadhyay V, Jain PK, Mehta NK (2013) In-process prediction of surface roughness in turning of Ti–6Al–4V alloy using cutting parameters and vibration signals. Measurement 46(1):154–160. https://doi.org/10.1016/j.measurement.2012.06.002 CrossRefGoogle Scholar
- 26.Karkalos NE, Galanis NI, Markopoulos AP (2016) Surface roughness prediction for the milling of Ti–6Al–4V ELI alloy with the use of statistical and soft computing techniques. Measurement 90:25–35. https://doi.org/10.1016/j.measurement.2016.04.039 CrossRefGoogle Scholar
- 27.Kermanshahi B, Iwamiya H (2002) Up to year 2020 load forecasting using neural nets. Int J Electr Power Energy Syst 24(9):789–797CrossRefGoogle Scholar
- 28.Tamang SK, Chandrasekaran M (2016) Integrated optimization methodology for intelligent machining of Inconel 825 and its shop-floor application. J Braz Soc Mech Sci Eng 39:1–13Google Scholar
- 29.Kohli A, Dixit US (2005) A neural-network-based methodology for the prediction of surface roughness in a turning process. Int J Adv Manuf Technol 25(1–2):118–129CrossRefGoogle Scholar
- 30.Hagan MT, Menhaj MB (1994) Training feedforward networks with the Marquardt algorithm. IEEE Trans Neural Netw 5(6):989–993CrossRefGoogle Scholar
- 31.Madić M, Radovanović M (2013) Modeling and analysis of correlations between cutting parameters and cutting force components in turning AISI 1043 steel using ANN. J Braz Soc Mech Sci Eng 35(2):111–121CrossRefGoogle Scholar