ANN Modeling of Cutting Performances in Spray Cooling Assisted Hard Turning

ANN Modeling of Cutting Performances in Spray Cooling Assisted Hard Turning

Available online at www.sciencedirect.com ScienceDirect Materials Today: Proceedings 5 (2018) 18482–18488 www.materialstoday.com/proceedings ICMPC_...

708KB Sizes 0 Downloads 2 Views

Available online at www.sciencedirect.com

ScienceDirect Materials Today: Proceedings 5 (2018) 18482–18488

www.materialstoday.com/proceedings

ICMPC_2018

ANN Modeling of Cutting Performances in Spray Cooling Assisted Hard Turning Ramanuj Kumar*, Ashok Kumar Sahoo, Purna Chandra Mishra, Rabin Kumar Das, Soumikh Roy School of Mechanical Engineering, Kalinga Institute of Industrial Technology, Deemed to be University, Bhubaneswar, Odisha, India

Abstract

Current work focused on prediction of wear at flank face and chip reduction coefficient in finish hard turning of heat treated AISI D2 (55 ± 1) HRC steel using uncoated tungsten carbide insert in spray cooling. Experimental data of flank wear (VBc) and chip reduction coefficient (CRC) are elaborated through surface plot. Speed is the most dominant variables for wear at flank face whereas depth-of-cut and speed acts as most valuable variables to affect chip reduction coefficient. Artificial neural network (ANN) concept is being implemented to predict flank wear and chip reduction coefficient. Feed forward back propagation network using Levenberg-Marquardt (L-M) algorithm is implemented for training the result data. Three network architectures namely 3-6-2, 3-7-2, and 3-8-2 are introduced for modelling purpose. The obtained modeling results are compared in terms of R-sq and average percentage error between actual and predicted results. Greatest R-Square value (0.9986) and minimum absolute error (VBc =0.9659 % and CRC = 0.3474 %) between experimental and model are found with 3-6-2 architecture when it was trained with 100000 epoch value. © 2018 Elsevier Ltd. All rights reserved. Selection and/or Peer-review under responsibility of Materials Processing and characterization. Keywords:Flank wear; Chip reduction coefficient; Hard turning; Spray cooling; ANN modelling

1. Introduction The cutting process concerning removal of material in the form of chips from discrete surface of the workpiece is termed as machining. The machining of hardened material above 45 HRC which is termed as hard machining has gained immense popularity in the field of machining as it has successfully replaced conventional grinding mechanism in various aspect of machinability [1]. *Corresponding author. Tel.: +0674 6540805 E-mail address: [email protected]

2214-7853© 2018 Elsevier Ltd. All rights reserved. Selection and/or Peer-review under responsibility of Materials Processing and characterization.

Kumar et al. / Materials Today: Proceedings 5 (2018) 18482–18488

18483

Use of coolant in proper procedure with the selected cutting insert can bring about compelling aid in terms of finish of the turned surface, wear at tools surface and life of tool [2]. In recent years, hard machining has been carried under various lubrication/cooling techniques in order to minimize the cutting temperature evolved during cutting process. Another vital role of lubrication is to flush away the chips from the cutting zone which results in providing better surface quality and clean machining environment. Various works were reported on different categories of cooling or lubricating techniques namely MQL, Mist lubrication, solid lubrication, compressed air cooling, liquid nitrogen in hard machining applications [3, 4]. Spray impingement cooling is a novel technique utilized in turning process and traced to be beneficial performance over dry cutting [5, 6]. To attain enhanced finish of turned surface and extended life of insert in hard turning relay on various appealing aspects such as machining time, axial speed, feed, depth-of-cut, job material selection, proper lubricant and discrete cooling techniques which could be selected prior to machining procedure. Literatures shows that the potential to do machining with cBN, PCBN, ceramic, cermet, PVD and CVD coated carbide tool is favorable but still invention are going on for low budget uncoated carbide tool in different cutting environments [4, 7, 8]. Numerous experimental trails and explorations has been conceded by researchers across the globe to figure out the inputs of factors on various machinability conditions, regarding wear at tool surface, cutting forces and surface integrity. The forecast of wear at tool surface, forces, temperature and finish of turned surface has been attain by utilizing different techniques like ANN, RSM, Fuzzy logic, etc. Artificial neural network (ANN) has utilized widely by numerous researcher for prediction of cutting responses due to its high rate of accuracy as compared to other soft computing techniques in hard machining. Utilizing neural networks a model is defined for every inputs variables and machining variables, which is then correlated for their forecast potential with the definite data’s [9]. Statistical and ANN prediction capability in hard turning of D2 steel was carried and ANN prediction for tool wear was traced to be more appropriate and favorable [10]. ANN modeling for forecasting tool wear as well as surface roughness in hard turning of D2 steel was carried and noticed very precise predicted results [11]. ANN coupled with image processing technique has been implemented to predict tool life in turning process. The outcomes revealed that the image processing software and ANN might be implementing in industries economically for tool life estimation [12]. Based on average absolute error in percentage, the capability of ANN based model was superior over RSM based prediction and displayed a strong bonding between input and output variables [13]. Based on literature studied, application of uncoated carbide insert under spray impingement cooling in turning of hardened steel (above 50 HRC) is lacking or almost nil. ANN modeling tool is very efficient to predict the turning responses. Majority of works reported the modeling of wear at tool surface, roughness of turned surface, cutting forces and cutting temperature in turning of hardened material. Modeling of chip reduction coefficient is not yet done using ANN. Mostly 3-2n-1 ANN architecture has been found in previous works. Very few works utilized 3-2n2 type of ANN architecture for predicting the responses in hard turning. However in this work, ANN modeling using different number of hidden layers with different epoch values is carried. 2. Experimentation details and procedure Cylindrical bar of AISI D2 (55±1 HRC) tool steel with machining length 200 mm and diameter 48 mm has been utilized for current experimental investigation. The three input variables namely speed (63-182 m/min), feed (0.04 0.16 mm/rev) and depth-of-cut (0.1-0.4 mm) are selected. Commercially available uncoated tungsten carbide insert is implemented to perform the turning operation. The insert geometry of ISO designation CNMG 120408 firmly clamped into aright handed tool holder of ISO designation PCLNR 2525 M12 with approach angle 950 and tip nose radius 0.8 mm. The entire experiments have been accomplished on a HMT made NH22 semi-automatic lathe of maximum spindle speed 2040 rpm under air-water mixed spray cooling surrounding as shown in Fig. 1. The compressed air with 1.5 bar of static pressure and 1 bar of water static pressure was utilized and this process is controlled by spraying machine. Spray nozzle has been placed vertically about 20 cm distance from cutting zone as shown in Fig. 1. Flank wear (VBc) are measured by Olympus made optical microscope. Digital caliper is used to measure chip thickness. Minitab-16 software has been utilized for 3 dimensional surface plots. Matlab R2013a is used for artificial neural network (ANN) modeling. After experiments, flank wear and chip thickness have been measured. Chip reduction coefficient (CRC) for each run has been computed by taking the ratio of chip thickness to uncut chip thickness [14].

18484

Kumar et al. / Materials Today: Proceedings 5 (2018) 18482–18488

AISI D2 steel

Cutting insert Fig. 1. Hard turning experimentation setup 3. Results morphology using surface plots 3.1 Flank wear analysis Impact of input variable specifically speed, feed and depth-of-cut on wear at flank face are analysed using 3dimensional surface plot as shown in Fig. 2 (a-f). From the surface plots it has been noticed that the wear at flank face enhances with rising speed, feed and depth-of-cut. Speed and depth-of-cut are more dominating term compare to feed for flank wear growth. . b

a

Surface Plot of VBc vs f, d

Surface Plot of VBc vs d, f

1.5 VBc 1.0 0.5 0.0 0.05 0.10 0.15 f

1.5 VBc 1.0 0.5 0.0 0.1 0.2 0.3 0.4 d

0.4 0.3 0.2 d 0.1

d

c

1.5 VBc 1.0 0.5 0.0 0.05 0.10 0.15 f

150 100 v 50

1.5 VBc 1.0 0.5 0.0 50

0.15 0.10 f 0.05

e Surface Plot of VBc vs v, f

Surface Plot of VBc vs f, v

100 150 v

0.15 0.10 f 0.05

f Surface Plot of VBc vs v, d 1.5 VBc 1.0 0.5 0.0 0.1 0.2 0.3 0.4 d

150 100 v 50

Fig. 2. (a-f) 3D surface plot for VBc

Surface Plot of VBc vs d, v 1.5 VBc 1.0 0.5 0.0 50

100 150 v

0.4 0.3 0.2 d 0.1

Kumar et al. / Materials Today: Proceedings 5 (2018) 18482–18488

18485

3.2 Chip-reduction coefficient analysis In hard machining, analysis of obtained metal chips with respect to variations in cutting parameters is highly needful to understand the overall machining performance. Chip reduction coefficient (CRC) is the proportion of chip thickness and uncut chip thickness and it indicates the machining forces behaviour during cutting process [14]. Herein, Chip reduction coefficient was analysed through surface plot as shown in Fig. 3 (a-f). From the 3D surface plot, it is clearly visible that with increase in feed and speed the CRC decreases subsequently whereas with increment in depth of cut the CRC rises. All the involved cutting variables significantly affects the CRC and depth of cut and feed are more dominant compare to speed. a

b

2.0

CRC

1.8 1.6 0.05 0.10 0.15 f

0.4 0.3 0.2 d 0.1

d

2.0 1.8 1.6 0.1 0.2 0.3 0.4 d

CRC

0.15 0.10 f 0.05

e Surface Plot of CRC vs v, f

CRC

2.0 1.6 0.05 0.10 0.15 f

150 100 v 50

2.0 1.8 1.6 50

100 v

0.15 0.10 f 0.05

150

f Surface Plot of CRC vs v, d

CRC

1.8

Surface Plot of CRC vs f, v

Surface Plot of CRC vs f, d

Surface Plot of CRC vs d, f

CRC

c

2.0

CRC

1.8 1.6 0.1 0.2 0.3 0.4 d

Surface Plot of CRC vs d, v

150 100 v 50

2.0 1.8 1.6 50

100 v

150

0.4 0.3 0.2 d 0.1

Fig. 3. (a-f) 3D surface plot for CRC

4. Artificial neural network modeling Neural network modeling works on human brain structure. It has three layers consisting of one or more neurons are interlinked with each other. First layer consists of input neurons which receives input numerical data. Here one neuron receives one variable values. Second layer as well termed as hidden layer resides of no. of neurons which gathers the information from the input layer. Here neuron of input layer is interlinked with each neuron of hidden layer. Further third layer also known as output layer which is interlinked with hidden layer by some defined weights and produced the output results. Amount of neurons in output layer is equivalent to number of response output. Modelling accuracy capability through ANN depends on training algorithm, type of architecture, number of hidden layers, weight factor, training variables like number of epoch, error tolerance, noise factor, slope parameter, learning rate, momentum factor etc. From the literature studied, the recommended neurons in hidden layer can be n/2, n, 2n, 2n+1, 2n+2 where n is no. of input variables or neurons [2, 15, 16]. In the current modelling work, Feed forward back propagation network using Levenberg-Marquardt (L-M) algorithm is implemented for training the result data. In hidden layer, three types 2n, 2n+1, and 2n+2 neurons are considered. However the three different network architectures namely 3-6-2 (Fig. 4a), 3-7-2 (Fig. 4b) and 3-8-2

18486

Kumar et al. / Materials Today: Proceedings 5 (2018) 18482–18488

(Fig. 4c) are introduced for modelling purpose. Three different epoch values (100000, 150000 and 200000) are utilized for all ANN architectures. b

a

c

Fig. 4. (a) 3-6-2 ANN architecture (b) 3-7-2 ANN architecture (c) 3-8-2 ANN architecture

ANN toolbox present in Matlab was utilized for correctly renewing the weights magnitude and biases of the Levenberg-Marquardt algorithm. All three ANN architectures have been trained for a constant number of cycles [17]. Predicted response results were generated with use of simulates command. The performances of predicted results are justified based on R-Square and average absolute error between actual and predicted data. After training networks the R-Square value are analysed based on 95 % of confidence level. The R-Square for all three ANN architectures are listed in Table 1. All R-Square results are more than 0.95 however the established models are significant. Similarly all R-Square values are very close to unity which confirmed that all the accomplished models are fit to predicting the cutting responses accurately. From Table 1, highest R-Square value among all conditions is noticed as 0.9986 at 3-6-2 network architecture with 100000 Epoch value and the interrelationship graph between experimental data and model data is displayed in Fig. 5. Table 1. R-square value of established models Epoch numbers 100000 150000 200000

3-6-2 0.9986 0.9903 0.9938

ANN architecture 3-7-2 0.9912 0.9887 0.9886

3-8-2 0.9958 0.9868 0.9888

The actual experimental response data have compared with the model result data and absolute average percentage error has been calculated using Eq. 1. Absolute average percentage error = ∑|

Experimental data

fitted data /Experimental data 100 |

Where, n denotes number of test runs.

(1)

Kumar et al. / Materials Today: Proceedings 5 (2018) 18482–18488

Data Fit Y=T

1.5

1

0.5

0.5

Output ~= 1*Target + -6.2e-05

Output ~= 0.99*Target + 0.028

2

Validation: R=0.99923

2

1

1.5

Data Fit Y=T

1.5

1

0.5

2

0.5

1

1.5

Target

Test: R=1

All: R=0.99932

Data Fit Y=T

1.5

1

0.5

0.5

2

Target

Output ~= 1*Target + 0.0059

Output ~= 1*Target + 0.0061

Training: R=0.99911

18487

1

Target

1.5

2

2

2

Data Fit Y=T

1.5

1

0.5

0.5

1

1.5

2

Target

Fig. 5.Interrelationship between experimental and model values for 3-6-2 ANN architecture at 100000 epoch value

Table 2. Absolute average error between experimental and model data Absolute average error in percentage Epoch numbers 100000 150000 ANN architecture VBc CRC VBc CRC 3-6-2 0.9659 0.3474 4.9009 1.5680 3-7-2 4.8781 1.3642 1.7289 0.9163 3-8-2 7.0534 1.4340 8.0862 1.5544

200000 VBc 2.5086 6.3140 8.2732

CRC 1.5970 1.6995 1.4365

From Table 2, least error for Flank wear (VBc) and chip reduction coefficient (CRC) is noticed to be 0.9659 % and 0.3474 % respectively. Similarly highest error for VBc and CRC is noticed as 8.2732 % and 1.6995 %. Least error represents highest efficiency to predict the response result whereas highest error represents worst capability to predict the responses among considered ANN architecture and epoch numbers. Least error for VBc and CRC are noticed when result data is trained by using 3-6-2 ANN architecture along with 100000 epoch value. However for accuracy and efficiency aspects, 3-6-2 ANN architecture with 100000 epoch values are more favourable than 3-7-2 and 3-8-2 architectures.

18488

Kumar et al. / Materials Today: Proceedings 5 (2018) 18482–18488

5. Conclusions The current work emphasized on application of ANN modeling to predict flank wear and chip-reduction coefficient with use of uncoated carbide insert in turning of heat treated AISI D2 grade steel (55 ± 1) HRC with spray cooling surrounding. The following outcomes have been made as follows:  Surface plots revealed that the wear at flank face gradually improve by means of rise in speed and depth-ofcut. Similarly chip reduction coefficient reduces with rise in speed and cutting feed whereas it improves with rise in depth of cut.  R-Square value for all ANN architecture is higher and significant at 95 % of level of confidence. Highest R-Square value (0.9986) was obtained with use of 3-6-2 architecture and 100000 epoch values.  Absolute average error between experimental data and model data for all ANN architectures at different epoch numbers has been estimated and lowest error is found with use of 3-6-2 ANN architecture with 100000 epoch values.  Overall application of 3-6-2 ANN architecture with 100000 epoch number is most favourable compare to 3-7-2 and 3-8-2 ANN architecture. Acknowledgement The authors express their gratitude and thank to Kalinga Institute of Industrial Technology, Deemed to be University, Bhubaneswar, Odisha, India for providing the experimental facility to execute the research work. Reference [1] S.A. Khan, M.A. Ahmad, M.Q. Saleem, Z. Ghulam, M.A.M. Qureshi, High Feed Turning of AISI D2 Tool Steel Using Multi Radii Tool Inserts: Tool Life, Material Removed and Workpiece Surface Integrity Evaluation, Materials and Manufacturing Processes 32(6) (2017) 670677 [2] M. Mia, N.R. Dhar, Prediction of surface roughness in hard turning under high pressure coolant using Artificial Neural Network, Measurement 92 (2016) 464–474. [3] P.J. Liew A. Shaaroni N.A.C. Sidik, J. Yan, An overview of current status of cutting fluids and cooling techniques of turning hard steel, International Journal of Heat and Mass Transfer 114 (2017) 380–394. [4] S. Chinchanikar, S.K. Choudhury, Machining of hardened steel—Experimental investigations, performance modelling and cooling techniques: A review, International Journal of Machine Tools and Manufacture 89 (2015) 95–109. [5] S.K. Sahu, P. C. Mishra, K. Orra, A.K. Sahoo, Performance assessment in hard turning of AISI 1015 steel under spray impingement cooling and dry environment, Proceeding Institution of Mechanical Engineering Part B: Journal of Engineering Manufacture 229(2) (2014) 251– 265. [6]P.C. Mishra, D.K. Das, S.K. Sahu, comparative performance in hard turning of AISI 1015 steel with carbide insert using orthogonal array design and grey relational analysis under spray impingement cooling and dry environment: A case study, International Journal of Manufacturing Materials and Mechanical Engineering 4(3) (2014) 1-32. [7] J. Sharma, B.S. Sidhu, Investigation of effects of dry and near dry machining on AISI D2 steel using vegetable oil, Journal of Cleaner Production 66 (2014) 619–623. [8] A. Das, A. Mukhopadhyay, S. K. Patel, B. B. Biswal, Comparative Assessment on Machinability Aspects of AISI 4340 Alloy Steel Using Uncoated Carbide and Coated Cermet Inserts During Hard Turning, Arabian Journal for Science and Engineering 41 (2016) 4531–4552. [9] V.S. Sharma, S. Dhiman, R. Sehgal, S. Sharma, Estimation of cutting forces and surface roughness for hard turning using neural networks, Journal of Intelligent Manufacturing 19 (2008) 473–483. [10] R. Quiza, L. Figueira, J.P. Davim, Comparing statistical models and artificial neural networks on predicting the tool wear in hard machining D2 AISI steel, International Journal of Advance Manufacturing Technology 37 (2008) 641–648. [11] T. Özel, Y. Karpat, L. Figueira and J. P Davim, Modelling of surface finish and tool flank wear in turningof AISI D2 steel with ceramic wiper inserts, Journal of Materials Processing Technology 189 (2007) 192–198. [12] T. Mikolajczyk, K. Nowicki, A. Bustillo, D. Yu Pimenov, Predicting tool life in turning operations using neural networks and image processing, Mechanical Systems and Signal Processing 104 (2018) 503–513. [13] M. Mia, N. R. Dhar, Response surface and neural network based predictive models of cutting temperature in hard turning, Journal of Advanced Research 7 (2016) 1035–1044. [14] A.B. Chattopadhyay, Machining and Machine Tools, 2nd Edition, Willey India Pvt. Ltd. [15] G. Zhang, B.E. Patuwo, M.Y. Hu, Forecasting with artificial neural networks: The state of the art, International Journal of Forecasting 14 (1998) 35–62. [16] B. A. Beatrice, E. Kirubakaran, P. R.J. Thangaiah, K. L. D.Wins, Surface Roughness Prediction using Artificial Neural Network in Hard Turning of AISI H13 Steel with Minimal Cutting Fluid Application, Procedia Engineering 97 ( 2014 ) 205 – 211. [17] A. K. Sahoo, A.K.Rout, D.K. Das, Response surface and artificial neural network prediction model and optimization for surface roughness in machining, International Journal of Industrial Engineering Computations, 6 (2) (2015 ) 229 – 240.