Prediction of surface roughness in hard turning under high pressure coolant using Artificial Neural Network

Prediction of surface roughness in hard turning under high pressure coolant using Artificial Neural Network

Accepted Manuscript Prediction of surface roughness in hard turning under high pressure coolant using Artificial Neural Network Mozammel Mia, Nikhil R...

1MB Sizes 4 Downloads 62 Views

Accepted Manuscript Prediction of surface roughness in hard turning under high pressure coolant using Artificial Neural Network Mozammel Mia, Nikhil Ranjan Dhar PII: DOI: Reference:

S0263-2241(16)30340-2 http://dx.doi.org/10.1016/j.measurement.2016.06.048 MEASUR 4168

To appear in:

Measurement

Received Date: Revised Date: Accepted Date:

4 February 2016 28 May 2016 22 June 2016

Please cite this article as: M. Mia, N.R. Dhar, Prediction of surface roughness in hard turning under high pressure coolant using Artificial Neural Network, Measurement (2016), doi: http://dx.doi.org/10.1016/j.measurement. 2016.06.048

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Title page Reference No: MEASUR 4168

TITLE: Prediction of surface roughness in hard turning under high pressure coolant using Artificial Neural Network

Authors: 1. Mozammel Miaa* 2. Nikhil Ranjan Dharb a

Mechanical and Production Engineering, Ahsanullah University of Science and Technology,

Dhaka 1208, Bangladesh Email: [email protected]; [email protected] Contact number: 8801689449864 b

Industrial and Production Engineering, Bangladesh University of Engineering and Technology,

Dhaka, Bangladesh Email: [email protected] Contact number: 8801711357885 *Corresponding author’s email: [email protected]; [email protected]

Cover letter TITLE: Prediction of surface roughness in hard turning under high pressure coolant using Artificial Neural Network The present work deals with the formulation of a predictive model of surface roughness in hard turning. The machining environments were dry and high pressure coolant jet. The artificial neural network (ANN) was chosen as the modeling technique. The hardened steel, owing to its extensive application in engineering products, has made its way to be the research material in this experiment. The selected work material (EN 24T), hardened at different levels, was employed in turning operation at various cutting speed and feed rate. The corresponding values of surface roughness was collected and later, used for modeling. The three different training algorithms i.e. Levenberg-Marquardt, Bayesian regularization and Scaled conjugate gradient algorithms were used to train the model. Furthermore, the performance of these algorithms was compared to the extent of the surface roughness prediction ability. Multiple ANN architectures were examined to select the best one for the modeling. Finally, the confirmation tests were performed to evaluate the error level of the model. The novelty of this work lies in the fact that, although a lot of ANN models exists to predict the surface roughness, mostly of which is developed for dry machining and very few is for flood cooling and no such ANN model is developed yet, for surface roughness prediction under high pressure coolant when the effects of high pressure coolant on surface roughness are well established. Hence, to economically and efficiently control the surface quality and adopt the HPC system in the hard machining, it is imperative to develop a model which will predict the surface roughness under dry as well as high pressure coolant condition.

*Corresponding email: [email protected]; [email protected]

Highlights 

Artificial neural network based predictive model of surface roughness



Dry and high pressure coolant (HPC) applied turning of hardened steels



Levenberg-Marquardt, Bayesian regularization, scaled conjugate gradient training



3-4-2 ANN structure trained by BR is recommended



Effective cooling and lubrication by HPC reduced roughness parameter

Graphical Abstract Experimental runs HPC – 48 data set Dry – 48 data set

Inputs and output Vc, So, H, CC Ra

Prediction model ANN for dry and HPC

Software MATLAB neural network toolbox 7

Data separation Training; Testing

Training of ANN model

Testing of ANN model

ANN Type Feed forward multilayer back propagation Network structures (a) 3-n-1: n = 20, 15, 12, 10, 7, 5, 4 (b) 3-n-2: n = 20, 15, 12, 10, 7, 5, 4 (c) 4-n-1: n = 20, 15, 12, 10, 7, 5, 4

ANN architecture 3-n-1, 3-n-2, 4-n-1

Transfer function ‘purelin’ & ‘tansig’ 2 tan sig n   1 1  e 2n



Performance function MSE

Training algorithm

Bayesian Regularization ‘trainbr’

Levenberg-Marquardt ‘trainlm’

Reject model

No



Scaled conjugate gradient ‘trainscg’

Lowest RMSE? Yes Accept model

Calculate RMSE

Manuscript

Title: Prediction of surface roughness in hard turning under high pressure coolant using Artificial Neural Network

Abstract: In this study, an artificial neural network (ANN) based predictive model of average surface roughness in turning hardened EN 24T steel has been presented. The prediction was performed by using Neural Network Tool Box 7 of MATLAB R2015a for different levels of cutting speed, feed rate, material hardness and cutting conditions. To be specific the dry and high pressure coolant (HPC) jet environments were explored as cutting conditions. The experimental runs were determined by full factorial design of experiment. Afterward the 3-n-1, 3-n-2 and 4-n-1 ANN architectures were trained by utilizing the Levenberg-Marquardt (LM), Bayesian regularization (BR) and scaled conjugate gradient (SCG) algorithms, and evaluated based on the lowest root mean square error (RMSE). The 3-10-1 and 3-4-2 ANN models, trained by BR, revealed the lowest RMSE. A good prediction fit of the models was established by the regression coefficients higher than 0.997. At last, the behavior of the surface roughness in respect of speed-feedhardness for dry and HPC conditions has been analyzed. The HPC reduced surface roughness by the efficient cooling and lubrication whereas the higher hardness of material induced higher average surface roughness due to higher restraining force against tool imposed cutting force.

Key words: Artificial neural network; Hard turning; Surface roughness; High pressure coolant.

1.

Introduction Steel is the highest used material in the production of engineering parts and products

around the world. The recent advancement in the material and tool development technology and increased exploration for better performance leads to the use of hardened steels having hardness value of 40-65 Rockwell hardness at C scale (HRC). Notable uses of hardened steels include extrusion and forging dies, landing gears, fuel injector nozzle, jet engine mounting, stamping and punching dies, bearing and its houses etc. At the same time, the best surface quality is required in these parts to function adequately. To manufacture these parts hard turning is gaining its popularity among the researchers and industrial practitioners due to its multi-various potential advantages, over conventional turning, such as reduced number of processing steps, machining cost and resources. Furthermore, the flexibility in producing complex parts is pushing hard turning to replace grinding operation, if and only if, the hard turning produces parts with a surface roughness level that is attained by grinding. Conventionally, no coolant is applied in hard turning but the use of coolant in the right technique with the right cutting tool can engender significant benefits in terms of surface finish, tool life and tool wear [1]. In this regard, the coated carbide tools which are used in more than 80% of machining operations, can be considered as the right cutting tool as it provides solid lubrication at the tool-work contact surface and produces better surface finish [2]. To achieve the best surface finish, it is essential to set the process parameters prior to machining operations. The quality of a surface produced by hard turning depends on factors such as cutting speed, feed rate, depth of cut, machining time, lubrication and cooling methods, and work materials. Experimental investigation and mathematical analysis have been performed by the researchers to evaluate the contributions of factors on the machinability aspect, specifically

on surface roughness. The prediction of surface roughness has been accomplished by using diverse techniques such as response surface method, artificial neural network, Taguchi method, fuzzy system [3-5]. Among different artificial intelligence based soft computing techniques, owing to the higher accuracy in prediction, the artificial neural network (ANN) has been employed extensively in machining operations. In hard turning, Sharma et al. [6] formed ANN model of surface roughness in terms of speed, feed, depth of cut and approaching angle, and found 76.4% accuracy. Karayel [7] derived surface roughness, by ANN, close to actual values. Gaitonde et al. [8] added time and cutting tools as inputs with cutting speed and feed; and acquired 20.57% prediction accuracy in surface roughness. Apart from turning, ANN has been adopted in boring operation to anticipate the surface roughness and by doing so, a 4.52% error rate was found [9], while in milling Zain et al. [10] found satisfactory results with ANN. Though there is evidence that in some cases ANN does not show adequate accuracy whereas other methods do, still in most of the cases ANN comes up with an accuracy that is greater than other techniques. For instance, Kumar and Chauhan [11] showed that ANN revealed higher error than RSM in surface roughness measurement, Sahoo et al. [12] revealed the supremacy of ANN over RSM. Furthermore, ANN revealed higher accuracy than linear and nonlinear regression [3, 13, 14] and Taguchi based surface roughness prediction [15]. It is prevalent that the cutting conditions directly afflict the machining performance by thermo-mechanical effect. In this regard high pressure coolant (HPC) is a viable coolant and lubrication technology. The superior performance of HPC over dry and conventional fluid is recognized by Naves et al. [16], while over cryogenic cooling is established by Bermingham et al. [17]. In general, the high pressure coolant jet ensures advantages like reduced tool wear,

prolonged tool life, better quality surface, reduced cutting temperature and force. Better chip breakability by high speed (i.e. high momentum) coolant jet of HPC is proved to favorably control the aforementioned merits. These benefits are more prominent in machining of hard and difficult-to-cut materials namely hardened and stainless steel, titanium and nickel based alloys [16-18]. After literature study, it is clear that although a significant amount of research is carried out regarding ANN modeling to predict surface roughness, mostly of which has been carried out in respect of cutting speed, feed rate and depth of cut. Some have incorporated tool geometry as input too [8]. Material hardness has been hardly taken into consideration during modeling as input [19, 20] while it is readily derivable that the surface characteristics after machining hard and soft material are not going to be same. At the same time, most of the works were performed for dry environment and very few is for flood cooling; no such ANN model is developed yet, for surface roughness prediction under high pressure coolant when the effects of high pressure coolant on surface roughness are long established [21, 22]. To fill this gap, the presented work possesses some novelties like – a) development of models for high pressure coolant assisted turning, besides dry cutting; b) inclusion of material hardness as input; c) neural network based model wherein all possible architectures were used (section 3.1); d) determination of the effect of hidden neuron number on the prediction accuracy; e) selection of the best training algorithm from Levenberg-Marquardt (LM), Bayesian Regularization (BR) and Scaled Conjugate Gradient (SCG) algorithms (section 3.2) which makes the work accurate and more reliable from other presented works. The aim of the present work is to employ artificial neural network to develop a model to predict average surface roughness parameter in hard turning of steels. The cutting speed and feed

rate are found to affect the surface roughness most, hence considered in this study. Furthermore, to meet the necessity aroused form literature the material hardness and cutting conditions were taken as the input variables. The developed model will help the hard machining industries to adopt the HPC and control the process parameters to attain a desired surface finish prior to actual machining.

2.

Material, Machine and Method

The experimented work material was hardened EN 24T round bar steels due to its wide application in gears, shafts, studs and bolts; even in hardened condition is has applicability in automobile, aircraft, connecting rods and landing gear components. The composition of the work material, before heat treatment, is shown in Table 1. Later, appropriate thermal treatment is performed to acquire desired hardness. As part of thermal treatment, three hollow round bars of length 200 mm, external diameter 120 mm, internal diameter 45 mm have been selected. The used induction furnace was equipped with a high heating element (RG-3000oC). The treatment consists of three steps i.e. austenizing, quenching, tempering. The austenizing process uses heating of materials and holds at temperature of 900 oC for 90 minutes. The quenching was executed by using oil to rapidly reduce the temperature and induce hardness. Finally, the tempering was performed at different temperatures to reduce the hardness level to 40 HRC, 48 HRC and 56 HRC, and release the induced stresses. The coated (with TiCN, WC, Co) carbide tools with ISO specification SNMG 120408, held by PSBNR 2525 M12 holder, were used due to its solid lubrication and favorable chip formation capability. According to the orthogonal rake system, the tool designation has an inclination angle

of -6o, orthogonal rake angle of -6o, orthogonal clearance of the principal flank of 6o, auxiliary orthogonal clearance of 6o, auxiliary cutting edge angle of 15o, principal cutting edge angle of 75o and a nose radius of 0.8 mm. The dry environment and high pressure coolant (HPC) jet were employed as cutting conditions. The reasons behind selecting HPC have already been discussed under ‘introduction’ section. In HPC system, the coolant was supplied at a constant pressure of 80 bar and flow rate of 6.0 l/min along the auxiliary cutting edge to the principal flank. The nozzle diameter was 0.5 mm. The environment friendly straight cut cutting oil (ISO grade VG-68) has been used as the cooling and lubrication agent and was impinged at angle 20o with the auxiliary cutting edge for effective penetration. The high pressure coolant system was attached in the experimental setup during the machining runs. The schematic of the experimental setup is shown in Fig. 1. The average surface roughness (Ra) was measured, after each machining run, using a Talysurf roughness checker (Surtronic 3+, Rank Hobson, UK). For each run, the surface roughness value was collected at three different locations which are ~120o apart along the circumference and the mean was recorded. Any unusual data was considered as outlier and excluded from mean value calculation. The experimental runs were conducted by using a reasonably rigid and powered centre lathe (7.5 kW, China). The depth of cut was kept fixed at 1.0 mm. The process variables with levels and the response are shown in Table 2. The number of experimental runs has been determined by using full factorial design of experiment (FFDoE). The FFDoE includes all possible combinations of inputs in planning the experiment, and hence the possibility of excluding significant runs becomes zero which in turn increases the accuracy of the developed model. Therefore, FFDoE endows with a total of 48 numbers of experimental runs for each of the

cutting conditions (dry and HPC). The experimental runs, corresponding results and usability type are shown in Table 3.

3.

Artificial neural network model Artificial neural network is a non-linear mapping system inspired by the functions of a

human brain. A total of three layers, each populated by one or more neurons, is the common structure of an ANN. Some numerical values are presented to the network through the neurons of the input layer. Each neuron of input layer can take only one input value and this value is transferred to hidden layers which are interconnected by synaptic weights to output layer in a way such that every neuron of hidden layer is connected to every single neuron of output layer. The output layer provides numerical values of responses. The training of ANN is accomplished by adapting the strengths or weights of the connections among the input, intermediate and output neurons which are capable of storing memory and information. By achieving the learning ability, ANN produces the desired responses according to the given decision variables. A network can, also, be constructed with more than one hidden layer or without any hidden layers which entirely depends on the input data, decision variables, relationship among the inputs and outputs, and the number of process parameters. The neuron numbers on a particular hidden layer can be varied but choosing an appropriate number, which avoids over-fitting due to too many neurons as well as under-fitting caused by too less neurons, is comparatively a difficult task and it is determined according to the input vector size and input-output vector space classifications [8]. The performance of ANN depends on the structure of the network, training algorithm, training time, training data size, learning function, transfer function, values of weights and

biases, testing data size and data representations. Since the ANN gradually originates its capability by training to find reasonable solutions for similar problems, it is becoming more popular and growing its contribution lately. In this work, a feed forward multi-layer neural network has been employed to predict surface roughness. The outline of the ANN modeling is demonstrated in Fig. 2.

3.1

ANN architectures

The number of layers, nodes and their relation in neural network model determines the accuracy level of prediction [10]. Hence, in developing the ANN model, it is imperative to form all possible architectures, test those for higher accuracy and then recommend one with the lowest root mean square error (RMSE). Table 2 shows that there exist two types of input variables i.e. categorical and numerical. Strategic orientation of these variables revealed the following three cases which are used in this study: I.

Two 3-n-1 structures wherein one for dry and another for HPC assisted turning; cutting speed, feed, hardness were inputs and average surface roughness was output; Fig. 3 shows these network structures.

II.

One 3-n-2 structure wherein two output neurons i.e. average surface roughness for dry and HPC, concurrently; cutting speed, feed, hardness were inputs; Fig. 4 shows this network structure.

III.

One 4-n-1 structure wherein cutting condition is also used as input variable with others and thus one average surface roughness is achievable either for dry or HPC based on selected input; Fig. 5 represents this network structure.

Irrespective of ANN network structures, the intermediate (hidden) layer consists of ‘n’ number of neurons at one layer. Multiple layer is not adopted as one layer was found sufficient to address low prediction error [10]. For each case, corresponding 34 data (70%) sets were used to train the network and 14 data (30%) sets for testing the model, separately. In Case-I, for dry and HPC model separately the 3×34 data for inputs (Vc, So, H) and 1×34 data for output (Ra) were used to train the model. In Case-II, the input parameters 3x34 and surface roughness 2x34 have been used to train the network. In case-III, 4x68 (including 34 data set for dry and 34 data set for HPC condition) matrix was formed and used as input for the network and 1x64 matrix as the target. After constructing networks with different configurations, different training algorithms were tested by using sample data (4x28). Simulation provided surface roughness on HPC when categorical value ‘0’ and on dry condition when ‘1’ has been used with the sample data.

3.2

Training algorithms

MATLAB R2015a ‘nnstart’ wizard has been used to create and train a network and afterward test the network. Neural network is trained by using Levenberg-Marquardt, Bayesian Regularization and Scaled Conjugate Gradient algorithms consecutively. These algorithms display competitive advantages over one another, hence it is needed to test which one divulges the lowest error and thus that one is recommended. The performance of these algorithms has been evaluated in term of RMSE of the predicted versus actual surface roughness. The Levenberg-Marquardt algorithm, denoted by ‘trainlm’, works faster when it trains a moderate-sized feed forward neural network that can hold up to several hundred weights [23] and supports the training with validation and test vectors. Bayesian regularization ‘trainbr’, proposed by MacKay [24], can overcome problems with imprecise noisy data and over-fitting as

well as under-fitting problems in neural network training. Scaled conjugate gradient is a supervised learning algorithm, denoted by ‘trainscg’, used in network training that updates weight and bias values according to the scaled conjugate gradient method. Though here the required number of iteration is more, number of computations per iteration is significantly low.

3.3

Transfer and performance function

During modeling, the hyperbolic tangent sigmoid function ‘tansig’ and pure linear function ‘purelin’ have been utilized as the transfer function for hidden and output layer, respectively. It is mentionable that the ‘tansig’ has been selected due to its symmetric nature [25]. The performance function was mean square error (MSE), defined in Eq. 1, that was used in evaluating and comparing errors with an objective of gaining the lowest error.

MSE 

4.

1 Ra actual   Ra  predicted 2  N

(1)

Results and Discussion In this work, different neural network based models were developed by considering three

different architectures and three different training methods. These models revealed accuracy to different degrees. Hence, to suggest the best architecture and best training algorithm, for this particular case of average surface roughness prediction in hard turning considering the aforementioned variables, a careful scrutiny of the predicted results and corresponding error analysis have been performed. Before doing so, for particular pair of NN architecture-training method, number of trials was made by altering the hidden neuron from 1 to 30, to train and test the network and achieve an acceptable predicted roughness with the least average RMSE.

Table 4 shows the RMSE in predicting average surface roughness for dry turning and HPC assisted turning by 3-n-1 model that is trained by Levenberg-Marquardt algorithm. It is well known that a model if trained multiple times, it generates different results each time due to different initial condition and sampling of data. Therefore, three consecutive trials were made for each configuration and the average value of RMSE was calculated. It is mentionable that one hidden layer was sufficient to draw low error; hence double hidden layer was not practiced. Coincidently, 3-4-1 revealed the lowest RMSE for both dry (0.0455) and HPC (0.0380) assisted turning. Table 5 reflects the RMSE for average surface roughness prediction model ‘3-n-1’ when trained by using Bayesian regularization. Notably, the 3-10-1 revealed the lowest average RMSE for dry (0.0341) and HPC (0.0284) assisted turning. Likewise, Table 6 lists the RMSE for 3-n-1 model, trained by scaled conjugate gradient algorithm. Herein, the 3-4-1 generated the lowest RMSE for dry (0.0816) and HPC (0.0516) applied hard turning. Upon doing the prediction, trials which showed unexpected and irregular results were ignored, because of the chances of the network to be over fitted. Presumably, different training algorithms recommended different hidden neuron numbers within 3-n-1 structure. Hence, it requires comparing the performance of the training algorithms based on the lowest average RMSE which is shown in Table 7. Although the Levenberg-Marquardt and scaled conjugate gradient algorithms agreed to 3-4-1 structure, the selection criteria of having the lowest average RMSE disclosed the Bayesian regularization as the acceptable training algorithm with 3-10-1 ANN structure. Fig. 6 illustrates the prediction capability of the 3-10-1 model, trained by BR algorithm, on the basis of correlation coefficient (R-value) on both cutting conditions, separately; wherein y-axis is the predicted average surface roughness parameter and x-axis represents targets (measured Ra). The solid line represents fit and dashed line represents identical actual and

predicted values. The R-value lies between 0 and 1; if it is 1 then it indicates the perfect correlation and in case of 0, it means no relationship exists between measured and predicted values. R-value close to 1 indicates good relationship and fit. Fig. 6(a) shows that the R-value is 0.99796 when the network was tested with 14 testing data sets under dry machining, and R-value is 0.99838 on HPC assisted machining which is illustrated in Fig. 6(b), hence establish the models as having very good fit. Basheer et al. [25] found R-value of 0.977 while predicting average surface roughness by using ANN. However, the fit is better when surface roughness parameter is predicted under high pressure coolant cutting condition. The possible reasons may be the generalization and measurement error of the data. Besides the regression plots of 3-10-1 NN model, a deviation graph is plotted in Fig. 7 wherein the difference between the actual and predicted average surface roughness parameter is shown in y-axis and the experimental runs, here which is 14 runs, along the x-axis. It is appreciable that the difference found in the prediction ability of ANN model lies within -0.05 to +0.05 µm with a random dispersion. Therefore, it is conclusive that the 3-10-1 ANN model produces good accordance between the predicted and actual Ra. Zain et al. [10] also concluded that the 3-n-1 (i.e. 3-1-1) structure revealed the minimum error in Ra prediction. Like the steps followed in 3-n-1 modeling, afterward to model by using 3-n-2 the identical steps were followed. The Bayesian regularization disclosed the least error compared to LM and SCG, thus only the results of BR trained ANN are listed in Table 8. Same network configurations were chosen in order to compare the results for all the approaches. It is appreciable that the lowest average RMSE is attainable at the 3-4-2 structure. The ‘3-n-2’ structure is capable of predicting average surface roughness on both HPC and dry condition

simultaneously, thus is less time consuming; yet less accurate than the ‘3-n-1’ network. The need for separate networks like ‘3-n-1’ for each cutting condition deemed less significant. Fig. 8 shows a regression curve corresponding to the first trial of 3-4-2 structure with RMSE 0.0279. Here, 28 predicted average surface roughness values (14 on dry and 14 on HPC condition) were compared to the measured data, and potted on the regression curve. The R-value is found to be 0.99837, an indication of good fit. The RMSE of surface roughness, using Bayesian Regularization, for ‘4-n-1’ structure is listed in Table 9. The lowest average RMSE is generated by the neural network with ‘4-15-1’ configuration. Three trials were made for each configuration ignoring over fitted networks. In this ‘4-n-1’ structure, the time to train was more and it showed less accuracy than the other two approaches. In this approach, Levenberg-Marquardt and scaled conjugate gradient algorithms were also being used during training and a larger deviation was observed. Fig. 9 shows the correlation between the predicted and measured average surface roughness of 28 testing runs. The R-value is found to be 0.99845 when the average surface roughness is predicted using ‘4-7-1’ neural network (RMSE = 0.0294). This R-value is greater than the R-value (0.95962) found by Beatrice et al. [26] when predicted Ra for minimum quantity lubrication assisted turning of steel. The points at the beginning have shown good fit then it slightly started deviating from the dashed line (actual value = predicted value). Mostly, the ‘4-n-1’ structure showed acceptable results, nonetheless, it possesses a lengthy set up process. From the discussion on the results of different models, done so far, it is determined that the Bayesian regularization algorithm is recommendable since it revealed the lowest average RMSE.

However, it is still unclear that which network structure should be accepted. To resolve this issue, two possible solutions are advised: a) use of 3-10-1 for only high pressure coolant assisted turning but not for dry turning, b) use of 3-4-2 wherein average surface roughness for both dry and HPC applied turning are expected. In addition, using ‘3-n-2’ provides less setup time with acceptable accuracy. As a result, the 4-15-1 has to be rejected due to its higher average RMSE. This conclusion is comparable with the suggestion made by Zhang et al. [27] regarding the number of hidden neurons as “3×n(=3)+1=10” that is also reflected in solution (a). However, the solution (b) does not agree to any of the recommended hidden neuron numbers i.e. 1, 3, 6, 7 rather indicates that 3-4-2 revealed the lowest error. This occurrence may be connected with the predictive number of outputs which here is 2 for same number of inputs (=3). To investigate how the average surface roughness parameter is related to the cutting speed and feed rate, three 3D plots are shown in Fig.10. Fig. 10(a-c) represents the roughness plots for dry turning and Fig. 10(d-f) shows same for HPC applied turning. The surface roughness parameter is found, in Fig. 10(a & d), to increase with increasing feed rate and decreasing cutting speed. The theoretical relation of average surface roughness with feed rate (

Ra  So ) also suggests that the average surface roughness is expected to increase as square of 2

feed rate for a constant nose radius of tool [28]. From the definition of feed in turning, it is understandable that a higher feed is reflected by a broader helicoidal groove over the machined surface which escalates the magnitude of surface roughness. Xavior and Adithan [29] found the similar pattern of surface roughness when turned AISI 304 steel under different cutting oils, as well. To be specific, the mean of average surface roughness is found in this study to be increased by ~1.4 times (for both dry and HPC) when the feed rate is augmented by 1.5 times, whereas a

double proportionate increment was obtained by Yallese et al. [30] in machining hardened bearing steel. On the other side, the average surface roughness parameter is reduced by the higher cutting speed. It is reported that the high cutting speed induces extra hardness within the material [31] and this hardness assisted in the reduction of Ra by producing favorable chips. However, if the material hardness is too high (~56 HRC) then different phenomenon prevail (Fig. 10 (b & e)). This outcome concurs with the finding of Ozel et al. [13]. This observation is valid for both dry and high pressure coolant assisted turning. A similar relation of speed and roughness parameters was presented in machining difficult-to-cut Ti alloy [32]. The prime cause may be that the high cutting speed tends to remove the built-up-edge and creates stability during machining with low chattering [33]. Furthermore, the frictional coefficient is low at the increased cutting speed which in turn congenially improves the average surface roughness quality. However, Yallese et al. [30] reported that, in their study, up to 120 m/min cutting speed the surface quality behaved favorably. But in present study, Ra exhibited a downward trend up to 161 m/min. Contrary to the gradual trends exerted by speed-feed, the change of material hardness value severely affects the roughness profile, explicitly when the hardness steps to 56 HRC from 48 HRC. This issue is addressed by the fact that material of high hardness (~56 HRC) when machined endures excessive compressive stress [34] imposed by the tool exerted cutting forces. This phenomenon gives rise to the reacting force by the hardened material thus the chip deformation becomes harder. As consequence, the tool edge experience accelerated wear and engenders higher surface roughness. In addition, the localized heat concentration induced by the intense cutting temperature in machining harder material, conversely soften the material and

assist chip particle adherence on the machined surface as roughness. However, Pal et al. [35] reported that higher hardness of steel was associated with lower roughness parameters. Presumably the high pressure coolant jet reduces the surface roughness by different scale in different speed-feed-hardness combination. The reduction of Ra by HPC was also indicated by Ezugwu et al. [1]. This is attributed to the thermo-mechanical effect triggered by pressurized jet that brings down friction and cutting temperature [32]. Furthermore, favorable chip formation by HPC [22] due the wedge effect created by the coolant jet between the chip and tool [36] prevents the chip to rub over the machined surface and thus ensures lower roughness value.

5.

Conclusions In present work, the artificial neural network has been embraced to develop a predictive

model of average surface roughness in turning of hardened EN 24T steel. Unlike other presented works of ANN, here the prediction of average surface roughness was carried out for different hardness of materials and cutting conditions. To be specific, the study covered the dry and high pressure coolant jet applied turning. The best predictive model was recommended only after evaluating the prediction capability of different neural network architectures and various training methods. The performance of the utilized training methods i.e. Levenberg-Marquardt, Bayesian regularization, scaled conjugate gradient technique was assessed in respect of the lowest root mean square error. The Bayesian regularization is found to offer the highest prediction accuracy. The determination of network structure was performed after determining Bayesian regularization as the training algorithm. For BR training, the possible structures i.e. 3-n-1, 3-n-2, 4-n-1 were formed and tested with unexposed data. The structure with the lowest RMSE has

been accepted. To predict average surface roughness only for turning with pressurized coolant 310-1 with BR training is appropriate, whereas 3-4-2 is recommended for average surface roughness in both cutting conditions for any specified speed-feed-hardness. The correlation coefficient has been found to be greater than 0.997 which reflects a good fit of the prediction models. Finally, the influence of input variables on the average surface roughness was analyzed by using 3D surface plots. Preferable cutting parameters are - higher cutting speed due to reduced BUE formation and chatter, lower feed rate due to deeper and wider tool impression on machined surface, lower to moderate hardness of material due to reduced particle adhesion originated by material softening, and lastly HPC as it ensures effective lubrication and cooling along with reduced friction by jet cushion.

6.

Acknowledgement

The authors are grateful to Directorate of Advisory Extension and Research Services (DAERS), BUET, Bangladesh for providing research fund, Sanction No. DAERS/CASR/R-01/2013/DR2103 (92) dated 23/08/2014 and the Department of Industrial and Production Engineering, BUET, Dhaka, Bangladesh for allowing laboratory facility to carry out the research work.

References [1] E.O. Ezugwu, J. Bonney, R.B. Da Silva, O. Cakir, Surface integrity of finished turned Ti–6Al– 4V alloy with PCD tools using conventional and high pressure coolant supplies, International Journal of Machine Tools and Manufacture, 47 (2007) 884-891. [2] A.K. Sahoo, B. Sahoo, Experimental investigations on machinability aspects in finish hard turning of AISI 4340 steel using uncoated and multilayer coated carbide inserts, Measurement, 45 (2012) 2153-2165. [3] G.M.A. Acayaba, P.M. de Escalona, Prediction of surface roughness in low speed turning of AISI316 austenitic stainless steel, CIRP Journal of Manufacturing Science and Technology, 11 (2015) 6267. [4] S. Ramesh, L. Karunamoorthy, K. Palanikumar, Measurement and analysis of surface roughness in turning of aerospace titanium alloy (gr5), Measurement, 45 (2012) 1266-1276. [5] A. Gok, A new approach to minimization of the surface roughness and cutting force via fuzzy TOPSIS, multi-objective grey design and RSA, Measurement, 70 (2015) 100-109. [6] 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) 473483. [7] D. Karayel, Prediction and control of surface roughness in CNC lathe using artificial neural network, Journal of Materials Processing Technology, 209 (2009) 3125-3137. [8] V.N. Gaitonde, S. Karnik, L. Figueira, J.P. Davim, Performance comparison of conventional and wiper ceramic inserts in hard turning through artificial neural network modeling, The International Journal of Advanced Manufacturing Technology, 52 (2011) 101-114. [9] K.V. Rao, B. Murthy, N.M. Rao, Prediction of cutting tool wear, surface roughness and vibration of work piece in boring of AISI 316 steel with artificial neural network, Measurement, 51 (2014) 63-70.

[10] A.M. Zain, H. Haron, S. Sharif, Prediction of surface roughness in the end milling machining using Artificial Neural Network, Expert Systems with Applications, 37 (2010) 1755-1768. [11] R. Kumar, S. Chauhan, Study on surface roughness measurement for turning of Al 7075/10/SiCp and Al 7075 hybrid composites by using response surface methodology (RSM) and artificial neural networking (ANN), Measurement, 65 (2015) 166-180. [12] A. Sahoo, A. Rout, D. Das, Response surface and artificial neural network prediction model and optimization for surface roughness in machining, International Journal of Industrial Engineering Computations, 6 (2015) 229-240. [13] T. Özel, Y. Karpat, L. Figueira, J.P. Davim, Modelling of surface finish and tool flank wear in turning of AISI D2 steel with ceramic wiper inserts, Journal of Materials Processing Technology, 189 (2007) 192-198. [14] T. Özel, Y. Karpat, Predictive modeling of surface roughness and tool wear in hard turning using regression and neural networks, International Journal of Machine Tools and Manufacture, 45 (2005) 467-479. [15] Ş. Karabulut, Optimization of surface roughness and cutting force during AA7039/Al 2 O 3 metal matrix composites milling using neural networks and Taguchi method, Measurement, 66 (2015) 139-149. [16] V. Naves, M. Da Silva, F. Da Silva, Evaluation of the effect of application of cutting fluid at high pressure on tool wear during turning operation of AISI 316 austenitic stainless steel, Wear, 302 (2013) 1201-1208. [17] M. Bermingham, S. Palanisamy, D. Kent, M. Dargusch, A comparison of cryogenic and high pressure emulsion cooling technologies on tool life and chip morphology in Ti–6Al–4V cutting, Journal of Materials Processing Technology, 212 (2012) 752-765. [18] E. Ezugwu, J. Bonney, Effect of high-pressure coolant supply when machining nickel-base, Inconel 718, alloy with coated carbide tools, Journal of Materials Processing Technology, 153 (2004) 1045-1050.

[19] A.P. De Paiva, J.H.F. Gomes, R.S. Peruchi, R.C. Leme, P.P. Balestrassi, A multivariate robust parameter optimization approach based on Principal Component Analysis with combined arrays, Computers & Industrial Engineering, 74 (2014) 186-198. [20] A.M. Hassan, M.T. Havajneh, Statistical analysis of the effects of machining parameters and workpiece hardness on the surface finish of machined medium carbon steel, Journal of Materials Engineering and Performance, 10 (2001) 282-289. [21] R.B. da Silva, Á.R. Machado, E.O. Ezugwu, J. Bonney, W.F. Sales, Tool life and wear mechanisms in high speed machining of Ti–6Al–4V alloy with PCD tools under various coolant pressures, Journal of Materials Processing Technology, 213 (2013) 1459-1464. [22] D. Kramar, P. Krajnik, J. Kopac, Capability of high pressure cooling in the turning of surface hardened piston rods, Journal of Materials Processing Technology, 210 (2010) 212-218. [23] M.T. Hagan, M.B. Menhaj, Training feedforward networks with the Marquardt algorithm, Neural Networks, IEEE Transactions on, 5 (1994) 989-993. [24] D.J. MacKay, Bayesian interpolation, Neural computation, 4 (1992) 415-447. [25] A.C. Basheer, U.A. Dabade, S.S. Joshi, V. Bhanuprasad, V. Gadre, Modeling of surface roughness in precision machining of metal matrix composites using ANN, Journal of Materials Processing Technology, 197 (2008) 439-444. [26] 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. [27] 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. [28] M. Mia, N.R. Dhar, Optimization of surface roughness and cutting temperature in highpressure coolant-assisted hard turning using Taguchi method, The International Journal of Advanced Manufacturing Technology, (2016) 1-15.

[29] M.A. Xavior, M. Adithan, Determining the influence of cutting fluids on tool wear and surface roughness during turning of AISI 304 austenitic stainless steel, Journal of Materials Processing Technology, 209 (2009) 900-909. [30] M.A. Yallese, K. Chaoui, N. Zeghib, L. Boulanouar, J.-F. Rigal, Hard machining of hardened bearing steel using cubic boron nitride tool, Journal of Materials Processing Technology, 209 (2009) 1092-1104. [31] G. Krolczyk, S. Legutko, P. Nieslony, M. Gajek, Study of the surface integrity microhardness of austenitic stainless steel after turning, Tehnicki Vjesnik-Technical Gazette, 21 (2014) 1307-1311. [32] T. Braham-Bouchnak, G. Germain, A. Morel, B. Furet, Influence of High-Pressure Coolant Assistance on the Machinability of the Titanium Alloy Ti555–3, Machining Science and Technology, 19 (2015) 134-151. [33] S. Dinesh, V. Senthilkumar, P. Asokan, D. Arulkirubakaran, Effect of cryogenic cooling on machinability and surface quality of bio-degradable ZK60 Mg alloy, Materials & design, 87 (2015) 10301036. [34] Y. Matsumoto, M. Barash, C. Liu, Effect of hardness on the surface integrity of AISI 4340 steel, Journal of Engineering for Industry, 108 (1986) 169-175. [35] A. Pal, S. Choudhury, S. Chinchanikar, Machinability assessment through experimental investigation during hard and soft turning of hardened steel, Procedia Materials Science, 6 (2014) 80-91. [36] M. Mazurkiewicz, Z. Kubala, J. Chow, Metal machining with high-pressure water-jet cooling assistance—a new possibility, Journal of Engineering for Industry, 111 (1989) 7-12.

Figures

Nozzle

Pressure Gauge Flow control

W/P

Filter Tool holder

Direction control valve

Relief valve

Block

Coolant Tank

Coolant Tank Supply pump

Fig. 1: Experimental setup attached with HPC system.

High pressure pump

Foot valve

Inputs and output Vc, So, H Ra

Experimental runs HPC – 48 data set Dry – 48 data set

Prediction model ANN for dry and HPC

Data separation Training; Testing

Software MATLAB neural network toolbox 7

Training of ANN model

Testing of ANN model

ANN Type Feed forward multilayer back propagation Network structures (a) 3-n-1: n = 20, 15, 12, 10, 7, 5, 4 (b) 3-n-2: n = 20, 15, 12, 10, 7, 5, 4 (c) 4-n-1: n = 20, 15, 12, 10, 7, 5, 4

ANN architecture 3-n-1, 3-n-2, 4-n-1

Transfer function ‘purelin’ & ‘tansig’ 2 tan sig n   1 1  e 2n



Performance function MSE

Training algorithm

Bayesian Regularization ‘trainbr’

Levenberg-Marquardt ‘trainlm’

Reject model

No



Scaled conjugate gradient ‘trainscg’

Lowest RMSE? Yes Accept model

Fig. 2: Outline of ANN modeling

Calculate RMSE

(a) Dry machining

(b) HPC machining Fig. 3: 3-n-1 architecture

Fig. 4: 3-n-2 architecture

Fig. 5: 4-n-1 architecture

(a) Dry condition [3-10-1, first trial]

(b) HPC condition [3-10-1, first trial]

Fig. 6: Regression curve between predicted and actual surface roughness

Actual Ra - Predicted Ra, m

0.075

Dry model HPC model

0.050

0.025

0.000

-0.025

-0.050

-0.075 0

2

4

6

8

10

12

14

16

Experimental (testing) runs

Fig. 7: Deviation graph for ‘3-10-1’ ANN model

Fig. 8: Regression curve between actual and predicted surface roughness [3-4-2, first trial]

Fig. 9: Regression curve between actual and predicted surface roughness [4-7-1, second trial]

a.

d.

b.

e.

c.

f.

Fig. 10: 3D plots of surface roughness – (a-c) dry turning, (d-f) HPC assisted turning

Tables Table 1

Chemical composition of EN-24T steel

Elements

C

Ni

Cr

Mo

Si

Mn

Content (%) 0.35-0.45 1.30-1.80 0.90-1.40 0.20-0.35 0.10-0.35 0.45-0.70

Table 2

P

S

Fe

0.05

0.05

Remaining

Independent variables, their level and the response

Notations Factors/Response

Unit

Type

Levels Level 1

Level 2

Level 3

Level 4

CC

Cutting condition

-

Categorical

2

Dry

HPC

-

-

Vc

Cutting speed

m/min Numerical

4

58

81

115

161

So

Feed rate

mm/rev Numerical

4

0.10

0.12

0.14

0.16

H

Hardness of workpiece

HRC

Numerical

3

40

48

56

-

Ra

Average surface roughness µm

Numerical

1

-

-

-

-

Table 3

Parameter orientations and experimental results

Trial Process variables

Response

Trial Process variables

Response

Trial

Process variables

Response

Ra

Type

No

CC

Vc

So

H

Ra

Type

No

CC Vc

So

H

Ra

Type

No

CC

Vc

So

1

Dry 58

0.1

40

0.92

Testing

16

Dry

81

0.12 40

1.02

Training

31

Dry

115

0.14

40

1.01

Testing

2

Dry 58

0.1

48

0.98

Training

17

Dry

81

0.12 48

1.08

Training

32

Dry

115

0.14

48

1.07

Training

3

Dry 58

0.1

56

2.02

Training

18

Dry

81

0.12 56

1.98

Testing

33

Dry

115

0.14

56

2.05

Testing

4

Dry 58

0.12 40

1.09

Training

19

Dry

81

0.14 40

1.08

Training

34

Dry

115

0.16

40

1.15

Testing

5

Dry 58

0.12 48

1.15

Training

20

Dry

81

0.14 48

1.14

Testing

35

Dry

115

0.16

48

1.21

Testing

6

Dry 58

0.12 56

2.15

Testing

21

Dry

81

0.14 56

2.15

Testing

36

Dry

115

0.16

56

2.15

Training

7

Dry 58

0.14 40

1.19

Training

22

Dry

81

0.16 40

1.32

Training

37

Dry

161

0.1

40

0.73

Training

8

Dry 58

0.14 48

1.25

Training

23

Dry

81

0.16 48

1.38

Testing

38

Dry

161

0.1

48

0.79

Training

9

Dry 58

0.14 56

2.25

Training

24

Dry

81

0.16 56

2.25

Training

39

Dry

161

0.1

56

1.64

Training

10

Dry 58

0.16 40

1.42

Training

25

Dry

115

0.1

40

0.75

Training

40

Dry

161

0.12

40

0.77

Testing

11

Dry 58

0.16 48

1.48

Training

26

Dry

115

0.1

48

0.81

Training

41

Dry

161

0.12

48

0.83

Training

12

Dry 58

0.16 56

2.35

Training

27

Dry

115

0.1

56

1.73

Training

42

Dry

161

0.12

56

1.78

Testing

13

Dry 81

0.1

40

0.83

Testing

28

Dry

115

0.12 40

0.92

Training

43

Dry

161

0.14

40

0.89

Training

14

Dry 81

0.1

48

0.89

Training

29

Dry

115

0.12 48

0.98

Training

44

Dry

161

0.14

48

0.95

Training

15

Dry 81

0.1

56

1.89

Training

30

Dry

115

0.12 56

1.85

Training

45

Dry

161

0.14

56

1.85

Training

H

Table 3

Parameter orientations and experimental results (contd.)

Trial Process variables

Response

Trial Process variables

Ra

Type

No

CC

No

CC Vc

So

46

Dry 161

0.16 40

1.05

Training

63

47

Dry 161

0.16 48

1.11

Testing

48

Dry 161

0.16 56

1.98

49

HPC 58

0.1

40

50

HPC 58

0.1

51

HPC 58

0.1

52

Response

Trial

Process variables

So

H

Ra

Type

No

CC

HPC 81

0.1

56

1.69

Training

80

64

HPC 81

0.12 40

0.87

Training

Training

65

HPC 81

0.12 48

0.95

0.77

Testing

66

HPC 81

0.12 56

48

0.85

Training

67

HPC 81

56

1.78

Training

68

HPC 58

0.12 40

0.94

Training

53

HPC 58

0.12 48

1.02

54

HPC 58

0.12 56

55

HPC 58

56

H

Vc

Vc

Response

So

H

Ra

Type

HPC 115

0.14

48

0.97

Training

81

HPC 115

0.14

56

1.82

Testing

Training

82

HPC 115

0.16

40

1.03

Testing

1.78

Testing

83

HPC 115

0.16

48

1.11

Testing

0.14 40

0.94

Training

84

HPC 115

0.16

56

1.95

Training

HPC 81

0.14 48

1.02

Testing

85

HPC 161

0.1

40

0.62

Training

69

HPC 81

0.14 56

1.92

Testing

86

HPC 161

0.1

48

0.7

Training

Training

70

HPC 81

0.16 40

1.2

Training

87

HPC 161

0.1

56

1.48

Training

1.92

Testing

71

HPC 81

0.16 48

1.28

Testing

88

HPC 161

0.12

40

0.67

Testing

0.14 40

1.03

Training

72

HPC 81

0.16 56

2.05

Training

89

HPC 161

0.12

48

0.75

Training

HPC 58

0.14 48

1.11

Training

73

HPC 115

0.1

40

0.66

Training

90

HPC 161

0.12

56

1.56

Testing

57

HPC 58

0.14 56

2.01

Training

74

HPC 115

0.1

48

0.74

Training

91

HPC 161

0.14

40

0.81

Training

58

HPC 58

0.16 40

1.27

Training

75

HPC 115

0.1

56

1.52

Training

92

HPC 161

0.14

48

0.89

Training

59

HPC 58

0.16 48

1.35

Training

76

HPC 115

0.12 40

0.81

Training

93

HPC 161

0.14

56

1.68

Training

60

HPC 58

0.16 56

2.15

Training

77

HPC 115

0.12 48

0.89

Training

94

HPC 161

0.16

40

0.92

Training

61

HPC 81

0.1

40

0.71

Testing

78

HPC 115

0.12 56

1.68

Training

95

HPC 161

0.16

48

1

Testing

62

HPC 81

0.1

48

0.79

Training

79

HPC 115

0.14 40

0.89

Testing

96

HPC 161

0.16

56

1.81

Training

Table 4

RMSE of surface roughness using Levenberg-Marquardt algorithm for ‘3-n-1’ architecture

Network

Dry turning

HPC applied turning

Configuration

First

Second

Third

Trial

Trial

Trial

3-20-1

0.1667

0.1464

0.1231

3-15-1

0.1008

0.1529

3-12-1

0.0927

3-10-1

First

Second

Third

Trial

Trial

Trial

0.1454

0.1676

0.0965

0.1278

0.1306

0.1303

0.1280

0.1124

0.0997

0.0939

0.1020

0.0799

0.0652

0.0793

0.0463

0.0488

0.0623

0.0524

0.1035

0.0918

0.0953

0.0969

0.0823

0.0636

0.0775

0.0744

3-7-1

0.0543

0.0682

0.0530

0.0585

0.0531

0.0523

0.0568

0.0540

3-5-1

0.0495

0.0496

0.0480

0.0490

0.0399

0.0358

0.0565

0.0440

3-4-1

0.0487

0.0433

0.0444

0.0455

0.0422

0.0387

0.0331

0.0380

Table 5

Average

Average

RMSE of surface roughness using Bayesian regularization algorithm for ‘3-n-1’ structure

Network

Dry turning

HPC applied turning

Configuration

First

Second

Third

Trial

Trial

Trial

3-20-1

0.0347

0.0342

0.0365

3-15-1

0.0347

0.0359

3-12-1

0.0371

3-10-1

Average

First

Second

Third

Average

Trial

Trial

Trial

0.0351

0.0275

0.0328

0.0340

0.0314

0.0331

0.0346

0.0388

0.0294

0.0334

0.0339

0.0331

0.0365

0.0356

0.0338

0.0273

0.0290

0.0300

0.0337

0.0346

0.0340

0.0341

0.0264

0.0295

0.0293

0.0284

3-7-1

0.0341

0.0350

0.0361

0.0351

0.0293

0.0338

0.0334

0.0322

3-5-1

0.0343

0.0377

0.0359

0.0360

0.0371

0.0358

0.0262

0.0330

3-4-1

0.0339

0.0349

0.0336

0.0341

0.0298

0.0303

0.0317

0.0306

Table 6

RMSE of surface roughness using scaled conjugate gradient algorithm for ‘3-n-1’ structure

Network

Dry turning

HPC applied turning

Configuration

First

Second

Third

Trial

Trial

Trial

3-20-1

0.2001

0.2646

0.2802

3-15-1

0.1148

0.1992

3-12-1

0.1853

3-10-1

First

Second

Third

Trial

Trial

Trial

0.2483

0.2394

0.2949

0.2787

0.2710

0.1004

0.1381

0.1544

0.1520

0.1752

0.1605

0.1909

0.0724

0.1495

0.1026

0.1381

0.1219

0.1209

0.1162

0.1262

0.0996

0.1140

0.0758

0.0654

0.0752

0.0721

3-7-1

0.1008

0.0744

0.1187

0.0980

0.1187

0.1488

0.1062

0.1246

3-5-1

0.2250

0.0863

0.1590

0.1568

0.1017

0.0839

0.0678

0.0845

3-4-1

0.0695

0.0935

0.0819

0.0816

0.0576

0.0494

0.0479

0.0516

Table 7

Average

Average

Performance comparison of training algorithms for 3-n-1 structure Levenberg-Marquardt

Bayesian Regularization

Scaled Conjugate Gradient

Avg. RMSE

Configuration

Avg. RMSE

Configuration

Avg. RMSE

Configuration

Dry turning

0.0455

3-4-1

0.0341

3-10-1

0.0816

3-4-1

HPC turning

0.0380

3-4-1

0.0284

3-10-1

0.0516

3-4-1

Table 8

RMSE of surface roughness using Bayesian Regularization algorithm for ‘3-n-2’ structure

Network Configuration

HPC and Dry Environment First Trial

Second Trial

Third Trial

Average

3-20-2

0.0362

0.0333

0.0368

0.0354

3-15-2

0.0323

0.0296

0.0329

0.0316

3-12-2

0.0345

0.0366

0.0344

0.0352

3-10-2

0.0317

0.0287

0.0313

0.0306

3-7-2

0.0333

0.0328

0.0323

0.0328

3-5-2

0.0390

0.0410

0.0374

0.0391

3-4-2

0.0279

0.0300

0.0316

0.0298

Table 9

RMSE of surface roughness using Bayesian regularization algorithm for ‘4-n-1’ structure

Network Configuration HPC/Dry Environment First Trial

Second Trial

Third Trial

Average

4-20-1

0.0323

0.0363

0.0356

0.0347

4-15-1

0.0321

0.0302

0.0314

0.0312

4-12-1

0.0328

0.0309

0.0330

0.0322

4-10-1

0.0308

0.0326

0.0321

0.0318

4-7-1

0.0338

0.0294

0.0334

0.0322

4-5-1

0.0335

0.0352

0.0303

0.0330

4-4-1

0.0332

0.0338

0.0325

0.0332

Graphical Abstract Experimental runs HPC – 48 data set Dry – 48 data set

Inputs and output Vc, So, H, CC Ra

Prediction model ANN for dry and HPC

Software MATLAB neural network toolbox 7

Data separation Training; Testing

Training of ANN model

Testing of ANN model

ANN Type Feed forward multilayer back propagation Network structures (a) 3-n-1: n = 20, 15, 12, 10, 7, 5, 4 (b) 3-n-2: n = 20, 15, 12, 10, 7, 5, 4 (c) 4-n-1: n = 20, 15, 12, 10, 7, 5, 4

ANN architecture 3-n-1, 3-n-2, 4-n-1

Transfer function ‘purelin’ & ‘tansig’ 2 tan sig n   1 1  e 2n



Performance function MSE

Training algorithm

Bayesian Regularization ‘trainbr’

Levenberg-Marquardt ‘trainlm’

Reject model

No



Scaled conjugate gradient ‘trainscg’

Lowest RMSE? Yes Accept model

Calculate RMSE

Highlights 

Artificial neural network based predictive model of surface roughness



Dry and high pressure coolant (HPC) applied turning of hardened steels



Levenberg-Marquardt, Bayesian regularization, scaled conjugate gradient training



3-4-2 ANN structure trained by BR is recommended



Effective cooling and lubrication by HPC reduced roughness parameter