Available online at www.sciencedirect.com
ScienceDirect Procedia CIRP 37 (2015) 205 – 210
CIRPe 2015 - Understanding the life cycle implications of manufacturing
Predictive Modelling for Energy Consumption in Machining using Artificial Neural Network Girish Kanta,*, Kuldip Singh Sangwana a
Department of Mechanical Engineering, Birla Institute of Technology and Science, Pilani-333031, India
* Corresponding author. Tel.: +91-1596-515837. E-mail address:
[email protected]
Abstract The energy efficiency is important evaluation criterion for new investment in machinery and equipment in addition to the classical parameters accuracy, performance, cost and reliability. Even the users in the automotive industry demand new acquisitions of energy consumed by a machine tool during machining. Large interrelated parameters that influence the energy consumption of a machine tool make the development of an appropriate predictive model a very difficult task. In this paper, a real machining experiment is referred to investigate the capability of artificial neural network model for predicting the value of energy consumption. Results indicate that the model proposed in the research is capable of predicting the energy consumption. The present scenario demands such type of models so that the acceptability of prediction models can be raised and can be applied in sustainable process planning during the manufacturing phase of life cycle of a machine tool. © 2015 2015 The The Authors. Authors. Published Published by by Elsevier Elsevier B.V. B.V.This is an open access article under the CC BY-NC-ND license © Selection and peer-review under responsibility of the International Scientific Committee of the “4th CIRP Global Web Conference” in the (http://creativecommons.org/licenses/by-nc-nd/4.0/). person of theunder Conference Chair of Dr.the John Ahmet Erkoyuncu. Peer-review responsibility organizing committee of CIRPe 2015 - Understanding the life cycle implications of manufacturing Keywords: Energy consumption; Artificial neural network; Predicitve Modelling;
1. Introduction Energy and materials are the two primary inputs required for the growth of any economy and these are obtained by exploiting the natural resources like fossil fuels and material ores. Pusavec et al. [1] indicated that the global environmental problems caused due to the consumption of natural resources and the pollution resulting during product life cycle (manufacture, use and end-of-life) have led to increasing political pressure and stronger regulations for the manufacturers and consumers of such products. According to Fang et al. [2] industrial sector accounts for about one-half of the world’s total energy consumption and the consumption of energy by this sector has almost doubled over the last 60 years. Manufacturing sector is one of them which are dependent on energy during processing of materials through various processes. According to Mani et al. [3] manufacturing sector alone accounts for 65% of the industrial sector’s energy consumption. Machine tool is one of the typical production
equipments widely used in the manufacturing industry. He et al. [4] reported that machine tools have efficiency less than 30% and Li et al. [5] claimed that more than 99% of the environmental impacts are due to the consumption of electrical energy used by the machine tools in discrete part manufacturing machining processes like turning and milling. Machine tools require energy during machining, build-up to machining, post machining and in idling condition to drive motors and auxiliary equipments. Salonitis [6] reported that the amount of energy consumed during the machining is approximately 15% of total energy; however, the design of a machine tool is based on the peak power requirement during machining of material which is very high as compared to nonpeak power requirement of the machine tool. This leads to higher inefficiency of energy in machine tools. Kant and Sangwan [7] proposed that the optimization of machining parameters for minimum energy requirement is expected to lead to the application of lower rated motors, drives and auxiliary equipments and hence save energy not only during
2212-8271 © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the organizing committee of CIRPe 2015 - Understanding the life cycle implications of manufacturing doi:10.1016/j.procir.2015.08.081
206
Girish Kant and Kuldip Singh Sangwan / Procedia CIRP 37 (2015) 205 – 210
machining but as well as during build-up to machining, post machining and idling condition. Therefore, the energy efficiency is an important evaluation criterion for new investment in machinery and equipment in addition to the classical parameters accuracy, performance, cost and reliability. Energy efficiency refers to technologies and standard operating procedures that reduce the volume of energy per unit of industrial production [8-9]. Kant and Sangwan [7] claimed that a generalized relationship between the machining parameters and the process performance is hard to model accurately mainly due to the nature of the complicated stochastic process mechanisms in machining. In the work of Arrazola [10] analytical, numerical, empirical and artificial intelligence methods have been proposed to predict the machining performance. Analytical models are generally developed using theoretical analysis of the material removal process. According to Sangwan et al. [11] numerical models are developed using finite element method, finite difference method, boundary element method etc. Empirical models are developed using conventional approaches such as factorial design, statistical regression, response surface methodology etc. On the other hand, Artificial intelligence based models are developed using nonconventional approaches such as the artificial neural network (ANN), Fuzzy logic(FL), Support Vector Regression (SVR) and Genetic Algorithm (GA). Machining process is very complex and does not permit pure analytical physical modeling [12]. The accuracy of numerical models is a major concern due to complex contact conditions, large plastic deformation, material property requirements and characteristics of friction mechanism at the tool-chip interface. Empirical models developed using conventional approaches may not describe the nonlinear complex relationship between machining parameters and machining performance [13-14]. Recently there has been a lot of interest to develop predictive models for investigating the influence of machining parameters on machining performance using artificial intelligence techniques as an alternative to conventional approaches [15–19]. This paper presents the following contributions. First, the experimental data of cutting energy as a performance characteristic is used to develop predictive model using Artificial Neural Network (ANN) technique during milling of medium carbon steel. The machining parameters are spindle speed (n), feed rate (f), depth of cut (ap) and width of cut (ae). Second, the ANN results are compared with the experimental data using relative error analysis. Next, the developed model is validated using some representative hypothesis testing. Finally, the influence of machining parameters on cutting energy is examined using 3D Plots. These 3D merged contour and surface plots are elaborated to estimate the cutting energy for any suitable combination of machining parameters. This paper is organized as follows. The experimental set up and design of experiment is presented in section 2. ANN methodology is presented in the section 3. The experimental data of section 2 is used to develop ANN model in section 4. The comparison of ANN results with experimental data using the relative error analysis, hypothesis testing to check the validity of model and the influence of machining parameters on cutting energy using contour plots is also discussed in the
section 4. Finally, conclusions are highlighted in section 5. 2. Case study 2.1. Experimental Set up To investigate the potential of ANN to predict the cutting energy, the work of Yan and Li [20] is undertaken as a case study. He developed the multi-objective optimization model to optimize the three machining performance characteristics surface roughness, material removal rate and cutting energy. To simplify the case study, in this work the focus was given only on cutting energy as a sustainable performance characteristic. He conducted face milling experiments on the medium carbon steel workpiece with 24 mm diameter, 3 flutes carbide tool in dry cutting environment. A CNC micromachining centre of 5.6 kW spindle power and maximum spindle speed of 6000 rpm was used to perform the experiments. A three phase power sensor was used to measure the power demand. The sensor was able to measure low voltage systems up to 380 V conductor to earth, and current up to 40 A. The measuring point was chosen at the main bus of the electrical cabinet to obtain the total energy consumption. Sensor was configured to record the active power of machine tool and the frequency was 10 Hz. 2.2. Design of experiments The Taguchi’s L27 standard orthogonal array was used to design the experimental. The measurements were carried out by varying four machining parameters spindle speed (n), feed rate (f), depth of cut (ap) and width of cut (ae). The combination of machining parameters and obtained corresponding cutting energy (CE) values obtained in the case study [20] are shown in Table 1. 3. Artificial Neural Network 3.1. Overview of ANN Artificial neural networks (ANNs) are inspired by the biological nervous systems — the brain, which consists of a large number of highly connected elements called neurons. Brain stores and processes the information by adjusting the linking patterns of the neurons [21]. In an ANN these neurons are connected together to form a network which mimics a biological nervous system. We can train a neural network to perform a particular function by adjusting the values of the connections (weights) between neurons. Neural networks are trained, so that a particular input leads to a specific target output. In the artificial intelligence learning, many input and target pairs are used to train a network. The network is adjusted based on a comparison of the output and the target, until the network output matches the target. 3.2. Methodology ANN as a computational model consists of three layers containing different neurons in an each layer. The three layers
Girish Kant and Kuldip Singh Sangwan / Procedia CIRP 37 (2015) 205 – 210
are input layer, hidden layers and an output layer. These layers are further interconnected to each other in such a way so that each neuron in one layer is connected to all neurons in the next layer. The diagram for a network with a single neuron is shown in Fig. 1. Table 1. L27 orthogonal array and cutting energy data [20] . Experiment
n
f
ap
ae
CE
No.
(rpm)
1
1000
(mm/min)
(mm)
(mm)
(kJ)
200
0.2
05
2
555.802
1000
200
0.3
10
204.929
3
1000
200
0.4
15
108.519
4
1000
250
0.2
05
446.109
5
1000
250
0.3
10
166.050
6
1000
250
0.4
15
89.823
7
1000
300
0.2
05
381.832
8
1000
300
0.3
10
142.976
9
1000
300
0.4
15
073.988
10
1500
200
0.2
10
357.042
11
1500
200
0.3
15
162.727
12
1500
200
0.4
05
319.031
13
1500
250
0.2
10
289.604
14
1500
250
0.3
15
133.648
15
1500
250
0.4
05
258.476
16
1500
300
0.2
10
233.559
17
1500
300
0.3
15
112.551
18
1500
300
0.4
05
213.109
19
2000
200
0.2
15
264.303
20
2000
200
0.3
05
445.797
21
2000
200
0.4
10
185.620
22
2000
250
0.2
15
213.939
23
2000
250
0.3
05
358.579
24
2000
250
0.4
10
151.343
25
2000
300
0.2
15
180.886
26
2000
300
0.3
05
306.850
27
2000
300
0.4
10
128.147
Inputs i1
Weights
w1k
i2
. . .
ij
w2k
Summation function
¦
Transfer function
f
Output
wjk bj
bias Fig. 1. Mathematical principal of a neuron.
The input layer does not perform any information processing. Each of its neuron takes the input from the actual environment. The input vector (multiple neurons) (ij) is transmitted using a connection that multiplies its strength by a
207
weight (w) to make the product (wi ). This neuron has a bias (bj). The corresponding output can be given into other interconnected neurons or directly into the environment. The output is produced by a summation function and activation function. Summation function calculates the net input from the processing neurons. The activation function converts the neuron’s weighted input to its output activation. An activation function consists of linear and non linear algebraic equations which make a neural network capable of storing nonlinear relationships between the input and the output. After being weighted and transformed by an activation function, neurons are then passed to other neurons. Output accepts the results of the activation function and present them either to the relevant processing neuron or to the outside of the network. As each input is applied to the network, the network output is compared to the target. The difference between the target output and the network output is known as error. Further, different network algorithms are applied to reduce the error. 4. Results and discussion The results of ANN used to predict the cutting energy based on input machining parameters in face milling process are shown and discussed below. 4.1. Prediction of cutting energy by ANN After a number of trials, it was found that the neural network structure 4-9-1 leads to best results. It consists of four input neurons in input layer (corresponding to four machining parameters n, f, ap, ae) one hidden layer with nine neurons and one output neuron in output layer (corresponding to one output CE). The experimental data shown in Table 1 is utilized as training data. Zhang [22] recommended that the ratio of training and testing samples could be given as percent, such as 90%:10%, 85%:15% and 80%:20% with a total of 100% for the combined ratio. The preferred ratio is selected as 80%:20% to fit in with the available experimental sample size of 27. The number of training and testing samples is 22 and 5. Data is normalized to a range of 0 and 1 before the training and testing process begins. The network is trained by using a random training data set. The training data is never used in the test data. The training is initialized by assigning some random weights and biases to all interconnected neurons. A feed forward back propagation algorithm has been used to train the network. The back propagation algorithm is based on gradient descent method which updates the weights iteratively until convergence to minimize the mean square error between network target values and training values. A logsig activation function in the hidden layer and a tansig activation function in the output layer are used to map the cutting energy values. traingdx is used as training function and learngd as learning function. The performance of the developed network examined on the basis of correlation coefficient (R value) between the output (predicted) values and the target (experimental) values for the test data (5) and entire data (27) is shown in Fig. 2. and Fig. 3. The R value is a measure of how closely the variation in output is explained by the targets. It lies in between 0 and 1. If it is 1 then it indicates the perfect
208
Girish Kant and Kuldip Singh Sangwan / Procedia CIRP 37 (2015) 205 – 210
values relative to the experimental values and is shown in Table 2. The mean relative error between the experimental and predicted values is 1.50%. It shows that the well trained network has good accuracy in predicting the cutting energy values.
MSE
correlation between the target values and output. Correlation coefficient of 0.99 was obtained between the entire data set (experimental data) and model predicted values which indicate good correlation.
Epochs
Fig. 2. Correlation between the predicted values and test data.
Fig. 4. Variation of mean square error (MSE) with number of Epochs. Table 2. Cutting energy predicted using ANN and relative error. Cutting Energy Experiment No.
Fig. 3. Correlation between the predicted values and entire data.
4.2. Comparison of ANN and experimental results for cutting energy The neural network model developed was trained using the selected parameters. The mean square error decreased with increasing iteration numbers until 324 iterations as shown in Fig. 4, but after this point it remained constant. The training of the algorithm was stopped at 324 iterations. After that, the ANN was tested for accuracy using the random test values selected from the experimental values which had not been used for the learning process. The predicted results of the entire data are shown in Table 2. Fig. 5. depicts the comparison of predicted results and the experimental values. It can be seen that the neural network prediction results are very close to the experimental values. The relative percentage error of the model prediction is also calculated as the percentage difference between the experimental and predicted
Relative Error
(kJ) Experimental
ANN
(%)
1
0.497
0.493
4.132
2
0.684
0.687
0.347
3
1.021
1.026
0.344
4
0.543
0.539
3.870
5
1.013
1.012
3.921
6
1.269
1.269
0.253
7
0.586
0.635
0.590
8
0.980
1.101
0.403
9
1.512
1.467
0.016
10
0.574
0.578
0.114
11
0.824
0.797
0.234
12
1.285
1.281
0.128
13
0.604
0.615
0.360
14
1.115
1.121
3.547
15
1.547
1.509
0.297
16
0.721
0.727
3.442
17
1.309
1.307
0.171
18
1.794
1.717
9.432
19
0.690
0.697
3.979
20
0.982
0.995
0.018
21
1.631
1.634
0.370
22
0.815
0.825
0.009
23
1.173
1.516
3.286
24
2.002
1.921
0.518
25
0.883
0.872
0.056
26
1.888
1.869
0.010
27
2.430
2.399
0.679
Girish Kant and Kuldip Singh Sangwan / Procedia CIRP 37 (2015) 205 – 210
209
Fig. 5. Comparison of ANN results with experimental values. Fig. 6. Surface and contour plot showing the influence of rpm and feed rate on cutting energy at 0.3 mm depth of cut and 10 mm width of cut.
4.3. Validation of ANN To compare the goodness of fit of the ANN model, some representative hypothesis tests are conducted and results are shown in Table 3. Table 3. Hypothesis testing to check the goodness of fit. 95 % Confidence Interval
P-value ANN
Mean paired t-test
0.573
Variance F-test
0.863
Levene’s test
0.937
Fig. 7. Surface and contour plot showing the influence of rpm and depth of cut on cutting energy at 250 mm/min feed rate and 10 mm width of cut.
These tests are t-test to test the means, f-test and Levene’s test for variance. In all these tests, the p-values are greater than 0.05, which means that the null hypothesis cannot be rejected. All the p-values in the Table 3 also indicate that there is no significant evidence to conclude that the experimental data and the data predicted from ANN differ to each other. Therefore, ANN as a prediction model has statistically satisfactory goodness of fit from the modeling point of view. 4.4. Influence of machining parameters on cutting energy The ANN model developed is used to analyze the influence of machining parameters (n, f, ae, ap) on cutting energy. The three dimensional merged contour and surface plots are created considering the two machining parameters at a time and fixing the other two parameters at level 2. The interactive effects of machining parameters on cutting energy are shown in Figs. 6-10. It can be observed from the Fig. 6. that the cutting energy decreases at higher feed rate and spindle speed. Same trend is observed in the Figs. 7-10. at higher values of other combinations of machining parameters. The higher values of machining parameters should account for higher loads on the machine tool and energy consumption should increase, but in this case the energy consumption reduces drastically with increase in value of machining parameters. It shows that the processing time dominates the energy consumption due to which the total energy consumption decreases even with increase in higher loads due to higher material removal rate.
Fig. 8.Surface and contour plot showing the influence of rpm and width of cut on cutting energy at 0.3 mm depth of cut and 250 mm/min feed rate.
Fig. 9. Surface and contour plot showing the influence of feed rate and depth of cut on cutting energy at 1500 rpm and 10 mm width of cut.
210
Girish Kant and Kuldip Singh Sangwan / Procedia CIRP 37 (2015) 205 – 210
Fig. 10. Merged surface and contour plot representing the influence of feed rate and width of cut on cutting energy at 1500 rpm and 0.3 mm depth of cut.
This is in agreement with the results of Diaz et al. [23] who also concluded that the processing time dominates the energy consumption and the energy consumed reduced to its one-third of its original value with increase in material removal rate. 5. Conclusions This paper presents the artificial neural network technique for the model development to predict the values of cutting energy as a sustainable performance characteristic. The predicted results are found to be very close to the experimental values. The mean relative error is 1.50% which shows that the developed model has good accuracy in predicting the cutting energy values. The hypothesis testing results validates that ANN as a prediction model has statistically satisfactory goodness of fit from the modeling point of view. The 3D plots shows that the processing time dominates the energy consumption due to which the total energy consumption decreases even with increase in higher loads. The constructed 3D merged contour and surface plots can be used at the shop floor to determine the cutting energy for any suitable combination of machining parameters. References [1] F. Pusavec, P. Krajnik, J. Kopac. Transitioning to sustainable production – Part I: application on machining technologies. J. Clean. Prod. 2010;18:174–184. [2] K. Fang, N. Uhan, F. Zhao, J.W. Sutherland. A new approach to scheduling in manufacturing for power consumption and carbon footprint reduction. J. Manuf. Syst. 2011;30:234–240. [3] M. Mani, J. Madan, J.H. Lee, K.W. Lyons, S.K. Gupta. Sustainability characterisation for manufacturing processes. Int. J. Prod. Res. 2014;1– 18. [4] Y. He, B. Liu, X. Zhang, H. Gao, X. Liu. A modeling method of taskoriented energy consumption for machining manufacturing system. J. Clean. Prod. 2012;23:167–174. [5] W. Li, A. Zein, S. Kara, C. Herrmann. Glocalized Solutions for Sustainability in Manufacturing, in: J. Hesselbach. C. Herrmann (Eds.). Glocalized Solut. Sustain. Manuf. Proc. 18th CIRP Int. Conf. Life Cycle Eng. Springer Berlin Heidelberg . 2011;268–275. [6] K. Salonitis. Energy efficiency assessment of grinding strategy. Int. J. Energy Sect. Manag. 2015;9:20–37.
[7] G. Kant, K.S. Sangwan. Prediction and optimization of machining parameters for minimizing power consumption and surface roughness in machining. J. Clean. Prod. 2014;83:151–164. [8] K. Salonitis., P. Ball. Energy efficient manufacturing from machine tools to manufacturing systems. Procedia CIRP. 2013;7:634–639. [9] K. Salonitis., P. Stavropoulos. On the integration of cax systems towards sustainable production. Procedia CIRP. 2009;7:115–120. [10] P.J. Arrazola, D. Umbrello, M. Davies, I.S. Jawahir. Recent advances in modelling of metal machining processes, CIRP Ann. - Manuf. Technol. 2013;62; 695–718. [11] K.S. Sangwan, G. Kant, A. Deshpande, P. Sharma. Modeling of Stresses and Temperature in Turning Using Finite Element Method. Appl. Mech. Mater. 2013;307:174–177. [12] C.A. van Luttervelt, T.H.C. Childs, I.S. Jawahir, F. Klocke, P.K. Venuvinod, Y. Altintas, et al. Present Situation and Future Trends in Modelling of Machining Operations Progress Report of the CIRP Working Group “Modelling of Machining Operations,”. CIRP Ann. Manuf. Technol. 1998;47:587–626. [13] I. Mukherjee, P.K. Ray. A review of optimization techniques in metal cutting processes. Comput. Ind. Eng. 2006;50:15–34. [14] G. Kant, V. Rao V, K.S. Sangwan. Predictive Modeling of Turning Operations Using Response Surface Methodology. Appl. Mech. Mater. 2013;307:170–173. [15] A.M. Zain, H. Haron, S. Sharif. Prediction of surface roughness in the end milling machining using Artificial Neural Network. Expert Syst. Appl. 2010;37:1755–1768. [16] G. Kant, K.S Sangwan. Predicitve Modelling for power consumption in machining using artificial intelligence techniques. Procedia CIRP. 2015;26:403–407. [17] G. Kant, K.S Sangwan. Predicitve modelling and optimization of machining parameters to minimize surface roughness using artificial neural network coupled with genetic algorithm. Procedia CIRP. 2015;31:448–453. [18] J.P. Davim, V.N. Gaitonde, S.R. Karnik. Investigations into the effect of cutting conditions on surface roughness in turning of free machining steel by ANN models. J. Mater. Process. Technol. 2008;205:16–23. [19] K.S Sangwan, S. Saxena, G. Kant. Optimization of machining parameters to minimize surface roughness using integrated ANN-GA approach. Procedia CIRP. 2015;29:305–310. [20] J. Yan, L. Li. Multi-objective optimization of milling parameters e the trade-offs between energy, production rate and cutting quality. J. Clean. Prod. 2013;52:462–471. [21] S.K. Pal, S. Mitra. Neuro-fuzzy pattern recognition: methods in soft computing. John Wiley & Sons Inc. New York: 1999. [22] G. Zhang, B.E. Patuwo, M.Y. Hu. Forecasting with artificial neural networks : The state of the art. 1998;14:35–62. [23] N. Diaz, E. Redelsheimer. D. Dornfeld. Energy consumption characterization and reduction strategies for milling machine tool use, In: Preceedings of the 18th CIRP International conference on life cycle engineering. Braunschweig. Germany. 2011:263–267.
