You are on page 1of 10

Int J Adv Manuf Technol DOI 10.

1007/s00170-011-3553-6

ORIGINAL ARTICLE

A hybrid algorithm to optimize cutting parameter for machining GFRP composite using alumina cutting tools
M. Adam Khan & A. Senthil Kumar & A. Poomari

Received: 21 January 2011 / Accepted: 25 July 2011 # Springer-V erlag London Limited 2011

Abstract In this paper, two different evolutionary algorithm-based neural network models were developed to optimise the unit production cost. The hybrid neural network models are, namely, genetic algorithm-based neural network (GA-NN) model and particle swarm optimizationbased neural network (PSO-NN) model. These hybrid neural network models were used to find the optimal cutting conditions of Ti[C,N] mixed alumina-based ceramic cutting tool (CC650) and SiC whisker-reinforced aluminabased ceramic cutting tool (CC670) on machining glass fibre-reinforced plastic (GFRP) composite. The objective considered was the minimization of unit production cost subjected to various machine constraints. An orthogonal design and analysis of variance was employed to determine the effective cutting parameters on the tool life. Neural network helps obtain a fairly accurate prediction, even when enough and adequate information is not available. The GA-NN and PSO-NN models were compared for their performance. Optimal cutting conditions obtained with the PSO-NN model are the best possible compromise compared with the GA-NN model during machining GFRP composite using alumina cutting tool. This model also proved that neural networks are capable of reducing uncertainties related to the optimization and estimation of unit production cost. Keywords GFRP composite . Ceramic . Hybrid algorithm . GA . PSO . Neural network
M. A. Khan (*) : A. S. Kumar : A. Poomari Department of Mechanical Engineering, Sethu Institute of Technology, Virudhunagar Dist., Tamil Nadu, India e-mail: adamkhanm@gmail.com

Abbreviations GA-NN Genetic algorithm-based neural network PSO-NN Particle swarm optimization-based neural network BP Back-propagation L/D Length-to-diameter ratio S/N Signal-to-noise ratio LB Lower is better NB Nominal is best HB Higher is better yi Response of cutting performance results MLP Multilayer perceptron LM LevenbergMarquardt TANSIG Tangent sigmoid MAE Mean absolute error ej Error signal at the output neuron j yj Output of neuron (n) Instantaneous error energy av Average squared error energy CO Operating cost CNO Non-operating cost CR Tool replacement cost CT Tool cost UC Unit production cost Ts Setup time Ti Idle time Tc Tool change time Tm Machining time TL Tool life KO Sum of overhead and labour cost KT Cutting edge cost XA Actual cutting parameter XU Cutting parameterupper bound XL Cutting parameterlower bound f(x) Fitness function

Int J Adv Manuf Technol

g(x) Pm C1, C2 w Vid Pid Pgd Vmax MRR

Minimum of cost value (objective function) Mutation probability Learning factors Weightage factor Particle velocity Best location about the particles Globally best particle Particle maximum velocity Material removal rate

1 Introduction In any machining operation, it is an important task to select cutting parameters for achieving high-quality cutting performance [27]. Usually, the desired cutting parameters are determined based on the experience or by the use of standard handbooks. Intelligent selection of cutting parameter achieves substantial savings in terms of money and time if it integrates an efficient automated process planning. Optimum selection of cutting conditions importantly contributes to the increase of productivity and the reduction of costs [32]. In the olden days, analytical-direct space search procedures were followed to optimise the machining parameters. They were dynamic programming, geometric programming, stochastic programming, etc. These techniques are not adequate to search the wide spectrum of the problem domain. Often, they are also not robust because the numerous constraints considered make the search complex [3, 18]. Even though they were previously used for many complex problems, the search is very complicated [15]. There are other suitable methods, such as hill climbing, tabu search, simulated annealing and genetic algorithm. The solutions found by these methods are often considered as near-optimal solutions [28]. Genetic algorithm (GA) exploits the idea of survival of fitness and an interbreeding population and creates a novel and innovative strategy [7]. Goldberg [5] describes that GA is the search algorithm based on the mechanics of natural selection and genetics. GAs are known as very efficient heuristic algorithms, surmounting problems of traditional optimization algorithms, and are applied to many engineering problems. GA starts with an initial set of random solution made of chromosomes and produces more number of solutions with the help of GA operators in a chromosome evolving method [4, 18]. GA provides a near-optimal solution for a complex problem having a large number of variables and constraints [26]. It contrast to traditional optimization algorithms, nontraditional methods such as particle swarm optimization (PSO), artificial bee colony, ant colony optimization and simulated annealing were used to compare the convergence rate of the algorithm. One among these algorithms, PSO,

accomplishes the same goal as GA optimization in a new and faster way. PSO possesses the features of straightforward logic, simple realization and underlying intelligence [30]. Particle swarm optimization has been successfully applied to some manufacturing processes such as pulsed laser micromachining, electrochemical machining, friction welding, boring, milling, etc., by various researchers [16, 19]. PSO and GA both work with a population of solutions; combining the searching abilities of both methods seems to be a very good approach. Researchers have attempted to compare the application and performance of GA and PSO techniques [15]. These techniques are inspired by nature and proved themselves to be effective solutions to optimization problems. Moreover, PSO utilizes a single way of information flow without the complex genetic operators of the GA, such as crossover and mutation; more differently, PSO adopts the current optimal solution as the mechanism for renewing the whole search process, in contrast to the GA, such that the PSO has the ability to quickly converge to a reasonably good solution [9]. From the point of parameter evolution, PSO is better than GA. The convergence rate of the PSO algorithm is very less than that of GA, and it requires only a small number of iterations for convergence of optimal solution. However, both PSO and GA consider more expanded ranges of solution space than other existing algorithms [6]. Artificial neural networks (ANN) are currently being used in a variety of applications with great success [1]. Neural network models play a major role in the prediction of response for linear or nonlinear problems. Experimental data required to model neural network are very few when compared with the other models [29]. Neural network models are free from the expression and equation for optimization [24]. Neural network learns the problem by example like the human brain. There are several learning algorithms in neural network. Back-propagation (BP) is one of the best and uses the unique learning principle which is called the delta rule [2, 25]. BP algorithm is used to calculate the error gradient of neural network with respect to modified weight. The main principle behind the BP algorithm is the minimization of errors in neural network output. Neural-based optimization algorithms can be employed to predict results with higher precision [3, 31]. These neural-based optimization algorithms ensure a simple, faster and efficient optimisation of all turning parameters. Neural network models are also capable of representing the performance of cutting parameters with proper training of data. One of the most promising techniques is the adaptation of network training with evolutionary algorithms (EA). The merging of EA and ANN will gain adaptability to dynamic environment and will lead to significantly better intelligent systems than relying on ANN, PSO, or GA. V reclusive ery

Int J Adv Manuf Technol

articles were published in a way adapting network model trained with EA on machining studies. The design of experiments, especially analysis of variance (ANOV is widely A), used to identify the critical process parameters which can be adapted to the optimization techniques [13, 14]. In this paper, a GA-based neural network (GA-NN) model and a PSO-based neural network (PSO-NN) model were framed to optimise the unit production cost of the turning process. The optimization was carried out for Ti[C,N] mixed alumina cutting tool and SiC whisker-reinforced alumina cutting tool whilst machining glass fibre-reinforced plastic (GFRP) composite. This alumina-based ceramic cutting tool is one of the best attractive cutting tools for hard machining process, and they produce good surface finish [11, 21, 22]. V few articles have been published to compare GA-NN ery and PSO-NN models and the utilization of the above techniques for the optimization of alumina cutting tools, whilst machining GFRP is rarely available in the literature. Hence, an attempt is made to optimise the alumina cutting tools on machining GFRP and the optimization techniques , GA-NN and PSO-NN are compared for their performance.

alumina cutting tool (CC650) and SiC whisker-reinforced alumina cutting tool (CC670). Machining studies were performed by varying the cutting speed, the feed rate and the depth of cut. The flank wear and the surface finish of the machined GFRP composite were observed. Flank wear is measured using Metzer Toolmakers microscope, and surface roughness was measured using TR200 surface profile meter.

3 Solution methodology 3.1 Contribution of cutting parameters The main objective of this paper was to minimise the unit production cost of the machining process. Selection of cutting parameter plays a major role with different machine constraints to optimise the unit production cost. There are many machining parameters, such as cutting speed, feed rate, depth of cut, tool geometry, cutting tool temperature, lubrication system, cutting tool treatment, etc. In consideration of these cutting parameters, the development of a mathematical model or to fix an objective function to predict the response of cutting operation will be very difficult and complex. To avoid such a situation, an ANOV A-based design of experiment model was developed to check the contribution of the cutting parameters. According to the Taguchi quality design concept, an L9 orthogonal array was chosen as a set of experimental work to predict the contribution of cutting parameters and presented in Table 1. The levels of cutting parameters are given in Table 2. ANOV was used to investigate the response of the A cutting parameters; the responses of the cutting parameter are further transformed into a signal-to-noise (S/N) ratio. There are several S/N ratios available depending on the type of characteristics, namely, lower is better (LB), NB or HB. Each performance characteristic may belong to a different category in the analysis of the S/N ratio [12, 17].
Table 1 Experimental layout using L9 orthogonal array Exp no. Cutting velocity (m/min) 1 1 1 2 2 2 3 3 3 Feed rate (mm/rev) 1 2 3 1 2 3 1 2 3 DOC (mm) 1 2 3 3 1 2 2 3 1

2 Experimental work 2.1 GFRP composite The GFRP composite rod with a diameter of 65 mm and a length of 400 mm, having a L/D ratio 6.15 [12], was prepared in our laboratory using filament winding process. Figure 1 shows the SEM micrograph of the GFRP composite prepared using the filament winding process. From the SEM picture, it can be noted that the glass fibres are strongly bonded and homogeneously impregnated with polyester matrix material. 2.2 Machining study Machining studies were carried out to machine the GFRP composite material in precision lathe using Ti[C,N] mixed

Fig. 1 SEM micrograph of the GFRP composite material

1 2 3 4 5 6 7 8 9

Int J Adv Manuf Technol Table 2 Level of cutting parameters for machining GFRP composite material Cutting parameter Cutting velocity (m/min) Feed (mm/rev) DOC (mm) Level I 150 0.06 0.1 Level II 190 0.08 0.2 Level III 300 0.12 0.3
Contribution % 70 60 50 40 30 20
Cutting Speed (m/min) Feed rate (mm/rev) DOC (mm) Error (%)

46.54

49.33 44.74

47.74

The lower surface roughness and the lower tool wear give an apparently better cutting performance. Therefore, the LB was selected in this study for both the tool wear and surface roughness. The response of cutting performance results yi of n repeated number is: S=NLB 10 log
n 1X 2 y n i1 i

10

1.60 2.53
0 GFRP + CC6 5 0 Cutting Parameter

3.13 4.39
GFRP + CC6 7 0

Fig. 3 Contributions of the cutting parameters to surface roughness

The S/N ratios of the tool wear and the surface roughness were calculated for each level of the cutting parameters. These findings are consistent with the experimental results. From the ANOV the contributions of A, cutting parameters are plotted in Figs. 2 and 3 for flank wear (in millimetres) and surface roughness (in micrometres). From these figures, the contribution of the cutting speed and the feed rate is much higher than that of the depth of cut. It shows that the cutting speed and the feed rate play major role on tool wear and surface roughness. The total contribution of cutting speed and feed rate is more than 86% and 91% for flank wear and surface roughness, respectively. 3.2 Problem formulation of hybrid algorithm For the optimization process, a multilayer perceptron (MLP) algorithm was adapted with the cutting speed and feed rate in the input layer having two hidden layers of nine neurons in each and the production cost as a single real80 Cutting Speed (m/min) 70 60 Feed rate (mm/rev) DOC (mm) Error (%)

time output. LM algorithms have been proposed with the hyperbolic tangent sigmoid (TANSIG) layers transfer function to train and to test the MLP algorithm. The algorithm is hierarchically connected and the network directed with varying numbers of the hidden layers with the help of the back-propagation of fully connected processing elements. The inputs from the continuous domain are mapped to outputs normalized to a range of 01.0. The basic goal in training any neural network is to minimise the overall MAE of the network. Training of ANN was made with the experimentally calculated cutting data. Expected training nodes in input layer and inputs in every other node are the sum of the weighted outputs of the previous layer. Each node is brought as an active case depending on the input of one node, the activation function and the threshold value of the node. The error signals at the output of neuron j at iteration n (i.e. presentation of the nth training examples) are defined by ej n dj n yj n where the neuron j is an output node and j is the output of neuron which is calculated by the function yj 1exp1v , with the locally induced field vj (i.e. weighted sum of all synaptic input plus the bias) of neuron j. This is commonly used, called as sigmoidal nonlinearity defined logistic function. Correspondingly, the instantaneous value (n) of the total error energy is obtained by summing 1 e2 n over all 2 j neurons in the output layer; these are the only visible neurons for which error signals can be calculated directly. We may thus write it as
j

51.08
Contribution % 50 40 30 20 10 0

5 1 .1 5 39.12

36.10

xn
4.24 5.57
GFRP + CC6 5 0 Cutting Parameter

1X 2 e n 2 j2c j

7.37

5 .3 9

GFRP + CC6 7 0

Fig. 2 Contributions of cutting parameters for flank wear

where set C includes all the neurons in the output layer of the network. Let N denote the total number of patterns contained in the training set. The average squared error energy is obtained by summing (n) over all n and then

Int J Adv Manuf Technol

normalizing with respect to the set size N, as shown by xav


N 1 X xn N j1

Table 4 Parameter bounds for optimization algorithm Parameters Cutting velocity Feed rate Depth of cut Unit m/min mm/rev mm Lower bound 150 0.06 Upper bound 500 0.12 0.2

The instantaneous energy error (n), and therefore the average error energy av, is a function of all the free parameters of the network. The arithmetic average of these individual weight changes over the training set is therefore an estimate of the true change that would result from modifying the weights based on minimising the objective function. The main objective function of the problem was framed with the basic elements of cost with respect to the machining parameters; the mathematical model was proposed in a previous research work [20]. This cost function is divided into four basic cost elements: operating cost (CO), non-operating cost (CNO), tool replacement cost (CR) and tool cost (CT). The cost function can be expressed as:
UC CO CNO CR CT UC Ko Tm Ko Ti Ts Ko TC Tm =TL Kt Tm =TL

In continuation, details of the elements of cost and the parameter bounds are given in Tables 3 and 4. In this approach, the solution space consists of all the different combinations of hidden layers and hidden neurons, i.e. of all the architecture. In this research paper, the GA and PSO are employed for such a complex problem to search the solution space for the best architecture, where best is defined according to predefined machining conditions. Practically, it means that the solution space is divided into subsets that are evaluated in order to find the best solution using the basic idea of the GA and PSO concepts. The mathematical model of these algorithms dictates the coding of solution in binary strings called chromosomes, the use of the objective function to evaluate how good the chromosomes (goodness of fit) are. 3.2.1 Genetic algorithm-based neural network Goldberg [5] described that genetic algorithms are search algorithms based on the mechanics of natural selection
Table 3 Details of costs, time and volume of material removal Details Setup time Idle time Tool change time V olume of material removal (GFRP) Sum of overhead and labour cost Cutting edge cost for CC650 Cutting edge cost for CC670

and natural genetics. The GA technique starts with an initial set of random solution called initial population which is made of group of chromosomes. Each chromosome in the population is real coded and contains the actual values of cutting conditions fed to the neural network model to obtain the accurate fitness value. A very small and a very large size of population leads to premature and slow rates of convergence of GA. Therefore, a population size of 20 chromosomes is selected for the present algorithm. The fitness value for each chromosome in the initial population is estimated using the fitness function f(x). The chromosomes with large fitness values in the population are selected by the reproduction operator for a second generation [23]. Parents are recombined to produce offspring. All offspring are mutated with a certain probability. The fitness of the offspring is then computed. The offspring are inserted into the population replacing the parents, producing a new generation. If the optimization criteria are not met, the creation of a new generation starts. Individuals are selected according to their fitness for the production of offspring. This cycle is performed until the optimization criteria are reached. Detailed procedure for the GA concept was explained in a previously published article [20]. 3.2.2 PSO-based neural network The particle swarm optimization algorithm is a relatively new computational intelligence technique to simulate a sort of social behaviour [8, 30]. In PSO, a point in the problem space is called a particle, which is initialized with a random position and search velocity. The particles, which are potential solutions in the PSO algorithm, fly around in the multidimensional search space and the positions of indiNotation Ts Ti Tc V KO KT(CC650) KT(CC670) Unit min min min mm3 US $/min US $/cutting edge US $/cutting edge V alue 0.6 0.2 0.5 65111 0.08 8 12

Int J Adv Manuf Technol

vidual particles are adjusted according to its previous best position, and the neighbourhood best or the global best. Since all particles in PSO are kept as members of the population throughout the course of the searching process, PSO is the only evolutionary algorithm that does not implement survival of the fittest. PSO has been successfully shown to optimise a wide range of continuous optimization problems. Similar to the GA-NN algorithm, individual particle positions from the PSO are fed to the neural network model to obtain the accurate best solution, continued to the number of iterations. The global best or local best solution in PSO is only reported for the other particles in a swarm. Therefore, evolution only looks for the best solution and the swarm tends to converge to the best solution quickly and efficiently, thus increasing the probability of global convergence. After a certain number of generations, the algorithm terminates. 3.3 Steps in hybrid model developed for optimization A flowchart of a hybrid GA-based neural network (GANN algorithm) and PSO-based neural network (PSO-NN algorithm) developed to optimise the cutting parameters is shown in Fig. 4a, b; the steps involved in the process are as follows.

1. Start: Random initial population is generated containing cutting information such as speed and feed rate as binary digits. The length of string assigned for each parameter is 10 bits. 2. Decoding and actual cutting parameters are calculated with respect to the fixed parameter bound. ! Xu XL XA XL Decode value 2n 1 3. Testing data: Decoded cutting parameters generated using the GA and PSO algorithms are normalized. Normalization of data means that all the data must have their values between 0 and 1. 4. Training data: From the experimental work done, the cost function and the fitness value were calculated in the data matrix. Data of the cutting speed, feed rate and cost value obtained from the experimental work are also normalized between 0 and 1, which are to be used as the training data for the neural network model. 5. Use of ANN: The neural network utilized in this work is based on a multilayer perceptron model which consists of input, hidden and output layers, and backpropagation algorithm scale conjugate gradient for data

START

Technological Data Base V, f, TL, MRR, etc...

START

Technological Data Base V, f, TL, MRR, etc...

Randomly Generated Population (Binary String)

Calculation of Cost Value

Initialize the random location & initial particle velocity

Calculation of Cost Value

Parameter Decoding & Data Normalization Testing Data

Parameter Decoding & Data Normalization Training Data Neural Network Model
(MATLab)

Parameter Decoding & Data Normalization Testing Data

Parameter Decoding & Data Normalization Training Data Neural Network Model
(MATLab)

Crossover & Mutation

Update Particle location

Update Particle velocity

Population Selection

Fitness Evaluation

Best location (Pbest)

Optimum condition

Globally best location (Gbest)

gen = gen+1

gen = max gen


gen = gen+1 gen = max gen

STOP
STOP

Fig. 4 a Flowchart of GAneural network algorithm for optimum cutting condition. b Flowchart of PSOneural network algorithm for optimum cutting condition

Int J Adv Manuf Technol

Fig. 5 Model neural network problem using MA TLAB software

learning [10]. With the above neural network basic concepts, the cost value for the initially generated cutting parameters was calculated. 6. Fitness: Based on the maximum fitness concept f(x)=1/ [1+g(x)] (i.e. minimum of cost value g(x)) for each individual, fitness was evaluated and optimal solution was stored for each set of chromosomes. 7. Genetic concept: Initially generated cutting parameters (string) were selected for reproduction of a new set of cutting parameters using rank selection method. These selected cutting parameters are made to swap at a single cut point for the crossover process. The mutation process starts immediately with the swapped parameters in random with the mutation probability of Pm =0.03. PSO concept: Initially generated particles are updated by calculating the particle velocity and particle position using the following equation.
New Old Vid w Vid C1 rand pid xold C2 rand pgd xold id id
New xNew xold Vid id id

where C1 =C2 =2 and w=(0.5+rand/2.0) [8]. Each particle is updated according to the equation (Vid), and the best location about the particles (Pid) and the globally best particle (Pgd) location are identified within the neighbourhood. A particle velocity in each dimension is clamped to a maximum velocity, Vmax, and the maximum velocity, Vmax, is set to a certain fraction of the range of the search space in each dimension. 8. Process is continued with steps 3 and 5 until it reaches the number of iteration required. 9. At each and every generation (iteration), the best solution was stored in a database and the globally best solution was identified. 10. Stop: Finally, all the global best solutions are plotted in a graph and the optimal solution was identified for each alumina cutting tool with the corresponding speed and feed rate along with tool life.

layers having nine neurons each for interpolation. The model developed using MA TLAB 7.0 is shown in Fig. 5. A set of input and output parameters is repeated with the neural network model as training to reduce the average error whilst predicting the unit production cost. Figure 6 shows the average error obtained whilst training the data for prediction of unit production cost with respect to 5,000 epochs (number of iterations). At the beginning, the error rating was observed in the rate of 101 when the iterative learning process reaches 5,000 epoch iterations; it was found to be at the rate of 106 and stabilized with minimum error. This is one of the main testimonies for error reduction in the prediction of production cost for a given set of cutting parameters. Overall, the neural network model exhibited good performance after training, and the output results indicate that the model is able to represent both the hybrid algorithms. The hybrid algorithm models, a GA- and PSO-based neural network, are used to obtain optimal cutting parameters having minimum unit production cost. The fitness value (for minimum unit production cost) obtained for alumina cutting tools using a GA-based neural network model (GA-NN) and a PSO-based neural network model (PSO-NN) result has been presented in Fig. 7a, b with respect to 50 generations.

4 Results and discussion 4.1 V alidation of the neural network model To predict an optimistic cutting parameter, a neural network model was developed with two inputs and two hidden
Fig. 6 V ariation of error in prediction of cost with respect to the number of iterations

Int J Adv Manuf Technol

0.08 Fitness Vs Generation for CC670 tool 0.07 0.06

0.05 0.04 0.03 0.02 0.01 0.00 0 10 20 30 No. of generation 40 50


GA based NN PSO based NN

0.08 0.07 0.06

Fitness Vs Generation for CC650 tool

0.05 0.04 0.03 0.02 0.01 0.00 0 10 20 30 No. of generation 40 50 GA based NN PSO based NN

Fig. 7 a Fitness vs. number of generations for alumina-based ceramic cutting tool (CC650) in machining the GFRP composite materials. b Fitness vs. number of generations for alumina-based ceramic cutting tool (CC670) in machining the GFRP composite materials

4.2 Comparison of GA-NN and PSO-NN models The optimal results obtained using the GA-NN and PSO-NN models for Ti[C,N] mixed alumina cutting tool

(CC650) and SiC whisker-reinforced alumina cutting tool (CC670) are tabulated for performance comparison (Table 5). The computational results (fitness value) obtained with respect to the number of generations shows that the GA-NN and PSO-NN models overlap each other with a micro-level variation (Fig. 7). PSO uses simple mathematical operators whereas the GA uses crossover and mutation mechanisms for population reproduction. It is well known that a simple mathematical operation runs much faster than other position changing operations and mechanisms for reproduction of a next generation. The rate of convergence in a PSO-based neural network model is better than the GA-based neural network model. It is evident that the convergence rate on predicting the optimal cutting condition with the PSO-NN model is much faster compared with the GA-NN model. Comparing the optimal cutting condition of alumina cutting tools during machining of the GFRP composite, the cutting speed is around 240 255 m/min for both the alumina cutting tool with a feed rate of 0.11 mm/rev using both hybrid models. The unit production cost of SiC whisker-reinforced alumina cutting tool (CC670) is US $16.69 for the GANN model and US $16.70 for the PSO-NN model. However, the unit production cost of Ti[C,N] mixed alumina cutting tool (CC650) is US $15.25 for the GANN model and US $15.36 for the PSO-NN model. The unit production cost on machining the GFRP composite using Ti [C,N] mixed alumina cutting tool (CC650) is lower than that of machining with SiC whisker-reinforced alumina cutting tool (CC670). However, the tool life of SiC whisker-reinforced alumina cutting tool is better than that of the Ti[C,N] mixed alumina cutting tool. As the tool cost plays a dominant role on machining the GFRP composite material, the cheaper CC650 is able to contribute a reduction in the unit production cost, although the performance is lower than that of the CC670 tool. However, the PSO model gave better results compared with the GA model in terms of accuracy as well as in convergence to the optimal solution.

Table 5 Details of optimised cutting parameters

Fitness Value

Fitness Value

Parameters

Units

GFRP+CC650 GA+NN PSO+NN 238 0.108 7.196 7.810 5141 15.23 0.0616

GFRP+CC670 GA+NN 254 0.108 8.946 6.109 5486 16.98 0.565 PSO+NN 255 0.111 8.828 6.212 5661 16.67 0.0566

Speed Feed Tool life Surface roughness MRR Cost Fitness

m/min mm/rev min m mm3/min US $

240 0.108 7.11 7.892 5184 15.27 0.0615

Int J Adv Manuf Technol

4.3 Application of GA-NN and PSO-NN models The proposed GA-NN and PSO-NN models have been successfully applied to various applications. The evolutionary techniques (EA)-based neural network models are highly recommended for complex and nonlinear optimization problems for their accurate results. To demonstrate the effectiveness and efficiency of the proposed models, machining studies were carried out to optimise the cutting condition of alumina cutting tool. Obviously, the PSO-NN model has been shown to outperform the GANN model in computational efficiency, rate of convergence, optimality and robustness. The PSO-NN model can be effectively used for large and complex problems like the optimization of flexible manufacturing machine cells and automatic-guided vehicles as it is faster and converges quickly.

mixed alumina cutting tool is lower than that of the SiC whisker-reinforced alumina cutting tool in machining the GFRP composite material.

References
1. Benardos PG, V osniakos G-C (2007) Optimizing feedforward artificial neural network architecture. Eng Appl Artif Intell 20:365382 2. Correa M, Bielza C, Pamies-Teixeira J (2009) Comparison of Bayesian networks and artificial neural networks for quality detection in a machining process. Expert Syst Appl 36:72707279 3. Cus F, Zuperl U (2006) Approach to optimization of cutting conditions by using artificial neural networks. J Mater Process Technol 173:281290 4. Gen M, Cheng R (1997) Genetic algorithm and engineering design. Wiley Interscience, Singapore, pp 141 5. Goldberg DE (2001) Genetic algorithm in search, optimization and machine learning. Pearson Education, Singapore, pp 121 6. Jamshidi SF, Amiri M, Karimi N (2010) Nonlinear continuous multi-response problems: a novel two-phase hybrid genetic based metaheuristic. Applied Soft Computing 10:12741283 7. Kalyanmoy D (1996) Optimization for engineering design: algorithms and example. Prentice Hall, New Delhi, pp 290318 8. Kao Yi-Tung, Zahara E (2008) A hybrid genetic algorithm and particle swarm optimization for multimodal functions. Appl Soft Comput 8:849857 9. Kuo H-C, Chang J-R, Liu C-H (2006) Particle swarm optimization for global optimization problems. J Mar Sci Technol 14 (3):170181 10. Lee ES (2001) Precision machining of glass fibre reinforced plastics with respect to tool characteristics. Int J Adv Manuf Technol 17:791798 11. Li L, He N, Wang M, Wang ZG (2002) High speed cutting of Inconel 718 with coated carbide and ceramic inserts. J Mater Process Technol 129:127130 12. Lima JG, Avila RF, Abrao AM, Faustino M, Paulo Davim J (2005) Hard turning: AISI 4340 high strength low alloy steel and AISI D2 cold work tool steel. J Mater Process Technol 169:388395 13. Mukherjee I, Ray PK (2006) A review of optimization techniques in metal cutting processes. Comput Ind Eng 50:1534 14. Muthukrishnan N, Paulo Davim J (2009) Optimization of machining parameters of Al/SiC-MMC with ANOV and ANN A analysis. J Mater Process Technol 209:225232 15. Panda S, Padhy NP (2008) Comparison of particle swarm optimization and genetic algorithm for FACTS-based controller design. Appl Soft Comput 8:14181427 16. Rao RV Pawar PJ, Shankar R (2008) Multi-objective optimization , of electro-chemical machining process parameters using a particle swarm optimization algorithm. J Eng Manuf 122:949958 17. Sahin Y (2009) Comparison of tool life between ceramic and cubic boron nitride (CBN) cutting tools when machining hardened steels. J Mater Process Technol 209:34783489 18. Saravanan R, Asokan P Vijayakumar K (2003) Machining , parameters optimizations for turning cylindrical stock into a continuous finished profile using genetic algorithm (GA) and simulated annealing (SA). Int J Adv Manuf Technol 21:19 19. Sathiya P Aravindan S, Narool Haq A, Paneerselvam K (2009) , Optimization of friction welding parameters using evolutionary computational techniques. J Mater Process Technol 209(5):2576 2584 20. Senthil Kumar A, Adam Khan M, Thiraviam R, Sornakumar T (2006) Machining parameters optimization for alumina based ceramic cutting tools using genetic algorithm. Mach Sci Technol 10:471489

5 Conclusion Two hybrid network models were developed to optimise the unit production cost of alumina cutting tools whilst machining the GFRP composite. These hybrid models are investigated in terms of convergence rate and accuracy of solution for optimum cutting conditions in selected machining strategies; the following conclusions have been drawn from study: 1. The hybrid algorithms GA-NN and PSO-NN models performed efficiently, and the optimal cutting conditions predicted by the above models are almost equal. However, the PSO-NN model produces the best optimal cutting condition in terms of convergence rate and accuracy compared with the GA-NN model. 2. The incorporation of a neural network into GA and PSO produces accurate optimal conditions due to the efficient training by neural network. 3. Unit production cost of Ti[C,N] mixed alumina cutting tool is lower (approx. US $15) than the SiC whiskerreinforced alumina cutting tool (approx. US $16) whilst machining the GFRP composite. 4. The important parameter that affects the flank wear and surface roughness, which in turn affects the unit production cost, was identified using design of experiments, and the cutting velocity and the feed rate are the dominant parameters in machining the GFRP composite material using alumina cutting tools. In conclusion, the PSO-NN model exhibits better performance when compared to the GA-NN model. The PSO-NN model produces efficient and faster results than the GA-NN model. The unit production cost of Ti[C,N]

Int J Adv Manuf Technol 21. Senthil Kumar A, Raja Durai A, Sornakumar T (2003) Machinability of hardened steel using alumina based ceramic cutting tools. Int J Refractory Met Hard Mater 21:109117 22. Senthil Kumar A, Raja Durai A, Sornakumar T (2006) Wear behaviour of alumina based ceramic cutting tools on machining steels. Tribol Int 39:191197 23. Srinivasu DS, Ramesh Babu N (2008) A neuro-genetic approach for selection of process parameters in abrasive waterjet cutting considering variation in diameter of focusing nozzle. Appl Soft Comput 8:809819 24. Tansel IN, Ozcelik B, Bao WY Chen P Rincon D, Yang SY , , , Y enilmez A (2006) Selection of optimal cutting conditions by using GONNS. Int J Mach Tools Manuf 46:2635 25. Tosun N, Ozler L (2002) A study of tool life in hot machining using artificial neural networks and regression analysis method. J Mater Process Technol 124:99104 26. V enkata Rao R, Pawar PJ (2010) Parameter optimization of a multi-pass milling process using non-traditional optimization algorithms. Appl Soft Comput 10:445456 27. Yang WH, Tarng YS (1998) Design optimization of cutting parameters for turning operations based on the Taguchi method. J Mater Process Technol 84:122129 28. Zain AM, Haron H, Sharif S (2010) Application of GA to optimize cutting conditions for minimizing surface roughness in end milling machining process. Expert Systems with Applications 37:46504659 29. Zain AM, Haron H, Sharif S (2010) Prediction of surface roughness in the end milling machining using artificial neural network. Expert Syst Appl 37:17551768 30. Zhou J, Duan Z, Li Y Deng J, Daoyuan Y (2006) PSO-based , u neural network optimization and its utilization in a boring machine. J Mater Process Technol 178:1923 31. Zuperl U, Cus F (2003) Optimization of cutting conditions during cutting by using artificial neural networks. Robotics Comput Integr Manuf 19:189199 32. Zuperl U, Cus F, Mursec B, Ploj T (2004) A hybrid analyticalneural network approach to the determination of optimal cutting conditions. J Mater Process Technol 157158:8290

You might also like