S Assarzadeh - Ph.D. Student, Specific Machinings Laboratory (S.M.L.), Department of Mechanical Engineering, K.N. Toosi University of Technology, P.O. Box: ۱۹۳۹۵-۱۹۹۹, Tehran, Iran,
M Shamsi - B.Sc. Graduate, Mashhad Technical College, P.O. Box: ۹۱۷۳۵-۱۷۱, Mashhad, Iran
M Ghoreishi - Associate Professor, Department of Mechanical Engineering, K. N. Toosi University of Technology,P.O. Box: ۱۹۳۹۵-۱۹۹۹, Tehran, Iran

In the present study, a unified coupled artificial neural network-genetic algorithm (ANNGA) based methodology is proposed for the modeling and optimal selection of process parameters involved in electrical discharge machining (EDM) process. Firstly, based onthe prior knowledge of the physics of process mechanism as well as the EDM machine characteristics, the controllable input variables were identified as the discharge current (I),pulse on-time (Ton) and pulse off-time (Toff). Having conducted a set of preliminary and screening experiments to specify the stable working domains of these inputs, a 43 full factorial design scheme was employed to plan the training experiments of the backpropagation neural network model. The experiments were done in the planing mode with both flat tool and work piece electrodes having equal diameters. Heat treated CK45 carbon steel up to 60 RC and commercial cupper were used as work piece and tool materials, respectively. Material removal rate (MRR) and surface roughness (Ra) were chosen as the response outputs. There-upon, a 3-8-6-2 size feed forward back propagation neural network was developed to establish the process model. Testing the generalization capabilities of the model by a set of new data not being used in the training phase, the second stage begins by combining the genetic algorithm technique with the neural model to find optimum machining conditions leading to the highest possible MRR.The multi-objective optimization problem was categorized as finishing (Ra ≤ 4μm), semifinishing (Ra ≤ 8μm), and roughing (Ra ≤ 12μm) regimes, from which a group of optimal input parameters were obtained in each case, satisfying the Ra constraint and yielding the maximum MRR, simultaneously. Finally, the modeling and traded-off selected optimal results were also interpreted and verified experimentally to validate the adopted approach. The observed errors are all in acceptable ranges (less than 10%) which show a good agreement with simulation results and confirm the feasibility and effectiveness of the implemented strategy.

Electrical Discharge Machining (EDM), Back propagation neural network (BPNN), Genetic algorithm (GA), Process modeling, Optimization