Modeling and Optimizing Powder Mixed Electrical Discharge Machining (PMEDM) Process using Neural Networks

In this research, a newly integrated neural network based methodology is proposed for modeling and optimal selection of process variables involved in powder mixed EDM process. Three different kinds of fine abrasive powders, namely, copper (Cu: metal), aluminum oxide (Al2O3: metal oxide), and silicon carbide (SiC: combination of metal and non-metal), having the same particle concentration and size of 2.5-2.8 gr/lit and 45-50 μm, respectively, were added separately into the kerosene dielectric liquid. To compare and investigate the effects caused by every powder of differently thermophysical properties on the EDM process performance with each other as well as the pure case, a series of experiments were conducted on a specially designed experimental setup developed in the laboratory. Discharge current (I), pulse period (T) and source voltage (V) were selected as the independent input parameters to evaluate the process performance in terms of material removal rate (MRR) and surface roughness (Ra). Generally, of the studied additives, Al2O3 produces the greatest MRR, esp., in high currents and low pulse periods, followed by SiC, with Cu powders producing the smallest. A 3-6-4-2 size feed-forward neural network with back-propagation (BP) learning algorithm is then developed to establish the process model of the best selected Al2O3 powder mixed EDM. Training and testing of the neural model are carried out using experimental data. Having established the process model, a second network, parallelizing the augmented Lagrange multiplier (ALM) algorithm, determines the corresponding optimal input parameters by maximizing MRR subject to appropriate operating and prescribed surface roughness constraints. The optimization procedure is implemented in each level of machining regimes (finishing, semi-finishing and roughing) and the obtained optimum machining parameters settings are also verified experimentally.

Keywords:

Powder mixed EDM (PMEDM) , Back-propagation (BP) neural network , Augmented Lagrange multiplier (ALM) algorithm , Modeling , Optimization