E. V. Parfenov, A. R. Fatkullin, D. M. Lazarev, A. L. Yerokhin
The increase of requirements to a workpiece treatment quality, continuous rise in cost of materials and energy carriers lead to intensive development of advanced technologies which allow receiving the required complex of surface properties directly during technological processes. In such conditions, the application of modern information technologies becomes an important factor in processes control and production quality management.
The technological process of electrolytic plasma treatment is characterized by nonlinear dependences between the input and output variables; therefore, it can be classified as a complex, nonlinear and multi input, multi output object which operates under uncertainty conditions in the parameters of the surface properties, the electrolyte composition and the power supply.
The models of plasma electrolytic oxidation of aluminium are considered in this research. These models are created using experimental data, on the basis of both regression analysis, and supervised neural networks of the various types. Two commonly used supervised network structures – a multilayer perceptron and a network with radial basis functions were chosen. The regression equations of the second and the third order with pair interactions were chosen for the regression modelling. The regression equation coefficients were calculated using least squares method.
Neural set structure having one hidden layer, 2 inputs and 2 outputs has been chosen for the multilayer perceptron models. The number of neurons in the hidden layer was varied from 4 to 30. The network was trained using Levengerg-Marquart backpropagation algorithm. General regression neural network having the similar structure was chosen for modelling with radial basal function networks. This type of networks was trained by increasing the number of the neurons in the hidden layer.
After the comparative analysis it was found out that for modelling the processes of electrolytic plasma treatment the application of neural networks with radial basis functions provides not only adequate approximation of the response surface with adjustable accuracy, but also smooth and monotonous interpolation. The degree of adequacy and monotony can be adjusted both by varying the spread of the radial basis function and by changing the number of training examples. The models obtained can be used as reference models within automated control systems for plasma electrolytic treatment processes.