Neural network approximation of deformation curves under uniaxial tension of steel and silumin specimens

Abstract

The purpose of the study is developing and testing of the new computational technique for approximation of deformation curves of steel and silumin specimens under uniaxial tension. A scheme of testing steel and silumin specimens for uniaxial tensile is presented. The experiment was carried out in the mechanical testing laboratory of the Department of Applied Mathematics and Mechanics of the Voronezh State Technical University. The experimental deformation curve of a steel specimen was approximated by P. Ludwig’s equation. Prediction of the true stress from the logarithmic strain using a pretrained artificial neural network with a multilayer perceptron architecture is discussed. The neural network model was trained using the RProp (resilient backpropagation) method. The software implementation of the neural network approximation was carried out in a framework of the open source for data analysis — Knime Analytics Platform. A scheme for the implementation of a multilayer perceptron that solves the approximation problem is considered. The simulation results are compared by the values of the mean squared error (MSE) of the approximation. The neural network approximation is turned out to be an order of magnitude more accurate for the steel specimen than the approximation by the P. Ludwig equation. The neural network approximation provided even a smaller MSE value for a silumin specimen than that or a steel specimen. It is revealed that changing the architecture of an artificial neural network affects the quality of modeling. With an increase in the number of hidden layers, the accuracy of the approximation increases. Neural network approximation is an effective approach to solving the problem of the analytical description of experimental deformation curves and leaves the possibility of using a universal technique for a variety of materials and different types of tests.