Comparison of Neural Network and Multilayered Approach to the Problem of Identification of the Creep and Fracture Model of Structural Elements Based on Experimental Data

Abstract

The paper considers the solution of one class of coefficient inverse problems connected with the identification of models describing the process of inelastic deformation of structural elements under creep conditions up to fracture moment. Systems of ordinary differential equations are used to describe the creep process. Some structural materials, such as metals, concrete and composite materials have creep properties at high and moderate temperatures. Wherein, the non-consideration of material creep can lead to significant errors in the structures deformation-strength characteristics determining, that leads to the emergencies. However, creep modeling involves some difficulties. There is no general creep theory describing all or most of the observed phenomena. Dozens of different creep theories have been developed applied to specific narrow classes of problems. Moreover, constitutive equations of each theory contain sets of material constants determined by the results of the experiment. Traditional methods of creep models identification depend both on the type of constitutive equations, and the conditions under which construction works. For identification of creep models parameters, the authors proposed a general unified method, which is based on the technique and principles of neural network modeling. A new multilayered method is developed in addition to using the traditional neural network approximations with activation functions of a special type. In the new approach, the problem is preliminarily discretized using known numerical schemes on a segment with a variable right boundary. The advantages of the proposed approaches compared with the traditional ones are shown by the example of creep model identification for a uniaxial tension of steel 45 cylindrical samples under creep conditions. The reliability of the obtained data is confirmed by a comparison with the experimental data and the results of other authors.