Study On The Establishment Of Concrete Failure Criterion Based On BP Neural Network

Research on the mechanical properties of concrete materials is of great significance to the safety of building structures.Both raw materials and production processes have complex effects on the mechanical properties of concrete.The traditional research method is to simplify the influencing factors,establish mathematical models,and obtain theoretical formulas or empirical formulas through complex derivations and calculations to provide guidance and reference for engineering applications.Neural network is an abstract and bionic algorithm for the human brain.It is good at fitting complex nonlinear mappings,but lacks interpretability.In practical applications,it is mainly to increase training samples,adjust parameters,and compare to evaluate its accuracy.The combination of mechanical analysis and neural network not only avoids the complicated analytical solution derivation and calculation process when establishing mathematical models,but also makes up for the inexplicability weakness of neural network to a certain extent.At present,there have been many successful cases of applying neural network to study the mechanical properties of concrete,which proves the research prospect of this method and has formed a fixed application mode.This paper takes the underlying logic of the neural network algorithm as the starting point,combines the mechanical analysis and the existing experimental data,and applies the neural network to the study of the mechanical properties of concrete,which provides a new thinking direction for the research methods of the mechanical properties of concrete.The main research contents of this paper are as follows:(1)Summarize the results of the existing neural network and the research on the mechanical properties of concrete,summarize the structure and algorithm principles of BP neural network,convolutional neural network,support vector machine,radial basis function network,and The application methods of these neural networks in concrete strength,fracture performance,crack identification and durability are summarized.(2)A method of applying the batch gradient descent method to fitting the parameters of the failure criterion of concrete under the multiaxial stress state is proposed.Due to the systematic error in the mathematical equation model of the simplified failure criterion,this method can obtain the failure surface with the smallest error to the maximum under the limitation of the function construction of the mathematical model.(3)Based on the analysis of the failure mechanism of concrete,two failure criterion models of BP neural network are established.These two failure criterion models are based on the assumption of the failure mechanism of concrete.One is the failure criterion model established based on the shear stress failure assumption,and the other is the failure criterion model established based on the failure morphology assumption.These two BP neural network damage criteria are approximately coincident with the classic Ottosen damage criteria,which proves the accuracy of the BP neural network damage criteria model and the effectiveness of this modeling method.