Research On Multiaxial Fatigue Life Prediction Model Based On Fully Connected Neural Network

Metal fatigue is one of the most common failure forms in the engineering field,and multiaxial fatigue is a common form of metal fatigue,because the components in actual fatigue service are usually in a multiaxial load state.Under the action of multi-axial fatigue load,the fatigue failure mechanism of metal materials and structures is complex,and it is difficult to accurately predict multi-axial fatigue life.In order to ensure the service safety of engineering structures,it is extremely critical and urgent to establish a multi-axis fatigue life prediction model.In order to solve the complex problem of multi-axial loading paths and establish an efficient and universal fatigue life prediction method,firstly,the relationship between multi-axial load variables and fatigue life is analyzed based on the multi-axial fatigue prediction model based on the physical mechanism and the characteristics of multi-axial loading paths.;Secondly,on the basis of the above load factor analysis,the input features of the neural network that can describe the multi-axis fatigue load characteristics are further determined,and the mapping relationship between the input features and the fatigue life is constructed based on the fully connected neural network.The neural network multi-axis fatigue life prediction model is trained,validated and generalized.The specific research results are as follows:1.Collect and organize multi-axial fatigue test data of 14 kinds of materials,with a total of 523 data points.Materials include ordinary carbon structural steel,alloy steel,aluminum alloy,titanium alloy and cobalt-based superalloy;test control methods include stress and strain control;test conditions include uniaxial tension and compression,torsion,and in-phase multiaxial proportional loading And out-of-phase multi-axis non-proportional loading,and includes complex non-proportional loading paths corresponding to 6 frequency ratios;test loading waveforms include sine wave,triangle wave,trapezoidal wave,square wave and triangle-trapezoidal combined wave.2.Combine the physical mechanism model of multi-axial fatigue life prediction and the multi-axial loading path characteristics to analyze the relationship between multi-axial load variables and fatigue life;determine the seven fatigue loads of the neural network model with the goal of building a model that can reflect the multi-axial load characteristics Input features(axial amplitude,tangential amplitude,axial mean,tangential mean,phase difference,loading waveform,and frequency ratio),among which,for the needs of neural network operations,specific numerical values ??are used to refer to a certain waveform The method is used to convert the data type,and the phase difference is converted into the data form according to the changing law of the loading path.Finally,a multi-axis fatigue life prediction model is established based on the fully connected neural network.3.Analyze the rationality and usability of the model from the four dimensions of residual statistics,fitted R value,prediction accuracy and generalization ability,and illustrate this model by comparing it with the prediction of another machine learning method,the support vector regression model.The data processing methods and sample selection are in line with the specifications.The results show that: the residual statistical graph of the model shows normal;the average fitted R value reaches 0.9;all the predicted values??fall within 2 times the error,and 92% of the predicted values ??fall within the 1.5 times error band;and the comparison results show that the neural network The fitting ability and generalization ability of the network model are better than the support vector regression model.4.Comparing with the four traditional life prediction models based on physical mechanism,it is concluded that the neural network model is superior to the traditional physical model in the four evaluation indicators of prediction stability,dispersion degree,error band ratio and relative error.