| Metal multiaxial fatigue is the most common and complex form of failure in the field of engineering,under the action of multiaxial fatigue load,the fatigue failure mechanism of metal materials and structures is complex,and it is difficult to predict multiaxial fatigue life.At present,the field of machine learning combined with the field of multi-axis fatigue alleviates the above problems to some extent,because the multi-axis loading path has an important and direct impact on life prediction,machine learning cannot directly accept input from the loading path,because the path is graph data or functional relationship,and insufficient multi-axis experimental data may cause problems when training the data of the built machine learning model.In view of this,in order to ensure the service safety of engineering structures,the above problems in the combination of machine learning and multi-axis fatigue are solved.In this study,a model based on image recognition and machine learning and a transfer learning model are built to solve the problem of insufficient loading path and data volume.The construction process is as follows: first,the influencing factors are analyzed to establish the features,the problem that the loading path cannot be directly used as features is solved through image recognition,and a two-layer random forest model with two-layer features is constructed from numerical and non-numerical data.Second,on the original model,the model in transfer learning fine-tunes the advantages of the migration source model,and the constructed model is adjusted to obtain the final model.The main research results are as follows:1.In order to build the machine learning prediction model to establish its feature input,the metal test data is sorted out,and the effects of strain amplitude,amplitude ratio,phase difference and loading frequency on fatigue life in the test data are analyzed.Finally,the normal strain,shear strain,and loading path are established as the input features.2.The multi-axis fatigue loading path cannot be directly used as input in machine learning,and the phase difference and loading frequency related to the loading path can only be approximated,which may affect the built machine learning model.Therefore,by establishing an image recognition model,the entire loading path is extracted and converted into one-dimensional data for input features.3.In order to realize the combination of numerical data features and non-numeric data features to form the feature input of the machine learning model,that is,the tangent amplitude and the loading path are united,a two-layer feature structure is proposed as the input feature of the random forest,the tangent amplitude can be directly input into the first layer of features,and the loading path is extracted and the second layer of features is entered.The trained model fits about 96%,and the prediction accuracy and generalization accuracy conform to the recognized two-fold error band range of multiaxial fatigue.In addition,due to the two-layer feature,the model has a certain ability to extrapolate the loading path,and the fatigue life of the loading path that is not in the model training and testing can be predicted within a certain range.4.In order to solve various problems that may occur in the model due to the small amount of data,a migration model is proposed.This model first transfers the knowledge of the model trained by the Smote algorithm generated by a large amount of data,and at the same time learns the knowledge in the experimental data training.The final results show that the transfer model is cross-verified under different training sets,indicating that the model is stable,and the prediction accuracy is improved from 2x error band to 1.5x error band. |
