| With the intensification of global energy issues,high-strength steel is gradually being widely used due to its excellent performance.The use of high-strength steel can ensure the overall strength of automobiles while reducing their own mass.The complexity of the microstructure of high-strength steel also leads to a more complex deformation behavior when plastic strain occurs.The evolution behavior of material damage is a multi-scale change process with obvious path correlation,which is non-uniform and dynamic on the time axis.In addition,advanced optical and other testing methods can provide more calibration information for damage testing,thereby utilizing machine learning methods to mine statistical features on the time axis from mechanical and optical information,achieving precise modeling and calibration of damage evolution,and incorporating multidimensional time series information to solve the uncertainty of damage evolution models and calibration caused by traditional information loss.Therefore,based on the feature that convolutional networks can automatically extract features,this article constructs three convolutional network structures using different data to achieve the calibration of material damage parameters,which is of great significance.This article uses finite element software to provide a dataset for convolutional networks using a GTN model that considers material micro pores,and designs three convolutional calibration network models that use data from different dimensions.The main content of this article is as follows:(1)Using finite element software to model GTN damage model,using sobol sequence to generate 4800 sets of damage parameters based on experimental data,and using Python language to write scripts,the finite element software can batch assign parameters to the pre processing of the simulation process,and use scripts to post process the result files to obtain a large amount of load data and strain cloud maps as datasets.(2)Three convolutional calibration network models were established using the Tensorflow framework,namely a load reverse calibration network using one-dimensional load data,an image reverse calibration network using three-dimensional strain cloud images,and a hybrid reverse calibration network using both types of data simultaneously.Design different network structures using data from different dimensions containing temporal information to predict damage parameters.(3)Visualize the convolutional calibration network and analyze it from two perspectives: the training process and the training results.Visualize the network attention mechanism of the training results using the Grad-CAM algorithm,and generate a class activation graph to analyze the impact of the temporal and spatial distribution of data on damage parameters.Analyze the loss values of the training results and evaluate the accuracy of three calibration networks in predicting damage parameters. |
