Prediction Of Crack Propagation Process Of The Pre-notched Plate Model Based On Neural Network

Peridynamics(PD)is a non-local theory that constructs equations of motion in the integral form to simulate cracks that originate in multiple locations and extend along arbitrary paths.However,the peridynamic theory has the problems of high computing complexity and long simulation time.With the current computing power of computers,it is difficult to visualize and predict the crack and damage state in real-time.However,with the development of theories deep learning,practical solutions can be provided for the above problems.In the research,A method is proposed for visual simulation and prediction of crack and damage states based on the GAN model,to solve the problem of long computation time for the peridynamic model.For purpose of calculating the damage value at any position in the model,a method based on BP(Back-propagation)neural network is proposed to fit the functional relationship between the R,G,B values and the damage value,and the effectiveness of the above method is verified based on the pre-notched plate model.The main content of the research is divided into three parts:(1)Construction of datasets based on the peridynamic algorithm.The dataset is divided into two parts,the first part is a damage map dataset that represents the crack or damaged state.The process of producing this dataset: based on the pre-notched plate model,numerical results are obtained by the discontinuous Galliakin finite element method,and the visualization software is used to transform the numerical results into a damage diagram,and then the damage map dataset is constructed by combining the corresponding loading conditions.The second part is the damage value dataset constructed by combining the R,G,B values in the damage diagram with the damage values.(2)Construction of the crack propagation prediction model,is divided into two parts,one is the crack prediction model and the other is the damage value prediction model.The crack prediction model is used to predict the crack or damage state of the pre-notched plate model.Also,to accelerate the convergence of the crack prediction model and improve the quality of the image results,algorithms such as WGAN(Wasserstein GAN),WGAN-GP(Gradient penalty),IN(Instance Normalization),and BN(Batch Normalization)are introduced into the crack prediction model for training.To achieve supervised learning,Euclidean distance is introduced to guide the direction of the crack prediction model generation.The crack prediction model obtains the image results and cannot obtain the damage value for numerical analysis.In the research,a damage value prediction model is constructed to calculate the damage value based on the R,G,B value information.(3)The results obtained by the method proposed in the research are compared with the results obtained by the peridynamic algorithm to verify the effectiveness of the method.The research compares the computational time and results of the crack propagation prediction model and the peridynamic algorithm to validate the effectiveness of the proposed method.The results show that:(1)According to the loading conditions,the crack propagation model can visualize and predict the crack expansion process and damage state of the pre-notched plate model in real-time.(2)The crack prediction model can accurately predict the crack propagation shape and damage state of the pre-notched plate model under different loading conditions,as well as the trend of the damage value in the model.(3)The damage value prediction model can accurately calculate the damage value from the R,G,B values information of the pixel.The model can be used to calculate the damage value at any location in the results of the crack prediction model.