Research And Application Of Mechanical Surrogate Model Based On Deep Neural Network

In various practical engineering problems,numerical simulation has become a common technology for structural mechanical property analysis and checking calculation.With the increasing complexity of research problems,the calculation method based on structural physical model can obtain more accurate simulation results,but it needs a lot of time;Moreover,in the tasks that need repeated simulation,such as optimization design and sensitivity analysis,the existing calculation results can not be reused,resulting in large amount of calculation and low efficiency.The deep neural network surrogate model is a data-driven approximate model,which can effectively reduce the amount of calculation,and realize the reuse of knowledge through transfer learning,which is helpful to solve complex mechanical problems quickly.Due to its strong data processing ability,deep neural network has attracted attention from all walks of life,especially for image,voice,text and other data.It can obtain the internal relationship between input and output through learning,and has good applicability for complex classification and regression problems.Compared with the traditional surrogate model,deep neural network can adaptively extract data features on different scales,and can effectively deal with the nonlinear mapping problem between high-dimensional data.Therefore,this paper proposes a method to establish the depth neural network proxy model for mechanical calculation,and puts forward the improvement scheme of loss function and network structure according to the test results of thin plate bending example.Then,it does some research on the construction technology of neural network model under small samples,and puts forward the migration learning and data enhancement methods for the model and data set,which are applied to the parameter analysis of gasket.The specific work is as follows:(1)Based on the basic theory of deep learning,combined with u-net network and residual network,a proxy model method of deep neural network is proposed.The feasibility of predicting structural response by deep neural network is verified by a plane structure example.Aiming at the local unsmooth phenomenon of the result,it is improved by increasing gradient loss and modifying the up sampling method;Combined with genetic algorithm,the performance of the model in the optimization problem is tested.The results show that the deep neural network model can meet the needs of calculation accuracy in structural optimization and greatly improve the optimization efficiency.(2)Because it is difficult to produce a large number of high-quality samples in practical application,this paper uses transfer learning and data enhancement technology to establish deep neural network on small sample data set.The model reuse is realized by using the transfer learning technology.By transferring the knowledge learned by the network model in the approximate task,the problem of low generalization ability of deep neural network model under small samples is effectively solved;In the absence of portable model,according to the sample characteristics of the example and combined with the existing data enhancement methods,this paper gives two data enhancement schemes,so that the accuracy of the model can be close to the network model established when there are enough samples.(3)Aiming at the influence of shield gasket manufacturing error on its waterproof performance,this paper uses the surrogate model based on depth neural network to predict the contact stress on the upper surface of the gasket;The Latin hypercube sampling method is used to generate the sequence of structural parameters,and combined with the Spearman rank correlation coefficient,the influence degree of each parameter on the waterproof performance index is calculated.The results show that the surrogate model can accurately predict the contact stress distribution on the upper surface of the gasket,and the sensitivity analysis results are in line with the actual situation,which has a certain reference significance for the design and processing of the gasket.