Structural Finite Element Model Updating Based On Deep Learning

Since the introduction of the finite element method,it has been widely used in various fields and has been highly valued by the engineering community.Building an accurate finite element model for structural response analysis is extremely important for structural design and evaluation.For example,in structural health monitoring,the structural response is often monitored at only a very few locations.However,the assessment and early warning of monitored structures often require a large amount of structural response information,and a model that can predict the structural response with sufficient accuracy is necessary.Finite element model updating is to correct the finite element model by the measured information,so that the finite element model can reflect the behavior of the actual structure more accurately.Along with the widely application of finite element analysis and the increasing attention to structural health monitoring,the need for this technology is becoming more and more urgent.In recent years,deep learning technology has developed quickly,but there is still a lack of research on the finite element model updating of deep learning,so it is necessary to carry out the research in this area.This paper adopts deep learning technology for finite element model updating research,which mainly includes: a simple recurrent neural network-based finite element model updating method and a Bi-LSTM network-based surrogate model for finite element models.Around the problem of finite element model updating using neural network,the selection of structural features is analyzed and using a method that considers multiple correlation coefficients to select the appropriate model updating parameters.Simple recurrent neural network is proposed for finite element model updating.The construction method of a simple recurrent neural network is described,and suitable activation and loss functions are selected according to the model updating problem.The viability of this method is demonstrated by a two-span continuous beam example.In addition,the tower crane structure is used as the finite element model updating object.The frequency and mode shape are used as structural features and the structural material and member geometry parameters are used as correction variables.Deep learning model for tower crane finite element model updating was trained by generating samples from finite element analysis.And the results of the finite element model updating of tower crane using simple recurrent neural network,BP neural network and one-dimensional convolutional neural network are compared to verify the effectiveness and superiority of the deep learning finite element correction model proposed in this paper.A Bi-LSTM network-based alternative to the complex finite element model updating process is proposed,which predicts the response of the location of interest directly from the measured response.Taking wind-induced vibration response prediction of tower crane structure as an example,a multidimensional pulsating wind speed simulation was implemented using AR model,and then the tower crane response under random wind field was calculated.Combined with the stochastic search method,a Bi-LSTM network code was written to automatically optimize the hyperparameters of the network,and the response prediction of the location points of interest was achieved by Bi-LSTM.The crane response prediction model based on Bi-LSTM network was obtained by training the single-point and two-point response samples of the tower crane,respectively.It is verified that the Bi-LSTM network has superior prediction accuracy and high computational efficiency for the response at any position of the tower crane.