Structural Reliability Analysis Method And Application Based On Deep Neural Network

Uncertainty exists widely in engineering structures.Reliability analysis is an important method to ensure the safe operation of structures.Surrogate model is an effective method for structural reliability analysis and is also a research hotspots in the field of structural reliability.The conventional surrogate model methods for structural reliability analysis have the problems of local optimal in experimental point selection and insufficient approximation ability of high-dimensional nonlinear limit state functions,which limit their application in large-scale engineering structure systems.In recent years,deep learning theory has been developed rapidly and has been successfully applied to many fields.In this thesis,study on deep neural networkbased structural reliability analysis method and application is carried out based on the ability of deep learning methods to learn complex problems and the approximation ability of deep neural networks for high-dimensional nonlinear functions.The main research contents in this thesis are as follows:(1)In this thesis,a deep reinforcement learning-based sampling method for structural reliability analysis is proposed,and a deep neural network surrogate model method for structural reliability analysis based on the deep reinforcement learningbased sampling method is established.By treating the random variable space and the experimental points as the input state and output actions of deep reinforcement learning,and using a deep neural network as the agent of deep reinforcement learning,the sampling framework for structural reliability analysis based on deep reinforcement learning is established.First,the optimization learning method for sampling in the important areas near the structural limit state surface based on the interactive learning between the agent and the environment is proposed.According to the purpose of the deep reinforcement learning-based sampling,a reward function based on the predicted limit state function value of the sample is proposed.Second,the deep neural network surrogate model method for structural reliability analysis is established based on the experimental points slected by the deep reinforcement learning-based sampling method.Finally,the reliability analysis of two numerical examples is carried out to test the efficiency and accuracy of the proposed method.(2)An active learning method for structural reliability analysis based on weighted sampling and deep neural network is proposed.First,an iterative learning framework is established to successively select candidate samples and experimental points to update the deep neural network surrogate model.Second,an iterative update method of the changing thresholds is developed to select important candidate samples from the Monte Carlo samples,and a weighted sampling algorithm is proposed to select uniformly distributed experimental points from the candidate samples.To ensure the uniformity of the selected experimental points in the random variable space,a calculation method of the weight coefficient based on the probability density function of the Monte Carlo sample is given.Finally,the efficiency and accuracy of the proposed method are tested by numerical examples and the reliability problem of an actual large-scale cable-stayed bridge structure.(3)An adaptive subset searching method is proposed,and a deep neural network surrogate model method for structural reliability analysis based on the adaptive subset searching method is established.First,an iterative searching strategy for the nested subsets based on the conditional probability is developed.An evaluation method of the prediction accuracy of the surrogate model on the current subset based on the predicted value of the conditional probability is proposed.Second,according to the prediction accuracy of the surrogate model on the current subset,an adaptive threshold update method for subset searching is proposed to balance the exploration and exploitation of the adaptive learning.Finally,numerical examples and the structural reliability problem of an actual cable-stayed bridge are used to test the ability of the method to deal with the local optimal sampling problem.(4)An adaptive filter sampling method is proposed,and a deep neural network surrogate model method for structural reliability analysis based on the adaptive filter sampling method is established.First,an adaptive filter sampling algorithm is proposed.The Latin hypercube sampling algorithm is used to generate a filter with many uniformly distributed filter points in the random variable space.Using the filter points,the uniformly distributed experimental points are then filtered out from the Monte Carlo samples based on a distance-based clustering method and an infinite norm-based filter region determination method.An adaptive filter sampling algorithm based on iteratively updating the number of filter points is proposed to select the required number of experimental points.Second,the deep neural network surrogate model method for structural reliability analysis based on the adaptive filter sampling is developed.Finally,the calculation efficiency of the proposed method in highdimensional and nonlinear problems are demonstrated using numerical examples,the structural reliability problem of an nonlinear a steel frame and the structural reliability problem of an actual cable-stayed bridge.(5)A deep neural network-based method for structural reliability analysis with hybrid uncertainties is proposed.First,the framework of the deep neural network for structural reliability analysis with hybrid uncertainties is proposed.Second,a deep neural network is constructed as the surrogate model of limit state function in the entire random variable space and interval variable space and another deep neural network is designed as the prediction model of failure probability in the interval variable space.The method for structural reliability analysis with hybrid uncertainties based on the prediction model of failure probability is further proposed.Finally,the calculation efficiency and accuracy of the proposed method are demonstrated using numerical examples of reliability analysis with hybrid uncertainties and the reliability problem of an actual cable-stayed bridge structure with hybrid uncertainties.(6)Considering an actual large-scale cable-stayed bridge structure with a health monitoring system,the application research of structural serviceability reliability analysis of the main girder of the cable-stayed bridge based on monitoring data is carried out.First,the corrosion prediction model of the cable steel wire is established based on the monitoring data of the cable stress and the percent relative humidity.The100-year maximum probability distribution model of the lane load is established based on the monitoring data of the weight-in-motion system.The on-site inspection data of the main girder cracks is used to establish the probability distribution models of the length of different kinds of cracks.Second,the reliability analysis method for the cable-stayed bridge structures based on monitoring data is proposed.Finally,the serviceability reliability analysis of the main girder of the cable-stayed bridge structure based on monitoring data is carried out.Three test conditions,including normal condition,with cracks and without corrosion,and without crack and with corrosion,are used to study the influence of the cable corrosion and the main girder cracks on the serviceability reliability of the main girder structure of the cable-stayed bridge.Considering the different failure modes caused by the displacement of different sections of the main girder exceeding the serviceability limit state,the failure probability of each failure mode and the failure probability of the main girder system are calculated,and the correlation between different failure modes of the main girder is analyzed.