Fatigue Damage Evolution And Reliability Analysis Method Of Orthotropic Steel Bridge Deck Based On Machine Learning

Orthotropic steel bridge decks are widely used in long-span steel bridges,but its severe fatigue problems have brought many difficulties to bridge maintenance.At present,most of the relevant researches focuses on single fatigue detail,which cannot characterize the fatigue state of the bridge deck system.This thesis systematically studies the uneven spatial distribution and sensitivity evaluation method of fatigue cracks on orthotropic steel bridge decks,the dynamic Bayesian network model for fatigue crack propagation analysis of full-bridge deck,the efficient reliability analysis method for local bridge deck based on Kriging model and subset simulation method,the reliability analysis method and maintenance decision-making method based on deep reinforcement learning for full-bridge deck system.A full-bridge fatigue reliability evaluation method is established from single to full-bridge fatigue cracks,from static distribution to dynamic propagation,and from local to global.Main contents of this thesis are as follows:A Bayesian network-based method for evaluating the uneven spatial distribution and sensitivity of fatigue cracks is proposed.The bridge deck is divided into finite research units.The state variables of units are defined according to the structural configuration,load environment,and fatigue state.Based on the Bayesian network,the joint probability distribution of the unit state variables is established.Parameter estimation,model validation and sensitivity analysis of the established Bayesian network are performed based on inspection results of fatigue cracks.A spatio-temporal analysis method of fatigue crack growth of full-bridge based on dynamic Bayesian network and hierarchical model is proposed.The multi-scale finite element model of long-span bridge and the stochastic traffic flow model based on monitoring data are established.Correlations of propagation parameters of different fatigue cracks are considered.Based on the hierarchical model and the dynamic Bayesian network,the stochastic process model of full-bridge crack propagation is established.The posterior probability distribution of equivalent initial crack size is obtained through the anlysis of filed inspection data.The effectiveness of the proposed method is verified through numerical simulation.An efficient structural reliability analysis method is proposed based on Kriging surrogate model and subset simulation method.An active learning sampling strategy of Kriging model is designed based on the misclassification probability of samples.A two-stage convergence criterion is designed to improve the analysis accuracy and efficiency,in which the hierarchical error and cumulative error of each conditional failure probability are controlled.The effectiveness of the proposed method is verified by four numerical examples.A deep reinforcement learning method for reliability analysis and maintenance decision-making of full-bridge orthotropic steel deck system is proposed.The bridge deck system is modeled as a two-dimensional redundant system,and an exact Markov chain based method is established for calculating the system reliability.A deep reinforcement learning method for maintenance decision-making is proposed with a reward function designed by cost and risk.The effectiveness of the proposed method is verified under the condition of linear and nonlinear degradation of local reliability.