Research On Machine Learning Method For Long-span Bridge Structural Health Diagnosis And Structural Maintenance Policy Making

Long-span bridge plays an import role in the transportation network among cities and regional economics.However,in the course of long-term service,long-span bridges are inevitably degraded due to the coupling effects of environmental erosion,reciprocating loads and sudden disasters.In extreme cases,they may even lead to catastrophic accidents,which seriously threaten our cities,which threatens the sustainable development of the urbanization process in China.Rational and scientific bridge lifecycle management has attracted worldwide concern.Structural Health Monitoring(SHM),which collects information of structural loads and responses via various types of sensors located on the structures,provides a method of online structural health diagnosis to assist the policy-making of bridge management.Currently,structural health diagnosis problems are commonly viewed as dynamic inversion problems,and system identification methods such as vibration-based modal identification and model update methods are frequently performed.However,the incomplete data collected from the limited quantity sensors makes the dynamic inversion equations ill-posed,and frequency-based features are usually not sensitive to local damage.Classic structural diagnosis methods encounter bottlenecks.Lots of significant bridges in China have installed SHM systems in the last two decades,and theses systems have collected big data that contains structural performance information of bridges.However,most data is wasted due to the lack of standard and efficient data-processing and data-mining methods.Therefore,how to efficiently mine the SHM data,how to discover the structural performance information from the data mining,and how to make structural health diagnosis based on the monitoring data for scientific bridge lifecycle management policy-making have been the most important and frontier topics.Therefore,this study presents machine learning methods for structural health diagnosis and lifecycle management policy-making based on the SHM data.Main contents involved are as follows:A statistical pattern recognition method for crack identification of the orthotropic steel deck(OSD)of long-span bridges is proposed.Firstly,the data feature of the strain response of OSD under vehicle loading is analyzed.The discrete wavelet transfer(DWT)method is employed to decouple the multi-source effect and the wheel-induced strain of high signal-to-noise ratio and the strong local property is separated.Then,the temporal-spatial correlation model of wheel-induced strain based on dynamic time warping(DTW)is established and the damage sensitive feature of the wheel-induced strain ratio is proposed.Finally,a statistical pattern recognition method for crack identification of OSD based on the deterioration of the statistical model of the wheel-induced strain ratio is proposed and is tested effectively on the SHM data collected from a long-span cable suspension bridge.A statistical pattern recognition method for condition assessment of stay cables of long-span cable-stayed bridges is proposed.Firstly,the data feature of the cable tension response is analyzed.Environmental effects like temperature are eliminated via the median value of cable tension history segmentations and vehicle-induced cable tension remains.Secondly,a linear model of the vehicle-induced cable tension of the upstream/downstream cables in the same cable pair is proposed based on the cable tension influence line,and damage sensitive feature of vehicle-induced cable tension ratio is proposed.Gaussian Mixture Model(GMM)is employed to model the distribution of the vehicle-induced cable tension ratio and expectation-maximization(EM)algorithm is employed for parameters estimation.Finally,the qualitative method and the quantitative method are proposed and tested the effectiveness based on 10 years of cable tension monitoring data collected from a long-span cable-stayed bridge.An LSTM-based self feature learning method for structural condition assessment of stay cable group is proposed.Considering that the performance of the statistical pattern recognition methods for structural health diagnosis highly rely on feature extraction,this study employs deep learning(DL)models that are able to learn the complex nonlinear feature of the correlation of the data in an unsupervised end-to-end learning way.LSTM is employed to model the temporal-spatial correlation of the cable tension of cables is the stay cable group and is trained on the monitoring cable tension data in the initial period.Predition error of the vehicle-induced cable tention of cables in the group is employed as the condition instructor of the cable group.Monitoring cable tension data of a long-span cable-stayed bridge is employed to test the effectiveness.A general deep reinforcement learning-based bridge lifecycle maintenance policy-making method is proposed.The lifecycle structural maintenance is viewed as the control problem that maintains the structural condition at some acceptable level by maintenance actions under some constraints(e.g.,structural risk limit,serving time limit,and maintenance budget limit).Based on the Markov decision model(MDP)a deep reinforcement learning method is proposed,in which the structural deterioration model is treated as the environment,the mai ntenance costs and structural condition improvement is treated as the reward,and the m aintenance actions are treated as actions.The proposed method is a general method that can be used in structures of various complexity with little change.Examples of a deck system of the simplified supported bridge and a cable-stayed bridge are conducted.