Machine Learning Method For Correlation Pattern And Health Diagnosis Based On Monitoring Quasi-static Responses Of Long-span Cable-stayed Bridges

Transportation is a solid guarantee for people’s safe travel and economic society operation among which bridges turn traffic gaps such as rivers and canyons into thoroughfares.However,long-span bridges carry heavy loads,serve in harsh environments,and bear complex force,which seriously threaten the service safety of bridges.Structural health monitoring is widely used in long-span bridges to perform health diagnosis by monitoring environment,loads and structural responses.Since initial residual stress and defaults,loads such as vehicle and temperature,and structure damage jointly affect the structural responses,monitoring data contain comprehensive information of bridge state,environment and loads.Bridge health state cannot be directly diagnosed from the structural responses.How to mine the features which are not related to environment and loads but related to structural parameters is one of the core scientific problem in structural health monitoring.Investigating correlation patterns between responses which are only related to structural parameters is a starting point.Based on the advantages of machine learning on mining data-related features,this paper explores the correlation patterns of monitoring data to eliminate loads and environmental impacts,extract structural information,discover new health indicators,and research health diagnosis methods based on correlation patterns of monitoring data.Main research contents of this paper are as follows:(1)Correlation patterns and health diagnosis method for vertical deformation of long-span cable-stayed bridge is proposed.Firstly,the cross-section of double-sided cable-supported bridge is simplified as a simply supported beam model,the correlation feature of the vertical deformation along the transverse direction is analyzed,the vehicleinduced vertical deformation ratio between upstream and downstream is proposed as an indicator,clustering and EM algorithm are adopted to recognize the vertical deformation ratio on different lanes as an indicator for bridge health diagnosis.Secondly,bridge influence lines are analyzed,and the integral ratio of absolute vertical deformation is proposed as an indicator,which is only related to the bridge structural parameters and independent of loads.Thus a new indicator for bridge health diagnosis is discovered.Finally,health diagnosis method based on control chart analysis of health diagnosis indicators is proposed and verified on structural health monitoring data of a long-span cable-stayed bridge.(2)Time-series correlation model between vehicle-induced vertical deformation and vehicle-induced cable force of cable-stayed bridge based on Bi LSTM network is established.Firstly,all available vertical deformation channels and cable force channels are taken as input and output,respectively,to establish the full-bridge global Bi LSTM model mapping time-series correlation between vertical deformation and cable force of cable-stayed bridge.Secondly,to address the problem that the broken of individual sensors will adversely affect the global model,the local Bi LSTM model based on Sobol’sensitivity is constructed.For the given cable force,the combination of only a few vertical deformation channels is selected as input based on Sobol’ sensitivity.Finally,the established models are verified with the structural health monitoring data of a long-span cable-stayed bridge,comparing the learning effect between the global model and local model,analyzing influences of temperature,noise and traffic flow on the performance of the proposed model.(3)Distribution correlation modeling and health diagnosis method between vehicleinduced vertical deformation and vehicle-induced cable force of cable-stayed bridges based on domain transformation network is proposed.Firstly,kernel density estimation on the time-histories data of vehicle-induced vertical deformation and vehicle-induced cable force is performed to obtain the PDFs of vertical deformation and cable force,which are used as input and output of domain transformation network.Intra-class reconstruction and inter-class transformation are realized by variational autoencoder and generative adversarial network,respectively.The mutual transformation ability between PDFs of vertical deformation and cable tension are ensured by shared latent space and the constraint of cycle consistency.Secondly,the Wasserstein distance is used to measure the similarity between the predicted PDF and the measured PDF,acting as an indicator of health diagnosis of the stay cables.Finally,the transformation effect of the proposed method and the health diagnosis result of the stay cables are verified with the structural health monitoring data of two long-span cable-stayed bridges,and finite element analysis on one of the cable-stayed bridges is performed to compare the sensitivities between cable force and vertical deformation to the damage of stay cables.(4)Correlation modeling and health diagnosis method between temperature and temperature-induced strain of cable-stayed bridges based on Transformer is proposed.Firstly,the sensitivities between trend term and vehicle-induced term to damage are compared by mechanical analysis,and the time-lag phenomenon between temperature and temperature-induced strain is analyzed.Secondly,the Transformer model for temperaturetemperature-induced strain is established.The temperature and trend term of strain are used as input and output of Transformer model,respectively,of which the encoder learns the deep representation of temperature,and the decoder learns the correlation of temperature-strain and strain-strain.Thirdly,bridge health diagnosis is performed by control chart analysis on reconstruction error of strain.Finally,the effectiveness of health diagnosis and model modification are verified with the structural health monitoring data of two long-span cable-stayed bridges.