Research On Bridge Damage Identification Methods Based On The Random Forest Algorithm And Bayesian Dynamic Models

How to reasonably identify bridge damage location and degree has become one of the key scientific problems in the field of structural damage identification.In this thesis,based on the acceleration response data,through the combination of machine learning and damage index methods including the Wold decomposition method,the weighted random forest algorithm,the dynamic linear models and the random forest dynamic coupled linear models,damage sensitive factors are constructed,further,structural damage assessment is made.The main research contents are described as follows:(1)Structural acceleration response data is adopted to build the weighted random forest models which can achieve the classification prediction of structural acceleration response data in the different damage conditions.The acceleration responses of the undamaged structure are considered as the reference signal.When the structure is damaged,there is the differences between the reference signal and acceleration data of the damaged structures.Based on the fitting degree among the accelerations in different damage conditions,the damage sensitive factor,which can characterize the structural damage states,is given,further,the relation between structural damage conditions and damage sensitive factors can be obtained.Considering the impact of noise,the supported beam with different damage conditions,is taken as an illustration example to demonstrate the effectiveness of the proposed method.(2)The prediction method of structural acceleration responses is presented based on the decoupled monitored data and multiple dynamic linear models by using Bayes method.With Wold decomposition method,the acceleration data is decoupled.With Pearson correlation coefficient,the correlation between decoupled data is analyzed.The decoupling data extracted from acceleration responses is utilized to build the corresponding multiple dynamic models,further,structural acceleration response prediction can be made with the fusion method.The supported beam is provided to illustrate the effectiveness and practicality of the proposed method.(3)Acceleration response data-based structural damage assessment method is presented with the random forest algorithm and Bayesian prediction method.The clustering analysis of the acceleration response data is made with random forest algorithm and agglomerative hierarchical clustering method.The different clustering data extracted from the acceleration responses is utilized to build the corresponding random forest dynamic coupled linear models,further,considering the impact of noise,the dynamic damage assessment can be made with the changes of damage sensitive factors.The finite element models for the supported beam and the continuous box girder are provided to illustrate the effectiveness and effectiveness of the proposed method.