Study On Damage Identification For Bridges Based On Machine Learning

As a part of infrastructure,bridge plays an important role.Large-scale bridges in my country are developing rapidly,the pressure of daily operation and management is increasing,and the damage to bridges without timely repairs can cause disastrous consequences.Containing a large amount of bridge structure information,the bridge monitoring data can help us solve these problems.The rational use of these monitoring data can timely understand the state of the bridge,identify bridge damage,and improve the efficiency of bridge operation.However,there is still no method to make good use of these data.How to extract damage information from bridge monitoring data and identify structural damage is still a problem,and the problem of multidamage identification is still lack of systematic research.Based on the two spans beam model,this paper studies the problem of bridge damage identification,and carries out the following work:(1)The response data sets of single damaged bridges and multi-damaged bridges are established for machine learning algorithms,and the influence of bridge response under damaged state is analyzed;(2)Considering the poor noise immunity of the model trained on the raw data,the eigenvectors from MPCA is used as input,as a result,recognition accuracy and noise immunity of model are improved.The damage recognition capabilities of the support vector machine(SVM)model,K-nearest neighbor model,decision tree(DT)model and multilayer perceptron(MLP)model is compared and it is showed that training DT model with data eigenvectors from moving principal components analysis(MPCA)and then identifying damage can get the best results;(3)The generalization performance of DT model is studied,and it is proved that DT model trained by MPCA eigenvector can recognize the damage combination not in training set including the data of the stage where the number of damages is larger than the samples in the training set.