Research On Damage Identification Of Arch Bridge Based On Dense Convolutional Network And Attention Mechanism

In recent years,with the in-depth development of China’s transportation construction to the western mountainous areas,arch bridges have become the first choice for mountain bridges due to their strong bearing capacity,high stiffness and low cost.With the continuous increase of service life,due to the combined action of load effect,material degradation,environmental erosion and other factors,the bearing capacity and normal use of in-service arch bridges inevitably degenerate.In order to grasp the health status of arch bridge in time,it is of great scientific significance and engineering application value to carry out research on damage identification of arch bridge.In this thesis,through theoretical analysis,algorithm development,numerical simulation and indoor test,an arch bridge damage identification algorithm based on dense convolutional neural network and attention mechanism is proposed.The indoor test verification of CFST arch structure and CFST rigid skeleton arch structure under external load is carried out,and the parameter analysis of arch bridge damage identification method is carried out.The main research work and achievements are as follows :(1)An arch bridge damage identification method based on dense convolutional neural network and attention mechanism is proposed.Firstly,based on the finite element model of arch bridge structure,the structural acceleration response data under different damage states are obtained by time history analysis.Then,the time-frequency analysis technique is used to convert the acceleration response into a time-frequency diagram to form a sample database.Secondly,a deep convolutional neural network model for arch bridge damage identification is constructed,which integrates Dense Convolutional Network(DenseNet)and Convolutional Block Attention Module(CBAM),and the training and testing of the model are completed based on the sample database.Finally,the accuracy,precision,loss value,recall rate and F1 value in the classification problem are used to evaluate the model,and the feature visualization analysis is realized by T-distributed Stochastic Neighbor Embedding(t-SNE)nonlinear dimensionality reduction technology.The results show that the method can accurately identify the damage of arch bridge,and the accuracy of single damage identification is 91.67%.The accuracy of multi-damage identification was 92.78%.The damage features have obvious clustering tendency,which proves that the wavelet time-frequency diagram has strong feature expression ability and the proposed method has strong feature extraction ability.(2)The indoor test verification of CFST arch structure and CFST stiff skeleton arch structure under external load is carried out.Firstly,the concrete filled steel tube arch structure and the concrete filled steel tube rigid skeleton arch structure were designed and manufactured.Then,the acceleration response data under different working conditions are obtained by tapping during the loading process.Secondly,the time-frequency diagram is obtained by continuous wavelet transform and the data set is divided.Finally,the network model is trained and tested.The research shows that this method can realize the identification of arch structure;for concrete filled steel tube arch structure,the accuracy of damage identification is 90.22%.For the concrete-filled steel tube stiff skeleton arch structure,the accuracy of damage identification is 92%.(3)The damage identification parameters of arch bridge are analyzed.Firstly,based on the multi-damage condition data of arch bridge,different sampling frequencies,different sample lengths and quantities,different noise levels and signal missing conditions are constructed.Secondly,the comprehensive parameter analysis of the arch bridge damage identification method is carried out,and the identification results under different parameters are obtained.Then,the internal reasons for the fluctuation of damage identification results under different parameters are analyzed.Finally,the variation law of damage identification accuracy under different sampling frequencies,different sample lengths and numbers,different noise levels and signal loss is revealed.The results show that in the case of small damage degree,the higher the sampling frequency,the more obvious the effect of damage identification,and the accuracy rate is reduced to 57.22%at 256 Hz sampling frequency.With the same sample length,appropriately increasing the number of samples and the same number of samples,increasing the sample length helps to improve the overall recognition accuracy of structural damage,and the effect of increasing the number of samples is more obvious.Noise will affect the identification of structural damage.The noise level within 30 d B can achieve good recognition effect,and the overall recognition accuracy is 91.11%.In the case of missing channels,the overall recognition accuracy of the model is uncontrollable,and the overall recognition accuracy is only 72.78% when the signal is missing as a channel.