Structural Damage Detection Based On Data Fusion And Convolution Neural Network

The traditional structural damage detection（SDD）method based on single data has low detection accuracy and each kind of signals has its limitation.In order to improve the SDD accuracy,a SDD method based on data fusion（acceleration and strain,mode and modal strain energy）is proposed in this thesis.As convolution neural network（CNN）has strong ability of feature extraction and data processing capacity,the single data and fusion data are combined with CNN for SDD.Two evaluation indexes of damage location detection and damage degree detection are established simultaneously,and the feasibility of this method is verified by comparing the predicted damage situation with the real damage situation.In addition,by comparing the detection effect between single data and fusion data,and the detection effect between the fusion data,the superiority of fusion signals for SDD is demonstrated,and the fusion type of data with best detection effect is obtained.Firstly,basing on the related theories of CNN and data fusion,the components of CNN for SDD,the functions and working principles of each component,the classification and the basic concept of data fusion,and the implementation process of the above research methods are introduced.Secondly,the bridge model established by ABAQUS is taken as the research object,and the elastic modulus of the bridge model member is reduced through the script program developed by our research group based on ABAQUS and Python,in order to simulate various damage conditions of the bridge model.Meanwhile,batch data is extracted for various damage conditions.Through the analysis of numerical simulation data,it is found that the detection effect of fusion signals is better than that of single signal.Among the three types of fusion data,the damage detection effect of decision level fusion data is the best,followed by data level fusion data,while the performance of feature level fusion data is inferior to the first two.In addition,based on the above detection effect,further study is carried out on the relevant data in data level fusion,mainly including the location of accelerator and the type of members for obtaining strain signal.The results show that the location of different accelerator and the type of members for acquiring strain signals will affect the damage detection accuracy of single signal.Finally,test samples of structural damage are obtained through vibration experiment,and the obtained test samples are input into the CNN trained with numerical data for SDD.It is shown that the CNN has high practicability due to its strong ability for SDD and data generalization ability for heterologous signals.