Damage Detection Method Based On Deep Learning And Structural Mode Shapes

This paper proposes an intelligent structural damage detection method based on the deep learning algorithm,which takes the structure mode as the index.As the number of buildings gradually increased,in order to ensure the security of people’s life,structural damage detection has gradually attracted more and more attention.Structural damage causes the change of the mass,stiffness and damping of structure,which leads to the change of the structure modes.Many scholars attempt to study the structure modes,in order to find out the features of structural damage.Compared with previous artificial intelligence algorithms,the deep learning algorithm has a powerful ability to extract data features automatically.Therefore,the deep learning algorithm can be applied into structural damage detection to extract damage features from structural modes data.In this paper,the finite element simulation method is used to obtain the structural mode data under various damage scenarios,and the structural mode data is combined with the convolutional neural network to carry out structural damage detection.The internal mechanism of the network is also analyzed carefully,the main research contents of this paper are as follows:（1）ABAQUS is used to simulate structures under various damage scenarios and get structure mode data under various damage scenarios.Based on the PYTHON interface to ABAQUS,batch analysis scripts were developed to get a lot of data of different damage scenarios,it saved a lot of manpower resources and provided the necessary technology for the realization of this method based on big data.（2）The main research objects of this paper are a simple steel beam model and a complex steel frame model,and the main tool is the deep learning toolbox in MATLAB.The simple beam model is mainly used to verify the feasibility of this method.The complex steel frame model is mainly used to explore the upper limit of the detection capability of the method and study the internal mechanism of the network.The results show that: as for the simple damage detection problem,the convolution neural network can maintain 100%accuracy even in the case of few data input,as for the complicated problem of damage detection,the damage detection effect can achieve a certain effect,but is very dependent on adequate amount of data,and the mode damage index which has been processed by human,such as modal curvature difference,its ultimate detection effect is lower than the original data（modal vibration mode）,which shows that the unprocessed data retain more damage information of the structure.（3）This paper also makes a detailed analysis of the CNN internal mechanism of damage detection,rather than just using it as a powerful "black box".In this paper,the intermediate data of each stage in the process of damage detection of neural network are extracted.The concrete mechanism of neural network is visulized.