Research On Structural Damage Detection Method Based On Deep Learning

In this paper,a structural damage detection method based on deep learning algorithm is proposed.The damage of a structure will cause the change of its mass,stiffness and damping,and then the change of its modal parameters.As an intelligent algorithm,a convolutional neural network(CNN)has the ability of automatically extracting features,while damage detection is actually to find the features of structural damage,so it has great potential to apply this algorithm to the field of damage detection.In this paper,modal strain energy and a CNN algorithm are used for damage detection.The method is applied to numerical models and a experimental model respectively.The results show that the method is feasible.The preparatory work mainly includes: the study of the basic principle of convolutional neural networks,the design of a CNN model and the acquisition method of network training samples.In this paper,we have developed the construction of the CNN framework based on MATLB and its application in damage detection and the Python scripts based on ABAQUS finite element software to obtain damage samples parametrically.We have established a method to obtain a large number of samples automatically,which saves human resources and a lot of computational time.For the proposed damage detection method,firstly,the numerical simulation is studied;the research object is a simply supported rectangular beam and a three-dimensional steel frame.The results show that the CNN has excellent recognition effect not only for simple models(simply supported beams with a rectangular section),but also for three-dimensional complex steel frame models.The CNN has ideal detection effect for single damage location and multiple damage locations,the detection of degree is of high precision.In order to simulate the complexity of real-world damage detection process,artificial noise and incomplete data are used to complicate the detection process.With the increase of noise intensity,the accuracy of damage location detection of the CNN has little change,but the error of damage degree detection is large.For the study of the influence of incomplete data,with the increase of missing data,the accuracy of damage location detection of the CNN decreases,and the error of damage degree detection also increases.In order to prove the advantages of the CNN,the damage detection process of normal data and disturbed data(including noise and incomplete data)is applied to a traditional neural network again,so as to compare the advantages and disadvantages of the two.For the normal data,the calculation speed of the CNN has been greatly improved,and the damage location detection has more advantages,and the fitting effect of damage degree is slightly lower than that of the traditional neural network.However,from both the aspect of calculation speed and detection effect,the CNN has more advantages than the traditional neural network;for the data withnoise,the CNN has more advantages After the network has a better anti-interference ability,in the damage location and damage degree has a better detection effect,the accuracy and accuracy are higher than the traditional neural network.Finally,in order to prove that the CNN can also identify the real model,a method of damage detection based on heterogeneous data for the real model is proposed.Taking a three-dimensional steel frame as the research object and the numerical simulation data as the training samples,the damage index adopts the change rate of the modal strain energy.After the network training,the experimental data is input into the network,which can accurately identify the damage of the real structure.Compared with the traditional detection method,the CNN has unique advantages.