| The traditional methods of structural health monitoring（SHM）have obvious disadvantages such as being time-consuming,laborious and late,and so on.This thesis presents a novel and efficient approach to detect structural damages from vibration signals（acceleration and strain）via a convolutional neural network（CNN）.When the structure is damaged,its inherent characteristics（stiffness,mass,damping）will change.As vibration signals reflect the structural response to the changes of the structural state,thus it carries almost all the unique information of the structure.As a powerful data processor,CNN can automatically extract data features.The CNN is used to extract the damage features from the vibration signals and map them to the structural state to realize the damage detection.In this thesis,the vibration signals and CNN are combined to study the damage detection.The feasibility of the damage detection method is demonstrated from two aspects of numerical simulations and vibration experiments.At the same time,the damage detection results（damage location ＆ damage degree）of acceleration,strain,acceleration and strain fusion are compared,and it is concluded that the damage detection effect of two signals fusion is better than that of single signal.This thesis studies CNN theory,sample data acquisition,numerical simulation and experimental verification.In this thesis,the theory of CNN is introduced firstly,and then the internal structure and function of a basic CNN are introduced,including input layer,convolution layer,pooling layer,activation layer,softmax layer,full connection layer and output layer,as well as the self-learning and updating process of CNN.The operation of the network is realized by MATLAB,and then the method of sample acquisition is studied.Based on ABAQUS and PYTHON,the automatic setting of damage condition and batch extraction of vibration signals under different damage conditions are developed,and a large number of samples required by the network are obtained by data augmentation and data fusion technology.Then the bridge model is established by using ABAQUS as the research object,and the acceleration and strain signals of the model under different damage conditions are obtained and the damage detection is carried out.The results show that the small damage condition is better than the large damage condition,and the rich damage sample set can provide accurate damage detection of convolution neural network.The accuracy of two signal fusion is higher than that of single signal.The thesis also analyzes the CNN in the case of abnormal signal damage detection.The results show that the abnormal data has little influence on the detection accuracy of the convolution neural network,but the noise has more obvious influence on the small damage condition.Therefore,the environmental impact of damage detection can not be ignored.Finally,the damage test samples are obtained through vibration experiments and input into the network trained by the signals obtained from numerical simulations.The results show that the CNN still has a good damage detection accuracy,which shows that the CNN has a strong generalization ability.In the aspect of heterogeneous signals,it also provides training samples for actual structures. |
