Application And Research Of Residual Learning-Based Convolutional Neural Networks In Damage Identification

With the development of engineering technology and many practical structure tends to be complicated,damage identification in structural health monitoring is an important field.Because of the complexity of the civil engineering structure,the damage distribution and damage degree has great randomness,the project is faced with many difficulties in the practical engineering,such as interference of measurement environment,and the limitation of the identification method.Therefore,there are higher requirements for damage identification technology.In recent years,with the rapid development of computing hardware and the method based on the deep learning for structural health monitoring system provides a new direction,the convolutional neural network is an important tool in deep learning,it has a strong processing capacity in the visual identification,convolution neural network also has a advantage on the processing of acoustic model.Acceleration signal,displacement signal are the most accessible and the largest number of data in health monitoring system and the two kinds of signals are main data source for damage diagnosis.Many scholars have also used them to structural damage detection and achieved good results.The main purpose of this paper is to apply the convolutional neural network in structural damage detection especially the determination of damage location.This paper takes the node damage of frame structure as the research object.A node damage localization method based on deep convolutional neural network for frame structure has been proposed in this paper.The method proposed in this paper is verified by numerical simulation and experiment.Firstly,the damage identification method based on convolutional neural network has been described and the shallow convolutional neural network model has been constructed in this paper.The finite element model has been established to obtain the time history signals of structural displacement under various load and damage conditions.The detection capability of the convolutional neural network is evaluated by using the results of finite element simulation as the input of the network.Secondly,the differences and connections between shallow and deep neural networks are introduced,and using residual neural networks to solve the problems of network degradation and gradient explosion/vanishing.The same data set is used to train set and test set,and the identification results of nodes damage of different networks are compared to evaluate their identification ability.White Gaussian noise with different noise levels is added into the original displacement signal to make the simulation result closer to the actual measurement,and to evaluate the anti-noise ability of various networks.Finally,the shaking table test of frame structure is designed,and the obtained data is used to evaluate the detection capacity of different networks.At the same time,a kind of data processing method is proposed as an improved algorithm for damage identification.The analysis results show that this method can improve the accuracy and convergence speed of damage identification.