Structural Health Status Identification Based On Depthwise Separable Convolutions

It is of great significance to effectively identify the structural health status of all kinds of buildings for the good long-term operation of buildings.With the development of sensor technology and the Internet of things,the structural health monitoring system is widely used.A large number of sensors deployed in the structure collect real-time data and analyze these data to obtain the health status of the structure.And a large number of sensors to collect data of high dimension and large amount of data,data collected by inevitably affected by noise,load and structure dynamic change constantly,the traditional methods need to extract the characteristics of the sensor data manually by human beings,not only time-consuming,and some areas also requires the researchers need to have knowledge of this field.How to identify the structural health state quickly and in real time under these factors is an urgent problem to be solved.Based on this,an end-to-end,simple and easy to implement method for structural health state recognition is proposed.This thesis presents a method for structural health state recognition based on depthwise separable convolution.The method takes the original time series signal as the input,uses the depthwise separable convolution to replace the standard convolution,and the global average pooling to replace the traditional full-connection layer,which reduces the number of parameters and calculation amount of the model and reduces the risk of overfitting.This method has good performance in the noise environment,and uses the large one-dimensional convolution kernel and the random damage training data to improve the anti-noise ability of the model.At the same time,this method has strong adaptive ability across the load domain,does not rely on any domain adaptive algorithm,also does not need the information of the target domain,and can achieve high accuracy when the structural load changes.Method proposed in this thesis,first of all,the two kinds of finite element model of the structure of ASCE benchmark experiments have been carried out to verify the data set,and compared with the existing one-dimensional convolutional neural network model,the two-dimensional convolution model and MLP neural network methods were compared,the results show that the proposed method has high accuracy,can accurate identification of the structure of the healthy state;At the same time,it can maintain high accuracy at different noise levels,and further improve the anti-noise ability of the model by training the convolutional neural network with randomly destroyed data,which indicates the effectiveness of the method presented in this thesis.To further verify the validity of the method in this thesis,again in bridge scale model experiment,the three data sets of results also show that the method performs better than other methods in this thesis,at the same time,this method was verified domain adaptive ability more outstanding than other methods,prove the feasibility and effectiveness of this method.