| Structural health monitoring is an important technology to obtain a healthy state of the structure,which is widely used in major engineering construction.How to obtain the damage information of the structure from massive data generated by the structural health monitoring system and accurately evaluate the working state of the structure is of great significance to the safe operation of the structure.In recent years,structural damage identification based on data-driven has become a research hotspot of structural health monitoring,which has motivated the research on the application of deep learning in this field.In this thesis,the structural health monitoring data are firstly processed based on the compressed sensing theory.Then a structural dynamic behavior surrogate model is established by the deep learning method,which is used for structural damage identification.The specific contents include:1.Focusing on the problems of the insufficient sampling rate of monitoring data,mixed noise,and data loss,three data processing methods are proposed in this thesis based on the theory of compressed sensing signal reconstruction,respectively.In terms of random under-sampling and sparse reconstruction,a sparse regularization method based on the lp norm is proposed,and iteratively updating the regularized weight parameters can be used to alleviate the problem of parameter selection.In an actual test,it was found that the SNR of the reconstruction of a certain mass-spring-damped system displacement by this method was more than 4.88 d B higher than that of the l1 method,which achieves better results than other methods.Secondly,a structure response data denoising method based on low-rank reconstruction is proposed in this thesis,and a general monitoring data denoising method based on the total variance as a sparse term is also given.The experiments indicate that with 80%noise intensity,the SNR is improved up to 449.5%by the low-rank reconstruction methods for noisy signals,and the denoising effect of both methods is better than of wavelet denoising method.Finally,a data loss reconstruction method based on compressive sensing theory is proposed in this thesis to reconstruct the lost data of stress monitoring data in Shaoxing Light Textile City Sports Center,and the reconstruction SNR is more than 36.36 d B.2.A structural dynamic behavior surrogate model is proposed based on deep learning.The dynamic response time series prediction model with a single structure measuring point is proposed based on LSTM neural network.The dynamic behavior of the structure under both smooth and non-smooth loads can be accurately predicted by this model,and the peak means square error of prediction less than 0.007.The spatiotemporal prediction model of dynamic response with multiple measurement points is established based on encoding Conv LSTM neural network,which also has a high prediction accuracy.The prediction mean square error is only 6.24×10-5 even with an input noise intensity as high as 30%.The structural dynamic response surrogate model established in this thesis appliesis applicable to different types of structures such as frames,cylinder lattice shells,and spherical latticed shells,as well as different types of dynamic loads and actions such as cyclic loads and seismic actions.3.Structural damage identification is identified by adopting the structural dynamic response surrogate model.Firstly,in response to the problem of noise and missing data in the monitoring data,the data processing method proposed in this paper is applied to purify and enhance the original training data set.Then,the LSTM is used as a surrogate model to obtain the structural state features,and then the state features are classified by Softmax regression to identify the structural damage.Finally,the damage to IASC-ASCE Benchmark structure and the stadium grandstand steel structure can be identified by the method in this paper.The experimental results show that the LSTM method improves the recognition accuracy by 1.3%after incorporating data processing.4.A structural damage identification method based on graph convolution and LSTM fusion network is proposed.Firstly,an undirected and weightless graph is used to represent the monitoring data of the structure.Then,the graph convolution with two layers are integrated into the LSTM network structure to extract the structural feature and identify the structural damage.Finally,the proposed method is applied to the damage identification of the IASC-ASCE benchmark structure and stadium grandstand steel structure.The results show that our method can identify the structural damage identification based on the collaborative analysis of irregular multiple measurement points monitoring data with higher accuracy than the deep learning methods such as1DCNN,2DCNN,LSTM,and deep residual networks.In the case of sensor damage at some of the measurement points,our method can still complete the damage identification of the structure based on the remaining measurement point data,which indicates that the method has good applicability. |
