Research On Structural Response Data Reconstruction Method Based On Advanced Deep Learning Technology

Long-term structural health monitoring systems can measure structural response and environmental data of civil engineering structures in real-time,which is important for the detection of structural damage,assessment of structural safety status,and prevention of catastrophic accidents in real time.The monitoring data is the bridge between the structure and engineers,and its integrity and validity are the basis for accurately analyzing the operation state of the structure.However,due to inevitable man-made and natural factors,such as extreme loads,sensor aging failures,and signal transmission interference,monitoring systems often lose measurement data during operation,resulting in the lack of current structural information,which further affects the continuity of structural health monitoring and the reliability of evaluation results.Currently,to compensate for the impact of data loss on the function of costly monitoring systems,data-driven data reconstruction algorithms are based on the inherent transferability of structural response,and the mathematical model describing the mapping between the responses of different positions of the structure is established by extracting data features directly from the monitoring data without the finite element model,and the lost data are recovered based on the existing measurement data,which has a wide application prospect.Therefore,this paper addresses the vulnerability of existing data-driven data reconstruction algorithms to structural state changes and external excitation,and model underfitting caused by the small data samples required for mapping learning.Based on the existing research,the reconstruction algorithm with high efficiency and robustness is developed by introducing and optimizing advanced artificial intelligence techniques such as self-attention mechanism and generative adversarial network.The main contents of this paper are as follows.1.A comprehensive description of deep learning techniques and reconstruction data evaluation methods relevant to this paper is presented.The main building blocks of the design model are presented,including the principles and advantages of convolutional neural networks,generative adversarial networks,self-attention mechanisms,densely connected neural networks,and network optimizers.Meanwhile,the signal processing and modal identification methods used to evaluate the reconstructed data are described in detail,including the fast Fourier transform,frequency domain decomposition method,and random subspace method.By comparing and evaluating the reconstructed data with the help of these methods,the effectiveness and applicability of the proposed methods in the field of structural monitoring can be grasped.2.For the existing response reconstruction algorithms based on deep learning with poor robustness and low feature learning efficiency,this paper proposes the self-attentionbased generative adversarial network and the attention-based U-shaped generative adversarial network.In this paper,the generators and discriminators of the generative adversarial networks are built by combining the advantages of the large sensory field of the self-attention mechanism and the high data mobility of densely connected networks.The study presents the structural components of the generator and discriminator of the network,as well as the flow of data and computation,respectively.Through empirical settings and comparative experiments,the configurations of the specific layers,sizes,locations,and training hyperparameters of the networks are optimized,and deep learning networks applicable to dynamic response reconstruction for structural health monitoring are constructed.3.To verify the effectiveness of the self-attention-based generative adversarial network,this paper carries out a practical engineering validation using the health monitoring data of Guangzhou Tower as an example and verifies the ability of the network model to reconstruct the structural response under daily operation and extreme excitation through two pre-defined working conditions.The reconstructed data have high accuracy in both time and frequency domains,indicating that the network has good robustness and can effectively cope with the changes in vibration response data characteristics caused by external excitation changes.Meanwhile,the study uses the reconstructed response for modal analysis of the structure,and the identified natural frequencies match exactly with the real measured values,and the coordinate modal assurance criterion value reaches 99.98%,indicating the high applicability of the reconstructed data in structural health monitoring problems.To explore more deeply the mechanism of the network reconstructed vibration response and the reason for the self-attention mechanism to reduce the reconstruction error,this paper uses the hidden layer visualization technique to derive the results of the internal matrix of the first convolutional layer and the self-attention mechanism layer of the network after training and confirms that the convolutional network and the attention mechanisms can effectively facilitate the learning of representative features of the structural response data and structural vibration characteristics.4.In this paper,the performance of the attention-based U-shaped generative adversarial network for data reconstruction with small sample data is verified based on finite element models of laboratory frame structures and experimental measurement data.The study is first based on the numerical simulation data generated from the finite element model,and two working conditions,random stiffness decay and local damage of the structure,are set.The results show that the network automatically learns and extracts representative and generalized data features from the structural response data through data convolution,feature compression,and self-attention weight learning.With only a small amount of structural complete state response data,the network accurately reconstructs the missing data under the structural random stiffness decay and structural local damage conditions,showing high applicability.In addition,the training network was used to reconstruct the missing measurements from shaking table seismic tests with a small amount of measured environmental excitation,and the relative reconstruction error was only 16% in the presence of measurement noise.In addition,the natural frequencies obtained from the modal analysis based on the reconstructed response data exactly match the real values,and the coordinate modal assurance criterion value reaches 99.96%,showing that the network is robust and has a high engineering application value.