Study On Structural Damage Identification Based On Response Time Frequency Diagram And Deep Learning

Structural damage is generally identified through the vibration response of the structure.Deep learning networks are often used as tools for damage identification due to their ability to automatically extract features.When vibration response signal and deep learning are used for structural damage identification,problems such as requiring more data information of measuring points,low accuracy of damage identification,and easy overfitting of the network will be encountered.To this end,the author conducts research from two aspects: signal feature extraction and deep learning network model combination.On the one hand,Generalized S Transform(GST)is used to analyze the acceleration response signal,and a two-dimensional time-frequency diagram with rich spatial local features is obtained,which is input into Convolutional Neural Network(CNN)for structural damage identification.On the other hand,CNN is used to extract spatially localized features from generalized S-transformed time-frequency maps,and Gated Recurrent Unit(GRU)network is used for extracting temporal features from onedimensional acceleration response signals.Then the above 2 features are fused and input to the subsequent fully connected layer and Softmax classifier for structural damage identification.The feasibility of the proposed method is verified by the measured test data of Benchmark in the second stage and the model test data of offshore wind power support structure.The specific research contents are as follows:(1)The research background and significance of structural damage identification is described,in which the application of time-frequency analysis techniques and deep learning in the field of structural damage identification and the current state of research are mainly introduced.(2)The time-frequency analysis technology and the principle of deep learning neural network model are introduced.As a signal processing method,time-frequency analysis technology aims to analyze and extract the time-frequency domain characteristics of signals,including GST,Short-Time Fourier Transform(STFT),Wavelet Transform(WT)and Wigner-Ville Distribution(WVD).The deep learning neural network models include CNN,Recurrent Neural Network(RNN),GRU,Bi-Directional Long Short-Term Memory(Bi-LSTM)and Long Short-Term Memory(LSTM).(3)A methodology for identifying structural damage using a combination of GST and Two Dimension(2D)CNN is proposed.As a time-frequency analysis technology for processing response signals,GST combines the advantages of STFT and WT to provide a clearer picture of the characteristics of the response signal in the time-frequency domain.The response signal is therefore generated as a two-dimensional time-frequency map by means of GST,which is fed into a 2D CNN network suitable for processing images for training and testing to obtain the damage recognition results.Compared with structural damage identification methods based on STFT-2D CNN,WT-2D CNN and WVD-2D CNN,it is found that the method proposed in this paper has better damage identification performance.(4)A structural damage identification method based on CNN-GRU parallel neural network is proposed.For extracting more abundant feature information related to structural damage from the response signal,GRU and CNN networks are used to extract temporal features and spatial local features from the response signal and time-frequency diagram,respectively.The extracted 2 features are then stitched in the feature stitching layer and input to the subsequent fully connected and Softmax layers for damage condition identification.Due to the comprehensive utilization of the features extracted by two different deep learning models,the proposed CNN-GRU parallel neural network has strong feature representation ability and can deeply mine the features contained in the response signal.Compared with the identification result using CNN,GRU,CNN-LSTM parallel neural network model,CNN-Bi-LSTM parallel neural network model,CNNLSTM tandem neural network model and CNN-GRU tandem neural network model,it is showed that the proposed method was effective.