Study On Structural Damage Identification Based On Modal Response And Deep Learning

In the research of damage identification based on vibration response and deep learning,the measured acceleration response signal not only contains components related to the natural vibration characteristics of the structure,but also contains noise and some components unrelated to structural damage,which affects the generalization ability of the model and the accuracy of damage identification.In order to solve the above problem,the author uses VMD to decompose the measured vibration response signal,selects the components related to the natural vibration characteristics of the structure as the effective components,and then takes the effective components and their time-frequency information as the input of the deep learning model respectively,and uses the deep learning model to identify the structural damage.The results of shaking table test of five-story steel frame model and IASC-ASCE SHM Benchmark secondstage structural test prove the effectiveness of the proposed method.The main contents are as follows:(1)The background and significance of this research are introduced,and the research status of structural damage identification based on time-frequency analysis and deep learning is emphasized.(2)Some basic theories used in this paper are briefly introduced,including Variational Mode Decomposition(VMD)algorithm,Convolutional Neural Networks(CNN),variations of Recurrent Neural Network(RNN)such as Long short-term Memory(LSTM)neural network,Bi-directional Long short-term Memory(Bi-LSTM)Neural Network and Gated Recurrent Unit(GRU)Network,Composite network model CNN-LSTM and some time-frequency analysis methods such as short-time Fourier Transform(STFT),wavelet transform(WT)and Hilbert-Huang Transform(HHT).(3)Study on structural damage identification based on VMD-CNN-LSTM.Firstly,FIR filter was used to filter the high-frequency noise in the measured acceleration response signal.VMD was used to decompose the filtered acceleration response signal into a series of IMF components,and the effective IMF components related to the natural vibration characteristics of the structure were selected for reconstruction.Then,the reconstructed signals were segmented to generate input samples,which were divided into training set,verification set and test set in the ratio of 8:1:1 and input to the CNN-LSTM network model for training and testing.Finally,the damage identification results were obtained.It has been verified by experiments that VMD-CNN-LSTM has the highest damage identification accuracy of 100%,and the training time is at least 16 s,which has high damage identification accuracy and efficiency.(4)Study on structural damage identification based on VMD-HT-CNN.Firstly,FIR filter was used to filter the high frequency noise in the measured acceleration response signal.Then,the acceleration response signal after filtering is decomposed by VMD,and the effective IMF components related to the natural vibration characteristics of the structure are selected from the decomposed series of IMF components,which are arranged in order to form the IMF component matrix.Then,Hilbert transform was used to conduct time-frequency analysis on IMF component matrix to generate timefrequency graph.The time-frequency graph was taken as the input sample and was divided into training set,verification set and test set in 8:1:1 ratio and input into CNN network model for training and testing.Finally,damage identification results were obtained.The experiment verified that VMD-HT-CNN method has the lowest damage recognition accuracy of 99.71% and the highest accuracy of 100%.The longest training time is only 69 s,which has excellent comprehensive damage recognition performance.