Research On Structural Damage Identification Method Based On Echo State And Multi-Scale Convolution Joint Model

In order to improve the service safety of major structural engineering,the real-time monitoring of possible damages is carried out.Structural Health Monitoring(SHM)technology is widely used in the domestic and overseas.Among them,structural damage identification is one of the important research tasks of SHM.It can judge whether the structure is damaged,the location where the damage occurred and the degree of damage.Accurating structural damage identification can prolong the service life of structural engineering.With the continuous development of sensor technology,the types of data acquired by SHM systems are increasing,and it is particularly critical to use these monitoring data to identify structural damage.Structural vibration response signals are selected as the research data for its high accuracy and mature technology.Therefore,the core of damage identification is to effectively extract the dependence and correlation between data.Based on the above characteristics,this thesis proposes a structural damage identification method based on a joint model of Echo State Network(ESN)and MultiScale Convolutional Neural Network(MSCNN).The main research contents are as follows:(1)Spatio-temporal correlation analysis of multivariable time series data for structural vibration response.From the time dimension,the time sequence of the same measuring point of the acceleration sensor is determined by the autocorrelation coefficient.From the spatial dimension,the correlation coefficient is used to determine the spatial correlation between different measuring points of the acceleration sensor,so as to make it clear that the vibration response information has the correlation of both spatial and temporal dimensions.(2)Data enhancement method for structural vibration response information characteristics.In a more scientific way partitioning samples used in the experiment,the structure vibration response information with fixed size sliding window moving from scratch with a fixed step length,capture all the data in turn,constitute the experimental data samples,and further study of the sliding window grew up in the window of the small and mobile step to model the influence of damage identification results.(3)A joint model based on ESN and MSCNN is proposed.The model extracts the time-series dependencies through the ESN and the spatial correlation through MSCNN.An improved Functional Echo State Network(FESN)model was proposed to optimize the superparameter of the joint model,and the optimal superparameter combination was searched by genetic algorithm.Furthermore,the grid search method is used to search the optimal combination of other superparameters of the joint model,and the combination values obtained by different algorithms are combined as the optimal superparameter combination of the joint model.(4)In order to verify the effectiveness of the proposed method for structural damage identification based on the joint model of ESN and MSCNN,the Benchmark finite element simulation model and the scale model of the continuous rigid frame bridge in the SHM field are studied experimentally.It is compared with the mainstream Deep Neural Network(DNN),Convolutional Neural Network(CNN),Long short-term Memory(LSTM),CNN-LSTM and Bi-directional Long Term Memory(Bi LSTM)models in Deep learning.According to the evaluation indexes of acuuracy,precision,recall and F1-score,the results of damage identification of ESN-MSCNN model are superior to the comparison model.