Structural Modal Parameters Identifcation Based On Machine Learning Under Ambient Exciation

The modal parameters(frequency,mode shape and damping ratio)are inherent characteristics of the structure,which are important for understanding the dynamic behaviors of the structure.The modal parameters are also the basis for structural damage detection and safety assessment.However,the traditional method of structural modal parameter identification often needs more manual intervention and parameter setting.It is difficult to realize the online identification of modal parameters in the structural health monitoring(SHM)system,which seriously affects the online warning of the SHM system.How to realize the intelligent identification of structural modals is a big challenge.Therefore,this paper combines machine learning theory to investigate the intelligent method of structural modal identification under ambient excitation.The main research contents of this article include:A machine learning method for structural modal identification based on modal independence is proposed.This method first transforms the modal identification problem into a deep learning problem.Then the uncorrelation and non-Gaussianity of the modal response are employed to form a loss function of the neural network to estimate the modal parameters.The numerical simulation examples and the SHM data of an actual cable-stayed bridge are carried out to illustrate the ability of the proposed method.The results show that the proposed method has similar modal identification accuracy with the commonly used frequency domain decomposition method(FDD),stochastic subspace identification(SSI),and NEx T + ERA methods.The results of the proposed methods are obviously better than the traditional independent component analysis(ICA)method.In addition,the proposed method has certain intelligent data processing capabilities,which can provide a new option for structural modal identification under ambient excitation.A machine learning method for structural modal identification based on the sparsity of structural response in time-frequency domain is proposed.The structural vibration response data has sparsity in the time-frequency domain.First,the structural vibration response data is transformed from the time domain to the time-frequency domain,and then the response single-source points corresponding to each order mode are selected.The selected response of single-source points are processed as symmetrical and normalization.Finally,the competitive neural network are designed and trained to obtain the cluster center of single source points,which is the value of each mode shape.The actual SHM data of a cable-stayed bridge are used to illustrate the modal identification ability of the proposedapproach and the results show the mode shapes can be well identified.A machine learning method for structural modal identification based on supervised multi-task learning is proposed.This method uses the structural system response data as inputs and the separated modal responses as the labels of the nueral network.Using the powerful learning capabilities of the deep neural network,the trained neural network can automatically separate the system responses into modal responses and mode shapes.The numerical examples are employed to verify the effectiveness proposed method.