Structural Damage Identification Based On Recurrence Plot And Convolutional Neural Network

In recent decades,civil engineering structures become more and more complex and large-scale.In order to ensure the long-term service and safety operation,structural health monitoring(SHM)and structural damage identification(SDI)have been drawn great attention in the field of civil engineering.Because of the complexity of operational environment,obvious non-stationarity can be found in the ambient excitations as well as the structural vibration responses.It further makes the traditional SDI methods based on the assumption that the excitation is stationary perform less than ideal.In this context,this study focuses on the SDI under non-stationary excitations.An SDI method was proposed based on recurrence plot(RP)and convolutional neural network(CNN).In this method,RP is utilized to reveal the non-stationarity of structural vibration responses and CNN is applied to automatically extract damage-sensitive features from RP and classify different damage scenarios.By this procedure,damage localization and damage degree identification can be achieved.The main contents of this thesis are listed as follows:(1)The current status of research on SDI is reviewed,mainly including structural dynamic characteristic-based methods,structural vibration signal processing-based SDI methods and finite element model updating-based methods.The advantages and disadvantages of these SDI methods are compared as well.(2)The unthresholded multivariate RP(UMRP),which utilizes structural multichannel vibration responses,is introduced to reveal the vibrational features under nonstationary excitation conditions.Additionally,the performance of conventional RP,multivariate RP(MRP)and UMRP in revealing non-stationarity of structural vibration responses are compared in this study.(3)Taking advantages of CNN in image processing and pattern recognition,a preliminary SDI method is proposed,which directly use RPs of different damage scenarios to train a CNN model.A finite element model of a simply-supported beam under non-stationary excitations is studied to verify this SDI method in detecting the location and degree of single-damage cases.(4)To solve the problem that the large number of scenarios and the huge quantity of data samples in multiple-damage cases,an improved SDI method is proposed.Multiple sets of labels are established for RPs plotted from structural vibration responses according to the damage locations.Then RPs are fed into several parallel CNN models to train them for classifying the damage degrees.Damage localization and detection of damage degree are decoupled in this way.The feasibility of improved method detecting multiple-damage cases is also verified using the numerical model of the simply supported beam.(5)The proposed preliminary and improved SDI methods are applied to the damage detection of a simply-supported aluminum alloy beam in the laboratory to further compare their practicability.Moreover,the influence of sample size on these two methods is discussed.The results of experimental test indicate that the improved method can achieve higher identification accuracy with small data size.