Complex Structure Flexibility Identification And Performance Assessment Based On Subspace Technology

To prolong the service life of infrastructures and other major structures,how to ensure the safety of the structures while controlling the life cycle cost(LCC)of the structures by effective operation and maintenance is a common problem that needs to be solved at home and abroad.Structural health monitoring(SHM)technology has attracted much attention from scholars and the industry because it is real-time and comprehensive.On the other hand,in the recent years,the data generation rate is significantly increasing,making the concept of “big data” a global prevalence.Big data technologies include data mining,machine learning,artificial intelligence,and etc.,all of which help people to improve life quality by extracting useful information from data everywhere around them.Because the data processed by SHM is always large,various and complex,it can be regarded as a typical “big data” problem.To improve the automaticity and accuracy of the monitoring system in SHM,a three-level structure analysis framework based on monitoring big data is proposed.At the first layer,data anomaly analysis and data prepossessing is studied by deep learning and anomaly detection techniques.At the second layer,structural performance assessment(flexibility and damage identification)is conducted by subspace method.The specific research work is as follows:(1)A three-level structure analysis framework based on monitoring big data: The framework supports two modes,batch analysis and real-time analysis to process the acquired big data.At the first level,the framework processes the streaming data in real-time to detect anomaly,and emphasizes for the first time the importance of distinguishing structural anomalies from system anomalies.At the second level,it applies a structural mathematical model to analyze the structural current state,and assess the safety and serviceablitiy under specific load case.At the third layer,the future state of structure is predicted by the structural mathematical model and the deterioration model.A response analysis based on the design load or the predicted load.Consequently,the remaining service life of the structure is evaluated to meet the requirement of structural safety,serviceablitiy and durability.The thesis focuses on the research on the first and second layer,while the third layer is shortly introduced.(2)A data anomaly analysis method for SHM based on deep learning at the first layer: In this thesis,a time series neural network,Long Short-Term Memory(LSTM),is used to model the structure response.A trained LSTM can accepts the historical response series to predict the response in the next time unit.This model presented high performance on the monitoring data of Jiangyin Yangtze Bridge.Next,the distance between the measured response and the predicted response is calculated by Dynamic Time Warping(DTW)as the anomaly of each time step.An anomaly detection model is trained with an unsupervised algorithm-isolation forest,to identify the abnormality of structure response.Finally,structural anomaly and system anomaly is distinguished by the logic correlation analysis and multi-dimensional anomaly detection without any human-defined threshold,which significantly improves the automaticity of the monitoring system and provides high-quality data for the analysis in upper layers.In addition,each kind of external load corresponds to a kind of response pattern in structural analysis.For ship collision,the response pattern must be different from that during normal operation.Next,the feature strain channel for the ship collision is found through the time-history analysis of the correlation coefficient of the cross-section strain data.Then,based on historical monitoring data,the normal range of feature strain pattern is built with the robust regression method.Finally,quasi-real-time identification of ship collision is achieved with the feature strain residual value,the lateral strain of the middle of the bridge and the rotation of the beam end.(3)Identification of structural long-gauge(LG)strain and displacement flexibility based on subspace method at the second layer: First,the approach improves the traditional state-space equations for the longgauge strain output,as well as the algorithm of identifying LG strain flexibility.Second,the relation between LG strain flexibility and displacement flexibility is derived,which proves the displacement flexibility can be calculated by the proposed method without other parameters.Hence,two kinds of flexibility matrix is identified respectively based on the impacting test and LG strain output.In addition,in the case of multireference multi-peak input,the identification result from the proposed algorithm outperforms that from a frequency-domain method called Complex Mode Indicator Function(CMIF).This comparison illustrates the proposed method is suitable to handle different kinds of the input force,which leads to a wide usage in practice.(4)A displacement-flexibility-based damage identification for a complex spatial structure: In this thesis,a double-arch rib tied arch bridge was simulated and analyzed through ANSYS.Both vertical and lateral flexibility matrix were accurately identified.It is founded that there are both vertical and lateral vibration modes in the low-frequency range for such a complex spatial structure so that the flexibility of each direction has a great effect on the structure performance.For two damage patterns at different locations of the crosssection with the same degree(the same thickness of damage),the global and local damaged locations can be confirmed with multi-direction damage indicators based on displacement flexibility.