Development Of A New Approach For Damage Identification In Structural Health Monitoring Using A Novel Wavelet Selection Method For Feature Extraction Combined With Artificial Neural Network

The common practice in designing civil engineering systems has inherent uncertainties.Moreover,these systems are exposed to various potential damages over their life cycles due to natural disasters,environmental variations,and human-induced utilizations,which may accelerate the spread and accumulation of damage.The uncertainties and damage may adversely affect the structures’ loadbearing capacity and useful life,causing gradual or unexpected failure and resulting in loss of lives and property.Given these inevitable issues,it is important to monitor and predict sudden changes in structural behavior by assessing and updating valid and reliable information about the structure’s performance so that any troubling signs of damage can be detected early.Thus,in recent decades there has been a considerable growth in the demand for monitoring the performance and safety of civil structures.Consequently,there have been advances in structural health monitoring tools.Due to rapid developments in sensing and data acquisition technologies,for example,dense sensor placement on structural systems with minimal installation and maintenance costs is now possible.However,extensive instrumentation and continuous monitoring inevitably produce large volumes of data affected by uncertainties associated with random sampling sensor faults and signal noise due to surrounding conditions.Hence,a fundamental issue is analyzing and interpreting the vast amounts of nonstationary contaminated data collected by sensors,extracting the essential components and the most appropriate damage features,reducing the high-dimensional features,and appropriately correlating the derived information to important information structural traits.In order to address these issues,this dissertation introduces several integrated damage detection techniques that utilize statistical pattern identification methods based on wavelet transform(WT)and artificial neural networks(ANN)as AI-based schemes.The algorithms introduced herein can be applied to stationary ambient vibration response,before and after damage,and nonstationary severe motion responses,such as dynamic response due to earthquake loadings.This latter case has not been properly addressed in the literature.The major contributions of this dissertation can be categorized into three parts.A detailed description of these contributions in each part is as follows.The first category of contributions in this dissertation includes two novel strategies which are original contributions to the literature on wavelet analysis: a)a computational framework that,for the first time,provides a methodical approach,instead of the currently used trial and error-based schemes,for choosing the most appropriate wavelet parameters,and b)a rule that can define the superiority of one wavelet base function over another base functions.The efficiency and robustness of the proposed techniques are validated using experimental data collected during a shaking table test of a six-story reinforced concrete frame building conducted in E-Defense,Japan,as well as El-Centro earthquake data.As demonstrated in the studies presented in this dissertation,the proposed algorithms that are based on the aforementioned novel strategies can i)determine which wavelet parameters are able to ensure effective wavelet analysis for time-varying responses and damage feature extraction(FE)in structural health monitoring of civil engineering systems,ii)the proposed original contributions also explain how wavelet analysis allows the system to reach its extreme performance,ensure effective denoising,and ensure that the recovered signal keep the necessary information of the original signal,and iii)the proposed approach allows to customize the wavelet approach for analyzing massive nonstationary data(NSD),to separate disturbances and extract damage-sensitive features(DSFs)from high-dimensional features in Civil engineering applications(CEAs).The second category of contributions is introducing a hyper wavelet-based damage diagnosis and condition assessment tool for ambient vibration response.The technique combines wavelets with novel qualitative and quantitative criteria to extract meaningful features from various domains.It can be utilized as damage indices for monitoring and assessing of dynamic response of bridges.Various damage-sensitive features(DSFs)are developed for numerous evaluation levels based on the retrieved wavelet coefficient and wavelet energies to demonstrate this capability.The efficiency and robustness of the proposed methods are validated using experimental data from the Tianjin Yonghe cable-stayed bridge,which was tested by the Center of Structural Monitoring and Control at the Harbin Institute of Technology.The key significance of this contribution is that it proposes an extensive list and combination of nonlinear damage indexes that detect hidden features that correlate well with damage and can be used for tracking the time-varying behavior of structures to identify the structure’s state according to different requirements of hierarchy levels of(SHM).The third category of contributions includes several novel frameworks based on a data-driven approach.The first framework is introduced to select the most important sensor data,from various sensor measurements,based on information theory,for accurate and reliable feature extraction and classification data analysis.Another framework is introduced by combining WT and deep learning algorithms(DLAs)capabilities.The technique is based on a windowed-two-dimensional(2D)convolutional neural network(CNN)for multiclass damage detection and classification using acceleration responses.These approaches are validated using acceleration data collected during a shaking table test of a six-story reinforced concrete frame building conducted in E-Defense,Japan,and experimental data from the Tianjin Yonghe cable-stayed bridge.The proposed scheme shows a more robust learning ability for complex tasks.It can also detect abstract features and complex classifier borders,which can discriminate various interesting aspects and allow various problem attributes to be distinguished.The contribution explores the capabilities of WT and AI and introduces a hyper technique for multiclass damage detection and classification based on 2D-CNN and WT.This strategy can also handle issues related to big data analysis in civil structures.It can automatically extract novel DSFs based on multi-sensor measurements to analyze the extracted features’ distributions to classify the structure’s health condition.