Data-Driven and Machine Learning Methods for Improving Infrastructure Performance and Health Assessments under Complex Environmental and Operational Conditions

Abstract

Infrastructure systems (IS) play a vital role in supporting the well-being of our society. A grand challenge confronting IS managers is existing asset management methods are falling short in ensuring safe and reliant IS components. Recent technological advances in the field of sensing and information technology have created opportunities to explore new approaches to managing IS components based on the use of data as quantitative evidence of structural performance and health. In tandem, advances in data science and machine learning (ML) have resulted in new data-driven analytical tools for efficiently processing large volumes of data. This dissertation explores the creation of data-driven analytical frameworks that extract information from the data generated by structural monitoring systems to help make better asset management decisions centered on structural performance and health. A challenge of assessing performance and health of IS components is the large variability such systems have in their environmental and operational conditions (EOCs). Hence, the overarching goal of the dissertation is to develop data-driven analytical frameworks that identify EOC for data normalization that improve structural performance assessments. First, the thesis explores new approaches to handling EOCs during the data normalization stage of structural health monitoring (SHM) algorithmic frameworks. The thesis proposes the extraction of EOC Sensitive Features (EOCSFs) from structural response data with EOCSFs used for data normalization. Unsupervised clustering of EOCSFs are used to establish EOC clusters during training to ensure damage sensitive features (DSFs) extracted from response data of the structure in an unknown state are fairly compared to DSFs of the healthy structure operating in the same EOC state. To normalize test data, a novel soft assignment approach is also proposed to account for the uncertainties associated with assigning an EOCSF to a given EOC cluster. These innovations are shown to outperform traditional hard assignment using EOCs inferred from measurements taken independently of the structure. Wind turbines that experience wide EOC variability are used as an illustrative example. Second, the dissertation challenges the assumption of independence and identical distribution traditionally applied to condition assessments of a structure over prolonged observation periods. A novel approach to training Hidden Markov Models (HMM) to track structural deterioration with state dependencies under varying EOCs is proposed. The Z24 Bridge SHM testbed is adopted to validate the efficacy of the method to assessing the condition of a structure based on past observations of structural condition. Finally, in the last part of this dissertation, two ML-based frameworks are applied for EOCs identification and response assessment of highway bridges. First, an encoder is trained to extract truck weight characteristics from bridge response to truck traffic using training data collected from a cyber-physical system (CPS) architecture that links bridge responses with measured vehicular weights. Next, a sequence-to-sequence (Seq2Seq) model is used to forecast the response of one bridge to a truck given the response of another bridge to the same load. The Seq2Seq model enables the estimation of bridge responses in a highway network by using the response of limited number of instrumented bridges. It is shown that by normalizing the input observations based on vehicle load type, the predictive performance of the Seq2Seq model is increased. In summary, the thesis breaks new ground in advancing data-driven frameworks that can automate the conversion of IS monitoring data into valuable information for a plethora of IS applications.