Tensor factorization in civil infrastructure systems

In today's world, advancements in data collection equipment and storage capabilities of computers have made it possible for researchers and asset managers to collect huge amounts of data generated from civil infrastructure systems. The conventional approach for analyzing structured data of this nature is to impose a matrix- structure on the data before analysis is done. Analysis tasks may include classification, clustering, regression and exploratory analysis. These are generally done with the aim of gaining in-depth knowledge about civil infrastructure systems that will be the basis of planning and future decision-making. However, data generated from civil infrastructure such as pavements and bridges may not necessarily have a matrix-structure. They may have an inherent multidimensional structure. Using analysis techniques that preserve this natural multidimensionality can provide additional knowledge about infrastructure systems previously unknown. In this dissertation, multiway data analysis, also known as tensor factorization are utilized to address the multiway nature of data from the National Bridge Inventory (NBI). The NBI is the largest bridge database with detailed information on the condition and general attributes of bridges nationwide. Interestingly, individual bridges have over a hundred items of data each. Given the fact that this data has been collected over several years, the NBI data base can be considered as a typical multidimensional data set in civil engineering. The two main tensor factorization techniques namely Tucker decomposition and PARAFAC decomposition are used to reduce the multidimensional NBI data sets into smaller sets that are used as a basis for clustering and prediction. Four applications on the NBI database are discussed in this work. The first application involves a study of structural deficiency of bridge design types across states in the US. The second application also seeks to cluster states based on structural deficiencies and functional obsolescence of bridges under different classifications of traffic volumes. The third analysis uses a time series prediction approach with the tensor factorization to predict future structural deficiencies of bridges in states. The final analysis is an exploratory analysis which focuses on the condition of bridges in the state of Oregon. Promising results yielded by each of the analysis will encourage civil engineering researchers to conceptualize and analyze data in various application areas as being naturally multidimensional leading to more accurate knowledge discovery in systems.