Machine Learning Tools for Handling Missing Data and Unbalanced Datasets in Engineering

Abstract

Communities rely on critical infrastructures that connect and power our urban systems. Due to the dramatic changes in the world climate scenarios, these systems are subject to harsh loads and severe environmental conditions not foreseen during their design stage. In this context, planning for resilient and sustainable infrastructures is vital, and it requires understanding and learning the complexities and dynamic interdependencies across all systems. As such, there is a need for continuous and reliable monitoring data to model and predict these complex interactions.

Artificial Intelligence (AI) and Data-Driven (DD) monitoring help make society's infrastructure systems more informative and adaptive through strategies that heavily rely on data availability. Despite technological advancements, in real applications, we are challenged with unreliable sensor networks and the unbalanced nature of datasets, which typically do not contain enough information from all the different conditions that need to be modeled. Consequently, the learning task carried on with AI and DD methods becomes less robust, case-dependent, and specialized for a particular system and a very limited number of applications.

This talk focuses on AI-driven methodologies that address the systemic lack of data and monitoring information crucial to a robust dynamic and generalized characterization of infrastructural systems. First, a virtual sensing predictive strategy based on classification and regression models is addressed to overcome the absence or malfunction of monitoring sensors. The methodology guarantees a reliable and real-time stream of environmental load information to properly assess the performance of energy systems such as off-shore wind turbines. Then, to overcome the limitations caused by unbalanced datasets, a unique approach for damage classification in structural systems is presented. It will be shown how the data shape of human voices is similar to the one of structural responses and how this information can be used to learn from a rich dataset of human voices, acquiring higher-level features characterizing vibration traits. The acquired knowledge is then transferrable to structural system domains, characterized by fewer data points, to classify their structural condition.

Bio:

Eleonora Maria Tronci is a Postdoctoral researcher at Tufts University in the Civil and Environmental Engineering Department. She received her second doctoral degree in 2022 from the Civil Engineering and Engineering Mechanics Department at Columbia University. Eleonora began the doctoral program in January 2018 after completing her bachelor’s, master’s, and first doctoral degree in Structural Engineering at the University of Rome “La Sapienza”.

Her research at Tufts focuses on developing machine learning and transfer learning integrated SHM frameworks for offshore wind turbines. Over the years, she investigated new parameters to be used as damage-sensitive features for civil and mechanical systems applications. She developed automated procedures based on unsupervised tools to extract modal parameters from the structural vibration response to pursue robust long-term monitoring campaigns. In her recent doctoral research, Eleonora carried out a pioneering and creative work focused on a cutting-edge transfer learning technique that takes advantage of speech signals to enrich the potential of damage classification procedures for civil applications.