UC Irvine

Critical infrastructure networks such as water and power systems serve as lifelines that sustain people and communities. Yet, they are becoming increasingly strained and fragile due to aging components, natural disasters, and rapid urbanization. The emergence of Internet-of-Things (IoT) technologies offers opportunities to enhance the operation of these infrastructures. This leads to the concept of cyber-physical human infrastructures (CPHIs): community-scale infrastructures that bridge sensing, communication, and computation with physical assets and people. However, enabling resilience within CPHIs is challenging: these systems operate under significant uncertainty and variability at scale, making it difficult to effectively monitor, analyze, and respond to adverse events. To address these issues, this thesis proposes a unified design methodology that jointly leverages available information about the network topology (structure), the physics-driven phenomena in the system (behavior), and the contextual knowledge that connects infrastructure events to their causes and consequences for people (semantics). By optimizing the balance between physics-based models and data-driven methods, our methodology constructs resilience solutions that remain faithful to infrastructure realities while scaling to community-sized networks. We illustrate this across diverse infrastructure use cases, each of which operates at a different spatial scale with different sensing capabilities, and faces different resilience issues. In particular, we examine: IoT deployment planning for real-time monitoring and root-cause failure analysis in stormwater networks; worst-case disruption from attacks and failures in electric power grids; and modeling human behavior in smart building CPHIs.First, we address the need for real-time monitoring and analysis to detect and respond to failures in CPHIs. We study this problem in the stormwater domain, in collaboration with a public water agency, Orange County Public Works (OCPW). In these networks, transient pollutant events are a recurring issue: such anomalies appear abruptly, cause significant harm, and dissipate before cleanup actions can be initiated. To enable rapid detection and response, we propose STEP, an optimization framework that designs practical sensor deployments, guided by community semantics such as land use, to maximize both coverage (detecting transient anomalies) and traceability (tracking anomaly origins). Then, to enable actionable decision support from sensor observations, we develop an efficient backwards inference algorithm that exploits physics-based computational models and fluid flow approximations to extend localized sensor measurements across the network and identify potential sources, supporting real-time human-in-the-loop intervention when such anomalies occur. Evaluations on six real-world stormwater networks in Southern California demonstrate improved anomaly detection and traceability over structural and behavioral baselines, and accurate reconstruction of anomalous events from sparse sensor observations.Second, we explore the modeling and analysis of complex failures in a network, and their impacts on the community. This problem is studied in the power grid domain, in collaboration with Los Alamos National Laboratory. Many grid networks depend on a few key components (generators and lines) to serve end consumers, but identifying them accurately and efficiently is difficult: doing so requires realistically modeling how such components fail, and capturing the physics-driven grid response. In this context, we leverage N−k interdiction models to analyze worst-case contingencies under: (i) sequential attacks, where components are disabled one at a time and the grid responds between failures; and (ii) cascading failures, where an initial attack can trigger secondary failures in the network. We develop a guided exploration framework that uses network science properties to efficiently navigate the combinatorial space of attack orderings, and a decoupled optimization approach that captures induced failures while remaining scalable to large networks. Evaluations on benchmark grid networks show that the order of failures can significantly alter the resulting disruption, and that accounting for induced failures substantially increases the identified worst-case impact, providing grid operators with decision support to identify components for ruggedization and reduce large service disruptions.Lastly, we examine the role of humans and their behavior as a central aspect of CPHI operation. We explore this in the context of smart building infrastructures, where data involving people and their interactions is difficult to obtain due to privacy constraints, limited instrumentation, and the inability to observe hypothetical scenarios. To this end, we present SmartSPEC, an event-driven approach to generate synthetic yet realistic smart space data that is required for the robust deployment and operation of smart spaces. SmartSPEC employs a two-phase architecture: a learning phase extracts semantic patterns from limited seed data to construct models of spaces, people, events, and sensors; and a generation phase produces synthetic trajectories and sensor observations that respect user-defined constraints. SmartSPEC also provides a structured assessment methodology to evaluate the realism of synthetic data, using customized similarity metrics for trajectories of people and space occupancy patterns; validated on two real-world settings, the generated trajectories better adhere to the underlying semantics of the space than baseline mobility models, even under hypothetical changes. Our tool has been leveraged to model many spaces, including UC Irvine campus buildings, a Navy ship for a DoD project, and senior care facilities for CareDEX, an NSF Civic Innovations Challenge project.Together, these contributions address core issues in supporting new, resilient CPHIs by optimizing the balance between physics-based infrastructure models and data-driven methods. While our work examines specific infrastructure domains, the developed techniques and frameworks are generalizable to other community-scale CPHIs that operate under uncertainty, resource constraints, and other disruptive conditions.