Ghorbani, Mohammadreza2025-09-162025-09-162025-09-162025-09-14https://hdl.handle.net/10012/22427Modern vehicle systems are increasingly reliant on accurate and robust state estimation to ensure safe and reliable operation of advanced control functionalities, such as driver-assistance and autonomous driving systems. However, the inherent complexity, along with various sensor faults, environmental disturbances, and model uncertainties, poses significant challenges to the resilience and reliability of vehicle state estimators—especially in safety-critical applications. This thesis addresses these challenges by developing a unified framework that enables real-time fault diagnosis and reliability-based reconfiguration of estimation architectures. The core idea is to first model the interconnected architecture of vehicle state estimations as a directed graph—termed the estimation graph—where each node represents a local estimator and edges capture structural dependencies. Within this graph, multiple redundant estimation paths may exist for a given state, enabled by sensory and model-based redundancies. However, faults introduce varying levels of estimation uncertainty across these paths. This research contributes a significant methodological advancement by introducing computationally efficient techniques for real-time reliability assessment, suitable for embedded implementation. These methods enable online selection of the most reliable estimation path based on a quantified reliability index, which reflects the uncertainty due to fault propagation and supports dynamic reconfiguration to the most reliable estimation topology. Complementing this, the thesis presents a unified and hybrid fault detection and isolation (FDI) methodology that integrates residual-based analysis with data-driven learning, supported by model-based quantified fault likelihoods to enhance diagnostic performance. Moreover, structural dependencies within the estimation architecture are encoded into graph-based representations—namely, the estimation graph and fault interaction graph—which enable structural analysis and scalable fault localization. Leveraging these structural insights, two distributed fault isolation strategies are proposed: a consensus-based approach that enables partial supervision through neighborhood-informed decision-making across fault sources, and a graph neural network (GCN)-based global classifier that incorporates structural priors to enhance diagnostic accuracy and reduce training cost—making it well-suited for large-scale dynamical systems. These structure-aware and computationally efficient designs improve diagnostic performance, reduce retraining overhead compared to centralized approaches, and ensure scalability in complex systems. The proposed framework is validated through high-fidelity vehicle simulations and experimental on-road data, demonstrating its effectiveness in isolating faults, quantifying uncertainty, and improving state estimation accuracy. Beyond vehicles, the methodologies developed here are applicable to a wide range of large-scale networked systems—including industrial automation and smart infrastructure—where fault tolerance, modularity, and real-time operation are paramount. This work lays a principled foundation for scalable and resilient state estimation in the face of uncertainty and faults, marking a significant step toward safer and more reliable autonomous systems.envehicle state estimationreal-time uncertainty quantificationKalman filterfault detection and diagnosisphysics-informed machine learninggraph neural networksestimation reliabilitynetworked dynamical systemsintelligent vehiclesDoctoral Thesis