Candidate: Ben Zhang
Date: May 12, 2025
Time: 11:00am
Location: online
Supervisors: Drs. Chris Nielsen and Derek Rayside
Abstract:
This thesis addresses the detection and mitigation of sensor faults in safety-critical systems through secure estimation techniques. Sensor faults, whether accidental or adversarial, pose significant risks in autonomous vehicles, aviation, robotics, and other domains heavily reliant on accurate sensor data for safe operation. Traditional fault tolerance methods typically depend on hardware redundancy or ad-hoc designs for specific systems, approaches that can be prohibitively costly, or simplified assumptions about fault conditions that may not hold in practice. Recent advances in secure estimation provide a general framework with provable guarantees against sensor faults; however, their central requirement—sparse observability—is a worst-case scenario analysis, limiting their applicability in practical systems.
To address this limitation, this thesis introduces the concept of sensor protection, explicitly modeling selected sensors as immune to faults. This serves as an initial step toward capturing the practical scenario where some sensors are more fault-tolerant than others.
Although prior studies have implicitly assumed sensor protection by restricting potential fault locations, explicit modeling of sensor protection and its theoretical implications for fault tolerance have not been formally explored. This thesis extends the sparse observability framework to include protected sensors, broadening the applicability of secure estimation's theoretical guarantees. Additionally, a metric termed the “safety factor" is introduced to quantify a system's resilience to sensor faults, enabling targeted enhancements in robustness under practical resource constraints.
Further, this thesis adapts secure state-reconstruction methods to develop a robust fault detection algorithm suitable for nonlinear systems through linearization. We validate our methods extensively through simulation studies, retrospective analysis of a real-world autonomous vehicle racing incident, and practical implementation on a skid-steer robot. Results demonstrate significant improvements in real-time fault detection and operational safety under diverse fault conditions.
Overall, this work bridges theoretical advances in secure estimation with real-world deployment considerations, providing a structured methodology to enhance the reliability and safety of autonomous systems.