Embedded Fault Detection for Low-Cost Safety-Critical Systems

ABSTRACT

Fault tolerance is vital to ensuring the integrity and availability of safety critical systems. Current solutions are based almost exclusively on physical redundancy at all levels of the design. The use of physical redundancy, however, dramatically increases system size, complexity, weight, and power consumption. Moreover, such systems are extremely expensive in terms of both the design and development, as well as the unit production costs. This research addresses the basic challenge of bringing high levels of reliability and integrity to other domains-domains that can afford neither the cost nor the extra power, weight, and size associated with physical redundancy. Our main research focus is on the development of algorithms and computing architectures which enable the detection of faults without relying on physical redundancy. In this context, we aim to address both detection of faults in the physical domain (sensor, actuator, and environment faults) as well as faults in the logical, cyber-domain (software and processor faults).

There is a significant opportunity to use model-based and data-driven monitoring techniques to reduce the reliance on physical redundancy. We are pursuing a three-pronged approach to fault detection. First, we are investigating the use of robust filters to estimate key signals using sensor measurements, actuator commands, and models of the physical system being controlled. We have developed a new algorithm to design optimal filters that are robust to model uncertainties. Faults in the physical (sensors, actuators, and environment) system can be detected by comparing the estimated quantities to the corresponding measured signals. Second, to detect faults in the cyber (software and hardware) domain, we are investigating the use of monitors derived from model-based software requirements. We have developed an efficient algorithm to non-intrusively measure structural coverage of an application during execution. The monitoring algorithm uses a bit set to describe the coverage obligations for an application to be monitored. Third, we are investigating the use data- driven anomaly detection methods to detect unexpected failure modes in the physical as well as cyber domains. We have constructed a hybrid algorithm capable of detecting many classes of faults by combining predicate logic satisfiability with existing statistical methods. Our algorithm detects contextual faults inaccessible to pure data-driven approaches using domain information to seed its knowledge base during training. We believe that these various fault detection methods are complementary schemes that are well-paired for reliably detecting a wide variety of faults. Implementing this three-prong approach to fault detection requires significant communication between the various software components to ensure the data sharing does not unduly affect the behavior of the application being monitored. This would be difficult, if not impossible, with existing software and computing architectures. We are currently developing extensions to multi-core processor architectures that enable efficient and non-intrusive monitoring algorithms for control software.

An important component to our research is the formulation of realistic benchmark problems that span across the boundary between the physical and cyber domains. We have successfully integrated and applied the model-based and data-driven fault detection algorithms to the UMN Aerospace Engineering and Mechanics Uninhabited Aerial Vehicle (UAV) test platform. The UMN UAV test platform consists of an Ultrastick radio- controlled aircraft with 1 meter wing span, a real-time operating system, inertial measurement unit, air data system, GPS, flight control computer and telemetry system embedded on the aircraft. Software- and hardware- in-the-loop simulations of the vehicle have been developed to validate flight control and fault detection algorithms prior to flight. The model-based and data-driven fault detection algorithms were initially validated on the UAV simulation model, and successful used to identify faults during flight tests. The UAV flight tests involved the injection of bias and ramp faults into an aileron actuator in the presences of a flight control system. Flight test validation of these algorithms is especially note worthy as it directly achieves the objective to develop embedded fault detection algorithms for low-cost, safety-critical systems.

Award ID: 0931931