Tutorial T2A: Safe Autonomous Systems: Real-Time Error Detection and Correction in Safety-Critical Signal Processing and Control Algorithms

摘要

While the last two decades have seen revolutions in computing and communications systems, the next few decades will see a revolution in the use of every-day robotics and artificial intelligence in broad societal applications. Examples of such systems include sensor networks, the smart power grid, self-driven cars and autonomous drones. Such systems are driven by signal processing, control and learning algorithms that process sensor data, actuate control functions and learn about the environment in which these systems operate. The trustworthiness and safety of such systems is of paramount importance and has significant impact on the commercial viability of the underlying technology. As a consequence, anomalies in system operation due to computation errors in on-board processors, degradation and failure of embedded sensors, actuators and electro-mechanical subsystems and unforeseen changes in their operation environment need to detected with minimum latency. Such anomalies also need to be mitigated in ways that ensure the safety of such systems under all possible failure scenarios. Many future systems will be selflearning in the field. It is necessary to ensure that such learning does not compromise the safety of all human personnel involved in the operation of such systems. To enable safe operation of such systems, the underlying hardware needs to be tuned in the field to maximize performance, reliability and error-resilience while minimizing power consumption. To enable such dynamic adaptation, device operating conditions and the onset of soft errors are sensed using post-manufacture and real-time checking mechanisms. These mechanisms rely on the use of built-in sensors and/or low-overhead function encoding techniques to detect anomalies in system functions. A key capability is that of being able to deduce multiple performance parameters of the system-under-test using compact optimized stimulus using learning algorithms. The sensors and function encodings assess the loss in performance of the relevant systems due to workload uncertainties, manufacturing process imperfections, soft errors and hardware malfunction and failures induced by electromechanical degradation. These are then mitigated through the use of algorithm-through-circuit level compensation techniques based on pre-deployment simulation and post-deployment self-learning. These techniques continuously trade off performance vs. power of the individual software and hardware modules in such a way as to deliver the end-to-end desired application level Quality of Service (QoS), while minimizing energy/power consumption and maximizing reliability and safety. Applications to signal processing, and control algorithms for example autonomous systems will be discussed.