Increasing the robustness of autonomous systems to hardware degradation using machine learning

Abstract

Autonomous systems perform predetermined tasks (missions) with minimum supervision. In most applications, the state of the world changes with time. Sensors are employed to measure part or whole of the world’s state. However, sensors often fail amidst operation; feeding as such decision-making with wrong information about the world. Moreover, hardware degradation may alter dynamic behaviour, and subsequently the capabilities, of an autonomous system; rendering the original mission infeasible. This thesis applies machine learning to yield powerful and robust tools that can facilitate autonomy in modern systems. Incremental kernel regression is used for dynamic modelling. Algorithms of this sort are easy to train and are highly adaptive. Adaptivity allows for model adjustments, whenever the environment of operation changes. Bayesian reasoning provides a rigorous framework for addressing uncertainty. Moreover, using Bayesian Networks, complex inference regarding hardware degradation can be answered. Specifically, adaptive modelling is combined with Bayesian reasoning to yield recursive estimation algorithms that are robust to sensor failures. Two solutions are presented by extending existing recursive estimation algorithms from the robotics literature. The algorithms are deployed on an underwater vehicle and the performance is assessed in real-world experiments. A comparison against standard filters is also provided. Next, the previous algorithms are extended to consider sensor and actuator failures jointly. An algorithm that can detect thruster failures in an Autonomous Underwater Vehicle has been developed. Moreover, the algorithm adapts the dynamic model online to compensate for the detected fault. The performance of this algorithm was also tested in a real-world application. One step further than hardware fault detection, prognostics predict how much longer can a particular hardware component operate normally. Ubiquitous sensors in modern systems render data-driven prognostics a viable solution. However, training is based on skewed datasets; datasets where the samples from the faulty region of operation are much fewer than the ones from the healthy region of operation. This thesis presents a prognostic algorithm that tackles the problem of imbalanced (skewed) datasets.