Learning-Based Fault-Tolerant Control With High-Order Control Barrier Functions

Fault-tolerant control can maintain acceptable performance in faulty systems but may violate safety constraints, posing potential risks. Additionally, model uncertainties present further challenges. In this paper, we propose a novel fault-tolerant probabilistic safe control framework that integrates Gaussian process (GP) regression, fault-tolerant control, and the high-order control barrier function (HOCBF) method. To handle inherent uncertainties, we adopt a probabilistic method that combines control Lyapunov function (CLF) with HOCBF using GP regression. Building on this foundation, we design a fault-tolerant GP-based CLF-HOCBF method to address unknown actuator faults. Furthermore, we establish two theoretical criteria to ensure the probabilistic safety and stability of the proposed control framework. To validate our method, we implement it in the autonomous driving simulator CARLA, demonstrating its effectiveness and competitiveness compared to existing approaches. Note to Practitioners—Control systems should maintain stability and performance in the presence of model uncertainties, especially by ensuring during faults until appropriate engineering measures are implemented. However, existing methods on addressing uncertainties will influenced by observation noise and imperfect measurements, and current fault-tolerant control strategies struggle to handle safety-critical control problems. Given this, this paper suggests a new approach using GP regression and HOCBF, considering the cases of unknown system dynamics and actuator bias faults. Accordingly, sufficient conditions are established for guaranteeing the probabilistic safety and stability. Our proposed method faces challenges when addressing unknown gain dynamics or unknown actuator gain faults. In future research, we will design an affine dot product compound kernel method to overcome these limitations.