Sensor Fault Detection and Isolation in Multi-Agent Systems

Abstract

Sensor fault detection and isolation in multi-agent systems (MAS) with uncertain dynamics and undirected, connected communication networks is addressed in this article. The proposed approach involves a two-step process: First, fault detection, and then likelihood-based fault isolation. A novel fault reconstruction technique is introduced by tuning the unscented Kalman filter (UKF) noise covariance matrices within the Q-learning framework. This adjustment helps reconstruct the uncertain states of the MAS and train the internal parameters of a neural network using historical measurements. This innovative method is referred to as Enhanced reinforced UKF (ERUKF). To reduce neural network approximation errors, a robust control term utilizing the hyperbolic tangent function is applied. The stability of ERUKF, when combined with the robust control method, is mathematically proven using the Lyapunov theorem. Simulations illustrate that ERUKF exhibits lower estimation errors compared to adaptive UKF, achieving a 96.67% success rate in fault isolation under Monte Carlo (MC) simulations.