Multi-modal Neural Adaptive Observer for Sensor and Actuator Fault Detection and Identification

This paper presents a multi-modal extension of neural adaptive observers (NAOs) to effectively handle the coupling between sensor and actuator effects in the context of fault detection and identification. The inherent coupling in the faults of sensors and actuators often leads to ambiguity in the process of fault identification, leading to high false alarm rates. This work incorporates the concept of probabilistic multi-modal estimation in the framework of neural adaptive observers to mitigate this coupling effect. The method features multiple NAOs representing distinct fault modes, and develops a posterior-mode probability update rule that takes both the observer stability and the fault identifiability. In the process, a scheme to enhance the convergence speed of the individual NAOs is also devised, and a Lyapunov stability analysis for the multi-modal NAOs is investigated. Case studies on fault detection and identification (FDI) for a quadcopter UAV demonstrate the superior fault identifiability and stability of the proposed scheme.