Fault diagnosis and recovery in MEMS inertial navigation system using information filters and Gaussian processes

An integrated navigation system (INS) on a vehicle platform such as a quadrotor UAV is an example of a multisensor system, wherein data streams coming from different sensors are combined to bring about improved situational awareness. This paper examines the implementation of two related approaches to distributed estimation and fault diagnosis in a multi-sensor INS: a centralised and decentralised (federated) Kalman filter in information form. We adapt a designated observer scheme (DOS), i.e., filter bank approach, to make use of local filter nodes coupled to individual inertial sensors in order to achieve detection and isolation of faults. The centralised filter implemented is based on the concept of adaptive measurement fusion, which allows for adaptive estimation of the measurement covariance. We extend this feature to the decentralised design and compare the two. A further contribution is the use of Gaussian processes (GPs) in tracking INS sensor deviations from model-predicted values and using this information for fault recovery in the case of the decentralised filter. Initial simulation results show that the decentralised filter is more robust in the face of multiple faults, even as the centralised information filter provides slightly higher quality estimation.