Fault diagnosis for MEMS INS using unscented Kalman filter enhanced by Gaussian process adaptation

Miniature unmanned aerial vehicles (UAVs) such as quadrotors are increasingly in demand due to their small size and cost. The base navigation solution for such systems is typically a micro electro mechanical system (MEMS) based strap-down inertial navigation system (INS). To allow safe operation, navigation instrument failures need to be robustly handled through effective fault diagnosis. A popular approach to fault diagnosis in non-linear systems is the extended Kalman filter (EKF), which may, however, prove sub-optimal in the presence of greater non-linearity. In this paper, we instead adopt an unscented Kalman filter (UKF), which relies on a more accurate stochastic approximation - the unscented transform - rather than a Taylor series expansion. A downside to MEMS inertial navigation is an attendant time-dependent drift, which can distort estimation quality. Hence, MEMS INS sensors characteristically result in large biases in the navigation solution. To mitigate this problem we employ Gaussian Processes to approximate a time-dependent offset which can be utilised during on-line operation in an adaptive fashion, as a compensatory mechanism. We apply the enhanced GP-UKF by means of a bank of dedicated observers within an analytical redundancy framework. The results are competitive with the EKF and represent arguably the first application of an enhanced GP-UKF filter in the context of fault detection and isolation.