Real-Time Adaptive Dynamics Based State Estimation Scheme for Unmanned Aircrafts

In this paper we present a state estimation scheme for Unmanned Aircrafts (UAs) utilizing dynamics based models and multi-sensor data fusion. Employing the UA dynamics in estimation can substantially enhance the estimator performance, but obtaining accurate dynamics parameters for each UA is computationally costly and complex. To eliminate these issues, we propose two decoupled Extended Kalman Filters (EKFs), namely the Rotational Decoupled Extended Kalman Filter (RDEKF) and the Translational Decoupled Extended Kalman Filter (TDEKF). The dynamics parameters in these filters are identified in real-time using the Deep Neural Network and the Modified Relay Feedback Test (DNN-MRFT) approach. This approach doesn't demand prior knowledge of the UA physical parameters, requiring only an Inertial Measurement Unit (IMU) and a positioning system for model classification. Our estimation scheme provides position, velocity and attitude estimates, in addition to smooth lag-free inertial acceleration estimates. We show experimentally the advantages of our approach on trajectory tracking problems that uses low rate position sensors. We also demonstrate how utilizing the estimated acceleration in feedback control can reduce the tracking error of an optimally tuned system by 43%. Moreover, the proposed estimator produces smooth estimates that leads to a reduction of controller action by 6.6%, when compared to kinematic based estimators. We compare the achieved results against other methods that require full prior knowledge of the UA parameters or the noise models, and show advantages in performance and real-time capability.