Adaptive Order-Switching Kalman Filter for Orbit Determination Using Deep-Neural-Network-Based Nonlinearity Detection

Abstract

This paper proposes an adaptive estimation algorithm for orbit determination, which consists of a deep-neural-network (DNN)-based nonlinearity detector combined with an adaptive order-switching procedure, to reduce the computational complexity while still maintaining the estimation accuracy. The DNN is trained to quickly evaluate the nonlinearity degree of the state equation. An adaptive order-switching strategy is designed based on the nonlinearity degree predicted by the DNN. The algorithm switches to a high-order method when the nonlinearity of the state equation is significant and uses a linear method when the nonlinearity degree is low. The proposed method is applied to estimate the orbit of a spacecraft in cislunar space. The sample forms in the inertial frame and rotating frame are investigated and compared to find the optimum one to train the DNN. Simulations show that the proposed method can deliver accurate state estimations comparable with the state estimations obtained by the second-order extended Kalman filter but with only half of the computational cost.