Robust attitude estimation of rotating space debris based on virtual observations of neural network

High precise estimation and prediction of the target's attitude motion are key technologies for capturing and removing rotating space debris. In this article, a neural-network-enhanced Kalman filter (NNEKF) is proposed to improve the precision and robustness of attitude estimation algorithm. The main innovation of the NNEKF is to utilize virtual observations of the inertia characteristics to improve the filter's performances. The virtual observations are obtained using a neural network, which is offline trained using simulation data. In order to decrease the number of nodes of the network, the input data are preprocessed using the discrete Fourier transformation method. Moreover, by involving the characteristic frequencies in the input vector, the neural network can extract information from all the past observations, so as to grasp long-term characteristics of the dynamical system. Therefore, the NNEKF can provide more precise estimation of the target's moment of inertia, and furthermore improve the accuracy and robustness of attitude estimation and prediction. Simulation results indicate that the NNEKF can reduce the estimation errors by 39% compared with the conventional EKF method when using the same measurement data. And the accumulation errors of prediction using estimates of the NNEKF is just as 24% as the conventional EKF.