Ai-aided parameters estimations for uncooperative space target with measurements failures

Parameters estimations for uncooperative space targets are vital important for the On orbit servicing operations. With limited prior information of the uncooperative space targets, good parameters estimations will not only provide rich knowledge of the target but also figure out whether the uncooperative space target can be captured and controlled. Since the uncooperative targets are tumbling in space, only remote sensors (i.e. stereo camera and LIDAR) can be utilized to measure some useful information about them. However, under the uncertain space environment, the remote sensors sometimes cannot output reliable measurements and will lead to failures of the parameters estimations. This paper focuses on parameters estimations of an uncooperative space target with measurements failures. In this paper, a LIDAR device is used to obtain the relative pose between the target and the servicing satellite, and long-term measurements failures of the LIDAR is considered. To enhance the estimating accuracies and efficiencies, the dual vector quaternions (DVQ) are utilized to model the system kinematics and dynamics. By the novel DVQ modeling technique, both the translational and rotational parameters can be estimated in the same time. Also, by utilizing the dynamic models, the inertial parameters (ratios of the moments of inertia of the target, and the location of the center of mass) can also be estimated. In the beginning, a DVQ based extended Kalman filter is designed to estimate the parameters of the uncooperative space target when the measurements of the LIDAR are available. Then, to overcome the estimating failures caused by the faults of the LIDAR measurements, a neural network aided extended Kalman filter (NNEKF) is designed. In the designed NNEKF, a BP neural network is designed in the measurement update procedure when the measurements from the LIDAR are failed. By using the historical time propagations of the states and the Kalman gain as inputs together with the relative correct estimations of the states as outputs, the designed BP neural network is trained off-line firstly. Then, when the measurements of the LIDAR is available, the trained BP neural network is retrained by the results from the DVQ-EKF in each estimating step to enhance its online estimating accuracy. As soon as the measurements failures happen, the designed BP neural network will output the estimations of the parameters instead of the ones made by DVQ-EKF. When the measurements of the LIDAR are available again, the estimations of the designed BP neural network will be used to reset the DVQ-EKF, which will be used under available LIDAR measurements situations. By the proposed DVQ-NNEKF, the measurements update procedure can switch between the designed BP neural network and the EKF to enhance its accuracy due to the availabilities of the LIDAR measurements and mitigate the computational burden. Finally, the proposed DVQ-NNEKF is validated by mathematical simulations to show its robust performances.