Hai Jiang, National Astronomical Observatories, Chinese Academy Sciences; Jing Liu, National Astronomical Observatories, Chinese Academy Sciences; Hao-wen Cheng, National Astronomical Observatories, Chinese Academy Sciences; Yao Zhang, National Astronomical Observatories, Chinese Academy Sciences
Keywords: Space debris, TLE Uncertainty, Neural Network
Abstract:
A growing number of space activities have created an orbital debris environment that poses increasing impact risks to existing space systems and human space flight. In order to avoid the on-orbit collision events, accurate orbital positions are needed, so does the need to improve the knowledge of orbital states and associated covariance. The covariance describing the accuracy of a satellite state is an important input for exercises, such as conjunction analysis and re-entry predictions, which are increasingly important for operating in todays space environment. Through error propagation, the probability of potential collisions can be calculated and a spread of impact locations and times anticipated. These efforts help significantly in managing and mitigating the problem of space debris.
Two-line elements (TLEs) present the most comprehensive and up-to-date source of Earth-orbiting objects and are key in many monitoring and analysis activities. Despite the importance of TLEs, they have many drawbacks: limited accuracy, miss maneuvers, and perhaps most importantly, lack covariance information. The lack of covariance information of TLEs has initiated numerous studies. A wide range of methods has been proposed to estimate the uncertainty information. These approaches differ greatly in complexity, accuracy and applicability. However, they can be divided into two classes: methods using only TLEs and methods relying on additional data. Unfortunately, methods relying on external data have many limitations, such as data are not available for the far majority of objects. Moreover, uncertainties derived for a few objects are hard to extrapolate across the population or time due to their dependency on object properties (shape, size, etc.), orbit type (semi-major axis, eccentricity, inclination, etc.), variability of the environment (solar, magnetic flux, etc.), and the models and determination routines of TLEs.
In this paper, a new short-term TLE uncertainty estimation method based on an artificial neural network model is proposed. An artificial neural network (NN) is a parallel distributed system consisting of massively interconnected simple processing units, also referred to as artificial neurons. It is a type of nonlinear model representation inspired by biological neural networks.Object properties, orbit type, space environment and prediction time-span are considered as the input of the network, the propagation errors in the direction of downrange, normal and conormald are as the output of the network. A multilayer perceptron (MLP) neural network with two layers of neurons is used in this work. The hidden layer consisting of 15 artificial neurons and the output layer has three artificial neurons. The number of neurons in each layer is fixed. For each set of input data (object size, semi-major axis, eccentricity, inclination, B-Star, solar, magnetic flux, the prediction time-span), the network provides a set of position uncertainties (the propagation errors in the direction of downrange, normal and conormald), which corresponds to a nonlinear function. Since the problem under investigation is a nonlinear process, the activation function applied to the hidden neurons is the hyperbolic tangent sigmoid function. For the output layer, a linear function was considered. The inputs are object properties, orbit type, space weather and prediction time-span. Usually, object properties can be calculated with SEM model. Three months TLE data are picked-up for orbit calculation, In order to assure the chosen orbit for training is not for an object using station-keeping, only debris and R/B are used. TLEs are used for covariance estimation analysis, which are mean values obtained by removing periodic variations in a particular way. In order to obtain good predictions, these periodic variations must be reconstructed in exactly the same way they were removed by the model suitable for TLEs. The corresponding space weather data are downloaded from the public website. We use pairwise-differencing method to generate the uncertainties as the corresponding outputs of the neural network model. We assumed that the state vector given is accurate and used it as a reference. We propagated the preceding element set to the epoch of the reference TLE data, and compared the position and its trivariate components: downrange, normal and conormal. We can get trivariate components errors. 90-percent of the data will be used for training and the left data will be used for network validation. The networks efficiency is also validated with some objects with high ephemeris data. Overall, the method proves accurate, computationally fast, and robust, and is applicable to any object in the satellite catalogue, especially for those newly launched objects.
Date of Conference: September 11-14, 2018
Track: Poster