Two-line elements are widely used for space operations to predict orbits with a moderate accuracy for 2-3 days. Local optimization methods can estimate a TLE as long as there exists an initial estimate, whereas global optimization methods are computationally intensive, and estimating a large number of them is prohibitive. In this paper, the feasibility of predicting the initial estimates within the radius of convergence of the actual TLEs using machine learning methods is investigated. First, a Monte-Carlo approach to estimate a TLE, when there is no initial estimate that is within the radius of convergence of the actual TLE, is introduced. The proposed Monte-Carlo method is leveraged for demonstrating the behavior of the fitting error between the realistic trajectory and the trajectory propagated by SGP4 theory during the TLE estimation processes and evaluating the unbiased performance of the proposed machine learning models. Second, gradient boosting decision trees and fully-connected neural networks are trained to map the orbital evolution of space objects to the associated TLEs using 9.5 million publicly available TLEs from the US space catalog. The desired precision in the mapping to estimate a TLE is achieved for one of the three test cases, which is a low area-to-mass ratio space object.

两行元素广泛用于太空操作，以2-3天的中等精度预测轨道。只要存在初始估计，局部优化方法就可以估计TLE，而全局优化方法则需要大量计算，并且估计大量这种方法是禁止的。在本文中，研究了使用机器学习方法在实际TLE的收敛半径内预测初始估计的可行性。首先，当没有初始估计值位于实际TLE的收敛半径之内时，采用蒙特卡洛方法估计TLE。提出的蒙特卡洛方法被用来证明在TLE估计过程中，真实轨迹与SGP4理论传播的轨迹之间的拟合误差行为，并评估了所提出的机器学习模型的无偏性能。其次，训练梯度增强决策树和完全连接的神经网络，使用来自美国太空目录的950万个公开可用TLE，将空间物体的轨道演化映射到相关的TLE。对于三个测试用例之一，这是一个低面积质量比的空间对象，在映射中估计TLE所需的精度得以实现。训练了梯度增强决策树和完全连接的神经网络，以使用来自美国空间目录的950万个公开可用的TLE，将空间物体的轨道演化映射到相关的TLE。对于三个测试用例之一，这是一个低面积质量比的空间对象，在映射中估计TLE所需的精度达到了。训练了梯度增强决策树和完全连接的神经网络，以使用来自美国空间目录的950万个公开可用的TLE，将空间物体的轨道演化映射到相关的TLE。对于三个测试用例之一，这是一个低面积质量比的空间对象，在映射中估计TLE所需的精度达到了。