Ángel Gallego, GMV; Adrian de Andrés, GMV; Carlos Paulete, GMV; Marc Torras, GMV; Alfredo M. Antón, GMV
Keywords: object characterization, data fusion, attitude estimation, light-curves, RCS, Machine Learning, SST data simulation
Abstract:
Space Surveillance and Tracking (SST) routine operations often involve performing orbit determination on a large catalogue of Resident Space Objects (RSOs), combining multiple data sources through data fusion to yield more accurate results. However, there is a growing interest in using the same and other sources of data to determine additional information about the objects in the catalogue. This object characterization goes beyond orbit determination, providing useful information for commercial, civil and military applications and advanced SST services. Among the possible gathered properties, perhaps the most valuable are those related to the attitude of the RSOs.
In this paper we propose a methodology for object characterization using data fusion in a broad sense, as explained later. It consists of several algorithms and methods, ranging from simple cross-checks of the information of different sources, up to the use of advanced Machine Learning (ML) techniques, also including the use of data simulators, such as light-curves or Radar Cross Section (RCS). The characterization obtained considers the classification of objects according to their type of stabilization, the determination of the apparent rotation period of rotating objects, and up to the full attitude mode estimation of an observed RSO. Even, when prior physical information (e.g. size, shape) is available, it is combined to determine which is the most accurate or is used as input data for the various algorithms. However, when no previous information exists, these methods also give an estimate of it.
In this study, data fusion is understood in its broadest sense, not only as the mixing of data of different types, but also as the mixing of data of the same type but at different times or sensors (e.g. RCS data from multiple radars). All sensor data typically provided by SST sensors (astrometry, radar range and range rate, laser ToF, passive TDoA/FDoA, light-curves, RCS…) are considered. Even higher-level processing data, such as direct information on size, shape, materials, attitude and rotation, which some sensors can provide, are now being considered in the study. However, one of the biggest problems in current SST operations and studies is the availability of such sensor data, either because of the need to reach different commercial agreements between entities, or because of restricted access to certain information. For the same reason, the methodology also considers the use of information from external public sources (Space-track.org, Celestrak, ESAs DISCOS ), to be included in the data fusion process.
In order to achieve the aforementioned characterization and enable full attitude estimation, we have designed an incremental approach that follows several steps and a decision tree. This architecture allows earlier steps to collect more general information about the attitude to estimate with limited data, while later stages are able to perform more accurate attitude determination. The first step involves classifying the objects based on their stabilization type (3-axis stabilized or in rotation) with Machine Learning techniques, following a previous work in this line conducted with RCS measurements only, and now extending with light-curves. Once classified, the apparent rotation of the corresponding objects is determined by means of Lomb-Scargle periodograms, which select candidate rotation rates, and epoch folding to find the optimum based on folding error. The results are combined with the a-priori information of the object (size, shape, materials ) and enter into the attitude estimation process, which determines the pointing law for the 3-axis stabilized objects or the rotation axis and rate for objects in rotation. This algorithm is based on iterative filters to minimize the residuals of real measurements, with respect to simulated measurements. It is at this point that advanced physical simulators are required. For each type of measurement used, a simulator is necessary to generate accurate synthetic observations of the same type. In the case of this study, attitude determination is done with light-curves, and it is intended to be extended to RCS. Hence, along with the algorithms for characterization and attitude determination, we have developed data simulators, based on OpenGL, of these measurements, which are presented in this same paper.
Our methodology is undergoing testing and validation using real data from reference cases. However, the primary issue encountered is the low availability of real data, which poses challenges in both training the ML algorithm and validating the algorithms. Moreover, even when real data is available, it is difficult to find cases where a real data fusion can be performed. In the case of the simulators, it has been possible to validate them using models of known satellites for which real measurements were available. The SMOS satellite was used as the base case for validation, producing very realistic light-curves in all cases analyzed.
Results obtained so far have been promising for each of the different algorithms and characterization steps. For stabilization classification, success rates of up to ~85% have been achieved using both RCS and light-curves separately, and improvements of up to ~90% are seen when data fusion can be performed or the number of tracks increases. Within the determination of rotation period the results are very accurate, particularly when using light-curves (alone or in combination with other data), leading to errors of less than 1%. However, attitude estimation with light-curves produces more variable results, depending on the initial knowledge of the RSO. The correct attitude mode is identified in a 90% of cases, even resolving the complete attitude in some of them, when there is prior knowledge of shape and size, and an initial attitude estimate is available, or when a large set of data is present. On the other hand, accuracy and convergence problems have been also encountered when there is little initial knowledge of the object, which also implies limited data fusion.
In conclusion, the methodology and simulators developed represent a significant step forward in the accurate characterization and data simulation of objects for SST applications. The results of this study suggest that the methodology has the potential to be highly valuable for a range of future applications.
Date of Conference: September 19-22, 2023
Track: Satellite Characterization