Research paperGaussian-Binary classification for resident space object maneuver detection
Introduction
The Earth’s space environment is getting more and more crowded. More than 34,000 orbital debris larger than 10 cm are known to exist. The estimated population of particles between 1 and 10 cm in diameter is approximately 900,000. The number of particles larger than 1 mm exceeds 128 million. As of May 20, 2021, the amount of material orbiting the Earth exceeded 9,400 metric tons [1]. Among these objects, about 7% of the tracked RSOs are operational satellites [2], implying 93% are space debris which are man-made objects that no longer serve useful purposes but pose risks to astronauts, satellites, and space exploration equipment (e.g., moon robotics).
To detect maneuvers for RSOs accurately and timely is critical to protect the long-term sustainability of space activities, including Space Situational Awareness (SSA) — the knowledge and characterization of space objects and their operational environment to support safe, stable, and sustainable space activities; and Space Traffic Management (STM) — the planning, coordination, and on-orbit synchronization of activities to enhance the safety, stability, and sustainability of operations in the space environment.
Researchers have proposed to detect maneuvers by using the classical estimation methods with different metrics. Kelecy and Jah [3] explored the detection and reconstruction of single low thrust in-track maneuvers by using the orbit determination method based on the batch least squared and Extended Kalman filter (EKF). Song, et al. [4] presented an analysis method called Semi-major Axis Change Method (SACM) by using the relationship between the change of orbit parameters and the change of velocity to detect maneuvers. Ko and Scheeres [5], [6], [7] proposed a method based on Thrust Fourier Coefficients with EKF to detect maneuvers. Jiang, et al. [8] used a residual normalized strong tracking filter (RNSTF) to detect the impulsive maneuvers. These methods are generally established on the standard spacecraft dynamic equations. Compared to the method proposed in this paper, such methods are generally computationally intensive and limited to the various assumptions introduced by the methods.
Rather than building the methods based on dynamic models, approaches based on “data-driven” is another type of the method to detect maneuvers by checking the differences of the observed values from their corresponding historical values [9], [10], [11]. Jia et al. [12] used historical orbit trajectories and asteroid light curves. The criteria for maneuver detection are often defined based on selected statistic measures including the mean and standard deviation of the changes [13]. These methods require different thresholds for different spacecrafts. Furthermore, it is difficult to detect the contextual anomalies when the values are smaller than the threshold limit around the noise level [14]. There are other maneuver detection methods that have been proposed, such as the principal component analysis (PCA) method [15], [16], [17] to reduce the data dimension and select the subspace feature to analyze. However, the traditional PCA method does not work well with nonlinear problems [18].
Recently, more advanced machine learning methods have been proposed. Gao [19] used normal behavior clustering based on the k-nearest neighbors (kNN) algorithm to detect whether an anomaly is presented or not, and also to label the type of anomaly. Gonzalez, et al. [20] proposed a binary classification method with a multi-layer neural network to detect the anomalous data. Shen, et al. have explored a variety of approaches for satellite behavior maneuver detection from tracking data including game theories [21], convolutional neural networks (CNNs) [22], and generative adversarial networks (GANs) [23]. Tariq [24] developed a multivariate convolution long short-term memory (LSTM) with mixtures of probabilistic principal component analysis. Pilastre [25] proposed the method based on sparse representation and dictionary learning. Pang et al. [26] proposed a method which combines Markov chains and a probability prediction method to detect the anomalies. Recent methods include time-delay Neural networks (TDNN) [27], clustering [28], and orbital control theory [29].
In an earlier study, Wang, et al. [30] proved that using AutoEncoder with binary classification can detect whether there is a maneuver or not between two tracks of the same RSO on different orbital paths for a predefined impulsive maneuver. In this paper, we explore a data-driven GBC model to detect maneuvers for three different cases: the maneuver is due to a small impulsive velocity change; the maneuver is due to a low thrust, and the maneuver is due to either an impulsive or a low thrust which is unknown a priori. The orbits and measurements are simulated from an in-house simulation-based space catalog environment. The simulation-based approach supports comparative methods over the sensor, environment, and target (SET) operating conditions such as scenarios of different RSO behaviors. Likewise, with the known “truth” values, performance metrics and bounds can be verified for effectiveness and efficiency. Furthermore, a controlled environment affords testing the robustness of the proposed methods on various scenarios that are presented in this paper.
This paper makes the following contributions. First, to the best of our knowledge, this is the first time that a GBC method is used for RSO maneuver detection. Second, the GaBRSOMD in this paper only needs 5 measurement data from each track to detect a maneuver change, whereas the classical orbit determination can require around 10 measurements due to the measurement errors [31]. Third, the GaBRSOMD can detect whether there are maneuvers between the two tracks of the same RSO on different orbital paths or not with both high accuracy and high reliability. Once trained the GaBRSOMD model can make quick detection decision without solving the orbit estimation problem. Furthermore, the model is robust with data with noise different from what are used in the training data.
The rest of this paper is organized as follows. Section 2 describes the methodology in detail, including the algorithm of GBC and the simulation environment. Sections 3 Impulsive maneuver detection, 4 Low thrust maneuver detection discuss two types of maneuvers, and present the simulation results. Section 5 investigates a hybrid case, compares the results to a principal components analysis (PCA) and an AutoEncoder, and compares the model robustness. Conclusions are presented in the last section.
Section snippets
Gaussian binary classification algorithm
For the space monitoring activity, assume there are observations: . to , and . is an input matrix with . is the number of input dimensions. is the corresponding response with value being either 1 or 1. The goal is to predict the probability of the response being either 1 or 1 for the test data , which is achieved by using a latent function to map the input into the unit interval by a sigmoid function. The general Gaussian Binary
Maneuver definition
To collect enough data to train the model, for a maneuver with a predefined magnitude, we simulate the orbits of a spacecraft for 280 days with the specified maneuver every 28 days, so there are ten orbits and 9 maneuvers in this process. The “DataSet” consists of the 280-day simulation results with the 9 maneuvers. For the DataSet, one orbit is defined as the satellite motion without maneuvers. For simplicity, we refer to these orbits as “orbit 1”, “orbit 2”,…, and “orbit 10” in the DataSet.
Maneuver definition
In this experiment, we keep the initial condition of the satellite orbits as described in the previous section, but change the impulsive maneuver to a low thrust maneuver at the end of every 28 days along the in-track direction and for a finite duration. We explore the low thrust maneuvers by referring to the information in Shafieenejad’s paper [44]: the initial mass of the satellite is 1000 kg, and the specific impulse of the engine is 1500 s. In Shafieenejad’s paper, they define the
Maneuver definition
In this section, we test the performance of the GaBRSOMD for a hybrid situation in which the maneuver can be either a small impulsive velocity change or a low thrust maneuver. Both the training data and test data are designed to include both cases. The training database is shown in Table 13, and the test database is shown in Table 14.
To maintain the consistency, the selection of the two databases is the same as the previous sections. The maneuver information contained in test database is
Conclusion
This paper proposes a GaBRSOMD model to detect whether there is a maneuver or not between two tracks of a resident space object on different orbit paths. The experiments includes three maneuvers: (1) small impulse change, (2) low thrust, and (3) impulse or thrust, with sensitivity tests to data noise; and three methods: Gaussian Processes, PCA, and AE.
With a well-trained model, different types maneuvers have been tested. When the maneuvers are impulsive velocity change with different
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was partially supported by the Air Force Research Laboratory, USA contract FA8750-19-C-1000. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the United States Air Force.
References (48)
- et al.
Detection and orbit determination of a satellite executing low thrust maneuvers
Acta Astronaut.
(2010) - et al.
A simple and valid analysis method for orbit anomaly detection
Adv. Space Res.
(2012) - et al.
New manoeuvre detection method based on historical orbital data for low earth orbit satellites
Adv. Space Res.
(2018) - et al.
Anomaly detection in mixed telemetry data using a sparse representation and dictionary learning
Signal Process.
(2020) - et al.
On-orbit recognition of resident space objects by using star trackers
Acta Astronaut.
(2020) - et al.
Improving orbit prediction accuracy through supervised machine learning
Adv. Space Res.
(2018) - et al.
Comparing and analyzing min-time and min-effort criteria for free true anomaly of low-thrust orbital maneuvers with new optimal control algorithm
Aerosp. Sci. Technol.
(2014) - Space debris by the numbers, http://www.esa.int/Safety_Security/Space_Debris/Space_debris_by_the_numbers,...
Orbital debris quarterly news
Natl. Aeronaut. Space Adm.
(2019)- H.C. Ko, D.J. Scheeres, Spacecraft orbit anomaly representation using thrust-fourier-coefficients with orbit...
Maneuver detection with event representation using thrust fourier coefficients
J. Guid. Control Dyn.
Tracking maneuvering satellite using thrust-fourier-coefficient event representation
J. Guid. Control Dyn.
Residual-normalized strong tracking filter for tracking a noncooperative maneuvering spacecraft
J. Guid. Control Dyn.
Satellite maneuver detection using two-line element (TLE) data
Space event detection method
J. Spacecr. Rockets
Space object classification using fused features of time series data
Adv. Maui Opt. Space Surveill. Technol. (AMOS)
An overview on trajectory outlier detection
Artif. Intell. Rev.
Athmos: Automated telemetry health monitoring system at GSOC using outlier detection and supervised machine learning
Anomaly detection of spacecraft based on least squares support vector machine
Anomaly detection for spacecraft using hierarchical agglomerative clustering based on maximal information coefficient
An unsupervised anomaly detection approach for spacecraft based on normal behavior clustering
Combining negative selection and classification techniques for anomaly detection
Cited by (14)
GEO spacecraft maneuver detection based on causal inference
2023, Advances in Space ResearchManoeuvre detection in Low Earth Orbit with radar data
2023, Advances in Space ResearchCitation Excerpt :For instance, Jaunzemis et al. (2016) consider Gaussian-mixture models and binary hypothesis testing; manoeuvre detection is developed using Mahalanobis distance and control distance metrics, and tested for geostationary satellites with optical measurements. In Wang et al. (2021), using an in-house simulated space catalog environment, the Gaussian Binary Classification method is used to detect several types of maneuvers; results are validated only by simulation. In Escribano et al. (2022), a Markov Chain Monte Carlo sampling scheme is used to test the post-maneuver observation in terms of control distance metric, the result is tested for geostationary satellites with optical measurements.
A hybrid data-driven framework for satellite telemetry data anomaly detection
2023, Acta AstronauticaA machine learning method for the orbit state classification of large LEO constellation satellites
2023, Advances in Space ResearchCitation Excerpt :Abay et al. (2018) proposed a maneuver detection method for space objects using Generative Adversarial Networks. Other machine learning methods such as the k-nearest neighbors, multi-layer neural network, convolutional neural networks, generative adversarial networks, long short-term memory, time-delay Neural networks, and Auto-Encoder with binary classification are adopted as well (Gao et al., 2012; Gonzalez, et al., 2002; Shen, et al., 2019; Shen et al., 2020a; Shen et al., 2020b; Tariq et al., 2019; Mortlock and Kassas, 2021; Wang et al., 2021a). This manuscript aims to develop a machine learning method for the orbit state classification of large LEO constellation satellites based on TLE segments.