Real-Time Telemetry-Based Recognition and Prediction of Satellite State Using TS-GCN Network

Abstract

With the continuous proliferation of satellites, accurately determining their operational status is crucial for satellite design and on-orbit anomaly detection. However, existing research overlooks this crucial aspect, falling short in its analysis. Through an analysis of real-time satellite telemetry data, this paper pioneers the introduction of four distinct operational states within satellite attitude control systems and explores the challenges associated with their classification and prediction. Considering skewed data and dimensionality, we propose the Two-Step Graph Convolutional Neural Network (TS-GCN) framework, integrating resampling and a streamlined architecture as the benchmark of the proposed problem. Applying TS-GCN to a specific satellite model yields 98.93% state recognition and 99.13% prediction accuracy. Compared to the Standard GCN, Standard CNN, and ResNet-18, the state recognition accuracy increased by 37.36–75.65%. With fewer parameters, TS-GCN suits on-orbit deployment, enhancing assessment and anomaly detection.

1. Introduction

The emergence of Mega Constellations and Intelligent Spacecraft marks the beginning of a new era in space exploration, presenting a significant challenge for the future development of intelligent satellite control and management systems (ISCM) [1,2,3]. Recognizing and predicting satellite states through telemetry data is an essential aspect of ISCM. The process of recognizing the state of a satellite involves the automated identification of its current operating condition, whereas predicting the state of a satellite refers to forecasting its subsequent operating condition. To a certain degree, these processes can mitigate abnormal circumstances during satellite system operations and offer decision support for control and management. For instance, fault detection, as an integral component of state identification, can facilitate the implementation of preemptive measures through the prediction module. By employing state recognition and prediction, intelligent control and management of satellites can be realized, thereby enhancing satellite systems’ autonomous operation efficiency and dependability while minimizing the risks and losses stemming from system failures [4]. Consequently, state recognition and prediction serve a crucial function for ISCM. Non-probabilistic time-dependent reliability, which assesses attitude states, is established using the interval process and first passage theory, as demonstrated by Yang et al. [5]. However, in previous engineering practices, telemetry data analysis had been solely confined to discerning the normal and fault states of satellites without providing an in-depth analysis of the working states of satellites during operation, such as attitude maneuvers [6]. Moreover, during the development stage, a prevalent issue is the inadequate simulation of the satellite’s actual on-orbit operational state [7,8]. Consequently, there has been an increasing focus on analyzing satellite telemetry data to determine operational modes [9]. The outcomes of telemetry data analysis can not only detect and predict satellite faults but also supply a more precise foundation for the satellite’s development, design, and testing processes [10]. These current research studies emphasize the importance of state recognition and prediction for intelligent control of satellites. However, there are two main issues in the field mentioned above: firstly, past engineering practices overlooked comprehensive analysis of satellites’ operational states during maneuvers; secondly, the issue of adequately utilizing real-time telemetry from on-orbit satellites, which is stored at ground stations and contains data on the satellites’ operational states, is widespread. To address these gaps, there is a growing emphasis on analyzing telemetry data to determine operational modes, which not only helps in fault detection and prediction but also enhances the satellite’s development, design, and testing processes.

With the advancements in deep learning technology, the potential to mine extensive telemetry data using neural networks is realized [11]. However, the current focus of satellite telemetry data recognition is primarily centered around fault diagnosis, anomaly detection, and related aspects. Pang et al. [12] proposed a telemetry data mining clustering framework based on salient shape features and spatiotemporal characteristics to cluster satellite telemetry data, validating the framework’s efficacy and applicability through real satellite telemetry data. Chen et al. [13] developed an unsupervised anomaly detection model employing Bayesian deep learning and domain knowledge, conducting experiments on an imbalanced satellite telemetry dataset, which demonstrated the model’s robustness. Wang et al. [14] proposed an unsupervised anomaly detection model based on variational transformers for monitoring satellites’ operational states. Yang et al. [15] integrated long short-term memory networks with multiscale anomaly monitoring strategies for spacecraft telemetry data anomaly detection tasks, verifying their method’s effectiveness using National Aeronautics and Space Administration (NASA) spacecraft data. Pan et al. [16] put forth a bidirectional long short-term memory neural network to capture abnormal satellite telemetry data. While the approaches mentioned above effectively detect anomalous signals in satellite telemetry data, facilitating prompt and accurate fault diagnosis, a research gap exists in the comprehensive identification of diverse satellite operational states. Moreover, these methods do not account for predicting satellite telemetry data and cannot thus forecast the future working states of satellites, rendering it challenging to offer effective guidance to ground control experts.

The application of traditional convolutional networks is greatly limited by their requirement for structured data and the transformation of one-dimensional time series data into two-dimensional image data [17]. This limitation restricts their ability to handle diverse types of data. In contrast, graph convolutional neural networks utilize unstructured data in graph form, which allows them to manage diverse types of data without the limitations of traditional convolutional networks. This makes them more compatible with one-dimensional telemetry time series data and dramatically expands the range of applications for graph convolutional networks. Xie et al. [18] proposed a satellite power system anomaly detection method based on graph neural networks and dynamic thresholds. Many studies have shown the incredible capability of GCN in the areas of fault diagnosis and anomaly detection [19,20,21,22,23,24,25]. Moreover, different satellite models have various threshold limits, which significantly limit the ability of ground stations to detect the states and manage the health of large-scale satellite clusters. Unlike traditional methods, neural networks can establish a mapping to the data mode by learning the data mode rather than being limited to fixed numerical values. As a result, neural network models can be easily migrated to another type of satellite, demonstrating high transferability [26]. In addition, in today’s era of satellite digitization and intelligence, analyzing historical satellite telemetry data can help create more realistic digital satellite models, driving the intelligent development of satellites and other spacecraft [27]. Additionally, this can also help optimize and improve satellite control algorithms, ultimately leading to high performance control and industrial applications [28,29,30,31,32,33].

In response to the above issues, this study aims to fill the research gap in the field of recognition and prediction of satellite states. We propose a two-step graph convolutional neural network for the recognition and prediction of the working states of the satellite attitude control system. The recognition of the operating states was realized in the first step, and in the second step, the prediction of the next time was realized. The main contributions are listed as follows:

• This study proposes a new satellite wheel motion state, the unloading and subsequent maneuver process, which addresses the issue of insufficient use of on-orbit telemetry data for evaluating the operating state of satellite attitude control systems in past satellite wheel designs. By analyzing the telemetry data of a certain model of an on-orbit satellite’s attitude system, this study fills the gap in using on-orbit telemetry data for analyzing the satellite attitude control system and fully utilizes on-orbit telemetry data for evaluating the operating state of the satellite attitude control system, which has important reference value;
• We propose an advanced recognition and prediction model based on a two-stage graph neural network model for the research gap in the state recognition and prediction of the satellite attitude control system. This paradigm aims to optimize the model’s feature extraction capability by aggregating different data features, thus achieving a significant improvement in the accuracy and robustness of recognition and prediction tasks and providing a baseline for this problem.

This paper is structured as follows: Section 2 gives a detailed description of the recognition and prediction of working states of satellite reaction wheels by analyzing the relevant telemetry from the on-orbit satellite. Section 3 introduces the proposed baseline two-step graph convolutional neural network for recognizing and predicting the working state of satellite reaction wheels, along with its workflow. Section 4 performs state recognition and prediction on a satellite telemetry dataset, compares it with some classical neural networks, and concludes that the two-step recognition-prediction graph network model is superior to some existing methods. Section 5 presents the conclusion of this study.

2. Problem Description

2.1. Satellite Attitude Control System and Its Telemetry

The normal operation of a satellite on orbit depends on each subsystem’s normal operation and coordination, especially the attitude control subsystem. The mission of the satellite’s attitude control subsystem is to control the satellite’s three-axis attitude to meet operational requirements, including high-precision attitude stabilization, tracking, and agile maneuvering. The active satellite attitude control system designed in this study is illustrated in Figure 1 [34].

The satellite telemetry data represent the encoded readings from onboard sensors broadcasted to the ground [35]. After being received and decoded by ground stations, these data provide insights into the actual operational state of the satellite. In contrast, simulation data are based on mathematical models and cannot encompass all actual situations. Therefore, when studying ISCM problems, it is necessary to use real telemetry data from satellites. This study utilized telemetry data from 2018 to 2022 for a specific satellite formation consisting of four small satellites as the experimental dataset. The dataset comprises telemetry parameters from various subsystems, including the orbit system, power system, and attitude control system. Five parameters related to the satellite’s attitude control system were selected for studying the features of the satellite’s operational status, including the angular momentum of the satellite’s yaw axis and pitch axis, the rotational speed of the reaction wheel, the control current of the reaction wheel, and the satellite’s yaw angle. The total length of the dataset was 52,700, and it was divided into a training set and a testing set, with 50% of the data in each set.

2.2. Model-Based Working States of Satellite Reaction Wheels

• Normal State:

Reaction wheels are mechanical devices that provide external torque by changing their angular momentum to resist atmospheric drag torque on low Earth orbit and maintain satellite attitude stability. In the past, when modeling reaction wheels, only three states, including normal state, maneuvering, and unloading, were considered [35]. Following is a detailed explanation of each state.

In the normal state, the reaction wheel runs within a certain range of speed and is subject to certain fluctuations caused by factors such as sensor measurement, as shown in Figure 2a. It is worth noting that, as shown in Figure 2b, even during the normal state of a satellite, its attitude angle slowly and gradually changes, mainly due to the interference of the disturbance moment.

• Maneuvering:

During the satellite attitude control, reaction wheels are required to provide control torque, which results in an instantaneous change in the reaction wheel speed and a corresponding change in the attitude angle and angular momentum of the satellite. This process is referred to as the reaction wheel maneuver. Due to the limitations of the actuators and satellite power, this process is brief compared to the orbital cycle and can be distinctly separated from telemetry data containing noisy signals in the reaction wheel speed, attitude angle, and angular momentum, as shown in Figure 3a.

It is worth noting that, as shown in Figure 3b, during satellite attitude maneuvers, the attitude angles undergo abrupt changes.

• Unloading:

Since the reaction wheel relies on changes in its speed to provide control torque, and there is a maximum speed limit for the reaction wheel design, it is necessary to unload it when it approaches the angular momentum saturation in order to ensure the working capability of the reaction wheel.

The typical unloading method used by satellites is to use thrusters, and the unloading process is reflected in the reaction wheel speed telemetry, as shown in Figure 4a. It is worth noting that, as shown in Figure 4b, unlike during attitude maneuvers, the attitude angle of the satellite reaction wheel does not experience drastic changes during unloading. However, its flywheel rotational speed telemetry will decrease.

2.3. Telemetry-Based Working State of Satellite Reaction Wheels

However, after studying the actual on-orbit satellite attitude control system real-time telemetry data, it was found that in addition to the above three states, there is a new reaction wheel motion state unloading followed by maneuvering. This state of the reaction wheel is distinct from the maneuvering and unloading. The speed, satellite attitude angle, and angular momentum undergo significant changes compared to normal state telemetry with noise over an extended period, contrasting with unloading and maneuvering states. This working state must be distinguished from single attitude maneuvers and reaction wheel unloading, as illustrated in Figure 5a.

It is worth noting that, as shown in the Figure 5b, like the attitude maneuvers process, the attitude angle of the satellite reaction wheel does experience drastic changes during unloading followed by maneuvering. Accurate identification and prediction of this state can effectively enhance the precision of the satellite attitude control system.

2.4. Problem of Recognition and Prediction of Working State of Satellite Reaction Wheels

Considering the aforementioned analysis, the problem of recognizing the working state of satellite reaction wheels can be defined as determining the reaction wheel’s state using available data. Furthermore, given the available data, the problem of predicting the working state of satellite reaction wheels can be framed as accurately forecasting the satellite reaction wheel’s future state for a specified period.

3. Methodology

3.1. Satellite Telemetry Data to Graph Data

This study utilized telemetry data from 2018 to 2022 for a specific satellite formation consisting of four small satellites on low Earth orbit as the experimental dataset. Based on the analysis of the four working states of the reaction wheel in Section 2, the telemetry data is segmented and organized into nodes and node features. The maximum information coefficient is used to analyze the correlation between different telemetry data, and the Maximum Information Coefficient (MIC) correlation is converted into edge structure features. Finally, the telemetry graph structure data are constructed by integrating nodes, node features, and edges.

The MIC method, proposed by Reshef et al. [36] in 2011, is a method for measuring the correlation between two variables. It has significant advantages in dealing with nonlinear relationships, especially in systems like satellite attitude systems that are highly distributed and involve multiple telemetry signals simultaneously. The formula is as follows:

$$mic\left(x,y\right)={\max }_{a*b<B}\frac{I\left(x;y\right)}{{\log }_2\min \left(a,b\right)}$$

(1)

where $a,b$ is the grid distribution, $B$ is a user-defined variable, usually set as 0.6, and $I\left(x;y\right)$ is the mutual information of variable, calculated by the following formula, where $p\left(x,y\right)$ represents the joint probability density.

$$I\left(x;y\right)={\displaystyle \int p\left(x,y\right){\log }_2\frac{p\left(x,y\right)}{p\left(x\right)p\left(y\right)}dxdy}$$

(2)

Figure 6 illustrates the process of transforming temporal telemetry data into graph-structured data.

3.2. Data Resampling

During satellite operation, the normal state constitutes over 90% of the total data, with the remaining three states collectively representing less than 10%, resulting in a substantial data imbalance issue that can impede the proper functioning of neural networks. In deep learning, resampling techniques are commonly used to address the data imbalance issue. Resampling techniques refer to the process of reducing the sampling frequency of data with more samples and increasing the sampling frequency of data with fewer samples during neural network training so that the number of samples in each class is balanced during training.

Resampling primarily involves two essential processes: down-sampling for the majority samples and over-sampling for the minority samples. During neural network training, the down-sampling process selectively eliminates a portion of the dominant samples at random.

When the total number of samples is relatively large, this method can balance the data well while retaining the features of the samples [37]. Over-sampling for minority samples refers to random sampling with the replacement for the minority samples during neural network training to increase their proportion. The distribution of the telemetry data used in this paper before and after resampling is shown in Figure 7.

3.3. Two-Step Graph Neural Network

The proposed method in this paper consists of a feature extraction module, a graph information aggregation module, and a classifier and a predictor. The feature module captures patterns, the graph module integrates contextual relationships, and the final module categorizes inputs and generates predictions. The dual-channel graph convolutional module extracts and fuses features using two channels simultaneously. The formula for graph convolution is as follows [22]:

$$x_j^{\prime }=W_1{x}_i+W_2{\displaystyle \sum e_{i,j}}·x_j$$

(3)

In which $x_j^{\prime }$ represents the features after being processed by the neural network, $W_1,W_2$ represents two learnable matrices, and $e_{i,j}$ represents the edge from node $j$ to node $i$.

Because the input data of graph networks are graph data expressed by nodes and edges, which are a multidimensional vector used to represent the telemetry parameters of a satellite over a period of time, while the working state of the satellite’s reaction wheel is a one-dimensional vector, it is necessary to aggregate the information of nodes on the graph and convert the multidimensional vector into a one-dimensional vector. The commonly used graph information aggregation methods are global graph average pooling, global graph sum pooling, and global graph max pooling. The method used in this paper is global graph max pooling, which preserves the maximum value in the features, as shown in the equation.

$$r_{}=ma{x}_{n=1}^{N_i}{x}_n$$

(4)

where $r$ represents the output feature and $N_i$ represents the length of feature vector $x_i$.

The state recognition module based on the graph convolutional network consists of a graph convolutional layer, which is used to collect useful feature information for state recognition from the features extracted by the feature extraction module. After passing through the graph information aggregation module and the linear classification layer, it outputs the recognition result of the current state of the satellite attitude control system. The state prediction module based on the graph convolutional network adopts the same structure as the state recognition module, and the graph convolutional layer is used to extract useful features for state prediction. Adding a graph convolutional network layer to the state recognition and prediction modules, respectively, is to extract useful features for state recognition and prediction separately, avoiding only extracting a single feature and affecting the overall accuracy.

Table 1. Parameters of TS-GCN.

Network Module	Layers	Input Size	Output Size
	GCN Layer 1-1	$5\times 13$	$5\times 100$
	GCN Layer 1-1	$5\times 13$	$5\times 100$
Dual-channel	GCN Layer 1-2	$5\times 100$	$5\times 60$
Feature Extraction	GCN Layer 2-1	$5\times 13$	$5\times 6$
Module State Recognition Module State Prediction Module	GCN Layer 2-2 GCN Layer 3 Classifier GCN Layer 4 Predictor	$5\times 6$ $5\times 60$ $5\times 12$ $5\times 60$ $5\times 16$	$5\times 60$ $5\times 12$ $1\times 4$ $5\times 16$ $1\times 4$

presents the relevant parameters of the proposed two-step prediction-recognition graph convolutional neural network in this paper.

3.4. Process of State Recognition and Prediction

The satellite attitude control system state recognition process is divided into four stages: telemetry data-graph structure data conversion, recognition model training, prediction model training, and model validation, as shown in Figure 8.

• Telemetry data to graph structure data conversion. Firstly, the telemetry data from the satellite attitude control system are reconstructed into a graph structure through normalization and resampling. Subsequently, the training graph set and the verification graph set for state recognition are generated;
• Recognition model training. The training graph set is input into the dual-channel graph convolutional network state recognition module, and feature extraction and feature fusion are performed in turn through dual-channel graph convolutional layers. The corresponding classification results, loss, and accuracy are output through the graph information aggregation layer and classification activation layer;
• Prediction model training. The training graph set is fed into the graph convolutional network fusion prediction module. Through the graph convolutional layers and a linear predictor, the module generates categorical predictions for the operating state of the satellite attitude control system at the next moment, along with the corresponding loss value with the cross-entropy loss function;
• Model validation. The test set is fed into the previously trained TS-GCN model. The model performs data recognition, classifies the data, and predicts the operating state of the satellite attitude control system for the next moment.

4. Experiments and Results

The proposed model’s performance is evaluated using telemetry data from several low orbit small satellite formations between 2019 and 2021, in terms of the accuracy of state recognition, state prediction, and the number of neural network parameters. The proposed two-step prediction-recognition graph convolutional neural network model is compared with classical neural network models. The model code is written in Python 3.9 and the deep learning framework PyTorch 1.13, using Ubuntu 22.04 LTS as the server operating system, Intel i9-13900k as the central processor, and NVIDIA RTX4090 as the graphics processor with 64 GB of memory. The hyperparameters of the neural network are listed in

Table 2. Neural network training hyperparameters.

Hyperparameters	Value
Learning rate	0.0001
Training epochs	100
Batchsize	128
Dropout probability	0.4

.

4.1. Result of State Recognition

To validate the effectiveness of the state recognition module of TS-GCN proposed in this paper, the proposed model was trained and compared with Standard CNN, Standard GCN, and ResNet-18 models using the same dataset. The experiment was repeated five times to reduce errors, and the final evaluation indicators of the models were obtained as shown in

Table 3. Recognition accuracy and number of parameters for different methods.

Methods	Average Accuracy	Maximum Accuracy	Minimum Accuracy	Parameters	Time
Standard CNN	23.28%	26.72%	21.46%	$767\times {10}^3$	18 s
Standard GCN	60.44%	65.01%	58.28%	$590\times {10}^3$	20 s
ResNet-18	61.56%	71.63%	60.69%	$11.2\times {10}^6$	24 s
Manually (one person)	100%	100%	100%	\	22 h
TS-GCN	98.93%	99.18%	98.58%	$72.9\times {10}^3$	16 s

, along with the accuracy curves of the four models with the number of iterations shown in Figure 9.

In this paper, we propose a TS-GCN model for satellite attitude control system state recognition, achieving an accuracy of 98.83%. As depicted in Figure 9, the accuracy of TS-GCN rapidly increases during the first 20 iterations and surpasses the other three neural network models (Standard CNN, Standard GCN, and ResNet-18) in terms of accuracy and stability. TS-GCN exhibits a substantial improvement of 38.49% in accuracy compared to Standard GCN, highlighting the superiority of the dual-channel graph convolution method in data mining. Furthermore, TS-GCN, which employs the graph convolution formula, demonstrates an accuracy improvement of 35–40% in satellite attitude control system state recognition in comparison to ResNet-18 and Standard CNN, which use traditional convolution formulas. This indicates that the graph convolution mode, based on graph-structured data, is highly adaptable to electro-mechanical systems with multiple telemetry parameters, such as satellites. Additionally, the parameter quantity of the proposed two-step graph convolutional neural network model is significantly smaller than that of the other three neural network models, showcasing the high performance of TS-GCN.

In order to evaluate the recognition accuracy of the trained TS-GCN model, the test set samples were tested and the confusion matrix of the recognition results was obtained, as shown in Figure 10. The diagonal elements of the confusion matrix represent the percentage of correctly predicted samples by the TS-GCN model. It can be seen from Figure 10 that the proposed neural network has a very high recognition accuracy for the working state of satellite reaction wheels by using telemetry data.

4.2. Result of State Prediction

After achieving high-precision recognition of the satellite attitude control system, the accuracy of TS-GCN in predicting the system’s state at the next moment was evaluated. The result of the confusion matrix is presented in Figure 11. As can be observed from Figure 11, the prediction module of TS-GCN exhibits extremely high accuracy in predicting the state of the satellite attitude control system, reaching 99.13%. This demonstrates that the proposed two-step prediction-recognition graph convolutional neural network model can not only accurately recognize the state of the satellite attitude control system but also predict its next state with high precision.

5. Conclusions

• This study introduces a novel operational state for the reaction wheel compared to the conventional state, incorporating an analysis of its performance using historical telemetry data. Building upon this framework, the problems of recognizing and predicting the operational modes of the satellite attitude control system are proposed;
• To address the problems of recognizing and predicting the operational states of the satellite attitude control system, this study proposes a two-step graph convolutional neural network (GCN) model. The model is designed based on the GCN framework and aims to achieve intelligent satellite management and control (ISMC) by providing a deep learning-based solution. In the first step, the proposed graph convolutional state recognition module utilizes satellite telemetry data for the task of recognizing the system’s operational states. Compared to traditional neural network models such as Standard CCN and ResNet-18, this module achieves an accuracy of 98.93% and has a smaller number of network parameters. In the second step, the model achieves an accuracy of 99.13% for the state prediction task. This research fills a research gap in the field of recognizing and predicting the operational states of satellite attitude control systems based on satellite telemetry data. Compared to traditional single-channel networks, multi-channel networks can reduce feature loss and have better performance. The model performs data recognition, classifies the data, and predicts the operating state of the satellite attitude control system for the next moment. This process enables high-precision state recognition, classification, and prediction of satellite attitude control system telemetry data.

This study primarily focuses on predicting and identifying the state of the attitude control system for low Earth orbit satellites. It aims to enhance the efficient use of in-orbit telemetry data, ultimately leading to the development of a spacecraft digital twin model. The overarching objective is to support aerospace industry development and advance high-performance control systems. With the methods proposed in this paper, there is potential for future research to expand into the recognition and prediction of the states of other satellite subsystems. Furthermore, these methods can also be extended to include medium and geostationary orbit satellites. Additionally, it is worth noting that further research into multi-step prediction is of significant academic value.