Satellite orbit maneuver detection method based on multichannel residual imaging

Abstract

The invention relates to the technical field of spaceflight measurement and control and space situation awareness, and discloses a satellite orbit maneuver detection method based on multichannel residual imaging, which comprises the following steps: acquiring an orbit element actual observation value, acquiring an orbit element theoretical value, calculating an orbit element residual sequence based on the orbit element actual observation value and the orbit element theoretical value, and calculating the orbit element maneuver according to the orbit element residual sequence. The method comprises the following steps of: extracting multiple feature channels, respectively carrying out Gramer angle field transformation on a residual sequence of each feature channel to obtain a multi-channel GAF image, constructing a classification model based on a deep neural network, and carrying out feature extraction and orbit maneuver judgment on the multi-channel GAF image. According to the method, the Gramb angle field transformation is applied to satellite orbit data processing for the first time, unequally spaced TLE time sequences are converted into regular two-dimensional images, the mature computer vision technology can be directly applied to maneuvering detection tasks, the time sequence correlation of the time sequences is reserved through GAF transformation, the time sequences are coded into space structure features of the images, and the space structure features of the images are obtained. And deep learning model extraction and learning are facilitated.

Description

Technical Field

[0001] This invention relates to the field of aerospace telemetry, tracking and command and space situational awareness technology, and in particular to a satellite orbital maneuver detection method based on multi-channel residual imaging. Background Technology

[0002] With the rapid development of human space activities, the number of artificial satellites in orbit has grown exponentially. As of 2025, there were more than 10,000 active satellites in Earth orbit, covering multiple application fields such as communication, navigation, remote sensing, and scientific experiments. Against this backdrop, space situational awareness has become a focus of attention for major spacefaring nations, and the timely and accurate detection of satellite maneuvering behavior is one of the core capabilities of space situational awareness.

[0003] Satellite maneuvers refer to the actions of satellites that actively change their orbital state by activating their propulsion systems. These maneuvers include various types such as orbit maintenance, orbit transfer, collision avoidance, and formation maneuvers. Accurate detection of satellite maneuvers is crucial for several key missions: in collision warning, timely detection of satellite orbital changes helps update collision risk assessments; in space safety, identifying abnormal maneuvers is an important basis for determining satellite intentions; and in orbit prediction, maneuvers can render historical orbit determination data invalid, necessitating timely updates to prediction models.

[0004] Currently, publicly available satellite orbit data primarily originates from published Two-Line Elements (TLEs). TLEs describe satellite orbits using a standardized format, including six classic Keplerian elements: semi-major axis, eccentricity, inclination, right ascension of the ascending node, argument of perigee, and mean perigee, along with corresponding auxiliary parameters such as epoch time and atmospheric drag coefficient. The update frequency of TLE data varies by satellite; low-Earth orbit (LEO) satellites are typically updated several times daily, while high-Earth orbit (GEO) satellites are updated relatively less frequently. Despite the limited accuracy of TLE data, its publicly available nature makes it an important data source for conducting satellite maneuvering detection research.

[0005] Existing satellite motion detection methods are mainly divided into three categories:

[0006] The first category is the traditional threshold-based method, which determines whether a maneuver has occurred by setting a threshold value for changes in orbital elements. For example, a maneuver is determined when the difference between the semi-major axes of two adjacent epochs exceeds a certain threshold. This type of method is simple in principle and easy to implement, but it suffers from problems such as difficulty in uniformly determining the threshold value and poor adaptability to different orbital altitudes and maneuver types. In addition, due to the orbit determination error inherent in TLE data itself, coupled with the influence of environmental perturbations such as atmospheric drag, solar radiation pressure, and the non-spherical gravity of the Earth, the fixed threshold method is prone to missed detections or false alarms.

[0007] The second category is model-driven methods based on orbital dynamics. These methods identify maneuvers by establishing accurate orbital prediction models and comparing the deviations between actual observations and model predictions. Commonly used prediction models include the SGP4 / SDP4 simplified perturbation model and numerical integration methods. These methods have clear physical meanings but require high model accuracy and struggle to handle the coupling between model errors and maneuver signals, especially when atmospheric drag modeling for low-Earth orbit satellites is inaccurate.

[0008] The third category is data-driven machine learning methods, which automatically learn maneuver features from historical data using statistical learning or deep learning techniques. In recent years, with the development of deep learning technology, models such as convolutional neural networks, recurrent neural networks, and Transformers have been attempted to be applied to the field of orbital anomaly detection. For example, Long Xi et al. proposed an adaptive deep learning method for GEO satellite maneuver identification in the Journal of Astronautics 2025(1), titled "Adaptive Deep Learning Method for GEO Satellite Maneuver Identification." This method models the orbital elements such as historical longitude, semi-major axis, and eccentricity, and outputs the longitude at the next moment, thus achieving automatic identification of satellite maneuver events. This method has achieved good detection results on specific datasets, but it still has the following shortcomings:

[0009] (1) Limitations of single feature representation. This method mainly uses a single or a few orbital elements as input features, failing to fully explore the correlation and complementarity between different orbital parameters. In fact, different types of maneuvers have different impact patterns on each orbital element, and relying solely on a single feature is insufficient to comprehensively characterize the maneuver features, resulting in insufficient detection capability for certain maneuver types. (2) Limited temporal modeling capability. This method uses a traditional LSTM network to process TLE sequences. Although it can capture certain temporal dependencies, it suffers from problems such as gradient vanishing and difficulty in modeling long-range dependencies when processing orbital data with long time spans. In addition, TLE data is sampled at unequal intervals, and directly inputting it into the sequence model will introduce the problem of inconsistent time scales. (3) Difficulty in removing environmental perturbation interference. This method does not incorporate environmental information such as atmospheric density, making it difficult to distinguish between maneuver signals and orbital changes caused by environmental factors such as atmospheric drag and solar activity, which can easily generate false alarms in low-Earth orbit satellite detection.

[0010] Therefore, existing technologies generally suffer from problems such as insufficient feature representation, inadequate fusion of multi-source information, and high dependence on annotation. A new method is needed that can effectively integrate multi-dimensional orbital information, reduce annotation costs, and improve detection accuracy. Summary of the Invention

[0011] To address the aforementioned problems, the present invention aims to provide a satellite orbital maneuver detection method based on multi-channel residual imaging, which effectively transforms unequal-interval TLE time-series data into image representations rich in physical information, thereby achieving accurate detection of satellite maneuvering behavior and reducing reliance on precise annotation.

[0012] This invention provides a satellite orbital maneuver detection method based on multi-channel residual imaging, comprising the following steps:

[0013] S1. Obtain the TLE orbital element sequence of the target satellite within the detection time window, perform time series resampling on the TLE orbital element sequence, obtain the actual observed values ​​of the orbital elements, and use the SGP4 simplified perturbation model to predict the theoretical values ​​of the orbital elements.

[0014] S2. Extract multiple feature channels with clear physical meaning. Based on the actual observed values ​​and theoretical values ​​of orbital elements of each feature channel, calculate the residual sequence. Each feature channel complements the other and jointly characterizes the changing characteristics of the satellite orbital state.

[0015] S3. Perform Gram angle field transformation on the residual sequences of each feature channel to obtain the GAF image of each feature channel;

[0016] S4. Stack the GAF images of each feature channel in multiple channels to obtain a multi-channel GAF image;

[0017] S5. Construct a classification model based on deep neural networks to extract features and determine orbital maneuvers in multi-channel GAF images.

[0018] In one embodiment of the present invention, the time series resampling of the TLE orbital element sequence in S1 includes:

[0019] The TLE orbital element sequence is resampled at equal intervals using cubic spline interpolation to convert the original irregular orbital element sequence into a standard time orbital element sequence with equal intervals.

[0020] In one embodiment of the present invention, the orbital element residual sequence calculated in step S2 is used to eliminate the normal evolution trend of the orbit and highlight the orbital change signal caused by abnormal events, as expressed as:

[0021]

[0022] in, For resampling Actual observed values ​​of orbital elements at time points. The perturbation model obtained based on SGP4 is simplified. Theoretical value of orbital elements at time.

[0023] In one embodiment of the present invention, the multiple feature channels with clear physical meaning in S2 include: a semi-major axis feature channel, an orbital inclination feature channel, a phase feature channel, and an atmospheric density feature channel.

[0024] In one embodiment of the present invention, step S3 involves performing Gram angle field transform on the residual sequences of each feature channel to obtain the GAF image of each feature channel, including the following steps:

[0025] S31. One-dimensional residual sequence of orbital elements for each characteristic channel. Perform MinMaxScaler normalization on each sample, mapping them to the [0, 1] interval, and obtain the normalized residual sequence. ;

[0026] S32. Apply the inverse cosine transform to the normalized residual sequence of each feature channel, as follows:

[0027]

[0028] in, for Polar coordinate angle values ​​of each channel at any given time;

[0029] S33. Calculate the sine value of the angle difference at different time points for each channel to obtain the Gram angle field matrix. , represented as:

[0030]

[0031] in, , for A two-dimensional matrix;

[0032] S34, Gram angular field matrix Visualized as a Gram angle field image.

[0033] In one embodiment of the present invention, when the satellite orbit is maneuvering, the GAF images of each feature channel exhibit characteristic visual patterns, including:

[0034] The semi-major axis characteristic channel exhibits a cross-shaped fault structure at the moment of maneuver. The position of the fault line corresponds to the time point when the maneuver occurs, and the gray difference on both sides of the fault line reflects the magnitude of the semi-major axis change caused by the maneuver.

[0035] The track inclination feature channel is sensitive to outward maneuvers and serves as an auxiliary criterion for maneuver type identification;

[0036] The phase characteristic channel exhibits a fan-shaped diffusion pattern after maneuvering, and the expansion of the fan-shaped region reflects the continuous accumulation of phase deviation;

[0037] The atmospheric density characteristic channel exhibits a gradually varying texture related to the space environment, which is not directly affected by maneuvers and is used to distinguish whether orbital changes originate from active maneuvers or environmental perturbations.

[0038] In one embodiment of the present invention, step S4 involves stacking the GAF images of each feature channel to obtain a multi-channel GAF image. , represented as:

[0039]

[0040] in, This is a GAF image of the semi-major axis residual channel. GAF image of orbital inclination residual channel. Phase residual channel GAF image, This is an image of the atmospheric density feature channel.

[0041] In one embodiment of the present invention, the classification model of the deep neural network in S5 includes: CNN and Transformer network models, and the classification layer outputs probability values ​​for two categories: orbital maneuvering and orbital non-maneuvering.

[0042] In one embodiment of the present invention, the classification model of the deep neural network in S5 is trained using a weakly supervised learning paradigm.

[0043] In one embodiment of the present invention, the classification model of the deep neural network adopts gradient weighted class activation mapping technology to generate an attention heatmap of the input image for the classification model, and determines the key image region based on the highlighted area indication of the heatmap.

[0044] The beneficial effects of this invention are:

[0045] 1. This invention is the first to apply Gram angular field transform (GAF) to satellite orbit data processing, converting unequal-interval Time-Like (TLE) time series into regular two-dimensional images, enabling mature computer vision technology to be directly applied to maneuver detection tasks. The GAF transform preserves the temporal correlation of the time series and encodes it as spatial structural features of the image, facilitating extraction and learning by deep learning models, thus innovating the technical path between orbit data and mature visual models.

[0046] 2. The four-channel representation constructed in this invention covers physical information in four dimensions: orbital energy, orbital plane attitude, phase evolution, and space environment. Each channel has a differentiated response to different types of maneuvers, complementing and verifying each other, significantly improving detection accuracy and anti-interference capability, while also supporting preliminary differentiation of maneuver types and physical-driven multi-channel information fusion.

[0047] 3. By calculating the residual between actual observations and orbit predictions, this invention effectively eliminates the background trend of normal orbit evolution, making the maneuvering signal stand out from the noise, improving the signal-to-noise ratio and detectability of maneuvering characteristics, and the residual processing enhances the significance of the signal.

[0048] 4. This invention only requires coarse labels at the image level to complete model training, without the need for precise annotation of movement moments, which greatly reduces the manpower cost of data preparation, facilitates model training and optimization using a large amount of historical data, and the weak supervision paradigm reduces annotation costs.

[0049] 5. This invention uses Grad-CAM attention heatmap visualization technology to intuitively present the model's judgment criteria, facilitating manual review and result verification, thereby enhancing the credibility and interpretability of the method in practical applications. Interpretability enhances the credibility of decision-making.

[0050] 6. The multi-channel representation designed based on physical principles in this invention has a natural generalization ability and can adapt to satellite targets with different orbital altitudes and different mission types. Through transfer learning, it can quickly adapt to new detection scenarios and has good generalization and adaptability. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the principle of the method of the present invention;

[0052] Figure 2 This is a schematic diagram illustrating the GAF image transformation principle of the present invention;

[0053] Figure 3 This is a comparison diagram showing the effect of using the original TLE data and dynamic residual data at the moment of maneuver in this invention;

[0054] Figure 4 This is an example image of the four-channel residual GAF feature image of the present invention;

[0055] Figure 5 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation

[0056] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0057] Existing satellite orbital maneuver detection technologies suffer from problems such as insufficient feature representation, inadequate fusion of multi-source information, high dependence on annotation, and susceptibility to environmental perturbations.

[0058] To address the aforementioned problems, this invention provides a satellite orbital maneuver detection method based on multi-channel residual imaging. To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description is provided in conjunction with specific embodiments.

[0059] Example 1:

[0060] This embodiment discloses a satellite orbital maneuver detection method based on multi-channel residual imaging, such as... Figure 1 As shown, the steps include:

[0061] TLE data acquisition and preprocessing: The TLE orbital element sequence of the target satellite within the detection time window is acquired, the TLE orbital element sequence is resampled over time to obtain the actual observed values ​​of the orbital elements, and the theoretical values ​​of the orbital elements are predicted using the SGP4 simplified perturbation model.

[0062] The TLE orbital element sequence of the target satellite within the detection time window was obtained from publicly available data sources such as Space-Track. Each TLE record contains six Kepler orbital elements at the corresponding time: including the semi-major axis. eccentricity Track inclination Right ascension of ascending node Perigeal argument Peace Angle .

[0063] Since TLE data is published at unequal intervals, the time interval between each record is not fixed. To facilitate subsequent processing, cubic spline interpolation is used to resample the orbital element sequence at equal intervals. A target sampling interval is set to convert the original irregular orbital element sequence into a standard time orbital element sequence with equal intervals, forming the actual observed values ​​of the orbital elements at each time point.

[0064] A simplified perturbation model using SGP4 is employed, with the TLEE orbital element data at the beginning of the detection window as the initial condition, to predict the theoretical values ​​of the orbital elements at subsequent times.

[0065] Multi-channel physical feature extraction: Extract multiple feature channels with clear physical meaning. Based on the actual observed values ​​and theoretical values ​​of orbital elements of each feature channel, calculate the residual sequence. Each feature channel complements the others and jointly characterizes the changing characteristics of the satellite orbital state.

[0066] The orbital element residual sequence is calculated to eliminate the normal evolution trend of the orbit and highlight the orbital change signal caused by anomalous events, and is represented as:

[0067]

[0068] in, For resampling Actual observed values ​​of orbital elements at time points. The perturbation model obtained based on SGP4 is simplified. Theoretical value of orbital elements at time.

[0069] This embodiment extracts four feature channels with clear physical meaning from the preprocessed data, including the semi-major axis feature channel, the orbital inclination feature channel, the phase feature channel, and the atmospheric density feature channel.

[0070] The semi-major axis characteristic channel (Sma) extracts the residual sequence of the semi-major axis to reflect anomalous changes in orbital energy. When a satellite performs orbital maneuvers, the thrusters work directly alter the orbital energy, causing a step-like abrupt change in the semi-major axis, which manifests as a significant jump signal in the residual sequence.

[0071] The orbital inclination feature channel (Inc) describes the angle between the orbital plane and the equatorial plane, reflecting the spatial orientation of the orbital plane. This channel extracts the inclination residual sequence to identify maneuvering behaviors involving orbital plane adjustments. Out-of-plane maneuvers cause significant changes in inclination, while in-plane maneuvers have little effect; this channel helps distinguish between different types of maneuvers.

[0072] The phase feature channel (Phase) defines the comprehensive phase feature parameter as the sum of the perigee argument and the mean perigee angle. The phase feature channel extracts the phase residual sequence to capture the cumulative effect after a maneuver. After a maneuver changes the orbital period, the deviation between the satellite and its originally expected position accumulates over time, forming a characteristic trend signal.

[0073] The Atmospheric Density Feature Channel incorporates atmospheric density information to distinguish orbital changes caused by active maneuvers and environmental perturbations such as atmospheric drag. Based on the NRLMSIE-00 atmospheric model, and combined with the satellite's orbital altitude and the prevailing space environment parameters (solar activity index, geomagnetic activity index, etc.), the atmospheric density at the satellite's location is calculated. The atmospheric density sequence is then logarithmically transformed and standardized before being used as the input to the Atmospheric Density Feature Channel, providing an environmental reference for determining the causes of orbital changes.

[0074] Gram angle field transform: Gram angle field transform is performed on the residual sequences of each feature channel to obtain the GAF image of each feature channel.

[0075] The Gramian Angular Field (GAF) transform is an encoding method that converts a one-dimensional time series into a two-dimensional image, preserving the temporal dependencies of the time series and transforming them into the spatial structural features of the image.

[0076] Gram angle field transform is performed on the residual sequences of each feature channel to obtain the GAF image of each feature channel, such as... Figure 2 As shown, the steps include:

[0077] S31. One-dimensional residual sequence of orbital elements for each characteristic channel (semi-major axis, orbital inclination, phase, and atmospheric density). Perform MinMaxScaler normalization on each sample, mapping them to the [0, 1] interval, and obtain the normalized residual sequence. This step eliminates dimensional differences and ensures the numerical stability of subsequent transformations.

[0078] S32. Apply inverse cosine transform to the normalized residual sequence of each feature channel to convert the one-dimensional data into polar coordinate angles, as follows:

[0079]

[0080] in, for The polar coordinate angle values ​​of each channel at each time point are mapped to angles to form a polar coordinate representation of the time series. This transformation preserves the relative order and dependencies of the original values.

[0081] S33. Calculate the sine value of the angle difference at different time points for each channel to obtain the Gram angle field matrix. , represented as:

[0082]

[0083] in, , for A two-dimensional matrix whose elements Encoded time points and The temporal correlation between them transforms the temporal dependency into spatial structural features.

[0084] S34, Gram angular field matrix Visualized as a Gram angle field image.

[0085] For example, the Gram matrix can be visualized as a grayscale image, where pixel values ​​correspond to... The size and spatial pattern of the image directly reflect the temporal characteristics. Gram angle field images are used to represent normal orbits and maneuvering events.

[0086] Normal track: Residual fluctuations are stable, and the image exhibits uniform texture.

[0087] Motor events: such as Figure 3 or Figure 4As shown, in the orbital inclination feature channel GAF image, the maneuvering time Clear vertical and horizontal dividing lines appear, forming a cross-shaped pattern.

[0088] like Figure 4 As shown, the phase feature channel GAF image presents a gradually changing fan-shaped region with the maneuvering point as the vertex.

[0089] like Figure 4 As shown, in the atmospheric density feature channel GAF image, if there is no geomagnetic storm, it presents a regular grid texture; if there is a geomagnetic storm, it presents a red striped texture.

[0090] GAF images encode temporal dependencies into image spatial structure, making it easier for models such as CNN to extract deep features; each channel image corresponds to different physical quantities (such as energy, tilt angle, and phase), which together improve detection robustness; the generated images only require window-level labels for training, reducing annotation costs.

[0091] Through the above steps, this embodiment realizes the conversion from irregular TLE orbital element sequences to structured images, providing a highly interpretable visual feature basis for satellite orbital maneuver detection.

[0092] Multi-channel GAF image construction: Gram angle field transform is performed on the residual sequences of each feature channel to obtain the GAF image of each feature channel.

[0093] For the semi-major axis feature channel, orbital inclination feature channel, phase feature channel, and atmospheric density feature channel, a Gram angle field transform is performed on each channel to generate four single-channel GAF images. The size of each image is equal to the length of the time series, i.e., if the series contains... If there are 1 sampling point, then generate A square array image.

[0094] Four single-channel GAF images are stacked along the channel dimension to form a four-channel multi-channel image, represented as:

[0095]

[0096] in, This is a GAF image of the semi-major axis residual channel. GAF image of orbital inclination residual channel. Phase residual channel GAF image, This is an image of the atmospheric density feature channel. This is a stacking operation.

[0097] This stacking method is similar to the RGB three-channel structure of a color image, but in this embodiment, the four channels carry different physical information, forming a multimodal representation with clear physical meaning.

[0098] Then, bilinear interpolation is used to uniformly scale the multi-channel images to a standard size (e.g., (pixels) to accommodate the input requirements of subsequent deep learning models. The scaling process maintains the relative feature structure of each channel unchanged.

[0099] Maneuver Feature Pattern Analysis: Based on the constructed multi-channel GAF images, the characteristic visual patterns of normal trajectories and maneuver events are analyzed in different channels.

[0100] Multi-channel GAF images can intuitively present the characteristic patterns of different orbital states, providing a basis for judgment in motion detection.

[0101] Image characteristics of a normal track: When no maneuvering occurs within the detection window, the track residual mainly consists of measurement noise and model error, exhibiting stable random fluctuations. The corresponding GAF image shows a relatively uniform background texture with no obvious structural features.

[0102] Image features of the maneuver track: When there is a maneuver within the detection window, the GAF images of each channel exhibit a characteristic visual pattern.

[0103] The semi-major axis channel exhibits a distinct cross-shaped fault structure during maneuvering. The position of the fault line corresponds to the time point when the maneuver occurs, and the difference in grayscale on both sides of the fault line reflects the magnitude of the semi-major axis change caused by the maneuver.

[0104] The tilt channel is sensitive to out-of-plane maneuvers and exhibits a fault-like feature similar to the semi-major axis channel during out-of-plane maneuvers. During in-plane maneuvers, the channel changes less and can be used as an auxiliary criterion for maneuver type identification.

[0105] The phase channel exhibits a "fan-shaped diffusion" pattern after the maneuver. The starting point of the fan corresponds to the moment of the maneuver, and the expansion of the fan-shaped region reflects the continuous accumulation of phase deviation.

[0106] Atmospheric density channels exhibit a gradually varying texture related to the space environment and are not directly affected by maneuvers. They are used to distinguish whether orbital changes originate from active maneuvers or environmental perturbations.

[0107] Deep learning model detection: Construct a classification model based on deep neural networks to extract features and determine orbital maneuvers in multi-channel GAF images.

[0108] A classification model based on deep neural networks is constructed to perform feature extraction and dynamic discrimination on multi-channel GAF images.

[0109] The model employs a hierarchical structure, including a feature embedding layer, multi-level feature extraction modules, and a classification output layer. The feature embedding layer uses an overlapped convolutional design. Convolutional kernels prevent key features from being cut off and lost at the boundaries.

[0110] Deep neural network classification models can use CNNs, Transformers, or state-space models for recognition. The classification layer outputs probability values ​​for two classes: maneuvering and non-maneuvering. The loss function uses Focal Loss or cross-entropy loss, and the optimizer is the AdamW optimizer, combined with a cosine annealing learning rate scheduling strategy.

[0111] The model is trained using a weakly supervised learning paradigm. During training, only image-level labels are needed, indicating whether each detection window contains a maneuver event; the precise time of the maneuver is not required. For example, cross-entropy is used as the loss function, expressed as...

[0112]

[0113] in, For real labels, This annotation method significantly reduces the workload and difficulty of manual annotation for predicting probabilities for the model, making it possible to train using large-scale coarse-labeled data.

[0114] For coarse localization of maneuver moments, Gradient Weighted Class Activation Mapping (Grad-CAM) is used to generate an attention heatmap of the input image. Highlighted areas in the heatmap indicate key image regions upon which the model bases its decisions. Since the row and column indices of the GAF image correspond one-to-one with the time series moments, analyzing the locations of high-response regions in the heatmap allows for the deduction of the approximate time window of the maneuver, thus achieving a rough localization of the maneuver moment.

[0115] The present invention also provides an electronic device, Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 5 As shown, the electronic device may include a processor, a communications interface, memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory, for example, to execute the following method:

[0116] S1. Obtain the TLE orbital element sequence of the target satellite within the detection time window, perform time series resampling on the TLE orbital element sequence, obtain the actual observed values ​​of the orbital elements, and use the SGP4 simplified perturbation model to predict the theoretical values ​​of the orbital elements.

[0117] S2. Extract multiple feature channels with clear physical meaning. Based on the actual observed values ​​and theoretical values ​​of orbital elements of each feature channel, calculate the residual sequence. Each feature channel complements the other and jointly characterizes the changing characteristics of the satellite orbital state.

[0118] S3. Perform Gram angle field transformation on the residual sequences of each feature channel to obtain the GAF image of each feature channel;

[0119] S4. Stack the GAF images of each feature channel in multiple channels to obtain a multi-channel GAF image;

[0120] S5. Construct a classification model based on deep neural networks to extract features and determine orbital maneuvers in multi-channel GAF images.

[0121] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0122] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments, including, for example:

[0123] S1. Obtain the TLE orbital element sequence of the target satellite within the detection time window, perform time series resampling on the TLE orbital element sequence, obtain the actual observed values ​​of the orbital elements, and use the SGP4 simplified perturbation model to predict the theoretical values ​​of the orbital elements.

[0124] S2. Extract multiple feature channels with clear physical meaning. Based on the actual observed values ​​and theoretical values ​​of orbital elements of each feature channel, calculate the residual sequence. Each feature channel complements the other and jointly characterizes the changing characteristics of the satellite orbital state.

[0125] S3. Perform Gram angle field transformation on the residual sequences of each feature channel to obtain the GAF image of each feature channel;

[0126] S4. Stack the GAF images of each feature channel in multiple channels to obtain a multi-channel GAF image;

[0127] S5. Construct a classification model based on deep neural networks to extract features and determine orbital maneuvers in multi-channel GAF images.

[0128] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0129] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A satellite orbital maneuver detection method based on multi-channel residual imaging, characterized in that, include: S1. Obtain the TLE orbital element sequence of the target satellite within the detection time window, perform time series resampling on the TLE orbital element sequence, obtain the actual observed values ​​of the orbital elements, and use the SGP4 simplified perturbation model to predict the theoretical values ​​of the orbital elements. S2. Extract multiple feature channels with clear physical meaning. Based on the actual observed values ​​and theoretical values ​​of orbital elements of each feature channel, calculate the residual sequence. Each feature channel complements the other and jointly characterizes the changing characteristics of the satellite orbital state. S3. Perform Gram angle field transformation on the residual sequences of each feature channel to obtain the GAF image of each feature channel; S4. Stack the GAF images of each feature channel in multiple channels to obtain a multi-channel GAF image; S5. Construct a classification model based on deep neural networks to extract features and determine orbital maneuvers in multi-channel GAF images.

2. The method according to claim 1, characterized in that, The time series resampling of the TLE orbital element sequence in S1 includes: The TLE orbital element sequence is resampled at equal intervals using cubic spline interpolation to convert the original irregular orbital element sequence into a standard time orbital element sequence with equal intervals.

3. The method according to claim 1, characterized in that, The orbital element residual sequence calculated in S2 is used to eliminate the normal evolution trend of the orbit and highlight the orbital change signal caused by abnormal events, and is expressed as follows: in, For resampling Actual observed values ​​of orbital elements at time points. The perturbation model obtained based on SGP4 is simplified. Theoretical value of orbital elements at time.

4. The method according to claim 1, characterized in that, The S2 contains multiple feature channels with clear physical meaning, including: semi-major axis feature channel, orbital inclination feature channel, phase feature channel, and atmospheric density feature channel.

5. The method according to claim 3, characterized in that, In step S3, Gram angle field transform is performed on the residual sequences of each feature channel to obtain the GAF image of each feature channel, including the following steps: S31. One-dimensional residual sequence of orbital elements for each characteristic channel. Perform MinMaxScaler normalization on each sample, mapping them to the [0, 1] interval, and obtain the normalized residual sequence. ; S32. Apply the inverse cosine transform to the normalized residual sequence of each feature channel, as follows: in, for Polar coordinate angle values ​​of each channel at any given time; S33. Calculate the sine value of the angle difference at different time points for each channel to obtain the Gram angle field matrix. , represented as: in, , for A two-dimensional matrix; S34, Gram angular field matrix Visualized as a Gram angle field image.

6. The method according to claim 5, characterized in that, When satellite orbits are maneuvering, the GAF images from each feature channel exhibit characteristic visual patterns, including: The semi-major axis characteristic channel exhibits a cross-shaped fault structure at the moment of maneuver. The position of the fault line corresponds to the time point when the maneuver occurs, and the gray difference on both sides of the fault line reflects the magnitude of the semi-major axis change caused by the maneuver. The track inclination feature channel is sensitive to outward maneuvers and serves as an auxiliary criterion for maneuver type identification; The phase characteristic channel exhibits a fan-shaped diffusion pattern after maneuvering, and the expansion of the fan-shaped region reflects the continuous accumulation of phase deviation; The atmospheric density characteristic channel exhibits a gradually varying texture related to the space environment, which is not directly affected by maneuvers and is used to distinguish whether orbital changes originate from active maneuvers or environmental perturbations.

7. The method according to claim 5, characterized in that, In step S4, the GAF images of each feature channel are stacked in multiple channels to obtain a multi-channel GAF image. , represented as: in, This is a GAF image of the semi-major axis residual channel. GAF image of orbital inclination residual channel. Phase residual channel GAF image, This is an image of the atmospheric density feature channel.

8. The method according to claim 1, characterized in that, The classification model of the deep neural network in S5 includes CNN and Transformer network models, and the classification layer outputs the probability values ​​of two classes: orbital maneuvering and orbital non-maneuvering.

9. The method according to claim 8, characterized in that, The classification model of the deep neural network in S5 is trained using a weakly supervised learning paradigm.

10. The method according to claim 8, characterized in that, The classification model of the deep neural network uses gradient-weighted class activation mapping technology to generate an attention heatmap of the input image. Based on the highlighted areas of the heatmap, the key image regions are determined.

Images