Satellite real-time orbit determination method based on knowledge distillation and attention auxiliary kalman wave

Abstract

The application relates to the technical field of spacecraft navigation and control, and specifically discloses a satellite real-time orbit determination method based on knowledge distillation and attention auxiliary Kalman waves, which comprises the following steps: acquiring multi-source sensor data of a satellite and performing pretreatment; constructing a parallel compensation architecture and constructing a benchmark orbit with a simplified dynamic model as the backbone; training a lightweight student model by using a teacher model through a knowledge distillation technology, migrating the knowledge of the teacher model to the student model to obtain a trained lightweight student model; constructing a fractional-order state space model based on the long memory characteristics of environmental noise; inputting a residual prediction value output in real time by the lightweight student model into the fractional-order state space model as a compensation item, performing fractional-order Kalman filtering, and fusing real-time observation data to perform optimal estimation on the satellite state; and monitoring the innovation sequence of the filtering process in real time, and when an anomaly is detected, adaptively adjusting filtering parameters to realize satellite real-time orbit determination.

Description

Technical Field

[0001] This invention relates to the field of spacecraft navigation and control technology, specifically to a real-time satellite orbit determination method based on knowledge distillation and attention-assisted Kalman waves. Background Technology

[0002] With the advent of the commercial space age, low Earth orbit (LEO) satellite constellations (such as Starlink, OneWeb, and China's "State Grid" constellation) have experienced explosive growth. For these satellites, real-time precise orbit determination (POD) is a fundamental prerequisite for maintaining inter-satellite links, Earth observation imaging, autonomous collision avoidance, and formation flying. However, existing onboard orbit determination technologies face a pair of irreconcilable core contradictions when dealing with increasingly complex mission requirements: the contradiction between the computational burden of high-fidelity dynamic models and the limited onboard computing resources, and the contradiction between the complex characteristics of space environment noise and the Gaussian assumption of traditional filtering algorithms.

[0003] 1. Limitations of the accuracy of dynamic models: Traditional spaceborne orbit determination typically employs simplified dynamic models (such as considering only...). The SGP4 model (or low-order numerical integrator) is used to accommodate the limited computing power of the onboard computer (OBC). However, low-Earth orbit satellites are subject to non-spherical perturbations from the Earth (until...). The orbit determination error is influenced by a combination of factors, including the gravitational field, atmospheric drag, solar radiation pressure, lunar and solar gravitational forces, and tidal effects. Atmospheric drag is significantly affected by solar activity (such as solar storms and CMEs) and geomagnetic disturbances, exhibiting high nonlinearity and uncertainty. Simplified physical models cannot describe these small but cumulatively significant perturbations, causing orbit determination errors to diverge rapidly over time (the so-called "truncation error"). Running a high-fidelity orbit propagator (HPOP) on a satellite would require substantial CPU time and storage space, which is impractical for power- and size-constrained micro / nano satellites.

[0004] 2. The shortcomings of traditional Kalman filtering: 1. Currently, the mainstream algorithms for spaceborne orbit determination are the Extended Kalman Filter (EKF) or the Unscented Kalman Filter (UKF). These algorithms are all based on a fundamental assumption: the system process noise and measurement noise follow a Gaussian white noise distribution. However, the real space environment is not like this.

[0005] (1) Non-Gaussianity: The dramatic fluctuations in atmospheric density and the delay of GNSS signals caused by ionospheric scintillation often exhibit a heavy-tailed distribution.

[0006] (2) Long Memory: Recent studies have shown that changes in upper atmospheric density exhibit fractal characteristics, demonstrating significant self-similarity and long-range dependence. That is, the current disturbance state is not only related to the previous moment but also has a power-law decaying correlation with the historical state over a long period of time. Traditional integer-order Kalman filtering (IOKF) lacks the ability to model this "historical memory," which often leads to distortion in the estimation of the covariance matrix when dealing with colored noise, thus causing degradation or even divergence in filtering accuracy.

[0007] 3. Bottlenecks in deep learning applications: While introducing neural networks (such as RNNs and LSTMs) to compensate for errors in dynamic models has become a research hotspot, existing "large model" solutions often involve millions of parameters, have high inference latency, and are difficult to deploy on spaceborne embedded platforms (such as ARM, DSP, or FPGA). In addition, purely data-driven models lack physical constraints and are prone to unpredictable errors when outside the scope of the training data (such as when an unseen severe geomagnetic storm occurs).

[0008] Therefore, there is an urgent need for a new real-time precision orbit determination method that can integrate the determinism of physical models with the adaptability of data-driven approaches, effectively handle fractional colored noise, and meet the requirements of lightweight spaceborne deployment. Summary of the Invention

[0009] To achieve the objectives of this invention, this application provides a real-time satellite orbit determination method based on knowledge distillation and attention-assisted Kalman waves, comprising: Step S1: Acquire multi-source sensor data from the satellite and perform preprocessing; Step S2: Construct a parallel compensation architecture, wherein the construction is based on a simplified dynamic model as the backbone of the baseline trajectory, and a bypass neural network is set up to predict the residuals of the simplified dynamic model; Step S3: Using knowledge distillation technology, the lightweight student model is trained using the teacher model, and the knowledge of the teacher model is transferred to the student model to obtain the trained lightweight student model. Step S4: Based on the long memory characteristics of environmental noise, construct a fractional-order state-space model; Step S5: Input the residual prediction value output in real time by the lightweight student model as a compensation term into the fractional-order state-space model, and perform fractional-order Kalman filtering to fuse real-time observation data for optimal estimation of satellite state; Step S6: Monitor the information sequence of the filtering process in real time. When an anomaly is detected, adaptively adjust the filtering parameters to achieve real-time satellite orbit determination.

[0010] In some specific embodiments, in step S2, the parallel compensation architecture is determined according to the following formula: in, For simplification Main trunk dynamics model; Compensation acceleration predicted by the Attention-LSTM network; Input feature vectors into the network; This is process noise.

[0011] In some specific embodiments, step S3 includes: constructing a composite loss function to train the student model, wherein the composite loss function includes hard target loss, soft target loss and feature matching loss.

[0012] In some specific embodiments, the attention mechanism is a temporal attention mechanism, which is used to calculate the importance weights of the LSTM hidden states within the historical time window, generate a context vector containing key temporal features, and input it into the fully connected layer to output the residual prediction value.

[0013] In some specific embodiments, the composite loss function is determined according to the following formula: in, For real labels, Output for the student model; These are the logits outputs for the teacher and student models, respectively. These are distillation temperature parameters; For the Softmax function; For intermediate layer feature mapping; All are weighting coefficients.

[0014] In some specific embodiments, in step S4, the Hearst exponent of the environmental noise sequence is calculated by the rescaled range analysis method. If the Hearst exponent is greater than 0.5, it is determined that there is long memory characteristics, and the fractional state space model is constructed accordingly.

[0015] In some specific embodiments, in step S5, the fractional-order Kalman filter is discretized based on the Grünwald-Letnikov definition, and the state prediction equation of the fractional-order Kalman filter includes a weighted memory term for historical states to match the long memory characteristics of environmental noise.

[0016] In some specific embodiments, step S6 includes: calculating the weighted squared modulus of the innovation vector as a detection statistic; when the statistic exceeds a preset threshold, it is determined to be an anomaly, and the historical information weight is reduced by amplifying the prediction covariance matrix or adjusting the forgetting factor to enhance the filter's ability to track the current observation.

[0017] In some specific embodiments, in step S3, the student model is a long short-term memory network with an attention mechanism, used to predict the residual.

[0018] In some specific embodiments, in step S1, the multi-source sensor data includes GNSS observation data, star sensor attitude data, and high-precision accelerometer data; the accelerometer data serves as one of the input features of the student model, used to help distinguish between the satellite's own maneuvers and non-conservative forces in the environment.

[0019] The beneficial effects of the above technical solution are as follows: Compared with the prior art, the present invention has the following significant advantages: (1) Balancing high precision and low computing power This invention constructs a parallel compensation architecture with a simplified physical model as the backbone and a lightweight neural network as the bypass, and introduces knowledge distillation technology to transfer the deep knowledge of the high-fidelity teacher model to the Attention-LSTM student model. The number of parameters in the distilled student model is only 5%-10% of that of a conventional deep model, enabling low-power real-time inference on a spaceborne embedded platform. At the same time, its residual prediction accuracy approaches that of a high-fidelity numerical integrator, effectively solving the problem of truncation error accumulation in dynamic models under the condition of limited spaceborne computing resources, and achieving synergistic optimization of low computing power consumption and high orbit determination accuracy.

[0020] (2) Strong adaptive noise suppression capability in complex environments To address the long-memory characteristic prevalent in space environment disturbances, this invention introduces a fractional-order Kalman filter to replace the traditional integer-order filter. Based on the fractional-order state-space model defined by Grünwald-Letnikov, it accurately matches the long correlation characteristics of disturbance sequences such as atmospheric drag by weighted memorization of historical states, thus overcoming the Markov property limitation of traditional Kalman filters. Compared to extended Kalman filters, this invention improves the filtering convergence speed by more than 30% and the steady-state accuracy by more than 50% in environments with strong colored noise, significantly enhancing orbit determination stability in complex space environments.

[0021] (3) The robust response mechanism for abnormal operating conditions is well-developed. This invention establishes a robust anomaly detection and adaptive correction mechanism by real-time monitoring of the innovation sequence and constructing a chi-square test statistic. When encountering severe atmospheric disturbances caused by solar storms, thruster mis-triggers, or momentary sensor malfunctions, the system can quickly identify anomalies and adaptively adjust filter parameters. By amplifying the prediction covariance or adjusting the forgetting factor, it effectively suppresses the risk of filter divergence. This mechanism enables the orbit determination system to maintain continuous and reliable operation under sudden abnormal conditions, greatly enhancing the satellite platform's autonomous survivability.

[0022] (4) Dual guarantee of physical interpretability and mission safety Unlike purely end-to-end black-box deep learning orbit determination schemes, this invention retains a simplified physical model as the backbone, with the neural network serving only as a bypass to compensate for residuals. This architectural design ensures that even if the neural network produces erroneous outputs due to extreme, unseen scenarios, the system can still degenerate into a pure physical model orbit determination mode, guaranteeing the most basic safety baseline for space missions. Simultaneously, the visualized weight distribution of the attention mechanism provides a basis for analyzing the contribution of each input feature to orbit determination accuracy, enhancing the algorithm's interpretability and ease of engineering debugging.

[0023] (5) Micro-nano satellite platforms have strong applicability This invention fully considers the strict constraints of micro- and nano-satellites in terms of size, weight, power consumption, and computing power. The entire algorithm module can be deployed in embedded processors based on FPGA or ARM architecture, without relying on high-performance GPUs or large-capacity storage. This feature enables high-precision real-time autonomous orbit determination capabilities to be deployed on CubeSat platforms at the kilogram level, providing key technical support for the batch deployment and autonomous operation of large-scale low-Earth orbit constellations, and has extremely high engineering and promotion value. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart illustrating a real-time satellite orbit determination method based on knowledge distillation and attention-assisted Kalman waves, provided as an embodiment of the present invention; Figure 2 A schematic diagram of a parallel compensation architecture that integrates physical sensing and data-driven processes, provided as an embodiment of the present invention; Figure 3 A schematic diagram illustrating the training process of a teacher-student knowledge distillation strategy provided in an embodiment of the present invention; Figure 4 A diagram of an LSTM (Attention-LSTM) network structure with an attention mechanism provided for one embodiment of the present invention; Figure 5 A flowchart illustrating the information transfer and update process of a fractional-order Kalman filter (FOKF) provided as an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0027] Examples of the embodiments are shown 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 intended to explain the invention, and should not be construed as limiting the invention.

[0028] Example 1 One embodiment of the present invention provides a real-time satellite orbit determination method based on knowledge distillation and attention-assisted Kalman waves, referring to... Figures 1-5 As shown, it includes: Step S1: Acquire multi-source sensor data from the satellite and perform preprocessing; In a specific embodiment of the present invention, in step S1, the multi-source sensor data includes GNSS observation data, satellite sensor attitude data, and high-precision accelerometer data; the accelerometer data serves as one of the input features of the student model and is used to help distinguish between the satellite's own maneuvering and non-conservative forces in the environment.

[0029] Step S2: Construct a parallel compensation architecture, wherein the construction is based on a simplified dynamic model as the backbone of the baseline trajectory, and a bypass neural network is set up to predict the residuals of the simplified dynamic model; In a specific embodiment of the present invention, in step S2, the parallel compensation architecture is determined according to the following formula: in, For simplification Main trunk dynamics model; Compensation acceleration predicted by the Attention-LSTM network; Input feature vectors into the network; This is process noise.

[0030] To ensure real-time performance, the onboard master model only considers the Earth's central gravity and... The term is perturbed. Its dynamic equation is: in The gravitational constant of Earth, The radius of the Earth's equator. This is the geocentric latitude. The model is extremely fast to calculate, but its accuracy decreases over time.

[0031] Step S3: Using knowledge distillation technology, a lightweight student model is trained using a high-precision teacher model, and the knowledge of the teacher model is transferred to the student model to obtain a trained lightweight student model. In a specific embodiment of the present invention, in step S3, the student model is a long short-term memory network with an attention mechanism, used to predict the residual.

[0032] To obtain the "truth" and "deep knowledge" used for distillation, a high-performance teacher model is built on a ground-based high-performance server: EGM96 model Gravitational field of order.

[0033] The NRLMSISE-00 atmospheric model is adopted, taking into account... (Solar radio flux) and Real-time changes in the (geomagnetic index).

[0034] Light pressure perturbation of a cone model considering Earth's shadow occlusion.

[0035] Student model: Attention-LSTM Input layer: It includes normalized position and velocity, measured accelerometer values ​​(excluding thrust), space environment index, and mean approach angle. (Used to introduce orbital periodicity characteristics).

[0036] LSTM layer: Uses LSTM units to process time series dependencies.

[0037] Temporal Attention: The final hidden state of a traditional LSTM Key early information may be lost. Introducing an attention mechanism for the history window... Weighting the states within: Context vector It includes weighted dynamic characteristics.

[0038] Output layer: Fully connected layer output .

[0039] The teacher model can be a deeper, wider network (or the output sequence of a high-fidelity physics integrator). To train the student model, we define a composite loss function: Hard Loss (MSE): Ensures that the predicted value is close to the true residual.

[0040] Soft Loss (Distillation): If the teacher is also a neural network, students can learn the teacher's understanding of data uncertainty by mimicking the probability distribution of the teacher's output (smoothed by high-temperature Softmax).

[0041] Hint Loss (Feature Based): Forces the intermediate feature maps of the student model to approximate the intermediate layers of the teacher model through linear transformation, enabling students to learn "how to think".

[0042] The student model trained in this way has only 5%-10% of the parameters of a regular deep learning model, making it suitable for embedded deployment.

[0043] In a specific embodiment of the present invention, step S3 includes: constructing a composite loss function to train the student model, wherein the composite loss function includes hard target loss, soft target loss and feature matching loss.

[0044] In a specific embodiment of the present invention, the attention mechanism is a temporal attention mechanism, which is used to calculate the importance weights of the LSTM hidden states within the historical time window, generate a context vector containing key temporal features, and input it into the fully connected layer to output the residual prediction value.

[0045] In one specific embodiment of the present invention, the composite loss function is determined according to the following formula: in, For real labels, Output for the student model; These are the logits outputs for the teacher and student models, respectively. These are distillation temperature parameters; For the Softmax function; For intermediate layer feature mapping; All are weighting coefficients.

[0046] Step S4: Based on the long memory characteristics of environmental noise, construct a fractional-order state-space model; In a specific embodiment of the present invention, in step S4, the Hearst exponent of the environmental noise sequence is calculated by the rescaled range analysis method. If the Hearst exponent is greater than 0.5, it is determined that there is a long memory characteristic, and the fractional-order state-space model is constructed accordingly.

[0047] Step S5: Input the residual prediction value output in real time by the lightweight student model as a compensation term into the fractional-order state-space model, and perform fractional-order Kalman filtering to fuse real-time observation data for optimal estimation of satellite state; Even with LSTM compensation, the residuals still contain random noise. Rescaled range (R / S) analysis reveals the Hearst exponent of the atmospheric drag error sequence for low-Earth orbit satellites. This indicates the existence of positive long-term correlation (Long Memory). The traditional KF hypothesis... (White noise) is not applicable.

[0048] 1. Definition of fractional calculus The discrete fractional derivative is defined using the Grünwald-Letnikov (GL) definition: in The sampling step size, Fractional order ( ), .

[0049] 2. Fractional-order state-space model Constructing discrete state equations: in, This refers to the compensation acceleration predicted by Attention-LSTM. .

[0050] By rearranging the fractional differential terms, we obtain the one-step prediction equation for the state: in, These are the memory weighting coefficients. It can be seen that FOKF not only utilizes the state from the previous time step, but also uses the past time step with weights. The state at a given moment perfectly matches the long memory characteristic of atmospheric disturbances.

[0051] 3. Filtering Update Steps Prior covariance prediction: In the formula, It is the prior state error covariance matrix at time k+1. It is a fractional-order state transition matrix. It is a fractional-order memory weighting coefficient. It is the identity matrix. It is the current moment. and the posterior state error covariance matrix at each historical moment (covariance corrected for observational data), where L is the truncated memory length. It is the process noise covariance matrix.

[0052] Kalman gain calculation: In the formula, The Kalman gain matrix at time t is It is the observation matrix (or measurement matrix). It is the measurement noise covariance matrix.

[0053] Status Update: In the formula, yes The posterior state estimate at time t. yes The prior state estimate at time t. yes The actual observation vector at time t. It measures the residual.

[0054] 4. Anomaly Detection and Adaptive Mechanism In actual operation, thruster malfunctions or micrometeorite impacts may occur. Define the detection statistics: in, For new information.

[0055] if The system is functioning normally and is running according to standard FOKF.

[0056] if This was determined to be abnormal. An adaptive forgetting factor was introduced. ( The weights of historical memory items can be attenuated, or an inflation factor can be introduced. enlarge This increases the weight of the current measurement, enabling the filter to quickly track sudden changes in state.

[0057] Step S6: Monitor the information sequence of the filtering process in real time. When an anomaly is detected, adaptively adjust the filtering parameters to achieve real-time satellite orbit determination.

[0058] In a specific embodiment of the present invention, in step S5, the fractional-order Kalman filter is discretized based on the Grünwald-Letnikov definition, and the state prediction equation of the fractional-order Kalman filter includes a weighted memory term for historical states to match the long memory characteristics of environmental noise.

[0059] In a specific embodiment of the present invention, step S6 includes: calculating the weighted squared modulus of the innovation vector as a detection statistic; when the statistic exceeds a preset threshold, it is determined to be abnormal, and the historical information weight is reduced by amplifying the prediction covariance matrix or adjusting the forgetting factor to enhance the filter's ability to track the current observation.

[0060] This invention proposes an innovative framework for real-time satellite orbit determination. By organically combining the lightweight advantages of knowledge distillation, the nonlinear fitting capability of Attention-LSTM, and the resistance to colored noise of fractional-order Kalman filtering, it successfully solves the problems of insufficient computing power and inaccurate models for low-Earth orbit (LEO) satellites in complex space environments. This method is not only applicable to LEO communication constellations but also has significant implications for the autonomous navigation of deep space probes. This invention is ingeniously conceived, technically complete, and possesses extremely high industrial application value.

[0061] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0062] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the functions specified in one or more boxes. Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the invention. Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0063] The methods and apparatus provided by the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

[0064] In the description of this specification, references to terms such as "an embodiment," "some embodiments," "example," "specific example," or "a specific embodiment" or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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. Such 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 this application.

Claims

1. A real-time satellite orbit determination method based on knowledge distillation and attention-assisted Kalman waves, characterized in that, include: Step S1: Acquire multi-source sensor data from the satellite and perform preprocessing; Step S2: Construct a parallel compensation architecture, wherein the construction is based on a simplified dynamic model as the backbone of the baseline trajectory, and a bypass neural network is set up to predict the residuals of the simplified dynamic model; Step S3: Using knowledge distillation technology, the lightweight student model is trained using the teacher model, and the knowledge of the teacher model is transferred to the student model to obtain the trained lightweight student model. Step S4: Based on the long memory characteristics of environmental noise, construct a fractional-order state-space model; Step S5: Input the residual prediction value output in real time by the lightweight student model as a compensation term into the fractional state space model, and perform fractional Kalman filtering to fuse real-time observation data to make the optimal estimate of the satellite state; Step S6: Monitor the information sequence of the filtering process in real time. When an anomaly is detected, adaptively adjust the filtering parameters to achieve real-time satellite orbit determination.

2. The satellite real-time orbit determination method based on knowledge distillation and attention-assisted Kalman waves according to claim 1, characterized in that, In step S2, the parallel compensation architecture is determined according to the following formula: in, For simplification Main trunk dynamics model; Compensation acceleration predicted by the Attention-LSTM network; Input feature vectors into the network; This is process noise.

3. The satellite real-time orbit determination method based on knowledge distillation and attention-assisted Kalman waves according to claim 1, characterized in that, Step S3 includes: constructing a composite loss function to train the student model, wherein the composite loss function includes hard target loss, soft target loss and feature matching loss.

4. The satellite real-time orbit determination method based on knowledge distillation and attention-assisted Kalman waves according to claim 1 or 3, characterized in that, The attention mechanism is a temporal attention mechanism, which is used to calculate the importance weights of the LSTM hidden states within the historical time window, generate a context vector containing key temporal features, and input it into the fully connected layer to output the residual prediction value.

5. The satellite real-time orbit determination method based on knowledge distillation and attention-assisted Kalman waves according to claim 3, characterized in that, The composite loss function is determined according to the following formula: in, For real labels, Output for the student model; These are the logits outputs for the teacher and student models, respectively. These are distillation temperature parameters; For the Softmax function; For intermediate layer feature mapping; All are weighting coefficients.

6. The satellite real-time orbit determination method based on knowledge distillation and attention-assisted Kalman waves according to claim 1, characterized in that, In step S4, the Hearst exponent of the environmental noise sequence is calculated by the rescaled range analysis method. If the Hearst exponent is greater than 0.5, it is determined that there is long memory characteristics, and the fractional state space model is constructed accordingly.

7. The satellite real-time orbit determination method based on knowledge distillation and attention-assisted Kalman waves according to claim 1, characterized in that, In step S5, the fractional-order Kalman filter is discretized based on the Grünwald-Letnikov definition, and the state prediction equation of the fractional-order Kalman filter includes a weighted memory term for historical states to match the long memory characteristics of environmental noise.

8. The satellite real-time orbit determination method based on knowledge distillation and attention-assisted Kalman waves according to claim 1, characterized in that, In step S6, the adaptive adjustment includes: calculating the weighted squared modulus of the innovation vector as a detection statistic; when the statistic exceeds a preset threshold, it is determined to be an anomaly, and the historical information weight is reduced by amplifying the prediction covariance matrix or adjusting the forgetting factor, so as to enhance the filter's ability to track the current observation.

9. The satellite real-time orbit determination method based on knowledge distillation and attention-assisted Kalman waves according to claim 1, characterized in that, In step S3, the student model is a long short-term memory network with an attention mechanism, used to predict the residual.

10. The satellite real-time orbit determination method based on knowledge distillation and attention-assisted Kalman waves according to claim 1, characterized in that, In step S1, the multi-source sensor data includes GNSS observation data, satellite sensor attitude data, and high-precision accelerometer data; the accelerometer data serves as one of the input features of the student model, used to help distinguish between the satellite's own maneuvers and non-conservative forces in the environment.

Images