Orbit forecasting method for short-time sudden maneuver of star chain satellite

Abstract

The invention discloses an orbit forecasting method for short-time sudden maneuver of a star chain satellite, and mainly solves the problems of updating lag and large forecasting error during sudden maneuver of the satellite in the prior art. According to the scheme, the method comprises the following steps: 1) performing initial orbit calculation based on TLE and SGP4 models; 2) ground observation data are received and differentiated with the predicted value to generate a residual sequence; (3) residual error abnormity is recognized through an abnormity detection algorithm, and a rapid orbit model fine adjustment mechanism is triggered when a sudden maneuvering event is detected; 4) updating antenna pointing parameters in real time through chained coordinate conversion of ECI, ECEF and ENU to realize tracking pointing; and 5) calculating transit starting and ending time by adopting Hermite interpolation based on the corrected satellite orbit and antenna position, and finishing orbit forecasting. According to the method, the orbit prediction precision in the sudden maneuvering state and the real-time response capability of a ground tracking system are remarkably improved, and the limitation that a traditional method depends on a TLE updating period is overcome.

Description

Technical Field

[0001] This invention relates to the fields of satellite orbital dynamics, space target management, and autonomous orbit prediction, specifically a method for orbit prediction of short-term sudden maneuvers of Starlink satellites. This method can quickly detect short-term sudden orbital maneuvers of Starlink satellites and make corresponding corrections, achieving accurate orbit prediction. It is applicable to scenarios such as low-Earth orbit multi-constellation management, intelligent pointing of ground receivers, and space-air networking. Background Technology

[0002] With the large-scale deployment of low-Earth orbit communication satellite constellations such as Starlink, higher demands are placed on the accuracy and response speed of orbit prediction for communication link establishment, interference avoidance, and airspace management. Most existing orbit prediction systems still rely on traditional two-parameter TLE (Two-Line Element set) and SGP4 (Simplified General Perturbations 4) models for orbit extrapolation. However, this method has significant limitations in the following aspects: 1) Low TLE data update frequency: Within the TLE data update interval, once an orbital maneuver occurs, traditional methods cannot reflect the orbital change in time, causing the position error to expand rapidly; 2) Difficulty in quickly detecting and responding to maneuvers: Existing systems lack the ability to automatically identify and process sudden orbital changes (such as avoiding space debris and fine-tuning formations), leading to pointing errors or communication interruptions.

[0003] Therefore, there is an urgent need for an orbit prediction scheme that combines short-term anomaly detection with a rapid model fine-tuning mechanism to monitor the orbital status of low-Earth orbit satellites such as Starlink in real time, thereby improving the system's responsiveness and prediction accuracy. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a new orbit prediction method for Starlink satellites undergoing sudden maneuvers. It primarily solves the problems of poor responsiveness and prediction accuracy in the face of abrupt orbital changes. First, the SGP4 model is used to obtain the Starlink satellite's ECI coordinates. These coordinates are then transformed using Earth rotation parameters and ground station position parameters to obtain the satellite's position in the local coordinate system of the ground station, enabling the calculation of azimuth and elevation angles. Next, an anomaly detection mechanism is introduced to address the lag in TLE orbital parameter updates. Upon detecting a sudden orbital change, ground observation data is fused, and a lightweight neural network or small-sample online learning technique is used to fine-tune the predicted orbit. Finally, piecewise Hermite cubic interpolation is used to model the satellite elevation curve, ensuring the continuity of the interpolation function and its first derivative, effectively improving prediction accuracy. This invention significantly improves system sensitivity, achieves stable tracking under sudden maneuvers, and enhances the accuracy of orbit prediction.

[0005] To achieve the above objectives, the technical solution of the present invention includes the following:

[0006] (1) Using publicly available two-planetary ephemeris data TLE as the basic orbit source, the SGP4 satellite orbit model is used to calculate the real-time position of the satellite in the geocentric inertial coordinate system ECI, and a pre-trained long short-term memory network LSTM model is called to identify sudden maneuvers using an anomaly detection algorithm.

[0007] (2) Satellite orbit observation data are acquired through ground-based telemetry and control equipment, and the observation data is compared with the predicted trajectory to obtain the orbital residual sequence ε at N time points for maneuver event detection and trajectory correction. t-N:t ;

[0008] (3) Calculate the mean μ and root mean square σ of the residuals within N time windows, and use an anomaly detection algorithm to identify residual anomalies and detect whether a maneuvering event has occurred in real time; the anomaly detection algorithm includes sliding window anomaly detection and LSTM model prediction;

[0009] (4) If a maneuvering event is detected, i.e. an abnormal trajectory, the rapid trajectory model fine-tuning mechanism is triggered, and then step (5) is executed; if no abnormal trajectory is detected, the process returns to step (1).

[0010] (5) Convert the adjusted orbit data and update the azimuth and elevation commands of the ground antenna to achieve tracking pointing;

[0011] (6) Based on the corrected satellite orbit and antenna position, Hermite interpolation is used to calculate the transit start and end times for antenna pointing scheduling;

[0012] (7) Complete the orbit prediction for short-term sudden maneuvers and obtain the satellite position and velocity vector.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] First, because this invention uses the SGP4 model to obtain the ECI coordinates of Starlink satellites, combines the Earth's rotation parameters (such as Greenwich Mean Time) to complete the ECI→ECEF conversion, and then uses the ground station position parameters to complete the ECEF→ENU conversion, the position of the satellite in the local coordinate system of the ground station can be obtained, enabling rapid and accurate calculation of azimuth and elevation angles. At the same time, to address the problem of lag in TLE orbital parameter updates, a statistical anomaly detection and time series prediction residual comparison mechanism is introduced, thereby enabling the identification of Starlink satellite orbital changes within seconds, improving the system's sensitivity.

[0015] Second, after discovering a sudden change in the orbit, this invention integrates ground observation data and uses lightweight neural networks or small-sample online learning technology to fine-tune and correct the predicted orbit. Combined with lightweight neural networks or incremental regression methods, it achieves real-time modeling of orbit changes without waiting for the next TLE cycle, thereby significantly shortening the response delay.

[0016] Third, because this invention uses piecewise Hermite cubic interpolation to model the satellite elevation curve, it can ensure the continuity of the interpolation function and its first derivative, thereby improving the accuracy of determining the transit time and significantly improving the prediction accuracy in scenarios with rapidly changing elevation angles.

[0017] Fourth, this invention integrates observation data for residual-driven correction, and is compatible with multi-source observation data input, including star sensor angle measurement, radar ranging, etc., and can realize adaptive orbit update based on trajectory residuals.

[0018] Fifth, this invention can output new azimuth-pitch commands in real time through the linkage calculation of orbit update and coordinate transformation, that is, it supports real-time pointing control and overpass determination update, which can ensure that the ground terminal can still track stably in the event of sudden maneuvering. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention. Detailed Implementation

[0020] The present invention will now be further described with reference to the accompanying drawings.

[0021] Example 1: Refer to Appendix Figure 1 The present invention proposes an orbit prediction method for short-term sudden maneuvers of Starlink satellites, the specific implementation steps of which include:

[0022] Step 1) Using publicly available two-planetary ephemeris data (TLE) as the basic orbit source, the SGP4 satellite orbit model is used to calculate the satellite's real-time position in the geocentric inertial coordinate system (ECI), and a pre-trained Long Short-Term Memory (LSTM) network model is invoked for anomaly detection algorithms to identify sudden maneuvers; the implementation process is as follows:

[0023] (1.1) Analyze the TLE orbital elements of Starlink satellites and calculate the real-time position of the satellites in the geocentric inertial coordinate system (ECI) using the SGP4 model:

[0024] r SGP4 (t)=SGP4(t;a,e,i,Ω,ω,M0)

[0025] The inputs to the SGP4 model include the semi-major axis a, eccentricity e, orbital inclination i, right ascension of the ascending node Ω, argument of perigee ω, and angle of parity M0.

[0026] (1.2) Call the pre-trained LSTM model, which learns the orbital residual sequence at N historical time points, and outputs the residual representation as follows:

[0027]

[0028] Where, ε t-N:t This represents the orbital residual sequence at N time points; f represents the residual orbital position predicted by the model at the next time step. LSTM θ represents the LSTM model function; θ represents the LSTM model parameters as a weight matrix.

[0029] Step 2) Acquire satellite orbit observation data through ground-based telemetry and control equipment, which includes at least ground radar and a radio telemetry and control system. In this embodiment, ground radar is preferably used to acquire the data. Perform differential calculation between the observation data and the predicted trajectory to obtain the orbital residual sequence ε at N time points for maneuver event detection and trajectory correction. t-N:t The orbital residual sequence is obtained by receiving measured orbital position data from ground-based telemetry and control equipment and aligning it with the SGP4 predicted position according to the timestamp; the residual ε(t) at time t is obtained according to the following formula, and an orbital residual sequence ε is constructed for N time points. t-N:t :

[0030] ε(t)=||r obs (t)-r SGP4 (t)||,

[0031] ε t-N:t =[ε t-N ,ε t-N+1 ,…,ε t ],

[0032] Where, r obs (t) represents the measured position of the spacecraft's orbit in the ECI coordinate system, r SGP4 (t) represents the predicted location by the SGP4 model; ε t-N =ε(t-NΔT), where ΔT is the step size.

[0033] Step 3) Calculate the mean μ and root mean square σ of the residuals within N time windows, and use an anomaly detection algorithm to identify residual anomalies and detect whether a maneuvering event has occurred in real time; the anomaly detection algorithm includes sliding window anomaly detection and LSTM model prediction.

[0034] In this embodiment, the mean residual μ and root mean square σ within N time windows are calculated according to the following formula:

[0035]

[0036] Where k represents the time offset index within the window, k = 1, 2, ..., N-1, ε t-k It represents the residual at time t-kΔT.

[0037] In this embodiment, sliding window anomaly detection is used to identify residual anomalies. This is achieved by determining whether the residual ε(t) at time t satisfies the following condition:

[0038] ε(t)>μ+3σ,

[0039] If the conditions are met, it is determined that a maneuvering event has occurred; otherwise, it is determined that no maneuvering event has occurred.

[0040] In this embodiment, LSTM model prediction is used to identify residual anomalies. This is achieved by loading a pre-trained LSTM model and judging the residual predicted by the LSTM model for the next time step. Does it meet the following conditions to be implemented:

[0041]

[0042] If the conditions are met, it is determined that a maneuvering event has occurred; otherwise, it is determined that no maneuvering event has occurred.

[0043] Step 4) If a maneuvering event, i.e., an abnormal trajectory, is detected, the fast trajectory model fine-tuning mechanism is triggered, and then step 5) is executed; if no abnormal trajectory is detected, the process returns to step 1). In this embodiment, the fast trajectory model fine-tuning mechanism triggered in this step specifically involves rapidly adjusting the pre-trained LSTM model by calling few-shot learning or online incremental learning methods to adapt it to the maneuvering trajectory situation. The steps are as follows:

[0044] (4.1) Update the model weights using the exponential moving average method:

[0045]

[0046] In the formula, θ new For the updated model weight parameters, θ old The current model weight parameters; α represents the moving average coefficient for lightweight fine-tuning model initialization, used for rapid weight updates, and its value ranges from 0 to 1; the closer α is to 1, the stronger the model's dependence on historical weights; the closer α is to 0, the more sensitive the model is to the current gradient. The loss function is used to measure the model's predicted residuals. With the observed residual ε t-N:t Differences; Let be the gradient of the loss function, representing the partial derivative of the loss function with respect to the model parameters θ;

[0047] (4.2) Feed the residual prediction values ​​back to the SGP4 orbit model to correct the orbit coordinates:

[0048]

[0049] The corrected orbital position is represented in ECI coordinates, u vel The velocity direction is a unit vector; Δt is the difference between the current moment and the next moment; through u vel The correction direction is determined and u is output synchronously by the SGP4 model. vel .

[0050] Step 5) Convert the adjusted track data and update the azimuth and elevation commands of the ground antenna to achieve tracking pointing. In this embodiment, the adjusted track data is converted to the ground-fixed coordinate system ECEF and the station-centered coordinate system ENU, the azimuth φ and elevation commands θ of the ground antenna are updated, and then φ and θ are converted into antenna control commands and sent to the antenna servo motor through the RS-485 bus. The response delay is <50ms, realizing real-time pointing control.

[0051] Step 6) Based on the corrected satellite orbit and antenna position, Hermite interpolation is used to calculate the transit start and end times for antenna pointing scheduling; this embodiment is specifically implemented as follows:

[0052] Calculate the corrected elevation sequence based on the corrected satellite orbit and antenna position. And construct a piecewise cubic Hermite interpolation function:

[0053] H(t) = h 00 (τ)E i +h 10 (τ)(t i+1 -t i )E′ i +h 01 (τ)E i+1 +h 11 (τ)(t i+1 -t i )E′ i+1 ,

[0054] Where n is the number of orbital extrapolations; E i For t i The observed elevation angle at time E′ i For t i The rate of change of elevation angle at time; in the Hermite interpolation of orbit prediction, four cubic basis functions together constitute the interpolation polynomial, where the basis functions are:

[0055] h00 (τ)=2τ 3 -3τ 2 +1,

[0056] h 10 (τ)=τ 3 -2τ 2 +τ,

[0057] h 01 (τ)=-2τ 3 +3τ 2 ,

[0058] h 11 (τ)=τ 3 -τ 2 ,

[0059] In the formula, τ is the normalized time parameter;

[0060] The root of H(t) = 10 is solved using Newton's iteration method to determine the start and end times of the border crossing.

[0061] Step 7) Complete the orbit prediction for short-term sudden maneuvers and obtain the satellite position and velocity vector.

[0062] Example 2: The overall implementation steps of the orbit prediction method proposed in this example are the same as in Example 1. Its core process includes orbit initialization, anomaly detection, model correction and pointing update, etc. The implementation process of this invention will be further described in detail with specific examples:

[0063] Step 1. System Hardware and Software Deployment

[0064] (1.1) Functional module deployment

[0065] Orbit calculation module: Deploys an orbit mechanics engine based on SGP4 / SDP4, supports NORAD format TLE parsing, and is configured with a multi-threaded parallel computing framework (a single node supports synchronous extrapolation of ≥1000 satellites).

[0066] Observation data receiving module: Building a multi-source sensor interface

[0067] Ground-based radar: Receives raw observation data via UDP protocol (sampling rate ≥ 1Hz, accuracy ≤ 20m)

[0068] Anomaly detection module: Deploys a sliding window statistical detector (window length N = 10 cycles) and an LSTM neural network.

[0069] (1.2) Distributed computing architecture construction

[0070] A microservice architecture is adopted, with modules communicating via the gRPC protocol. The ground antenna control interface is bound to a dedicated industrial control computer to ensure real-time command transmission (jitter <1ms).

[0071] Step 2. System Parameter Initialization and Calibration

[0072] (2.1) Spatiotemporal reference configuration

[0073] 1) Time System: Load the IAU 2006 precession-nutation model from the SOFA library to achieve multi-time system conversion from UTC to TT to TDB. Calculation formula:

[0074] TT = UTC + δT

[0075] In the formula: TT is Earth Time, a high-precision time base, in seconds (s); UTC is Coordinated Universal Time, the international standard time; δT is the time difference between UTC and TT, which is dynamically updated through IERS bulletins.

[0076] 2) Coordinate system: Initialize ECI (J2000), ECEF (ITRF2014) and the station center coordinate system transformation matrix.

[0077] (2.2) Model weight loading

[0078] The pre-trained LSTM model is downloaded from the cloud, its hash value is verified, and then it is loaded into memory. The output residual of this model is represented by the following formula:

[0079]

[0080] In the formula: f represents the residual of the orbital position predicted by the model for the next time step, i.e., the deviation between the SGP4 extrapolated value and the true position. LSTM For the LSTM model function. ε t-N:t Given the input residual sequence, ε t-N:t =[ε t-N ,ε t-N+1 ,…,ε t ], where θ represents the LSTM model parameters, expressed as a weight matrix. The LSTM model predicts the residual value at the next time step by learning the orbital residual sequence from N historical time steps. The prediction results are used to correct the extrapolated orbit of SGP4 in real time.

[0081] Step 3. Align the orbit extrapolation with the observation data.

[0082] (3.1) TLE orbit extrapolation

[0083] Analyze the TLE orbital elements and calculate the ECI coordinates for the next 2 hours using the SGP4 model:

[0084] r SGP4(t)=SGP4(t;a,e,i,Ω,ω,M0)

[0085] In the formula: SGP4 is the orbital dynamics model. The inputs to SGP4 are: semi-major axis a, eccentricity e, orbital inclination i, right ascension of the ascending node Ω, argument of perigee ω, and mean perigee M0. The SGP4 extrapolation time step Δt = 1 s. This model is mainly based on TLE orbital elements, and the time labels in these TLE data are UTC time. When calling the SGP4 model to calculate the orbit, the input prediction time should also be in UTC time. Through this step, the position and velocity information of the orbit at different UTC times are output.

[0086] (3.2) Spatiotemporal alignment of observation data

[0087] Radar measured data r obs (t) and predicted value r SGP4 (t) Align with UTC timestamps and calculate residuals:

[0088] ε(t)=||r obs (t)-r SGP4 (t)||

[0089] In the formula, r obs (t) represents the measured position of the spacecraft's orbit in the ECI coordinate system, with units of km; r SGP4 (t) represents the SGP4 predicted position in ECI coordinate system, in km.

[0090] Construct a residual sequence ε of length N and step size ΔT. t-N:t ,Right now

[0091] ε t-N:t =[ε(t-NΔT),…,ε(t-ΔT),ε(t)]

[0092] Step 4. Sudden maneuver detection and model fine-tuning;

[0093] (4.1) Sliding window anomaly detection

[0094] Calculate the residual statistics within the window:

[0095]

[0096] Where μ is the mean of the residuals, σ is the root mean square of the residuals, and N is the window length (default 10 periods). Trigger condition: If the current residual ε(t) > μ + 3σ, it is determined to be a maneuver event.

[0097] (4.2) LSTM Residual Prediction

[0098] Input the residual sequence into the neural network to predict the residual at the next time step:

[0099]

[0100] If the predicted value Early warning of orbital deviation.

[0101] (4.3) Lightweight model online fine-tuning

[0102] The model weights are updated using the exponential moving average method:

[0103]

[0104] In the formula, θ old θ represents the current model weight parameters (vector or matrix form). new These are the updated model weight parameters. The lightweight fine-tuning model initializes the moving average coefficient α, with a value between 0 and 1, for rapid weight updates. The closer α is to 1, the stronger the model's dependence on historical weights, meaning a more conservative update; the closer α is to 0, the more sensitive the model is to the current gradient, and it will start immediately once an anomaly is detected. The loss function is used to measure the model's predicted residuals. With the observed residual ε t-N:t The differences. The gradient of the loss function represents the partial derivative of the loss function with respect to the model parameters θ, pointing in the direction of the fastest decrease in error, and is calculated in real time through automatic differentiation.

[0105] Step 5. Track correction and transit window calculation.

[0106] (5.1) Track parameter correction

[0107] The residual predictions are fed back into the orbital model to correct the ECI coordinates.

[0108]

[0109] In the formula: the corrected orbital position is represented in ECI coordinates, u vel The velocity direction is a unit vector. Δt is the difference between the current moment and the next moment. Residual. To provide the error magnitude, you need to use the unit vector u in the velocity direction. vel Determine the direction of correction, u vel Output synchronously from the SGP4 model.

[0110] (5.2) Hermite interpolation transit window

[0111] Input the corrected elevation angle sequence Construct a piecewise cubic Hermite interpolation function:

[0112] H(t) = h 00(τ)E i +h 10 (τ)(t i+1 -t i )E′ i +h 01 (τ)E i+1 +h 11 (τ)(t i+1 -t i )E′ i+1

[0113] Where n is the number of orbital extrapolations; E i For t i The observed elevation angle at time E′ i For t i The rate of change of elevation angle at time; in the Hermite interpolation of orbit prediction, four cubic basis functions together constitute the interpolation polynomial, where the basis functions are:

[0114] h 00 (τ)=2τ 3 -3τ 2 +1, h 10 (τ)=τ 3 -2τ 2 +τ,h 01 (τ)=-2τ 3 +3τ 2 h 11 (τ)=τ 3 -τ 2

[0115] The parameter τ is a normalized time parameter, which is dimensionless. The root of H(t) = 10 is solved using Newton's iteration method to determine the start and end times of the transit.

[0116] Step 6. Antenna pointing scheduling and data logging

[0117] (6.1) Real-time pointing control

[0118] The corrected azimuth angle φ and elevation angle θ are converted into antenna control commands and sent to the antenna servo motor via RS-485 bus, with a response delay of <50ms.

[0119] (6.2) Data storage and backtracking

[0120] Structured storage to database:

[0121] a) Orbit data table: Satellite ID, timestamp, ECI coordinates (x,y,z), residual value.

[0122] b) Event Log Table: Exception Type, Trigger Time, Correction Amount, Operator ID.

[0123] c) Time-series data is written to InfluxDB with a 30-day retention policy and supports SQL queries.

[0124] Step 7: Anomaly Handling and System Disaster Recovery

[0125] (7.1) Observation Interruption Degradation Mode

[0126] If radar data is lost, switch to hybrid prediction mode:

[0127] r hybrid (t)=r SGP4 (t)+λε hist (t)

[0128] In the formula: r hybrid (t) represents the mixed prediction location, indicating the downgraded prediction result when observations are interrupted. It is expressed in ECI coordinate system, with units of km. λ is the historical residual decay coefficient, typically 0.7; ε hist (t) represents the historical residual mean.

[0129] (7.2) Model self-checking and retraining

[0130] Automatic model validation at fixed times daily: If the prediction error is >10km for 3 consecutive times, full model retraining is triggered, taking ≤30 minutes.

[0131] The parts of this invention not described in detail are common knowledge to those skilled in the art.

[0132] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Obviously, those skilled in the art, after understanding the content and principle of the present invention, may make various modifications and changes in form and detail without departing from the principle and structure of the present invention. However, these modifications and changes based on the concept of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. An orbit prediction method for short-time burst maneuvers of Starlink satellites, characterized by, The method comprises the following steps: (1) taking the disclosed two-line element data TLE as a basis orbit source, calculating the real-time position of the satellite in the earth-centered inertial coordinate system ECI by using the SGP4 satellite orbit model, and calling a pre-trained long short-term memory network LSTM model for an abnormality detection algorithm to identify a sudden maneuver; (2) Obtain satellite orbit observation data through ground tracking equipment, and calculate the difference between the observation data and the predicted trajectory to obtain an orbit residual sequence ε at N time points for maneuver detection and trajectory correction t-N:t ; (3) calculating the residual mean μ and root mean square σ in the N-time window, identifying residual abnormalities by using an abnormality detection algorithm, and detecting whether a maneuver event occurs in real time; the abnormality detection algorithm comprises a sliding window abnormality detection and an LSTM model prediction; (4) if a maneuver event, i.e., an abnormal trajectory, is detected, triggering a rapid orbit model fine-tuning mechanism, and then performing step (5); if no abnormal trajectory is detected, returning to perform step (1); (5) converting the adjusted orbit data, updating the azimuth and elevation angle instructions of the ground antenna, and realizing tracking pointing; (6) based on the corrected satellite orbit and the antenna position, calculating the transit start and end times by using Hermite interpolation, and using the transit start and end times for antenna pointing scheduling; (7) completing the orbit prediction of the short-time sudden maneuver, and obtaining the satellite position and velocity vector.

2. The method of claim 1, wherein: In step (1), the real-time position of the satellite in the earth-centered inertial coordinate system ECI is calculated by using the SGP4 satellite orbit model, and a pre-trained long short-term memory network LSTM model is called, and the implementation of step is as follows: (1.1) analyzing the TLE orbit root number of the Starlink satellite, and calculating the real-time position of the satellite in the earth-centered inertial coordinate system ECI by using the SGP4 model: r SGP4 (t) = SGP4(t; a, e, i, Ω, ω, M0) wherein the input of the SGP4 model comprises a semi-major axis a, an eccentricity e, an orbit inclination i, an ascending node right ascension Omega, an argument of perigee omega and a mean anomaly M0; (1.2) calling a pre-trained LSTM model, which learns the orbit residual sequence of N historical times, and the residual is represented as follows: where ε t-N:t represents the orbit residual sequence at N time instants; is the model predicted next time instant orbit position residual; f LSTM is the LSTM model function; θ is the LSTM model parameters represented by weight matrix.

3. The method of claim 1, wherein: The ground measurement and control equipment in step (2) comprises a ground radar and a radio measurement and control system.

4. The method of claim 3, wherein: The orbit residual sequence in step (2) is obtained by receiving the measured orbit position data of the ground tracking equipment, and aligning it with the SGP4 predicted position according to the time stamp; the residual error ε(t) at time t is obtained according to the following formula, and the orbit residual sequence ε of N time points is constructed in this way t-N:t : ε(t) = ||r obs (t) - r SGP4 (t) ||, ε t-N:t = [ε t-N , ε t-N+1 ,..., ε t ], where r obs (t) is the measured position of the spacecraft orbit in the ECI coordinate system, r SGP4 (t) is the SGP4 model predicted position; ε t-N = ε(t - NΔT), ΔT is the step size.

5. The method of claim 1, wherein: In step (3), the residual mean μ and root mean square sigma in the N-time window are calculated according to the following formula: where k represents a time offset index within the window, k = 1, 2,..., N-1, ε t-k denotes the residual at time t - kΔT.

6. The method of claim 5, wherein: In step (3), the residual abnormality is identified by using the sliding window abnormality detection, which is realized by judging whether the residual error epsilon(t) at time t satisfies the following condition: epsilon(t) > mu + 3 * sigma, If yes, it is determined that a maneuver event occurs; otherwise, it is determined that no maneuver event occurs.

7. The method of claim 5, wherein: In step (3), the residual anomaly is identified by using the LSTM model for prediction. The pre-trained LSTM model is loaded to determine whether the LSTM model predicts the residual at the next time point The following conditions are met: If yes, it is determined that a maneuver event occurs; otherwise, it is determined that no maneuver event occurs.

8. The method of claim 1, wherein: In step (4), the rapid orbit model fine-tuning mechanism is specifically realized by calling a small sample learning or online incremental learning method to quickly adjust the pre-trained LSTM model to adapt to the post-manipulation orbit situation, and the implementation of step is as follows: (4.1) updating the model weight by using the exponential moving average method: In the formula, θ new is the updated model weight parameter, θ old is the current model weight parameter; α represents the lightweight fine-tuning model initialization moving average coefficient, which is used for weight fast updating, and the value is 0-1; the closer α is to 1, the stronger the model depends on the historical weight; the closer α is to 0, the more sensitive the model is to the current gradient; is the loss function, which is used to measure the difference between the model prediction residual and the observation residual ε t-N:t ; is the loss function gradient, which represents the partial derivative of the loss function with respect to the model parameter θ. (4.2) feeding back the residual prediction value to the SGP4 orbit model to correct the orbit coordinates: wherein the corrected orbital position is expressed in ECI coordinates, u vel is the unit vector of velocity direction; Δt is the difference between the current time and the next time; u vel is determined by the SGP4 model synchronous output vel .

9. The method of claim 1, wherein: Step (5) converts the adjusted track data, updates the azimuth and elevation angle instructions of the ground antenna, specifically converts the adjusted track data to the ECEF and ENU coordinate systems, updates the azimuth φ and elevation angle instructions θ of the ground antenna, and then converts φ and θ into antenna control instructions, which are sent to the antenna servo motor through the RS-485 bus, with a response delay of <50 ms, thereby achieving real-time pointing control.

10. The method of claim 1, wherein: Step (6) calculates the transit start and end times using Hermite interpolation, specifically as follows: Based on the corrected satellite orbit and antenna position, a corrected elevation angle sequence is calculated And build piecewise cubic Hermite interpolation function: H(t) = h 00 (τ)E i +h 10 (τ)(t i+1 -t i )E′ i +h 01 (τ)E i+1 +h 11 (τ)(t i+1 -t i )E′ i+1 , where n is the number of extrapolation of the orbit; E i is the elevation observation at time t i is the elevation observation at time t i is the rate of change of the elevation at time t i is the rate of change of the elevation at time t In the Hermite interpolation of the orbit prediction, four cubic base functions jointly constitute the interpolation polynomial, where the base functions: h 00 (τ) = 2τ 3 -3τ 2 +1, h 10 (τ) = τ 3 -2τ 2 +τ, h 01 (τ) = -2τ 3 + 3τ 2 , h 11 (τ) = τ 3 -τ 2 , In the formula, τ is a normalized time parameter; The transit start and end times are determined by solving the root of H(t) = 10 using the Newton iteration method.

Images