Target track maneuver passive detection method and system based on double attention contrast learning

Abstract

The application discloses a target orbit maneuver passive detection method and system based on double attention contrast learning, and belongs to the technical field of space situation awareness. The application proposes a technical scheme combining square root cubature Kalman filtering and DC detector double attention contrast learning neural network by only using right ascension and declination angle measurement data under the conditions of small impulse maneuver and long-distance passive angle measurement of non-cooperative spacecraft in the geostationary orbit. By establishing an orbit dynamics and angle measurement model, filtering output normalizing innovation, constructing a sample database and training a double attention network to obtain an abnormal threshold, whether the target has an orbit maneuver and the maneuver time point are finally determined according to the abnormal score and output. The application does not require range and speed measurement data, reduces the requirements of spaceborne equipment, can accurately capture small impulse maneuvers of 0.5-2 m / s, the detection success rate is increased by more than 10% compared with traditional methods, the time error is reduced by more than 100 s, and is suitable for real-time passive detection of non-cooperative targets in long-distance, high-orbit, near-circular and near-coplanar orbits, thereby providing key support for spacecraft safety warning.

Description

Technical Field

[0001] This invention relates to the field of space situational awareness technology. Specifically, it is a passive detection method for non-cooperative target orbital maneuvers based on dual attention contrastive representation learning for time series anomaly detection (DCdetector). This method is applicable to non-cooperative spacecraft maneuvering in geostationary orbit (GEO) using pulse thrust systems, enabling accurate detection of their maneuvering time points and providing key technical support for spacecraft to safely avoid threats. Background Technology

[0002] With the vigorous development of human space activities, the number of space debris, rocket debris, failed satellites, and active satellites in low Earth orbit has surged. Due to the limitations of ground-based detection in terms of station location and number, it is impossible to achieve full coverage, all-weather, and all-time detection and early warning. Therefore, the development of onboard autonomous detection and early warning is of great significance.

[0003] Orbital maneuvers typically cause significant changes in satellite orbital parameters, so monitoring changes in orbital characteristic parameters is often used to identify whether a satellite has undergone a maneuver. Currently, research on satellite orbital maneuver detection can be mainly divided into two categories. The first category is maneuver detection methods based on historical data, primarily using two-line orbital elements (TLE) data for maneuver detection; the second category is orbital maneuver detection methods based on observational data, generally involving real-time detection, with observational data typically sourced from ground-based or space-based radar, optical cameras, and other equipment.

[0004] The relative motion of near-circular and near-coplanar high-orbit orbits changes slowly, making it extremely difficult to detect orbital maneuvers of distant targets using passive measurements via spaceborne sensors. Existing technologies have significant limitations: when the magnitude of pulse maneuvers is small (typically 0.5-2 m / s) and the relative distance at the moment of maneuver is large, the measured angle changes slowly, the anomaly characteristics are weak, and they are easily drowned out by measurement noise, making them difficult to effectively capture by traditional detection methods. Currently, most methods use TLE data for maneuver detection, which cannot achieve real-time detection. Observational data-based detection methods suffer from low success rates, large detection time errors, and even missed or false detections, failing to meet the needs of space situational awareness for refined monitoring of maneuvering targets.

[0005] To address the shortcomings of the existing technologies, this invention proposes a passive detection method for non-cooperative target orbital maneuvers based on DCdetector. This method utilizes only angle measurement data and constructs a square root capacitive Kalman filter and a DCdetector neural network model to achieve accurate detection of maneuvering targets, filling the gap in existing technologies in the field of maneuver detection. Summary of the Invention

[0006] The core objective of this invention is to provide a passive detection method for non-cooperative target orbital maneuvers based on DCdetector, which solves the problems of low success rate and large detection time error in existing technologies, and achieves high-precision real-time maneuver detection of maneuvers using only angle measurement data.

[0007] Technical solution: A passive detection method for target orbital maneuvers based on dual-attention contrastive learning, comprising the following steps: S1. Establish a spacecraft motion dynamics model and an observation star line-of-sight angle measurement model for the target star based on the geocentric inertial coordinate system and the LVLH coordinate system; S2. Introducing square root occult Kalman filtering, the mean and covariance of the nonlinear random function are directly calculated based on the numerical integration of the occultation principle. The square root occult Kalman filtering takes the observed data as input, including setting the occultation points and weights, calculating the state prediction and variance prediction, then calculating the observation prediction and covariance, and outputting the measurement innovation and normalized innovation. S3. Based on the normalized innovation obtained in step S2, construct a normalized innovation database. The normalized innovation database can generate normalized innovation sequences by setting different maneuvering conditions or different target star orbits. According to the window size, take the normalized innovation sequence before the maneuver as neural network training data and store it in the database. S4. Construct a DCdetector-based mobile detection model, which includes a data reconstruction and encoding module, a dual-attention contrast module, and a representation difference module: The reorganization and encoding module performs block division, embedding, and positional encoding; The dual-attention comparison module includes an intra-block attention layer and an inter-block attention layer, which share weights; The differential representation module uses KL divergence to calculate loss and outlier scores, enabling asynchronous training. S5. Train the DCdetector neural network model using sample data so that it can calculate the anomaly score threshold. The anomaly score threshold is used to determine whether the target spacecraft has performed an orbital maneuver, including outputting the maneuver time point.

[0008] Furthermore, the orbital coordinate system mentioned in step S1 includes a geocentric inertial coordinate system and an LVLH coordinate system. The geocentric inertial coordinate system, or simply the inertial frame, has its origin at the Earth's center. The axis points north along the Earth's axis of rotation. shaft and The axis lies in the equatorial plane, and The axis points to the J2000 vernal equinox. shaft and axis, The axis forms a right-handed system.

[0009] The LVLH coordinate system has its origin at the center of mass of the observed star. The axis is opposite in direction to the satellite's position vector. The axis is located in the orbital plane and is parallel to it. The axis is perpendicular to and in the same direction as the velocity. shaft and axis, The axis forms a right-handed system, perpendicular to the plane of the track.

[0010] Considering the non-spherical gravitational perturbations of the Earth, it mainly includes , and (The perturbation term is the main non-gravitational disturbance source of the GEO spacecraft's motion and needs to be included in the model to ensure dynamic accuracy.) A dynamic model of the spacecraft's motion in the inertial frame is constructed, and the constructed dynamic equations of the spacecraft's motion in the inertial frame are as follows:

[0011] in, It is the velocity vector of the spacecraft in the inertial frame. This represents the position vector of the spacecraft in the inertial frame. For the central gravitational acceleration, For the acceleration caused by Earth's non-spherical gravitational perturbation, adding a dot to the variable indicates the first derivative with respect to time; Let be the motion state quantity of the spacecraft in the inertial frame, and be the dimension. ; It is the position vector of the spacecraft in the inertial frame; It is the velocity vector of the spacecraft in the inertial frame; For the central gravitational acceleration, It is the Earth's gravitational constant.

[0012] Furthermore, in the LVLH system, the equation for measuring the right ascension of the line of sight between the observed star and the target star is used. This indicates that the equation for measuring the declination of the line of sight between the observed star and the target star is used... express: In the formula, , , These represent the relative positions of the observed star and the target star on the three axes of the LVLH system (the orbit of the observed star is known and can be accurately obtained through its own orbit measurement data). , To separate the measurement noise of right ascension and declination, and considering the measurement noise of actual observation equipment, the model is zero-mean Gaussian white noise, i.e. and , and These are the standard deviations of right ascension and declination, respectively.

[0013] Further, step S2 involves constructing a square root capacitive Kalman filter. The filter input is the measurement data, and the output is the normalized information. The filter calculation process is as follows: The time update process is as follows: Calculate the volume point: , for Posterior state estimation at time 10:00. express Time error covariance matrix The square root of the equation satisfies , The third spherical radial rule generated by the third order One volume point; Propagation volume point: , The system state transition function is based on the spacecraft orbital dynamics equations and needs to complete the conversion from relative state to geocentric inertial frame, orbit recursion, and inverse conversion of the recursion result to LVLH frame. Calculate the mean of state predictions: , For the first The weight of each volume point This represents the total number of volumetric points. Calculate the centered volume point matrix: ; Calculate the square root of the predicted covariance: In the formula, The process noise covariance matrix is... For the Cholesky decomposition operator, For QR decomposition operators; The measurement update process is as follows: Calculate and measure the volume point: In the formula This is the measurement model function, corresponding to the right ascension and declination angle measurement model; Calculate the predicted mean of the measurement: ; Calculate the centered volume point matrix of the measurement: ; Calculate the square root of the cross covariance matrix: ; Calculate the square root of the measurement covariance: In the formula, The measurement noise covariance matrix; Calculate the Kalman gain: ; Updated state estimate: In the formula, yes The actual measurement value at that moment. To measure new information; Update the square root of the covariance: .

[0014] Furthermore, the construction of the normalized information database in step S3 is as follows: S31. Based on the measurement update completed by square root occultation Kalman filtering, extract the measurement innovation vector at each time step and construct the original innovation sequence. ,in for The actual measurements obtained from the observed stars at all times, including right ascension and declination; The filtering model is based on the state prediction of the previous time step. Time measurement value; S32. Calculate the normalized innovation by measuring the innovation and the square root of the covariance:

[0015] in for Information on the measurement of time for The square root of the measurement covariance at time. for The normalization of information in time; S33. Obtain the real timestamp of the target star's orbital maneuver. The effective time period before the maneuver occurred was selected as the positive sample set, including: Extracting sample fragments: by fixed length Slide segments from the normalized innovation sequence to generate a sample matrix. :

[0016] in , This refers to the length of the pre-mobilization time period; Operating condition classification and labeling: The samples are classified and labeled according to different track movement operating conditions. Samples under the same operating condition are grouped into the same sub-database to facilitate balanced training in the future. S34. Store the processed normalized information fragments and their corresponding working condition labels and timestamp information into the database, and establish a multi-dimensional index structure. , , For the first A sample fragment of new information prior to the maneuver. For sample labels, These are the driving parameters corresponding to this sample. This represents the total number of samples in the database.

[0017] Furthermore, in the DCdetector-based motion detection model in step S4, the recombination and encoding module first obtains time-series normalized information from the normalized information database. Preprocessing is performed, and the preprocessing is expressed as follows: When the target does not maneuver, the normalized innovation value typically remains below 0.01. The preprocessing method used not only helps maintain the non-negativity of the data but also effectively reduces the order-of-magnitude differences between normal data points while significantly amplifying outliers. At this point, the data is updated to... The training and validation sets are divided according to a certain ratio. The test set is generated by applying a sliding window to a specific track, and then the test set is divided according to the set block size and number. It is divided into intra-block structure and inter-block structure; Set the block size to The number of blocks is ,satisfy The data is divided into the following intra-block structure based on the block operation. and inter-block structure :

[0018] Through embedding and positional encoding operations, the data is transformed into a high-dimensional representation suitable for attention mechanisms, with the embedding dimension set to [value]. The corresponding intra-block data and inter-block data are:

[0019] The dual-attention comparison module inputs the intra-block data into the intra-block attention layer, first initializing the query matrix, key matrix, and value matrix:

[0020] In the formula: These are the query matrix, key matrix, and value matrix, respectively. The corresponding weight matrix is ​​the same for both inter-block and intra-block weight matrices, and the two attention layers share weights. Number of heads in the attention layer: Then, the attention of each head in the attention layer is calculated, and the intra-block attention is obtained by concatenating the attention. The inter-block data is input into the inter-block attention layer. Similarly, the query matrix, key matrix and value matrix are initialized first, then the attention of each head is calculated, and the inter-block attention is obtained by concatenating the attention. Calculate the attention of each head:

[0021] Then, attention within the block is obtained through attention splicing. (Inter-block attention) ):

[0022] In the formula: Output the weight matrix for in-block attention.

[0023] Finally, after processing by the linear layer and the flattening layer, the data maintains the same structural format as before the partitioning.

[0024] The final intra-block attention value is obtained by averaging the intra-block attention and inter-block attention outputs of all modules. Inter-block attention value ; The representation difference module obtains two different attention values ​​from the dual attention comparison module, measures the similarity between the two different outputs using KL divergence, and calculates the loss and anomaly score.

[0025] The loss function is calculated as follows:

[0026] In the formula: For KL divergence calculation, To stop gradient operations, used for asynchronous training of two branches.

[0027] Furthermore, the anomaly score is used to determine whether the data is abnormal. An anomaly score is calculated for each set of time series data. The anomaly score is compared with the anomaly score threshold obtained during the training phase. If the anomaly score is greater than the threshold, the data is considered abnormal, that is, the target has performed a maneuver. The formula for calculating the anomaly score is as follows: ,in Calculate the KL divergence.

[0028] On the other hand, the present invention provides a passive target orbital maneuver detection system based on dual-attention contrastive learning, which is configured to perform the steps of the above-described passive target orbital maneuver detection method. The system includes: The model building unit is used to establish a spacecraft motion dynamics model and an observation star line-of-sight angle measurement model for the target star based on the geocentric inertial coordinate system and the LVLH coordinate system. The filtering calculation unit is used to perform square root commensurate Kalman filtering, taking right ascension and declination angle measurement data as input and outputting measurement innovation and normalized innovation sequences; The sample database unit is used to build and store a normal sample database before maneuvering based on normalized information, and supports the generation and indexing of samples under multiple operating conditions and multiple tracks. The network model unit is used to construct a DCdetector mobile detection model that includes a data recombination and encoding module, a dual attention contrast module, and a representation difference module. The training decision unit is used to train the DCdetector model using normal samples to obtain the anomaly score threshold, calculate the anomaly score in real time, determine whether the target has performed an orbital maneuver, and output the maneuver time point.

[0029] Furthermore, the filtering calculation unit includes: The time update module is used for volume point calculation, state propagation, state prediction, and calculation of the square root of the prediction covariance. The measurement update module is used for measurement volume point calculation, observation prediction, innovation covariance calculation, Kalman gain calculation, state update, and covariance update. The innovation normalization module is used to perform normalization calculations on the measured innovation to obtain the normalized innovation.

[0030] Furthermore, the anomaly score is calculated using the following formula: ,in Let KL divergence be the KL divergence. Indicates the attention value within the block. Indicates the inter-block attention value; The training decision unit is configured as follows: DCdetector was trained using normal samples before the maneuver to determine the anomaly score threshold; Real-time anomaly judgment: when When the threshold is reached, it is determined that the target has undergone orbital maneuvering, and the current time is output as the maneuvering time point.

[0031] Beneficial effects: The method provided by this invention includes the following substantial features and significant advancements: (1) This invention can detect maneuvering targets using only angle measurement data (right ascension and declination). This reduces the requirements for observation equipment and makes it easy to deploy in existing space observation systems.

[0032] (2) The present invention can accurately detect the maneuvering time point of space maneuvering targets through the algorithm, with a high detection success rate and low detection time error, providing key support for space situational awareness.

[0033] (3) The present invention has a wide range of applications and has broad requirements for target trajectory type, target maneuver time and distance, target maneuver size and direction. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the algorithm logic of one embodiment of the present invention; Figure 2 This is a flowchart of the mobile detection model based on DCdetector; Figure 3 This is a statistical analysis of the detection success rate of target satellites 1-10 in 100 target tests according to an embodiment of the present invention. Figure 4 This is a statistical analysis of the detection time error of target stars 1 to 10 in 100 shooting tests according to an embodiment of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] This invention provides a passive detection method for non-cooperative target orbital maneuvers based on dual-attention contrastive learning (DCdetector). By utilizing angle measurement data, a square-root calculus Kalman filter and a DCdetector neural network model are constructed. Anomaly score thresholds are obtained through neural network training. Anomaly detection is achieved by calculating the anomaly score and comparing it to the threshold, and the maneuver time point is output. Specific implementation steps are combined with… Figure 1 As shown.

[0037] A passive detection method for non-cooperative target orbital maneuvers based on DCdetector includes the following steps: Step S1: Establish a spacecraft motion dynamics model Model for measuring the line-of-sight angle between the observed star and the target star .

[0038] For spacecraft motion dynamics model By differentiating the motion state variables of the spacecraft in the inertial frame, the mathematical representation of the model is constructed as follows: This indicates the calculation of the derivative of the spacecraft's position vector in the inertial frame.

[0039] The observation model for right ascension and declination of the target relative to the tracker's line of sight is as follows:

[0040] use This represents the state of motion of a spacecraft in an inertial frame of reference. and These represent the spacecraft's position vector and velocity vector in the inertial frame, respectively.

[0041] Step S2: Construct a square root capacitive Kalman filter, input the measurement data, and output the normalized information.

[0042] Step S3: Construct a normalized information database.

[0043] Step S4: Preprocess the database data and divide it into training set, validation set and test set.

[0044] Step S5: Construct the DCdetector neural network model, such as... Figure 2 As shown.

[0045] Step S6: Input the training set and validation set to train the neural network model and obtain the anomaly score threshold.

[0046] Step S7: Input the test set, calculate the anomaly score, compare it with the threshold to achieve anomaly detection and output the result.

[0047] Furthermore, this embodiment sets the following calculation conditions and technical parameters: 1) The initial orbital parameters of the observed stars and target stars 1-10 are shown in the table below; Table 1. Initial orbital root numbers for observed stars and target stars 1-10

[0048] 2) Simulation parameter settings are shown in the table below: Table 2 Simulation Parameters

[0049] 3) Initial position of the target star in the LVLH system. The corresponding errors for the three axes are 40km, 4km, and 40km, with initial speeds... The errors corresponding to the three axes are 3 m / s, 3 m / s, 3 m / s, and the standard deviations of the right ascension and declination angle measurements are... and All are 0.005°.

[0050] The passive detection method for non-cooperative target orbital maneuvers based on the present invention, utilizing a square root capacitive Kalman filter combined with a DCdetector neural network, was simulated and verified against the aforementioned computational parameters. The simulation duration was 80,000 s. A pulse velocity of 0.5 m / s was applied to the target when the relative distance between the two satellites was 350 km. The maneuver direction was random in the inertial frame, and then the method of the present invention was used to detect the target maneuver time points. Figure 3 and Figure 4 The figures show the detection success rate statistics for targets 1-10 in 100 firing attempts and the detection time error statistics for targets 1-10 in 100 firing attempts. As can be seen from the figures, the method presented in this paper achieves a high success rate in detecting maneuvering targets and a small detection time error. Under the method presented in this paper, the detection success rate for targets 6-10 is higher than 95% when the time error is ≤400s, and the average time error under successful detection does not exceed 300s; for targets 1-5, the detection success rate is higher than 85% when the time error is ≤600s, and the average time error under successful detection does not exceed 300s.

[0051] Therefore, the method of this invention breaks through the space target maneuver detection method based on the passive angle measurement condition of spaceborne spacecraft. It can accurately detect the maneuver time point of space maneuvering targets using only angle measurement data. The detection success rate is improved by more than 10% compared with the traditional method, and the detection time error is reduced by more than 100 seconds compared with the traditional method.

[0052] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A passive detection method for target orbital maneuvers based on dual-attention contrastive learning, characterized in that the steps include: include: S1. Establish a spacecraft motion dynamics model and an observation star-to-target star line-of-sight angle measurement model based on the geocentric inertial coordinate system and the LVLH coordinate system, including the right ascension and declination observation models for the target's line of sight relative to the tracker. S2. Introducing square root occult Kalman filtering, the mean and covariance of the nonlinear random function are directly calculated based on the numerical integration of the occultation principle. The square root occult Kalman filtering takes the observed data as input, including setting the occultation points and weights, calculating the state prediction and variance prediction, then calculating the observation prediction and covariance, and outputting the measurement innovation and normalized innovation. S3. Based on the normalized innovation obtained in step S2, construct a normalized innovation database. The normalized innovation database can generate normalized innovation sequences by setting different maneuvering conditions or different target star orbits. According to the window size, take the normalized innovation sequence before the maneuver as neural network training data and store it in the database. S4. Construct a DCdetector-based mobile detection model, which includes a data reconstruction and encoding module, a dual-attention contrast module, and a representation difference module: The reorganization and encoding module performs block division, embedding, and positional encoding; The dual-attention comparison module includes an intra-block attention layer and an inter-block attention layer, which share weights; The differential representation module uses KL divergence to calculate loss and outlier scores, enabling asynchronous training. S5. Train the DCdetector neural network model using sample data so that it can calculate the anomaly score threshold. The anomaly score threshold is used to determine whether the target spacecraft has performed an orbital maneuver, including outputting the maneuver time point.

2. The passive detection method for target trajectory maneuvering according to claim 1, characterized in that, The orbital coordinate system mentioned in step S1 includes the geocentric inertial coordinate system and the LVLH coordinate system; The spacecraft motion dynamics model considers the J2, J3, and J4 terms of the Earth's non-spherical gravitational perturbation; The line-of-sight angle measurement model is a model for measuring the right ascension and declination of the observed star relative to the target star, and the measurement noise is zero-mean Gaussian white noise.

3. The passive detection method for target trajectory maneuvering according to claim 1 or 2, characterized in that, The spacecraft motion dynamics equations in the inertial frame constructed in step S1 are as follows: in, It is the velocity vector of the spacecraft in the inertial frame. This represents the position vector of the spacecraft in the inertial frame. For the central gravitational acceleration, For the acceleration caused by Earth's non-spherical gravitational perturbation, adding a dot to the variable indicates the first derivative with respect to time; In the LVLH system, the equation for measuring the right ascension of the line of sight between the observed star and the target star is used. This indicates that the equation for measuring the declination of the line of sight between the observed star and the target star is used... express: In the formula, , , These represent the relative positions of the observed star and the target star along the three axes in the LVLH system. , The measurement noise for right ascension and declination is modeled as zero-mean Gaussian white noise.

4. The passive detection method for target trajectory maneuvering according to claim 1, characterized in that, The square root volume Kalman filter calculation in step S2 includes two stages: time update and measurement update. The time update process is as follows: Calculate the volume point: ， for Posterior state estimation at time 10:

00. express Time error covariance matrix The square root of the equation satisfies , The third spherical radial rule generated for the third order One volume point; Propagation volume point: , The system state transition function is based on the spacecraft orbital dynamics equations and needs to complete the conversion from relative state to geocentric inertial frame, orbit recursion, and inverse conversion of the recursion result to LVLH frame. Calculate the mean of state predictions: , For the first The weight of each volume point This represents the total number of volumetric points. Calculate the centered volume point matrix: ; Calculate the square root of the predicted covariance: In the formula, The process noise covariance matrix is... For the Cholesky decomposition operator, For QR decomposition operators; The measurement update process is as follows: Calculate and measure the volume point: In the formula This is the measurement model function, corresponding to the right ascension and declination angle measurement model; Calculate the predicted mean of the measurement: ; Calculate the centered volume point matrix of the measurement: ; Calculate the square root of the cross covariance matrix: ; Calculate the square root of the measurement covariance: In the formula, The measurement noise covariance matrix; Calculate the Kalman gain: ; Updated state estimate: In the formula, yes The actual measurement value at that moment. To measure new information; Update the square root of the covariance: .

5. The passive detection method for target trajectory maneuvering according to claim 1, characterized in that, The construction of the normalized information database in step S3 is as follows: S31. Based on the measurement update completed by square root occultation Kalman filtering, extract the measurement innovation vector at each time step and construct the original innovation sequence. ,in for The actual measurements obtained from the observed stars at all times, including right ascension and declination; The filtering model is based on the state prediction of the previous time step. Time measurement value; S32. Calculate the normalized innovation by measuring the innovation and the square root of the covariance: in for Information on the measurement of time for The square root of the measurement covariance at time. for The normalization of information in time; S33. Obtain the real timestamp of the target star's orbital maneuver. The effective time period before the maneuver occurred was selected as the positive sample set, including: Extracting sample fragments: by fixed length Slide segments from the normalized innovation sequence to generate a sample matrix. : in , This refers to the length of the pre-mobilization time period; Operating condition classification and labeling: The samples are classified and labeled according to different track movement operating conditions. Samples under the same operating condition are grouped into the same sub-database to facilitate balanced training in the future. S34. Store the processed normalized information fragments and their corresponding working condition labels and timestamp information into the database, and establish a multi-dimensional index structure. , , For the first A sample fragment of new information prior to the maneuver. For sample labels, These are the driving parameters corresponding to this sample. This represents the total number of samples in the database.

6. The passive detection method for target trajectory maneuvering according to claim 1, characterized in that, In step S4, in the DCdetector-based motion detection model, the recombination and encoding module first preprocesses the time-series normalized information obtained from the normalized information database, and then divides the normalized information into intra-block structure and inter-block structure according to the set block size and number. The dual attention comparison module inputs the data within the block into the attention layer within the block. First, it initializes the query matrix, key matrix, and value matrix. Then, it calculates the attention of each head of the attention layer and obtains the attention within the block by concatenating the attention. Inter-block data is input into the inter-block attention layer. The query matrix, key matrix, and value matrix are initialized first, then the attention of each head is calculated, and the inter-block attention is obtained by concatenating the attention values. The final intra-block attention value is obtained by averaging the intra-block and inter-block attention outputs from all modules. Inter-block attention value ; The representation difference module obtains two different attention values ​​from the dual attention comparison module, measures the similarity between the two different outputs using KL divergence, and calculates the loss and anomaly score.

7. The passive detection method for target trajectory maneuvering according to claim 6, characterized in that, The aforementioned anomaly score is used to determine whether the data is abnormal. An anomaly score is calculated for each set of time series data. The anomaly score is compared with the anomaly score threshold obtained during the training phase. If the anomaly score is greater than the threshold, the data is considered abnormal, that is, the target has performed a maneuver. The formula for calculating the anomaly score is as follows: ,in Calculate the KL divergence.

8. A passive target trajectory maneuver detection system based on dual-attention contrastive learning, characterized in that, The system is configured to perform the steps of the passive detection method for target orbital maneuvers as described in any one of claims 1-7, the system comprising: The model building unit is used to establish a spacecraft motion dynamics model and an observation star line-of-sight angle measurement model for the target star based on the geocentric inertial coordinate system and the LVLH coordinate system. The filtering calculation unit is used to perform square root commensurate Kalman filtering, taking right ascension and declination angle measurement data as input and outputting measurement innovation and normalized innovation sequences; The sample database unit is used to build and store a normal sample database before maneuvering based on normalized information, and supports the generation and indexing of samples under multiple operating conditions and multiple tracks. The network model unit is used to construct a DCdetector mobile detection model that includes a data recombination and encoding module, a dual attention contrast module, and a representation difference module. The training decision unit is used to train the DCdetector model using normal samples to obtain the anomaly score threshold, calculate the anomaly score in real time, determine whether the target has performed an orbital maneuver, and output the maneuver time point.

9. The passive detection system for target trajectory movement according to claim 8, characterized in that, The filtering calculation unit includes: The time update module is used for volume point calculation, state propagation, state prediction, and calculation of the square root of the prediction covariance. The measurement update module is used for measurement volume point calculation, observation prediction, innovation covariance calculation, Kalman gain calculation, state update, and covariance update. The innovation normalization module is used to perform normalization calculations on the measured innovation to obtain the normalized innovation.

10. The passive detection system for target trajectory maneuvering according to claim 8, characterized in that, The difference representation module calculates the anomaly score according to the following formula: ,in Let KL divergence be the KL divergence. Indicates the attention value within the block. Indicates the inter-block attention value; The training decision unit is configured as follows: DCdetector was trained using normal samples before the maneuver to determine the anomaly score threshold; Real-time anomaly judgment: when When the threshold is reached, it is determined that the target has undergone orbital maneuvering, and the current time is output as the maneuvering time point.

Images