Spatial adaptive filtering positioning method and device based on hybrid network model

Abstract

The invention relates to the technical field of space target tracking, in particular to a space adaptive filtering positioning method and device based on a hybrid network model, and the method comprises the steps: determining an initial relative orbit state based on angle measurement observation data; iteratively executing a loop: inputting to a pre-trained CNN-LSTM hybrid network model, and outputting an adaptive state transition matrix and an adaptive model error factor at a next moment; state prediction is carried out, and prior estimation and a prior error covariance matrix of the relative orbit state at the next moment are obtained; updating is carried out, and the optimal estimation of the relative orbit state at the next moment and a posterior error covariance matrix are determined; outputting the optimal estimation of each time as a space target positioning result of the next moment; and subsequent Kalman filtering circulation is carried out. According to the method, the model is introduced, the adaptive parameters of Kalman filtering are dynamically predicted according to the current motion state, and the state estimation precision and the algorithm robustness in a maneuvering scene are remarkably improved.

Description

Technical Field

[0001] This disclosure relates to the field of space target tracking technology, specifically to a space adaptive filtering localization method and apparatus based on a hybrid network model. Background Technology

[0002] In the field of precise orbit determination and tracking of space targets, space-based optical observation is a crucial means of achieving continuous monitoring of space targets. Traditional state estimation methods, such as Kalman filtering and its improved algorithms, heavily rely on accurate system dynamics models and noise statistics. However, in practical applications, the unknown maneuvering behavior of space targets can lead to model mismatch, and the complex space environment makes it difficult to accurately model noise statistics. This results in a significant decrease in positioning accuracy of traditional filtering methods during target maneuvers, and may even cause filter divergence.

[0003] To address model uncertainty, deep learning technology has been introduced for state prediction. However, it still suffers from inherent limitations such as a lack of ability to extract spatial features or difficulty in handling long-term temporal dependencies, making it difficult to meet the requirements of high-precision positioning for model adaptability and robustness. Therefore, designing a filtering method that can adaptively adjust filtering parameters while taking into account the spatiotemporal characteristics of motion state modeling in complex scenarios involving unknown maneuvers of spatial targets and uncertain model parameters, in order to significantly improve the tracking accuracy and algorithm robustness of maneuvering targets, has become a pressing technical challenge in this field. Summary of the Invention

[0004] In view of this, this disclosure provides a spatial adaptive filtering localization method and apparatus based on a hybrid network model to solve the problem of how to design a filtering method that can adaptively adjust filtering parameters and take into account the spatiotemporal characteristics of motion in complex scenarios where space targets undergo unknown maneuvers and model parameters are uncertain, so as to significantly improve the tracking accuracy and algorithm robustness of maneuvering targets.

[0005] This disclosure provides a spatial adaptive filtering localization method based on a hybrid network model, the method including: Based on angular measurement data from at least two observation platforms, the initial relative orbital state of a space target with respect to the observation platforms is determined through a joint positioning algorithm. Starting from the initial relative orbital state, the Kalman filter loop is executed iteratively, wherein the following operations are performed at each time step of the Kalman filter loop: The current relative orbit state is input into a pre-trained CNN-LSTM hybrid network model, which outputs an adaptive state transition matrix and an adaptive model error factor corresponding to the next time step based on the current relative orbit state. The adaptive state transition matrix is ​​used to predict the state, and the prior estimate of the relative orbit state and the prior error covariance matrix are obtained. Using the angle measurement observation data at the next moment, the prior estimate is updated to determine the optimal estimate of the relative orbital state and the posterior error covariance matrix at the next moment; The optimal estimate obtained after each update is output as the spatial target positioning result for the next moment; and the optimal estimate is used as the relative orbital state at the new current moment for subsequent Kalman filter loops. The pre-trained CNN-LSTM hybrid network model is configured to be trained on orbital dynamics data containing the maneuvering behavior of space targets, and is used to establish a nonlinear mapping relationship from relative orbital state to Kalman filter model parameters.

[0006] This disclosure also provides a spatial adaptive filtering localization device based on a hybrid network model, the device comprising: The initial state determination module is used to determine the initial relative orbital state of a space target with respect to the observation platforms based on angular measurement observation data from at least two observation platforms and through a joint positioning algorithm. The filter loop execution module is used to iteratively execute the Kalman filter loop, starting from the initial relative orbital state. The filtering loop execution module includes: The adaptive prediction unit is used to input the relative orbit state at the current moment into the pre-trained CNN-LSTM hybrid network model. The model outputs the adaptive state transition matrix and adaptive model error factor corresponding to the next moment based on the relative orbit state at the current moment. The adaptive state transition matrix is ​​used to predict the state and obtain the prior estimate of the relative orbit state at the next moment and the prior error covariance matrix. The state update unit is used to update the prior estimate using the angle measurement observation data at the next time step, and to determine the optimal estimate of the relative orbital state and the posterior error covariance matrix at the next time step. The result output and state transfer unit is used to output the optimal estimate obtained after each update as the spatial target positioning result for the next moment; and to use the optimal estimate as the new relative orbital state for the current moment, which is used for subsequent Kalman filter loops. The pre-trained CNN-LSTM hybrid network model is configured to be trained on orbital dynamics data containing the maneuvering behavior of space targets, and is used to establish a nonlinear mapping relationship from relative orbital state to Kalman filter model parameters.

[0007] This disclosure also provides an electronic device, including: a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the above-described spatial adaptive filtering localization method based on a hybrid network model.

[0008] This disclosure also provides a computer-readable storage medium storing computer instructions for causing a computer to implement the above-described spatial adaptive filtering localization method based on a hybrid network model.

[0009] This disclosure also provides a computer program product, including computer instructions for causing a computer to execute the above-described spatial adaptive filtering localization method based on a hybrid network model.

[0010] The spatial adaptive filtering localization method and apparatus based on a hybrid network model in the above embodiments of this disclosure effectively overcomes the model mismatch problem caused by target maneuvering in traditional filtering algorithms by introducing a pre-trained CNN-LSTM hybrid network model and dynamically predicting the adaptive parameters of Kalman filtering according to the current motion state. This significantly improves the state estimation accuracy and algorithm robustness in maneuvering scenarios.

[0011] Furthermore, by deeply integrating the deep learning model with the Kalman filter framework, a complete closed loop was constructed, from state feature extraction and adaptive prediction of model parameters to state estimation update. This enabled intelligent perception and compensation of nonlinear maneuvering behavior, thereby maintaining a continuous and stable high-precision positioning capability in complex spatial environments. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the specific embodiments or related technologies of this disclosure, the accompanying drawings used in the description of the specific embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the spatial adaptive filtering localization method based on a hybrid network model provided in this embodiment of the disclosure; Figure 2 This is a schematic diagram of the CNN network model structure of the spatial adaptive filtering localization method based on a hybrid network model provided in this embodiment of the disclosure; Figure 3 This is a schematic diagram of the LSTM network model structure of the spatial adaptive filtering localization method based on a hybrid network model provided in this embodiment of the disclosure; Figure 4 This is a schematic diagram of the training process of the CNN-LSTM hybrid network model for the spatial adaptive filtering localization method based on the hybrid network model provided in this embodiment of the disclosure; Figure 5This is a schematic diagram of the localization result of the spatial adaptive filtering localization method based on a hybrid network model provided in this embodiment of the present disclosure; Figure 6 This is a schematic diagram of the spatial adaptive filtering positioning device based on a hybrid network model provided in an embodiment of this disclosure; Figure 7 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0015] With the development of emerging missions such as on-orbit servicing, space debris removal, and space traffic management, extremely high demands are placed on the real-time, accurate tracking and orbit determination capabilities of non-cooperative space maneuvering targets. To achieve continuous monitoring of space targets, related technologies often utilize multiple observation satellites equipped with optical cameras to acquire angular information such as the azimuth and elevation angles of the space targets, and combine this information with state estimation filtering algorithms to ultimately determine their precise orbits.

[0016] In practical applications, accurate trajectory determination is highly dependent on the performance of the state estimation filter, i.e., how to accurately predict and update the target's trajectory state using noisy angle measurement data. Currently, the state estimation methods commonly used in related technologies are mainly based on the Kalman filter framework and its improved algorithms. Although these methods have optimal estimation characteristics in linear Gaussian systems, they often have the following limitations when dealing with nonlinear, non-Gaussian problems such as maneuvering space targets: 1. Traditional Kalman filtering algorithms heavily rely on accurate system dynamics models. When a space target undergoes unknown maneuvers, a significant mismatch occurs between the pre-set linear or near-linear model (such as the CW equation) and the actual nonlinear maneuvering behavior. This model error leads to inaccurate state predictions, and the fixed process noise covariance cannot effectively characterize the uncertainties brought about by the maneuver, ultimately causing a decrease in filtering accuracy or even divergence.

[0017] 2. To address the target maneuvering problem, nonlinear filtering methods such as extended Kalman filtering and unscented Kalman filtering have been proposed. However, these methods still require accurate model structures and noise statistical priors. In complex space environments, target maneuvering patterns are highly variable, and the dynamic models exhibit significant uncertainties, making it difficult for these methods to adaptively adjust model parameters, thus limiting their tracking performance in continuous maneuvering scenarios.

[0018] 3. In recent years, deep learning models such as Long Short-Term Memory (LSTM) networks have been introduced for state prediction, enabling them to learn maneuver patterns from historical data. However, LSTM excels primarily at capturing temporal dependencies and is insufficient in extracting spatial structural features from state vectors; while Convolutional Neural Networks (CNNs), although effective at extracting spatial features, struggle to model the dynamic evolution of long-term time series. Single network structures have inherent limitations in characterizing the spatiotemporal complexity of maneuvering behavior of spatial targets, failing to fully utilize the complete features of state information.

[0019] To address the aforementioned problems, various embodiments of this disclosure provide a spatial adaptive filtering localization method based on a hybrid network model. The method includes: determining the initial relative orbital state of a space target relative to the observation platforms using joint localization algorithms based on angular measurement observation data from at least two observation platforms; iteratively executing a Kalman filter loop starting from the initial relative orbital state, wherein at each time step of the Kalman filter loop, the following operations are performed: inputting the current relative orbital state into a pre-trained CNN-LSTM hybrid network model, which outputs an adaptive state transition matrix and an adaptive model error factor corresponding to the next time step based on the current relative orbital state; and using the adaptive state transition matrix to... The system predicts the relative orbital state to obtain a prior estimate and a prior error covariance matrix for the next time step. Using angle measurement observation data from the next time step, the prior estimate is updated to determine the optimal estimate and posterior error covariance matrix for the relative orbital state at the next time step. The optimal estimate obtained after each update is output as the spatial target localization result for the next time step. This optimal estimate is then used as the new relative orbital state for subsequent Kalman filter loops. The pre-trained CNN-LSTM hybrid network model is configured to be trained on orbital dynamics data containing spatial target maneuvering behavior, and is used to establish a nonlinear mapping relationship from the relative orbital state to the Kalman filter model parameters.

[0020] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a spatial adaptive filtering localization method based on a hybrid network model provided in this disclosure. The method may include the following steps: Step S101: Based on the angle measurement observation data from at least two observation platforms, determine the initial relative orbital state of the space target with respect to the observation platforms using a joint positioning algorithm.

[0021] In this embodiment, the implementation of this scheme is based on a collaborative observation network consisting of at least two observation platforms. These observation platforms are distributed in different spatial locations, and under unified temporal control, at each discrete observation time (i.e. Simultaneous detection of the same space target to obtain its azimuth angle ( ) and pitch angle ( Angle measurement observation data, such as angle measurement data, are used to form an observation dataset over a time series.

[0022] This step aims to utilize the initial time (such as the first three time points) Using observation data from [the observation platform], and through a specific joint positioning algorithm, the initial relative orbital state (i.e., the initial position and velocity) of the space target relative to a designated observation platform in the Vertical Velocity Location Heading (VVLH) coordinate system is calculated. This initial state will serve as the accurate starting point for subsequent Kalman filter loops.

[0023] In this context, an observation platform can refer to a space or ground-based device that carries observation payloads for detecting space targets and can operate in a specific orbit or location. For example, an observation platform can include, but is not limited to, space detection satellites and spacecraft in orbit.

[0024] Space targets can refer to natural or man-made objects located in the space environment that are of interest to and tracked by observation platforms. For example, space targets can include, but are not limited to: artificial satellites that are operating normally in orbit, defunct spacecraft, mission waste, rocket final stages, and space debris.

[0025] Angle measurement data can be angular information about the direction of a space target obtained by the observation platform. For example, angle measurement data can include azimuth and elevation angles.

[0026] Preferably, at least one of the observation platforms is designated as a reference observation platform, serving as the origin of the reference frame throughout the positioning process. The selection of the reference observation platform can be based on factors such as its orbital stability, payload accuracy, or communication capabilities.

[0027] In a specific example, an observation platform is constructed, each equipped with an observation camera. and a space target The "2-to-1" observation scenario, assuming the initial time The sampling time was 8:00:00 AM Beijing time on September 16, 2025. The total sampling time was 48 hours, and the sampling step size was 1 second, resulting in 172,800 sets of angle measurement data. Step S101 mainly uses the data from the first few moments to determine the initial state.

[0028] Furthermore, a joint positioning algorithm can refer to a set of algorithms that use angle measurement observation data to calculate the three-dimensional position and velocity of a space target through geometric relationships. The initial relative orbital state can refer to the position and velocity vector of the target relative to the reference platform in the Vertical Velocity Location Heading (VVLH) coordinate system at a specified moment (such as the intermediate moment).

[0029] Step S102: Starting from the initial relative orbital state, iteratively execute the Kalman filter loop.

[0030] In this embodiment, this step constitutes the core processing flow of this disclosure, which aims to initiate a continuously recursive "prediction-update" loop process based on the initial relative orbital state provided in step S101, thereby achieving continuous, adaptive, and high-precision estimation of the relative orbital state of the space target.

[0031] The Kalman filter loop is a recursive optimal estimation algorithm framework. The loop starts from the first formal moment corresponding to the initial relative orbital state and performs the following operations sequentially at each discrete time step: Step S102a: Input the relative orbital state at the current moment into the pre-trained CNN-LSTM hybrid network model. The model outputs the adaptive state transition matrix and adaptive model error factor corresponding to the next moment based on the relative orbital state at the current moment. Use the adaptive state transition matrix to predict the state and obtain the prior estimate of the relative orbital state at the next moment and the prior error covariance matrix.

[0032] In this embodiment, this step realizes the intelligent prediction link in the filtering process.

[0033] It is understandable that the relative orbital state at the current moment... It is a state vector, specifically a physical quantity used to fully describe the instantaneous motion state of a spatial target in the VVLH coordinate system. Preferably, From position vector and velocity vector Together they constitute.

[0034] Current moment relative orbital state The input is fed into a pre-trained CNN-LSTM hybrid network model, which can dynamically predict the next time step based on the complex nonlinear mapping relationships learned during its pre-training. Adaptive state transition matrix Error factor of adaptive model .

[0035] Among them, the adaptive state transition matrix It can describe the relative orbital state from the current moment. To the next moment The dynamic evolution law; adaptive model error factor It can be used to quantify from the current moment. To the next moment Uncertainty in the prediction process.

[0036] Furthermore, using the model parameters obtained from the above predictions, the following two core calculations are performed: Using the adaptive state transition matrix State prediction is performed to obtain a priori estimate of the relative orbital state at the next moment, and its calculation expression is as follows: (1) in, Indicates the current time The relative orbital state, This represents the next moment predicted by the model. Prior estimates of the relative orbital state.

[0037] Based on the adaptive state transition matrix, the error covariance matrix at the current time step, and the process noise covariance matrix constructed from the adaptive model error factors, the prior error covariance matrix at the next time step is calculated. The calculation is shown in the following formula (2): (2) in, Indicates the error factor of the adaptive model The process noise covariance matrix is ​​the coefficient. It is an identity matrix.

[0038] Step S102b: Using the angle measurement observation data at the next moment, update the prior estimate to determine the optimal estimate of the relative orbital state and the posterior error covariance matrix at the next moment.

[0039] In this embodiment, this step implements the measurement update stage of the filtering process.

[0040] Specifically, the update process is implemented through the update phase of the Kalman filter, which mainly includes the following sub-steps: Step b1, Kalman gain calculation.

[0041] Here, the Kalman gain matrix is ​​calculated based on the prior error covariance matrix, the observation matrix, and the observation noise covariance matrix. The gain matrix determines the weights of the predicted and observed values ​​during state updates.

[0042] The Kalman gain is calculated as shown in formula (3): (3) in, The calculated Kalman gain matrix; Let be the prior error covariance matrix for the next time step; The observation matrix is ​​used to map the state space (i.e., velocity and position in the VVLH coordinate system) to the observation space (i.e., azimuth and pitch angles). The observation noise covariance matrix is ​​used to characterize the uncertainty of the angle sensor's measurement values ​​and is usually preset according to the sensor's performance.

[0043] Step b2, state vector update.

[0044] Here, Kalman gain is used to fuse the prior state estimate with the new observation data to calculate the updated state vector, i.e., the optimal estimate. The calculation is shown in the following formula (4): (4) in, The observation residual is the difference between the predicted angle measurement data and the actual angle measurement observation data.

[0045] Step b3: Update the error covariance matrix.

[0046] Here, the posterior error covariance matrix corresponding to the state estimate is updated synchronously to reflect the reduction in state estimate uncertainty after fusing observation information. The calculation is shown in the following formula (5): (5) Step S103: The optimal estimate obtained after each update is output as the spatial target positioning result for the next moment; and the optimal estimate is used as the new relative orbital state for the current moment, which is then used in subsequent Kalman filter loops.

[0047] In this embodiment, this step realizes the output of the positioning result and the recursive transfer of the status.

[0048] Specifically, the optimal estimate (i.e., the posterior state estimate) obtained in step S102b is updated. As The final spatial target positioning result at time t is output. This result represents the most reliable estimate of the spatial target's relative orbital state at that time after optimal filtering.

[0049] Simultaneously, this optimal estimate is used as the new relative orbital state at the current moment and assigned to... The process is then recursively pushed to the next time step, restarting a new round of Kalman filtering from step S102a. This cycle repeats until all data from all observation times has been processed, thereby achieving continuous, smooth, and high-precision tracking and positioning of the space target trajectory.

[0050] The pre-trained CNN-LSTM hybrid network model is configured to be trained on orbital dynamics data containing the maneuvering behavior of space targets, and is used to establish a nonlinear mapping relationship from relative orbital state to Kalman filter model parameters.

[0051] In this embodiment, the pre-trained CNN-LSTM hybrid network model is configured to be trained based on orbital dynamics data containing the maneuvering behavior of space targets, and is used to establish a nonlinear mapping relationship from relative orbital state to Kalman filter model parameters.

[0052] In this embodiment, the pre-trained CNN-LSTM hybrid network model is the core component for realizing the adaptive filtering capability of this method. This model is built through offline training, and its training process is specifically designed to learn the motion characteristics of space targets in complex dynamic environments, especially their maneuvering behavior.

[0053] Specifically, the training process of a CNN-LSTM hybrid network model can include the following key steps: First, a training dataset containing rich maneuver scenarios is constructed. Based on the initial orbital states of the observation platform and space target, long-term predicted trajectories are generated using a high-precision orbital dynamics model (considering various perturbations such as Earth's non-spherical gravity, solar radiation pressure, atmospheric drag, and third-body gravity) and a numerical integrator (such as the Adams-Cowell integrator).

[0054] During trajectory generation, random pulses conforming to a specific statistical distribution (such as Gaussian distribution) are added to the relative velocity components at specific times or continuously to simulate various maneuvering behaviors that space targets may undergo in actual missions, thereby ensuring that the training data can fully cover the complexity and uncertainty of target motion.

[0055] Secondly, training samples are extracted from the generated trajectory data. For each time step, the relative position and relative velocity of the space target with respect to the benchmark observation platform in the VVLH coordinate system are calculated, forming the input features of the model. Simultaneously, the precisely calculated state transition matrix and model error factor for that time period are recorded as the target output for model learning. In this way, a large number of paired data samples from "relative orbital state" to "optimal filter parameters" are constructed.

[0056] Then, the CNN-LSTM hybrid network model is trained under supervision. This network model typically adopts a cascaded structure, with a CNN module at the front end responsible for extracting local spatial features and patterns from the input state sequence; and an LSTM module at the back end responsible for capturing the dynamic evolution and long-term dependencies of these features over time. Finally, a fully connected layer maps the high-level temporal features output by the LSTM to the target output space, namely the parameters of the adaptive state transition matrix and the adaptive model error factor.

[0057] Furthermore, during training, loss functions such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination (DC) can be used. The network weights are continuously adjusted through backpropagation and a gradient descent optimizer (such as Adam), enabling the model to minimize the difference between its predicted filter parameters and the true values. The trained model then possesses the ability to intelligently and adaptively predict the Kalman filter model parameters best suited to the current motion state based on the input relative orbital state sequence.

[0058] The MSE is shown in the following formula (6): (6) The RMSE is shown in the following formula (7): (7) DC is shown in the following formula (8): (8) in, This represents the true model error factor or state transition matrix. This represents the model error factor or state transition matrix used for prediction. This represents the true model error factor or the average value of the state transition matrix over all time steps.

[0059] Here, the closer the values ​​of MAE and RMSE are to 0, the smaller the prediction error of the model. The closer the value of DC is to 1.0, the stronger the model's ability to explain the data variance and the better its prediction performance.

[0060] As an example, please refer to Figure 2 , Figure 2 This is a schematic diagram of the CNN network model structure of the spatial adaptive filtering localization method based on a hybrid network model provided in this embodiment of the disclosure, wherein: The input layer of a CNN network model takes the relative position and velocity of a spatial object as input. After passing through a multi-layered feature extraction structure, convolutional layers detect local patterns in the input data through convolutional operations. Pooling layers, introduced after the convolutional layers, reduce data dimensionality and enhance feature robustness. After extracting rich spatial features through multiple convolutional and pooling units, these distributed features need to be integrated and mapped to the target output space. Fully connected layers act as classifiers, non-linearly combining the learned high-level features to prepare for the final output. Finally, the output layer outputs intermediate parameters—the refined time series—which are then fed into the next stage, an LSTM, for temporal modeling.

[0061] Please refer to further information. Figure 3 , Figure 3 This is a schematic diagram of the LSTM network model structure of the spatial adaptive filtering localization method based on a hybrid network model provided in this disclosure embodiment, wherein: At every current moment The LSTM network at the previous time Hidden state With memory units As an internal memory carrier, and related to the current moment Input feature vector They work together. The internal data processing flow involves three gating functions and one state update in sequence: 1. Forgot Gate: Receive and The sigmoid activation function generates a forgetting factor between 0 and 1 to determine the memory unit from the previous time step. The degree of retention of each component.

[0062] 2. Input Gate: Two computations are performed in parallel. First, the Sigmoid function determines the state components that need to be updated; second, the Tanh function generates new candidate state values. Both work together to determine how much new information will be stored in the cell state.

[0063] 3. Cell state update: The old state that has passed the forgetting gate will be updated. With the new candidate value after input gate regulation Add them together to generate the updated current cell state. This state contains filtered long-term memory information.

[0064] 4. Output Gate: First, the Sigmoid function is used to determine which parts of the cell state will be output based on the current input and the previous hidden state. Then, the current cell state is... The hidden state output is generated by scaling using the Tanh function and multiplying it with the result of the output gate. .

[0065] This hidden state output On one hand, it serves as the output of this unit for subsequent calculations (such as predicting the state transition matrix and model error factor); on the other hand, it is used in conjunction with the updated cell state. These processes are then passed to the next LSTM unit, forming a recursive loop for sequence processing. Through this series of ordered gating and state transition operations, the LSTM network can effectively capture long-term dynamic dependencies in the input feature sequence.

[0066] Furthermore, training with a CNN-LSTM hybrid network model can achieve convergence.

[0067] For example, the training process can be as follows: Figure 4 As shown, Figure 4 This diagram illustrates the training process of a CNN-LSTM hybrid network model based on a spatial adaptive filtering localization method using a hybrid network model, as provided in this embodiment. The horizontal axis represents the number of training epochs, where one training epoch indicates that the model has been trained once using the entire training set. The vertical axis represents the loss value; a smaller loss value indicates more accurate predictions. As the number of training epochs increases, the loss value curve shows a rapid decrease followed by a gradual stabilization, eventually converging to a lower value. This indicates that the model is learning effectively. It extracts features and patterns from the training data, adjusts its parameters to reduce its prediction error on known data, and finally, the model is sufficiently trained and converges. At this point, the model has found an optimal or near-optimal parameter state, the loss value no longer decreases significantly, and training can stop.

[0068] The spatial adaptive filtering localization method and apparatus based on a hybrid network model, as described in the above embodiments of this disclosure, effectively overcomes the model mismatch problem caused by target maneuvering in traditional filtering algorithms by introducing a pre-trained CNN-LSTM hybrid network model to dynamically predict the adaptive parameters of the Kalman filter based on the current motion state. This significantly improves the state estimation accuracy and algorithm robustness in maneuvering scenarios. By deeply integrating the deep learning model with the Kalman filtering framework, a complete closed loop is constructed from state feature extraction, adaptive prediction of model parameters to state estimation update, achieving intelligent perception and compensation for nonlinear maneuvering behavior, thereby maintaining a continuous and stable high-precision localization capability in complex spatial environments.

[0069] In one possible implementation of step S101 above, based on angular measurement data from at least two observation platforms, a joint positioning algorithm is used to determine the initial relative orbital state of the space target with respect to the observation platforms, including: Designate at least one of two observation platforms as the baseline observation platform; Based on angle measurement observation data at least three consecutive time points, the relative position of the space target with respect to the reference observation platform at each time point is determined using a dual-line least squares positioning algorithm. Based on the relative positions at least three consecutive moments, the relative velocity at the middle moment among the at least three consecutive moments is calculated using the central difference method; The relative position and relative velocity at the intermediate moment are used together as the initial relative orbital state.

[0070] In this embodiment, the implementation provides a specific and robust method for determining the initial relative orbital state. First, one of multiple observation platforms is explicitly designated as the reference observation platform to establish a unified reference coordinate system, providing a benchmark for the representation and calculation of all subsequent relative states.

[0071] The specific positioning process is as follows: using at least three consecutive moments (e.g.) Using the angular observation data, the target position is calculated using the dual-line least squares positioning algorithm.

[0072] For any given moment, the algorithm utilizes angular measurement data from at least two observation platforms to construct and solve an overdetermined system of equations based on line-of-sight geometry, thereby determining the three-dimensional relative position of the space target with respect to the reference observation platform in the VVLH coordinate system at that moment. Repeating this process yields the relative position sequence for each of the three consecutive moments.

[0073] After obtaining the relative position sequence, the central difference method can be used to calculate the intermediate time (e.g., The relative velocity of ( ). The central difference method uses the position information at different times to perform numerical differentiation.

[0074] Choosing the state at the intermediate moment as the initial value helps reduce the error caused by the endpoint effect and provides a more accurate and reliable starting point for the Kalman filter loop.

[0075] For example, the line-of-sight least squares localization algorithm is used to calculate... The three-dimensional relative positions of the space target with respect to the reference observation platform at these three moments are used to determine the initial moment using the central difference method. The relative velocity of the space target with respect to the observed star.

[0076] As a specific example, the specific calculation process of the dual-line least squares localization algorithm and the central difference method is as follows: First, for each time step (with...) (For example), the unit direction vector pointing from each observation platform to the space target is calculated in the ECI coordinate system based on angle measurement data: For the observation platform Based on its angular observation data (azimuth angle) Pitch angle ), its unit direction vector As shown in the following formula (9): (9) in, express The rotation matrix from the camera (CAM) coordinate system to the geocentric inertial coordinate system at any time.

[0077] Similarly, for observation platforms Based on its angular observation data (azimuth angle) Pitch angle ), its unit direction vector As shown in the following formula (10): (10) Observation platform The position in the ECI coordinate system is as follows:

[0078] Observation platform The position in the ECI coordinate system is as follows:

[0079] Assuming space target The position in the ECI coordinate system is an unknown vector. According to spatial geometric relationships, vectors ( - ) and direction vector Parallel, vector ( - ) and direction vector If they are parallel, then the cross product of parallel vectors is 0, corresponding to the following formulas (11) and (12): (11) (12) Expanding the cross product equations above by components yields a system of six linear equations, as shown in formulas (13) to (18) below: (13) (14) (15) (16) (17) (18) Rewrite the above formula as The unknowns are represented by the following formulas (19) to (24): (19) (20) (twenty one) (twenty two) (twenty three) (twenty four) Constructing a matrix sum vector ,available The matrix and vector are systematically constructed from the above linear equations (19) to (24), and the corresponding matrix and vector results are shown in the following formulas (25) to (26): (25) (26) The number of equations in this linear system (6) exceeds the number of unknowns (3), forming an overdetermined system of equations. The least squares method is used to solve the system, yielding the optimal estimate of the position of the space target in the ECI coordinate system, as shown in formula (27): (27) Subsequently, the absolute position in the ECI coordinate system was converted relative to the reference observation platform. The relative position in the VVLH coordinate system is shown in the following formula (28): (28) in, Let be the rotation matrix from the ECI coordinate system to the VVLH coordinate system at time t=0.

[0080] By repeating the above process, we can obtain the results sequentially. and Relative position of time and .

[0081] After obtaining the relative positions at three consecutive moments, the central difference method is then used to calculate... Speed ​​of time : (29) in, This represents the sampling time interval.

[0082] Finally, the initial time can be obtained. Space targets Relative reference observation star Position and velocity in VVLH coordinate system .

[0083] The spatial adaptive filtering positioning method and apparatus based on a hybrid network model, as described in the above embodiments of this disclosure, ensures the consistency of the spatial reference system throughout the positioning process by specifying a single reference observation platform, avoiding the complex transformations and errors caused by multiple reference systems. Employing a dual-line least squares positioning algorithm, it comprehensively utilizes the geometric information of multiple observation platforms, effectively overcoming the ambiguity of single-line positioning. Furthermore, by optimizing the solution of the overdetermined equations, observation noise is adjusted, resulting in a more stable and accurate initial position estimate than that obtained with a single platform or algorithm. Based on the position sequence at multiple time points, numerical differentiation is performed using the central difference method. Compared to simple forward or backward differencing, this method has second-order accuracy and can more accurately reflect the instantaneous motion trend of the target, thus obtaining a more reliable initial velocity value.

[0084] In one possible implementation of step S102 above, the pre-trained CNN-LSTM hybrid network model is obtained in the following way: Based on the initial orbital state of the observation platform and space target, a predicted trajectory containing simulated maneuvering behavior is generated through a dynamic model and a numerical integrator. Based on the predicted trajectory, the relative position and relative velocity of the space target with respect to the observation platform in the vertical velocity position heading coordinate system are calculated, and the state transition matrix and model error factor are calculated and recorded accordingly to construct the training dataset. The angle measurement observation data used to construct the training dataset was generated through the following simulation steps: Based on the predicted trajectory, the position of the space target in the geocentric inertial coordinate system at each moment is successively converted into the position in the vertical velocity and heading coordinate system, the body coordinate system, and the camera coordinate system relative to the observation platform. Based on the relative position in the camera coordinate system, the azimuth and elevation angle sequences are calculated, and random errors following a Gaussian distribution are added to the sequences. A CNN-LSTM hybrid network model is trained using a training dataset to establish a nonlinear mapping relationship from relative position and relative velocity to state transition matrix and model error factor, thus obtaining a pre-trained CNN-LSTM hybrid network model.

[0085] In this embodiment, the implementation aims to illustrate the complete process of obtaining a pre-trained CNN-LSTM hybrid network model, which can be completed offline, providing a fully trained intelligent model for the core adaptive filtering algorithm.

[0086] The acquisition of the model begins with the preparation of training data. Specifically, based on the initial orbital states of the observation platform and the space target, a high-precision orbital dynamics model incorporating various perturbations such as Earth's non-spherical gravity, solar radiation pressure, and third-body gravity is used. An Adams-Cowell numerical integrator is then employed for orbit extrapolation to generate long-term predicted trajectories. To simulate the uncertain maneuvers of targets in real space environments, random pulses conforming to a Gaussian distribution (e.g., with a standard deviation of 1 m / s) are injected into the relative velocity components at specific times or continuously during trajectory generation, thereby constructing trajectory data covering various typical and complex maneuver scenarios.

[0087] Furthermore, based on the generated predicted trajectory, the relative position and relative velocity of the space target with respect to the specified observation platform in the VVLH coordinate system are calculated. During this process, the precise state transition matrix derived from the orbital dynamics model and the model error factor used to quantify model uncertainty are recorded simultaneously. Relative position and relative velocity serve as input features, while the state transition matrix and model error factor serve as labels for supervised learning, together forming the basic samples of the training dataset.

[0088] To enhance the completeness and authenticity of the training data, the angle measurement observation data is generated through a high-fidelity simulation process.

[0089] Here, this step transforms the position of the space target in the geocentric inertial coordinate system at each moment into its position relative to the observation platform in the VVLH coordinate system, its body coordinate system, and its camera coordinate system through a rigorous coordinate transformation chain. Based on the relative position vector in the camera coordinate system, the azimuth and elevation angle sequences are calculated through geometric relationships. To simulate the measurement errors of actual optical sensors, a Gaussian distributed random error with zero mean and a preset standard deviation is added to the generated ideal angle measurement sequence.

[0090] Wherein, at time t, the space target and observed stars The position and velocity in the geocentric inertial coordinate system are respectively and . relatively Position in the VVLH coordinate system As shown in the following formula (30): (30) Formula (30) is used to represent space targets. Relative observation platform Position in the VVLH coordinate system ; This is the position rotation matrix from the ECI coordinate system to the VVLH coordinate system; Indicates the space target at the observation time The absolute position vector in the ECI coordinate system; Indicates the first Each observation platform at the observation time The absolute position vector in the ECI coordinate system.

[0091] Relative position from VVLH to the BODY coordinate system The conversion is shown in the following formula (31): (31) Formula (31) is used to represent space targets. Relative observation platform Position in the BODY coordinate system ; , and These are rotation matrices about the x, y, and z axes, respectively; , and These represent the installation rotation angles of the BODY coordinate system of the observation platform relative to the VVLH coordinate system around the x, y, and z axes, respectively.

[0092] Relative position of body to CAM coordinate system The conversion is shown in the following formula (32): (32) Formula (32) is used to represent space targets. Relative observation platform Position in CAM coordinate system ; , and These are rotation matrices about the x, y, and z axes, respectively; , , These represent the installation rotation angles of the CAM coordinate system relative to the BODY coordinate system of the observation platform around the x-axis, y-axis, and z-axis, respectively.

[0093] The transformation from the CAM coordinate system to angle measurement is shown in formulas (33) and (34), and the azimuth angle is obtained by solving. and pitch angle : (33) (34) in, Represents the components of the vector along the x-axis in the CAM coordinate system; This represents the y-component of the vector in the CAM coordinate system. This represents the component of the vector along the z-axis in the CAM coordinate system.

[0094] Here, the azimuth angles obtained by formulas (33) and (34) and pitch angle These are ideal angle measurement data.

[0095] Furthermore, in order to simulate the measurement noise present in the sensor in the real observation environment, random errors that conform to specific statistical characteristics can be added to the ideal angle measurement data obtained by formulas (33) and (34).

[0096] For example, random error can have a mean of zero and a standard deviation of . Gaussian white noise. Preferably, the standard deviation is... A value of 0.1 milliradians (mrad) can be used.

[0097] After obtaining the training dataset, perform necessary preprocessing, such as data normalization, and divide the dataset into training and validation sets.

[0098] Preferably, the dataset is divided into two parts: the first 80% of the data is used as the training set, mainly for optimizing the CNN-LSTM hybrid network model; the last 20% of the data constitutes the validation set, mainly for evaluating the hybrid network model.

[0099] This dataset was used to supervise the training of a CNN-LSTM hybrid network model. During training, the CNN module was responsible for extracting local spatial features from the input state sequence, while the LSTM module was responsible for learning the long-term dependencies of these features over time. The network parameters were optimized using backpropagation to minimize the differences (such as mean squared error) between the model's predicted state transition matrix, the model error factor, and the true labels. The trained model then possesses the ability to dynamically and accurately predict the key parameters required for Kalman filtering based on the relative orbital state sequence of the input, thus yielding a pre-trained CNN-LSTM hybrid network model.

[0100] The spatial adaptive filtering positioning method and apparatus based on a hybrid network model, as described in the above embodiments of this disclosure, introduces a high-precision orbital dynamics model containing various precise perturbations and actively injects random pulse maneuvers that conform to statistical laws. The generated training trajectory can highly simulate the complex motion behavior of space targets in real environments, especially various uncertain maneuvers. This ensures that the training dataset can comprehensively cover potential maneuver scenarios, fundamentally guaranteeing the generalization ability and robustness of the subsequently trained model when facing real situations. Through a rigorous coordinate transformation chain and angle measurement solution model consistent with physical reality, and by artificially adding Gaussian white noise that conforms to the performance of actual sensors, angle measurement data that is statistically indistinguishable from real observations is generated. This high-fidelity simulation process allows for the acquisition of massive, precisely labeled training samples without the need for real on-orbit data, greatly reducing data acquisition costs and project risks. The pre-trained model obtained through offline training is a prerequisite for realizing online high-precision adaptive filtering. It enables the Kalman filter to dynamically sense and adapt to the target's maneuver changes, adjusting its internal model in real time, thereby significantly overcoming the inherent defects of traditional filtering algorithms that lead to decreased estimation accuracy or even divergence due to model mismatch during target maneuvers.

[0101] In one possible implementation of step S102a above, the relative orbital state at the current moment is input to a pre-trained CNN-LSTM hybrid network model, which outputs an adaptive state transition matrix and an adaptive model error factor corresponding to the next moment based on the relative orbital state at the current moment, including: Construct a relative orbital state sequence that includes multiple consecutive historical moments, with the current moment as the endpoint; The relative orbital state sequence is input into a pre-trained CNN-LSTM hybrid network model. The pre-trained CNN-LSTM hybrid network model extracts spatial features and models temporal dependencies of the relative orbital state sequence, and outputs an adaptive state transition matrix and an adaptive model error factor corresponding to the next time step.

[0102] In this embodiment, the specific input format and internal processing logic of the CNN-LSTM hybrid network model are clearly defined. That is, the model does not make predictions based solely on isolated states at a single moment, but rather uses a historical state sequence to capture the dynamic evolution trend of motion states.

[0103] Specifically, construct a system based on the current moment. The endpoint is defined as including multiple consecutive moments preceding it (e.g., ...). The relative orbital state sequence of the target. The state at each moment in the sequence contains the relative position and relative velocity components of the target in the VVLH coordinate system. Therefore, the entire sequence forms a two-dimensional data matrix with dimensions of (sequence length × state dimension).

[0104] After inputting this state sequence into the pre-trained CNN-LSTM hybrid network model, the model performs the following calculations in sequence: In step a1, the CNN module performs convolution and pooling operations on the "time-state" grid. Its core function is to detect and extract the implicit local spatial patterns in the sequence, such as the coupling relationship and cooperative change law between different state components (e.g., x-direction position and y-direction velocity) within a short time span.

[0105] In step b1, the high-level feature sequence rich in spatial information output by the CNN module is fed into the LSTM module. LSTM, with its gating mechanism, recursively processes this feature sequence to learn the long-term dynamic dependencies of state features over time. It can identify long-term trends, periodic or complex maneuvering patterns, thereby understanding "how past state evolution will affect future state transitions."

[0106] Step c1: The hidden state output by the LSTM module at the end of the sequence contains summary information about the dynamic patterns of the entire historical sequence. This information is nonlinearly transformed through a fully connected output layer and mapped to the next time step. The required adaptive state transition matrix and adaptive model error factor.

[0107] Through this cascaded structure of spatial feature extraction and time-dependent modeling, the hybrid network model can intelligently predict the optimal filtering parameters that best match the current and recent motion state based on a historical trajectory.

[0108] The spatial adaptive filtering localization method and apparatus based on a hybrid network model, as described in the above embodiments of this disclosure, utilizes historical state sequences as input. The model can perceive motion trends, acceleration changes, and potential maneuvering intentions from the data, overcoming the limitations of relying solely on a single-moment state and providing rich contextual information for accurate prediction. By leveraging CNNs to capture complex local spatial correlations between state components and combining them with LSTMs to learn long-term temporal dependencies in state evolution, this hybrid architecture fully utilizes the strengths of both networks, enabling a more comprehensive understanding and representation of nonlinear and non-stationary maneuvering patterns.

[0109] In one possible implementation of the above embodiments, when constructing the training dataset, the relative position, relative velocity, state transition matrix, and model error factor are preprocessed using normalization; wherein, the normalization preprocessing is based on the following formula:

[0110] in, For variables to be normalized, and These are the maximum and minimum values ​​of the variable in the training dataset, respectively.

[0111] In this embodiment, a data normalization preprocessing step is introduced to optimize the training process and final performance of the CNN-LSTM hybrid network model. Since different physical quantities (such as position and velocity) and different components of the same physical quantity in the training dataset typically have different dimensions and numerical ranges, this difference can cause the gradient descent direction to deviate from the optimal path during model training, thus affecting training efficiency and model convergence.

[0112] In practice, each feature variable in the training dataset (including the components of the relative position vector, the components of the relative velocity vector, the elements of the state transition matrix, and the model error factor) is normalized separately. First, the maximum value of each variable is calculated across the entire training set. and minimum value Then, according to the above formula, the original values ​​of each variable are... A linear transformation is performed to obtain its normalized value within the interval [0,1]. .

[0113] in, .

[0114] Through the spatial adaptive filtering localization method and apparatus based on a hybrid network model as described in the above embodiments of this disclosure, the preprocessing constrains all input features and output labels to a similar numerical range. This eliminates the impact of dimensional differences on model training, enabling the updating of each connection weight in the network to be based on relatively fair gradient information, thereby significantly accelerating the convergence speed of the model training process. Furthermore, the normalization process improves the numerical stability of the model optimization process, effectively avoiding problems such as gradient vanishing, gradient exploding, or ill-conditioned weight matrices caused by excessively large or small feature values, ensuring that the model can stably find a better solution.

[0115] In one possible implementation of the above embodiment, the prior error covariance matrix is ​​obtained based on the following steps: Based on the adaptive state transition matrix, the error covariance matrix at the current time step, and the adaptive model error factor, the prior error covariance matrix at the next time step is calculated; whereby the adaptive model error factor is used to construct the process noise covariance matrix.

[0116] In this embodiment, this implementation constitutes a key step in the Kalman filter prediction stage for recursively predicting the uncertainty of state estimation. This step is performed synchronously with state prediction, ensuring that the filter can dynamically quantify the reliability of its prediction results.

[0117] Specifically, at the current moment Given the following parameters: the adaptive state transition matrix predicted by the CNN-LSTM hybrid network model. Current moment posterior error covariance matrix And the adaptive model error factor predicted by the CNN-LSTM hybrid network model. .

[0118] The next time step is calculated using formula (2). Prior error covariance matrix ,Right now:

[0119] In this calculation, the adaptive model error factor Its core function is to construct the process noise covariance matrix. In traditional Kalman filtering, the process noise covariance matrix... Typically, this is a fixed value that needs to be preset. However, in this method, it is dynamically constructed. This design allows the intensity of process noise to be matched in real time with the uncertainty of the prediction model: when the target is highly maneuverable and the model prediction error is large, the network prediction... An increase in the value indicates that a greater degree of uncertainty is assigned to the prediction result, and vice versa.

[0120] This step ensures that the filter's assessment of its own prediction accuracy is accurate and adaptive. The calculated prior error covariance matrix... This will be directly used for Kalman gain calculation in subsequent update stages, thus affecting the weight of the observed data in state updates.

[0121] The spatial adaptive filtering localization method and apparatus based on a hybrid network model, as described in the above embodiments of this disclosure, dynamically constructs the process noise covariance by introducing an adaptive model error factor based on network prediction. This allows the filter's estimation of prediction uncertainty to respond in real time to changes in target maneuverability and model error. Compared to a fixed noise model, the quantification of uncertainty is more accurate and realistic. This step ensures that the corresponding confidence level is predicted synchronously and adaptively during state prediction, providing a crucial basis for the filter to make optimal data fusion decisions during the update phase. This effectively prevents filter divergence and ensures long-term operational stability.

[0122] In one possible implementation of the above embodiments, the relative orbital state sequence is a sliding window sequence containing multiple consecutive historical moments, and the feature dimension of the sliding window sequence corresponds to the relative position and relative velocity components in the vertical velocity position heading coordinate system.

[0123] In this embodiment, the implementation method clarifies the specific data structure and organization of the historical state sequence, providing a standardized and regularized input for the CNN-LSTM hybrid network model.

[0124] Specifically, the sliding window sequence can be a fixed-length data buffer that slides along a time axis. At each time step... This window captures from continuous to t A historical moment (e.g.) The relative orbital states of ) constitute a space of size A two-dimensional data matrix.

[0125] in, The sliding window size, or sequence length, determines the duration of historical information that the model relies on for prediction. This is the feature dimension, specifically corresponding to the number of state components in the VVLH coordinate system. Preferably, These correspond to three relative position components. and three relative velocity components .

[0126] Therefore, at any time The data input to the CNN-LSTM hybrid network model is a dimensional... A tensor. This tensor has a time dimension. The dynamic evolution of the state is recorded on the top, and the motion state at each moment is fully described on the feature dimension (6).

[0127] By employing a sliding window mechanism, the system enables the processing of continuous data streams, i.e., when a new time frame is obtained... When in a certain state, the window updates automatically, discarding the earliest time. The state, and incorporate the new The state at any given moment, thus always keeping the window display the latest continuous state. Data from a specific time point. This design allows the model to make predictions based on a fixed and controllable recent historical context, ensuring both the continuity of time-series information and controlling the complexity of the model's input.

[0128] As a preferred embodiment, the hyperparameter values ​​of the CNN-LSTM hybrid network model can be as shown in Table 1 below: Table 1. Hyperparameter values ​​of the CNN-LSTM hybrid network model

[0129]

[0130] The hyperparameters listed in Table 1 can be combined to form a CNN-LSTM hybrid network architecture with moderate depth, sufficient capacity, and stable training. Its design fully considers the characteristics of space target orbital state data (i.e., temporal, nonlinear, and maneuverable), aiming to optimally balance the model's fitting ability, generalization ability, and computational efficiency. This ensures its reliable integration into the adaptive Kalman filter framework, ultimately achieving high-precision space target localization.

[0131] The spatial adaptive filtering localization method and apparatus based on a hybrid network model, as described in the above embodiments of this disclosure, provides a well-structured and information-rich input for the deep learning model by constructing a standardized sliding window sequence, thus laying a data foundation for the model to effectively extract spatiotemporal features. The sliding window design ensures that the model always focuses on the most recent motion state, making it more sensitive to the latest maneuvering behavior of the target and enabling it to quickly respond to sudden changes in motion state, thereby improving the dynamic response speed and tracking accuracy of the adaptive filter.

[0132] In one possible implementation of the above embodiments, the CNN-LSTM hybrid network model includes a convolutional neural network module and a long short-term memory network module connected in sequence; The convolutional neural network module is configured to perform convolution and pooling operations on the input state sequence to extract local spatial features from the relative orbital state sequence. The Long Short-Term Memory (LSTM) network module is configured to receive local spatial features output by the convolutional neural network module and model the long-term dependencies of local spatial features in the time dimension through a gating mechanism to output an adaptive state transition matrix and an adaptive model error factor.

[0133] In this embodiment, the implementation method clarifies the core architecture of the hybrid network model, namely, adopting a serial hybrid deep learning model, and realizing deep intelligent analysis of relative orbital state sequences through clear functional division and collaboration.

[0134] Specifically, the CNN module, acting as a front-end feature extractor, typically contains multiple alternating stacked convolutional and pooling layers. When a relative orbital state sequence is input, the convolutional layers perform sliding convolution operations on a two-dimensional "time-state" grid of the sequence. Activation functions (such as ReLU) introduce non-linearity, thereby detecting and generating local spatial feature maps that characterize the coupling relationships between state components within a short timeframe. The subsequent pooling layers (such as max pooling) downsample this feature map, preserving the most salient features while reducing data dimensionality and expanding the receptive field of subsequent layers, thus enhancing the robustness of the features.

[0135] The LSTM module acts as a backend temporal modeler, receiving the high-level feature sequence after it has been processed and flattened by the CNN module. LSTM, with its sophisticated internal gating mechanism (including forget gate, input gate, and output gate), recursively processes this feature sequence. This process enables it to learn the evolution of features over time, capturing dynamic patterns and long-term dependencies across long time windows, thereby understanding the potential impact of historical maneuvers on future state transitions.

[0136] Ultimately, the hidden state output by the LSTM module at the end of the sequence carries a summary of the dynamic characteristics of the entire input sequence. This state undergoes nonlinear transformation and dimensionality mapping through one or more fully connected layers, regressing to a fixed-size output vector, and is then reconstructed into an adaptive state transition matrix and an adaptive model error factor.

[0137] The spatial adaptive filtering localization method and apparatus based on a hybrid network model, as described in the above embodiments of this disclosure, achieves hierarchical and collaborative intelligent analysis of state sequences from local spatial patterns to long-term temporal dynamics by constructing a hybrid model that cascades CNN and LSTM. This avoids the difficulty of a single model processing two types of features simultaneously, enabling deeper and more accurate feature learning. The CNN module, as a powerful feature extractor, provides the LSTM module with higher-level and more informative input than the original data. This allows the LSTM to focus more efficiently and intently on learning the long-term dynamic dependencies between high-level features, eliminating the need for the arduous search for temporal patterns in the original noisy data, resulting in a synergistic enhancement effect.

[0138] In one embodiment, Figure 5 This is a schematic diagram of the localization result of the spatial adaptive filtering localization method based on a hybrid network model provided in this disclosure embodiment, as shown below. Figure 5 As shown, the positioning results The components in all three directions start from a relatively high initial value in a short period of time, then drop rapidly and sharply, and finally remain stable within a relatively low numerical range (e.g., within 10 meters) with slight fluctuations, indicating that the filter has completed convergence and entered a stable tracking state.

[0139] In one embodiment, a spatial adaptive filtering localization device 600 based on a hybrid network model is provided, which corresponds one-to-one with the spatial adaptive filtering localization method based on a hybrid network model in the above embodiments. For example... Figure 6 As shown, the spatial adaptive filtering localization device 600 based on a hybrid network model includes: The initial state determination module 601 is used to determine the initial relative orbital state of a space target with respect to the observation platforms based on angular measurement observation data from at least two observation platforms and through a joint positioning algorithm. The filter loop execution module 602 is used to iteratively execute the Kalman filter loop starting from the initial relative orbit state; The filtering loop execution module 602 includes: The adaptive prediction unit 602a is used to input the relative orbit state at the current moment into the pre-trained CNN-LSTM hybrid network model. The model outputs the adaptive state transition matrix and the adaptive model error factor corresponding to the next moment based on the relative orbit state at the current moment. The adaptive state transition matrix is ​​used to predict the state and obtain the prior estimate of the relative orbit state at the next moment and the prior error covariance matrix. The state update unit 602b is used to update the prior estimate using the angle measurement observation data at the next moment, and to determine the optimal estimate of the relative orbital state and the posterior error covariance matrix at the next moment. The result output and state transfer unit 603 is used to output the optimal estimate obtained after each update as the spatial target positioning result for the next moment; and to use the optimal estimate as the new relative orbital state for the current moment for subsequent Kalman filter loops. The pre-trained CNN-LSTM hybrid network model is configured to be trained on orbital dynamics data containing the maneuvering behavior of space targets, and is used to establish a nonlinear mapping relationship from relative orbital state to Kalman filter model parameters.

[0140] In one embodiment, the initial state determination module 601 is specifically used to designate one of at least two observation platforms as the reference observation platform; Based on angle measurement observation data at least three consecutive time points, the relative position of the space target with respect to the reference observation platform at each time point is determined using a dual-line least squares positioning algorithm. Based on the relative positions at least three consecutive moments, the relative velocity at the middle moment among the at least three consecutive moments is calculated using the central difference method; The relative position and relative velocity at the intermediate moment are used together as the initial relative orbital state.

[0141] In one embodiment, the pre-trained CNN-LSTM hybrid network model is obtained in the following way: Based on the initial orbital state of the observation platform and space target, a predicted trajectory containing simulated maneuvering behavior is generated through a dynamic model and a numerical integrator. Based on the predicted trajectory, the relative position and relative velocity of the space target with respect to the observation platform in the vertical velocity position heading coordinate system are calculated, and the state transition matrix and model error factor are calculated and recorded accordingly to construct the training dataset. The angle measurement observation data used to construct the training dataset was generated through the following simulation steps: Based on the predicted trajectory, the position of the space target in the geocentric inertial coordinate system at each moment is successively converted into the position in the vertical velocity and heading coordinate system, the body coordinate system, and the camera coordinate system relative to the observation platform. Based on the relative position in the camera coordinate system, the azimuth and elevation angle sequences are calculated, and random errors following a Gaussian distribution are added to the sequences. A CNN-LSTM hybrid network model is trained using a training dataset to establish a nonlinear mapping relationship from relative position and relative velocity to state transition matrix and model error factor, thus obtaining a pre-trained CNN-LSTM hybrid network model.

[0142] In one embodiment, the adaptive prediction unit 602a is specifically used to construct a relative orbital state sequence that includes multiple consecutive historical moments, with the current moment as the endpoint; The relative orbital state sequence is input into a pre-trained CNN-LSTM hybrid network model. The pre-trained CNN-LSTM hybrid network model extracts spatial features and models temporal dependencies of the relative orbital state sequence, and outputs an adaptive state transition matrix and an adaptive model error factor corresponding to the next time step.

[0143] In one embodiment, the apparatus further includes a normalization processing module 604, wherein: The normalization module 604 is used to perform normalization preprocessing on relative position, relative velocity, state transition matrix, and model error factor when constructing the training dataset; wherein, the normalization preprocessing is based on the following formula:

[0144] in, For variables to be normalized, and These are the maximum and minimum values ​​of the variable in the training dataset, respectively.

[0145] In one embodiment, the adaptive prediction unit 602a is specifically used to calculate the prior error covariance matrix for the next time step based on the adaptive state transition matrix, the error covariance matrix at the current time step, and the adaptive model error factor; wherein the adaptive model error factor is used to construct the process noise covariance matrix.

[0146] In one embodiment, the relative orbital state sequence is a sliding window sequence containing multiple consecutive historical moments, and the feature dimension of the sliding window sequence corresponds to the relative position and relative velocity components in the vertical velocity position heading coordinate system.

[0147] In one embodiment, the CNN-LSTM hybrid network model includes a convolutional neural network module and a long short-term memory network module connected in sequence; The convolutional neural network module is configured to perform convolution and pooling operations on the input state sequence to extract local spatial features from the relative orbital state sequence. The Long Short-Term Memory (LSTM) network module is configured to receive local spatial features output by the convolutional neural network module and model the long-term dependencies of local spatial features in the time dimension through a gating mechanism to output an adaptive state transition matrix and an adaptive model error factor.

[0148] It should be noted that the spatial adaptive filtering localization device based on a hybrid network model provided in the above embodiments is only illustrated by the division of the above program modules when implementing the corresponding spatial adaptive filtering localization method based on a hybrid network model. In practical applications, the above processing can be assigned to different program modules as needed, that is, the internal structure of the above system can be divided into different program modules to complete all or part of the processing described above. In addition, the system provided in the above embodiments and the corresponding Figure 1 The embodiments of the methods shown belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0149] This disclosure also provides an electronic device having the above-described features. Figure 6 The spatial adaptive filtering positioning device based on a hybrid network model is shown.

[0150] Figure 7 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure.

[0151] The following is a detailed reference. Figure 7 The diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present disclosure. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 701, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 702 or a program loaded from memory 708 into random access memory (RAM) 703. The RAM 703 also stores various programs and data required for the operation of the electronic device. The processor 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0152] Typically, the following devices can be connected to I / O interface 705: input devices 706 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 707 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 708 including, for example, magnetic tapes, hard disks, etc.; and communication devices 709. Communication device 709 allows electronic devices to exchange data via wireless or wired communication with other devices. Although Figure 7 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0153] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 709, or installed from memory 708, or installed from ROM 702. When the computer program is executed by processor 701, it performs the functions defined in the network data stream hardware offloading method for heterogeneous descriptor unified processing of embodiments of this disclosure.

[0154] Figure 7 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0155] This disclosure also provides a computer-readable storage medium in which the methods described in this disclosure can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the network data stream hardware offloading method for unified processing of heterogeneous descriptors shown in the above embodiments is implemented.

[0156] A portion of this disclosure can be applied to computer program products, such as computer program instructions, which, when executed by a computer, can invoke or provide methods and / or technical solutions according to this disclosure through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, and installation package files. Accordingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions; the computer compiling the instructions and then executing the corresponding compiled program; the computer reading and executing the instructions; or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0157] Although embodiments of the present disclosure have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present disclosure, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A spatial adaptive filtering localization method based on a hybrid network model, characterized in that, The method includes: Based on angular measurement data from at least two observation platforms, the initial relative orbital state of a space target with respect to the observation platforms is determined through a joint positioning algorithm. Starting from the initial relative orbital state, a Kalman filter loop is executed iteratively, wherein the following operations are performed at each time step of the Kalman filter loop: The current relative orbit state is input into a pre-trained CNN-LSTM hybrid network model, which outputs an adaptive state transition matrix and an adaptive model error factor corresponding to the next moment based on the current relative orbit state. The adaptive state transition matrix is ​​used to predict the state, and the prior estimate of the relative orbit state at the next moment and the prior error covariance matrix are obtained. Using the angle measurement observation data at the next moment, the prior estimate is updated to determine the optimal estimate of the relative orbital state and the posterior error covariance matrix at the next moment; The optimal estimate obtained after each update is output as the spatial target positioning result for the next moment; and the optimal estimate is used as the relative orbital state at the new current moment for subsequent Kalman filter loops. The pre-trained CNN-LSTM hybrid network model is configured to be trained based on orbital dynamics data containing the maneuvering behavior of space targets, and is used to establish a nonlinear mapping relationship from relative orbital state to Kalman filter model parameters.

2. The method according to claim 1, characterized in that, The determination of the initial relative orbital state of the space target with respect to the observation platforms, based on angular measurement observation data from at least two observation platforms and using a joint positioning algorithm, includes: Designate one of the at least two observation platforms as the reference observation platform; Based on angle measurement observation data at least three consecutive time points, the relative position of the spatial target with respect to the reference observation platform at each time point is determined using a dual-line least squares positioning algorithm. Based on the relative positions at the at least three consecutive moments, the relative velocity at the middle moment among the at least three consecutive moments is calculated using the central difference method; The relative position and relative velocity at the intermediate moment are used together as the initial relative orbital state.

3. The method according to claim 1, characterized in that, The pre-trained CNN-LSTM hybrid network model is obtained in the following way: Based on the initial orbital state of the observation platform and space target, a predicted trajectory containing simulated maneuvering behavior is generated through a dynamic model and a numerical integrator. Based on the predicted trajectory, the relative position and relative velocity of the space target with respect to the observation platform in the vertical velocity position heading coordinate system are calculated, and the state transition matrix and model error factor are calculated and recorded accordingly to construct a training dataset. The angle measurement observation data used to construct the training dataset is generated through the following simulation steps: Based on the predicted trajectory, the position of the space target in the geocentric inertial coordinate system at each moment is sequentially converted into the position in the vertical velocity and heading coordinate system, the position in the body coordinate system, and the position in the camera coordinate system relative to the observation platform. Based on the relative position in the camera coordinate system, the azimuth and elevation angle sequences are calculated, and random errors following a Gaussian distribution are added to the sequences. A CNN-LSTM hybrid network model is trained using the training dataset to establish a nonlinear mapping relationship from the relative position and the relative velocity to the state transition matrix and the model error factor, thereby obtaining the pre-trained CNN-LSTM hybrid network model.

4. The method according to claim 1, characterized in that, The process of inputting the current relative orbital state into a pre-trained CNN-LSTM hybrid network model, which then outputs an adaptive state transition matrix and an adaptive model error factor corresponding to the next time step based on the current relative orbital state, includes: Construct a relative orbital state sequence that includes multiple consecutive historical moments, with the current moment as the endpoint; The relative orbital state sequence is input into the pre-trained CNN-LSTM hybrid network model, which extracts spatial features and models temporal dependencies of the relative orbital state sequence, and outputs an adaptive state transition matrix and an adaptive model error factor corresponding to the next time step.

5. The method according to claim 3, characterized in that, When constructing the training dataset, the relative position, relative velocity, state transition matrix, and model error factor are preprocessed using normalization; wherein, the normalization preprocessing is based on the following formula: in, For variables to be normalized, and These are the maximum and minimum values ​​of the variable in the training dataset, respectively.

6. The method according to claim 1, characterized in that, The prior error covariance matrix is ​​obtained based on the following steps: Based on the adaptive state transition matrix, the error covariance matrix at the current time step, and the adaptive model error factor, the prior error covariance matrix at the next time step is calculated; wherein, the adaptive model error factor is used to construct the process noise covariance matrix.

7. The method according to claim 4, characterized in that, The relative orbital state sequence is a sliding window sequence containing multiple consecutive historical moments. The feature dimension of the sliding window sequence corresponds to the relative position and relative velocity components in the vertical velocity position heading coordinate system.

8. The method according to claim 3, characterized in that, The CNN-LSTM hybrid network model includes a convolutional neural network module and a long short-term memory network module connected in sequence; The convolutional neural network module is configured to perform convolution and pooling operations on the input state sequence to extract local spatial features from the relative orbital state sequence. The Long Short-Term Memory (LSTM) network module is configured to receive local spatial features output by the convolutional neural network module and model the long-term dependencies of the local spatial features in the time dimension through a gating mechanism, so as to output the adaptive state transition matrix and the adaptive model error factor.

9. A spatial adaptive filtering positioning device based on a hybrid network model, characterized in that, The device includes: The initial state determination module is used to determine the initial relative orbital state of a space target with respect to the observation platforms based on angular measurement observation data from at least two observation platforms and through a joint positioning algorithm. The filter loop execution module is used to iteratively execute the Kalman filter loop starting from the initial relative orbit state; The filtering loop execution module includes: The adaptive prediction unit is used to input the relative orbit state at the current moment into the pre-trained CNN-LSTM hybrid network model. The model outputs an adaptive state transition matrix and an adaptive model error factor corresponding to the next moment based on the relative orbit state at the current moment. The adaptive state transition matrix is ​​used to predict the state and obtain the prior estimate of the relative orbit state at the next moment and the prior error covariance matrix. The state update unit is used to update the prior estimate using the angle measurement observation data at the next moment, and to determine the optimal estimate of the relative orbital state and the posterior error covariance matrix at the next moment. The result output and state transfer unit is used to output the optimal estimate obtained after each update as the spatial target positioning result for the next moment; and to use the optimal estimate as the new relative orbital state for the current moment, for subsequent Kalman filter loops. The pre-trained CNN-LSTM hybrid network model is configured to be trained based on orbital dynamics data containing the maneuvering behavior of space targets, and is used to establish a nonlinear mapping relationship from relative orbital state to Kalman filter model parameters.

10. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the spatial adaptive filtering localization method based on a hybrid network model as described in any one of claims 1 to 8 when executing the computer program.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, it implements the steps of the spatial adaptive filtering localization method based on a hybrid network model as described in any one of claims 1 to 8.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the spatial adaptive filtering localization method based on a hybrid network model as described in any one of claims 1 to 8.

Images