A method for space space debris monitoring

Abstract

This invention discloses a method for monitoring space debris, belonging to the field of space debris monitoring technology. This invention significantly improves the reliability of debris monitoring by low-Earth orbit satellites in maneuvering states through the deep integration of physical models and deep learning. A plume diffusion physical field model accurately characterizes the interference generation mechanism, a generative adversarial network (GAN) enables dynamic stripping of interference features, a multi-scale residual network enhances the micro-motion characteristics of debris, and optical flow alignment suppresses temporal noise. An online learning framework overcomes the bottleneck of satellite-ground collaboration, enabling autonomous evolution and updating of model parameters. This invention systematically solves the false alarm problem caused by plume interference without changing existing onboard hardware, while simultaneously improving the detection sensitivity for non-metallic debris and ensuring the autonomous monitoring capabilities of the satellite constellation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of space space debris monitoring, in particular to a space space debris monitoring method. BACKGROUND

[0002] With the large-scale deployment of low-orbit satellite constellations, spaceborne optical sensors have become an important part of the space debris monitoring system; these monitoring payloads carried on communication and remote sensing satellites can achieve near real-time debris tracking through networked observation; however, when satellites frequently implement orbit maintenance and attitude adjustment, the plume generated by hydrazine propellant forms a dispersed particle cloud in the vacuum environment, and its optical reflection characteristics are highly similar to those of small debris; especially during the maneuvering stage before the satellite enters the shadow area, the plume particles will produce dynamic flicker noise in the sensor field of view due to the influence of the solar irradiation angle, which will seriously interfere with the identification and trajectory tracking of effective debris.

[0003] In recent years, the mainstream solution has adopted multi-modal data fusion to improve anti-interference capability, integrating millimeter wave radar outside the optical payload to cross-verify the Doppler characteristics of metal debris, but it cannot identify non-metal debris; or an association model between the working state of the propulsion system and optical noise is established to predict the plume interference mode through fuel consumption, nozzle temperature and other parameters, but the complex flow field changes during satellite maneuvering lead to cumulative prediction errors of the model.

[0004] The current federated learning framework can update the interference feature library through multi-satellite collaboration, but it is difficult to achieve high-frequency model iteration due to the limited inter-satellite communication bandwidth; some commercial constellations attempt to deploy lightweight YOLO-v7 detection networks, and their attention mechanisms can suppress static background noise, but they are too sensitive to time-varying plume interference; this leads the system to still rely on manual intervention by ground stations for false alarm filtering when dealing with sudden maneuvering tasks, which severely restricts the timeliness of monitoring. SUMMARY

[0005] In view of the above existing problems, the present application is proposed.

[0006] The present application provides a space space debris monitoring method to solve the problems of existing solutions relying on multi-modal fusion and transfer learning, such as non-metal debris missed detection, plume interference modeling error and low efficiency of satellite-ground collaboration.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] The present application provides a space space debris monitoring method, which comprises,

[0009] Step S1, acquiring real-time attitude angle data, propellant pressure parameters and temperature sensor data of a target satellite in a low-orbit satellite constellation;

[0010] Step S2: Establish the plume diffusion physical field model at the current moment based on the attitude angle data and propellant pressure parameters;

[0011] Step S3: Spatiotemporally align the output of the physical field model with the original image sequence acquired by the spaceborne optical sensor;

[0012] Step S4: Construct a dynamic interference feature extraction module based on generative adversarial networks to generate an interference mask that matches the physical field model;

[0013] Step S5: Use a multi-scale residual network to enhance the fragment features of the masked image sequence;

[0014] Step S6: Update the weight parameters of the feature extraction module through the online learning framework, and deploy the updated model to other satellite nodes within the constellation.

[0015] As a preferred embodiment of the space debris monitoring method described in this invention, establishing a plume diffusion physical field model specifically includes:

[0016] The initial particle distribution was calculated based on the satellite nozzle geometry and propellant injection velocity.

[0017] Simulation of the three-dimensional diffusion trajectory of particles in a vacuum environment based on simplified Navier-Stokes equations;

[0018] Temperature sensor data is used to correct the dynamic change parameters of particle surface reflectivity.

[0019] As a preferred embodiment of the space debris monitoring method of the present invention, in step S2, the three-dimensional diffusion trajectory of particles is simulated based on the simplified Navier-Stokes equations in a vacuum environment. To obtain the velocity field in the propellant wake and track the particle motion accordingly, the gas flow field is first solved, and then the particle trajectory is integrated within the Lagrange framework. The process includes:

[0020] A simplified flow field equation is established and solved. Taking into account incompressibility, isotherm, and neglecting the influence of body forces, the flow field satisfies:

[0021] ,

[0022] in, Represents the gas velocity vector. Indicates time, Represents the spatial gradient operator, Indicates the kinematic viscosity of the gas. Represents the Laplace operator.

[0023] The boundary condition is taken as the velocity distribution at the nozzle exit, and the far field velocity is set to zero.

[0024] The velocity field has been obtained Then, the Lagrange method is applied to a single particle, and its kinetic equation is:

[0025] ,

[0026] in, Indicates the mass of a single particle. Represents the particle velocity vector. Indicates the dynamic viscosity of the gas. Indicates the diameter of the particle. Indicates photothermal driving force;

[0027] The particle position update equation is:

[0028] ,

[0029] in, Represents the particle position vector;

[0030] The mass of a particle in the kinetic equation is determined by its density and volume, expressed as:

[0031] ,

[0032] in, Indicates particle density,

[0033] The relationship between the kinematic viscosity, dynamic viscosity, and density of a gas is as follows:

[0034] ,

[0035] in, Indicates gas density;

[0036] During the simulation:

[0037] The time integration uses the second-order Runge-Kutta method.

[0038] The grid resolution should be at least ten times that of the nozzle feature scale.

[0039] The convection term uses an upwind scheme, and the diffusion term uses a central difference scheme.

[0040] Photothermal driving force Linear correction is performed based on temperature sensor data.

[0041] As a preferred embodiment of the space debris monitoring method of the present invention, the dynamic interference feature extraction module in step S4 includes:

[0042] The generator network receives the diffusion parameters output by the physical field model and generates multiple frames of continuous interference prediction images;

[0043] The discriminator network compares the frequency domain energy distribution differences between real optical images and predicted images;

[0044] During adversarial training, the parameters of the discriminator network are frozen, and the generator network is optimized through backpropagation.

[0045] The generator network described in step S4 adopts a U-Net architecture, in which the skip connection layer injects diffusion parameters as conditional inputs, and the output layer ensures the consistency of optical flow between consecutive frames through a temporal convolution module.

[0046] As a preferred embodiment of the space debris monitoring method of the present invention, in step S5, the multi-scale residual network performs the following operations:

[0047] Establish parallel feature extraction channels containing 5×5, 3×3, and 1×1 convolutional kernels in the spatial dimension;

[0048] Perform motion-compensated alignment on three consecutive frames of images in the time dimension;

[0049] The enhanced fragment candidate region is output by fusing multi-scale feature maps through a channel attention mechanism.

[0050] The motion compensation alignment described in step S5 adopts a global registration algorithm based on optical flow field. Taking the first frame image as a reference, pixel-level offset correction is achieved by minimizing the gray-level gradient difference between adjacent frames, and the allowed residual offset is no more than 2 pixels.

[0051] As a preferred embodiment of the space debris monitoring method of the present invention, in step S5, during the process of performing motion compensation alignment on three consecutive frames of images in the time dimension, a global optical flow registration algorithm is used to estimate pixel displacement by minimizing the gray-level gradient difference between adjacent frames, and then resampling to obtain the aligned image.

[0052] For frames Middle adjacent frames Optical flow field between Solve and construct the energy functional:

[0053] ,

[0054] in, Indicates the first The gradient vector of the frame image, Indicates the first frame image gradient vector, Represents pixel coordinates, and These represent the displacement components of the optical flow field in the horizontal and vertical directions, respectively. Represents the image domain. Indicates the smoothing regularization weights. and These are the gradients of the displacement components in the spatial dimension;

[0055] For the above functionals respectively and Variational calculus yields partial differential equations:

[0056] ,

[0057] ,

[0058] in, express right Jacobian matrix, The Laplacian operator is represented by an iterative update using a multi-scale pyramid and Gaussian filter. until convergence;

[0059] frame Through the obtained optical flow field Perform reverse mapping:

[0060] ,

[0061] in, Indicates alignment to frame The Frame image;

[0062] Add a restriction, limiting the maximum displacement to no more than :

[0063] ,in, This indicates the maximum allowed pixel displacement.

[0064] As a preferred embodiment of the space debris monitoring method of the present invention, in step S6, the online learning framework is implemented in the following ways:

[0065] The ground station pre-trains the teacher model and extracts prior knowledge of feature distribution;

[0066] The teacher model parameters are compressed to 15%-20% of their original size using knowledge distillation technology.

[0067] The satellite in orbit receives false detection sample data transmitted back from other nodes in the constellation and updates the student model using a sliding window method;

[0068] In step S6, the knowledge distillation technique selects the KL divergence of the intermediate feature map as the knowledge transfer target. The compression ratio is dynamically adjusted according to the real-time memory usage of the onboard processor, and the adjustment step does not exceed 5% of the original parameters.

[0069] As a preferred embodiment of the space debris monitoring method of the present invention, the transmission method of the false detection sample data is as follows:

[0070] The false detection image region is segmented into 32×32 pixel blocks and then compressed using JPEG2000;

[0071] Transmit compressed data blocks during idle periods of the inter-satellite laser communication link;

[0072] The receiving satellite performs a weighted average fusion of the data blocks and then stores them in its local training cache.

[0073] In a preferred embodiment of the space debris monitoring method of the present invention, in step S6, after a false detection target is detected by the onboard node, the corresponding pixel region is divided into blocks of fixed size and efficiently compressed. The specific process includes:

[0074] Define the false detection image region The pixel size is The side length of the block is set as:

[0075] Divided into rows and columns respectively A complete block;

[0076] No. line, number The pixel submatrix of a column block is defined as:

[0077] ,

[0078] in, Indicates the first line, number The pixel block matrix of the column, This represents the pixel matrix of the false detection image region. , , This indicates the block's side length is 32 pixels. This indicates the floor function. and These represent the height and width of the region, and the number of pixels, respectively.

[0079] For each block Perform the following operations in sequence:

[0080] 1) Wavelet Transform:

[0081] ,

[0082] in, This represents the output of a two-dimensional discrete wavelet transform operator. The wavelet coefficient matrix,

[0083] 2) Quantization: Select the quantization step size Perform scalar quantization on the coefficient matrix:

[0084] ,

[0085] in, This represents the quantized coefficient matrix. Represents the coordinates within the coefficient matrix. This indicates the quantization step size, which is dynamically adjusted based on the real-time memory usage of the onboard processor. For floor operations,

[0086] 3) Entropy encoding, for the quantization coefficient matrix Perform EBCOT entropy encoding according to the JPEG2000 standard to generate a bitstream:

[0087] ,

[0088] in, This represents the embedded block coding operator in JPEG2000. Indicates the first The compressed bitstream output of the block;

[0089] An overall compression ratio assessment was performed, with a total number of bits transmitted:

[0090] ,

[0091] Actual compression ratio:

[0092] ,

[0093] in, Indicates bitstream length and number of bits. This represents the average number of bits per pixel and is a feedback indicator for quality control.

[0094] As a preferred embodiment of the space debris monitoring method described in this invention, the space debris monitoring method employs the following optimization measures during deployment:

[0095] Configure the onboard processor with dual buffered storage areas to alternately execute model inference and parameter updates;

[0096] The number of training iterations for the generative adversarial network is dynamically adjusted based on the satellite's remaining battery power.

[0097] When the attitude control system is detected to have entered maneuver mode, it automatically switches to high-sensitivity monitoring mode.

[0098] The beneficial effects of this invention are as follows: By deeply integrating physical models with deep learning, this invention significantly improves the reliability of debris monitoring for low-Earth orbit satellites in maneuvering states; the plume diffusion physical field model accurately characterizes the interference generation mechanism, generative adversarial networks achieve dynamic stripping of interference features, multi-scale residual networks enhance the micro-motion characteristics of debris, and optical flow alignment suppresses temporal noise; the online learning framework breaks through the bottleneck of satellite-ground collaboration and realizes the autonomous evolution and updating of model parameters; this invention systematically solves the false alarm problem caused by plume interference without changing the existing onboard hardware, while improving the detection sensitivity of non-metallic debris and ensuring the autonomous monitoring capability of the satellite constellation. Attached Figure Description

[0099] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0100] Figure 1 This is a flowchart illustrating the space debris monitoring method in Example 1. Detailed Implementation

[0101] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0102] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0103] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0104] Example 1, referring to Figure 1 This embodiment provides a method for monitoring space debris, including the following steps:

[0105] Step S1: Obtain real-time attitude angle data, propellant pressure parameters, and temperature sensor data of the target satellite in the low-Earth orbit satellite constellation;

[0106] Step S2: Establish the physical field model of the plume diffusion at the current moment based on the attitude angle data and propellant pressure parameters;

[0107] Step S2, establishing the physical field model for plume diffusion specifically includes:

[0108] The initial particle distribution was calculated based on the satellite nozzle geometry and propellant injection velocity.

[0109] Simulation of the three-dimensional diffusion trajectory of particles in a vacuum environment based on simplified Navier-Stokes equations;

[0110] Temperature sensor data is used to correct the dynamic change parameters of particle surface reflectivity.

[0111] In step S2, the three-dimensional diffusion trajectory of the particles is simulated in a vacuum environment based on the simplified Navier-Stokes equations. To obtain the velocity field in the propellant wake and track the particle motion accordingly, the gas flow field is first solved, and then the particle trajectory is integrated within the Lagrangian framework. The process includes:

[0112] A simplified flow field equation is established and solved. Taking into account incompressibility, isotherm, and neglecting the influence of body forces, the flow field satisfies:

[0113] ,

[0114] in, Represents the gas velocity vector. Indicates time, Represents the spatial gradient operator, Indicates the kinematic viscosity of the gas. Represents the Laplace operator.

[0115] The boundary condition is taken as the velocity distribution at the nozzle exit, and the far field velocity is set to zero.

[0116] The velocity field has been obtained Then, the Lagrange method is applied to a single particle, and its kinetic equation is:

[0117] ,

[0118] in, Indicates the mass of a single particle. Represents the particle velocity vector. Indicates the dynamic viscosity of the gas. Indicates the diameter of the particle. Indicates photothermal driving force;

[0119] The particle position update equation is:

[0120] ,

[0121] in, Represents the particle position vector;

[0122] The mass of a particle in the kinetic equation is determined by its density and volume, expressed as:

[0123] ,

[0124] in, Indicates particle density,

[0125] The relationship between the kinematic viscosity, dynamic viscosity, and density of a gas is as follows:

[0126] ,

[0127] in, Indicates gas density;

[0128] During the simulation:

[0129] The time integration uses the second-order Runge-Kutta method.

[0130] The grid resolution should be at least ten times that of the nozzle feature scale.

[0131] The convection term uses an upwind scheme, and the diffusion term uses a central difference scheme.

[0132] Photothermal driving force Linear correction is performed based on temperature sensor data;

[0133] Specifically, this method first obtains the velocity field distribution of the propellant wake in vacuum by simplifying the Navier-Stokes equations, and then tracks the particle motion in this field using a Lagrange framework. The simplified model retains the convection and viscous diffusion characteristics while ignoring pressure gradients and volume forces. The numerical scheme and high-order time integrals take into account both computational stability and accuracy. By combining temperature-corrected photothermal driving forces, the surface thermal effect of the wake particles can be reflected. This method maintains the rationality of the Navier-Stokes framework under extreme vacuum conditions and fully captures the multi-physics coupling process.

[0134] Step S3: Spatiotemporally align the output of the physical field model with the original image sequence acquired by the spaceborne optical sensor;

[0135] Step S4: Construct a dynamic interference feature extraction module based on generative adversarial networks to generate an interference mask that matches the physical field model;

[0136] The dynamic interference feature extraction module in step S4 includes:

[0137] The generator network receives the diffusion parameters output by the physical field model and generates multiple frames of continuous interference prediction images;

[0138] The discriminator network compares the frequency domain energy distribution differences between real optical images and predicted images;

[0139] During adversarial training, the parameters of the discriminator network are frozen, and the generator network is optimized through backpropagation.

[0140] In step S4, the generator network adopts the U-Net architecture, where the skip connection layer injects diffusion parameters as conditional inputs, and the output layer ensures the consistency of optical flow between consecutive frames through a temporal convolution module.

[0141] Step S5: Use a multi-scale residual network to enhance the fragment features of the masked image sequence;

[0142] In step S5, the multi-scale residual network performs the following operations:

[0143] Establish parallel feature extraction channels containing 5×5, 3×3, and 1×1 convolutional kernels in the spatial dimension;

[0144] Perform motion-compensated alignment on three consecutive frames of images in the time dimension;

[0145] The enhanced fragment candidate region is output by fusing multi-scale feature maps through a channel attention mechanism.

[0146] In step S5, motion compensation alignment adopts a global registration algorithm based on optical flow field. Taking the first frame image as the reference, pixel-level offset correction is achieved by minimizing the gray-level gradient difference between adjacent frames, and the allowed residual offset is no more than 2 pixels.

[0147] In step S5, during the motion compensation alignment of three consecutive frames of images in the time dimension, a global optical flow registration algorithm is used to estimate pixel displacement by minimizing the gray-level gradient difference between adjacent frames, and then resample to obtain the aligned image.

[0148] For frames Middle adjacent frames Optical flow field between Solve and construct the energy functional:

[0149] ,

[0150] in, Indicates the first The gradient vector of the frame image, Indicates the first frame image gradient vector, Represents pixel coordinates, and These represent the displacement components of the optical flow field in the horizontal and vertical directions, respectively. Represents the image domain. Indicates the smoothing regularization weights. and These are the gradients of the displacement components in the spatial dimension;

[0151] For the above functionals respectively and Variational calculus yields partial differential equations:

[0152] ,

[0153] ,

[0154] in, express right Jacobian matrix, The Laplacian operator is represented by an iterative update using a multi-scale pyramid and Gaussian filter. until convergence;

[0155] frame Through the obtained optical flow field Perform reverse mapping:

[0156] ,

[0157] in, Indicates alignment to frame The Frame image;

[0158] Add a restriction, limiting the maximum displacement to no more than :

[0159] ,in, Indicates the maximum allowed pixel displacement;

[0160] Specifically, through global optical flow registration, the optimal balance between gradient consistency and displacement smoothness is iteratively solved on the multi-resolution pyramid, preserving real motion details and suppressing artifacts introduced by noise. The pyramid strategy and Gaussian filtering can effectively overcome the non-convexity problem caused by large displacement. The variational solution of the Euler-Lagrange equation combined with smoothing regularization makes the optical flow field both accurate and smooth. Inverse mapping resampling ensures the brightness continuity during the interpolation process, while displacement threshold limitation ensures that the residual after alignment is controllable and meets the accuracy requirement of no more than two pixels.

[0161] Step S6: Update the weight parameters of the feature extraction module through the online learning framework, and deploy the updated model to other satellite nodes within the constellation;

[0162] In step S6, the online learning framework is implemented in the following ways:

[0163] The ground station pre-trains the teacher model and extracts prior knowledge of feature distribution;

[0164] The teacher model parameters are compressed to 15%-20% of their original size using knowledge distillation technology.

[0165] The satellite in orbit receives false detection sample data transmitted back from other nodes in the constellation and updates the student model using a sliding window method;

[0166] In step S6, the knowledge distillation technique selects the KL divergence of the intermediate feature map as the knowledge transfer target, and the compression ratio is dynamically adjusted according to the real-time memory usage of the onboard processor, with the adjustment step not exceeding 5% of the original parameters;

[0167] The transmission method for false positive sample data is as follows:

[0168] The false detection image region is segmented into 32×32 pixel blocks and then compressed using JPEG2000;

[0169] Transmit compressed data blocks during idle periods of the inter-satellite laser communication link;

[0170] The receiving satellite performs a weighted average fusion of the data blocks and stores them in its local training cache.

[0171] In step S6, after the onboard node detects a false target, the corresponding pixel region is divided into blocks of fixed size and compressed efficiently. The specific process includes:

[0172] Define the false detection image region The pixel size is The side length of the block is set as:

[0173] Divided into rows and columns respectively A complete block;

[0174] No. line, number The pixel submatrix of a column block is defined as:

[0175] ,

[0176] in, Indicates the first line, number The pixel block matrix of the column, This represents the pixel matrix of the false detection image region. , , This indicates the block's side length is 32 pixels. This indicates the floor function. and These represent the height and width of the region, and the number of pixels, respectively.

[0177] For each block Perform the following operations in sequence:

[0178] 1) Wavelet Transform:

[0179] ,

[0180] in, This represents the output of a two-dimensional discrete wavelet transform operator. The wavelet coefficient matrix,

[0181] 2) Quantization: Select the quantization step size Perform scalar quantization on the coefficient matrix:

[0182] ,

[0183] in, This represents the quantized coefficient matrix. Represents the coordinates within the coefficient matrix. This indicates the quantization step size, which is dynamically adjusted based on the real-time memory usage of the onboard processor. For floor operations,

[0184] 3) Entropy encoding, for the quantization coefficient matrix Perform EBCOT entropy encoding according to the JPEG2000 standard to generate a bitstream:

[0185] ,

[0186] in, This represents the embedded block coding operator in JPEG2000. Indicates the first The compressed bitstream output of the block;

[0187] An overall compression ratio assessment was performed, with a total number of bits transmitted:

[0188] ,

[0189] Actual compression ratio:

[0190] ,

[0191] in, Indicates bitstream length and number of bits. It represents the average number of bits per pixel and is a feedback indicator for quality control;

[0192] Specifically, through a fixed size Pixel block segmentation ensures that each block obtains consistent transform and quantization characteristics during JPEG2000 encoding, which is beneficial for balancing coding performance. Two-dimensional wavelet transform provides multi-resolution representation and quantization step size. The adjustability allows the compression ratio to be dynamically optimized according to onboard storage and bandwidth conditions, while EBCOT entropy coding achieves efficient packing while ensuring lossless reconstruction, resulting in a high average bit count. It can serve as a basis for adaptive control, enabling high-quality data transmission during link idle periods. The process is complete and coherent, taking into account the constraints of limited onboard computing resources and communication delays.

[0193] The following optimization measures are adopted when deploying space debris monitoring methods:

[0194] Configure the onboard processor with dual buffered storage areas to alternately execute model inference and parameter updates;

[0195] The number of training iterations for the generative adversarial network is dynamically adjusted based on the satellite's remaining battery power.

[0196] When the attitude control system is detected to have entered maneuver mode, it automatically switches to high-sensitivity monitoring mode.

[0197] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for monitoring space debris, characterized in that, include, Step S1: Obtain real-time attitude angle data, propellant pressure parameters, and temperature sensor data of the target satellite in the low-Earth orbit satellite constellation; Step S2: Establish the plume diffusion physical field model at the current moment based on the attitude angle data and propellant pressure parameters; Step S3: Spatiotemporally align the output of the physical field model with the original image sequence acquired by the spaceborne optical sensor; Step S4: Construct a dynamic interference feature extraction module based on generative adversarial networks to generate an interference mask that matches the physical field model; Step S5: Use a multi-scale residual network to enhance the fragment features of the masked image sequence; Step S6: Update the weight parameters of the feature extraction module through the online learning framework, and deploy the updated model to other satellite nodes within the constellation.

2. The space debris monitoring method as described in claim 1, characterized in that, Step S2, establishing the physical field model for plume diffusion specifically includes: The initial particle distribution was calculated based on the satellite nozzle geometry and propellant injection velocity. Simulation of the three-dimensional diffusion trajectory of particles in a vacuum environment based on simplified Navier-Stokes equations; Temperature sensor data is used to correct the dynamic change parameters of particle surface reflectivity.

3. The space debris monitoring method as described in claim 2, characterized in that, In step S2, the three-dimensional diffusion trajectory of the particles is simulated in a vacuum environment based on the simplified Navier-Stokes equations. To obtain the velocity field in the propellant wake and track the particle motion accordingly, the gas flow field is first solved, and then the particle trajectory is integrated within the Lagrangian framework. The process includes: A simplified flow field equation is established and solved. Taking into account incompressibility, isotherm, and neglecting the influence of body forces, the flow field satisfies: ， in, Represents the gas velocity vector. Indicates time, Represents the spatial gradient operator, Indicates the kinematic viscosity of the gas. Represents the Laplace operator. The boundary condition is taken as the velocity distribution at the nozzle exit, and the far field velocity is set to zero. The velocity field has been obtained Then, the Lagrange method is applied to a single particle, and its kinetic equation is: ， in, Indicates the mass of a single particle. Represents the particle velocity vector. Indicates the dynamic viscosity of the gas. Indicates the diameter of the particle. Indicates photothermal driving force; The particle position update equation is: ， in, Represents the particle position vector; The mass of a particle in the kinetic equation is determined by its density and volume, expressed as: ， in, Indicates particle density, The relationship between the kinematic viscosity, dynamic viscosity, and density of a gas is as follows: ， in, Indicates gas density; During the simulation: The time integration uses the second-order Runge-Kutta method. The grid resolution should be at least ten times that of the nozzle feature scale. The convection term uses an upwind scheme, and the diffusion term uses a central difference scheme. Photothermal driving force Linear correction is performed based on temperature sensor data.

4. The space debris monitoring method as described in claim 1, characterized in that, The dynamic interference feature extraction module in step S4 includes: The generator network receives the diffusion parameters output by the physical field model and generates multiple frames of continuous interference prediction images; The discriminator network compares the frequency domain energy distribution differences between real optical images and predicted images; During adversarial training, the parameters of the discriminator network are frozen, and the generator network is optimized through backpropagation. The generator network described in step S4 adopts a U-Net architecture, in which the skip connection layer injects diffusion parameters as conditional inputs, and the output layer ensures the consistency of optical flow between consecutive frames through a temporal convolution module.

5. The space debris monitoring method as described in claim 1, characterized in that, In step S5, the multi-scale residual network performs the following operations: Establish parallel feature extraction channels containing 5×5, 3×3, and 1×1 convolutional kernels in the spatial dimension; Perform motion-compensated alignment on three consecutive frames of images in the time dimension; The enhanced fragment candidate region is output by fusing multi-scale feature maps through a channel attention mechanism. The motion compensation alignment described in step S5 adopts a global registration algorithm based on optical flow field. Taking the first frame image as a reference, pixel-level offset correction is achieved by minimizing the gray-level gradient difference between adjacent frames, and the allowed residual offset is no more than 2 pixels.

6. The space debris monitoring method as described in claim 5, characterized in that, In step S5, during the process of performing motion compensation alignment on three consecutive frames of images in the time dimension, a global optical flow registration algorithm is used to estimate pixel displacement by minimizing the gray-level gradient difference between adjacent frames, and then resampling to obtain the aligned image. For frames Middle adjacent frames Optical flow field between Solve and construct the energy functional: ， in, Indicates the first The gradient vector of the frame image, Indicates the first frame image gradient vector, Represents pixel coordinates, and These represent the displacement components of the optical flow field in the horizontal and vertical directions, respectively. Represents the image domain. Indicates the smoothing regularization weights. and These are the gradients of the displacement components in the spatial dimension; For the above functionals respectively and Variational calculus yields partial differential equations: ， ， in, express right Jacobian matrix, The Laplacian operator is represented by an iterative update using a multi-scale pyramid and Gaussian filter. until convergence; frame Through the obtained optical flow field Perform reverse mapping: ， in, Indicates alignment to frame The Frame image; Add a restriction, limiting the maximum displacement to no more than : ,in, This indicates the maximum allowed pixel displacement.

7. The space debris monitoring method as described in claim 1, characterized in that, In step S6, the online learning framework is implemented in the following ways: The ground station pre-trains the teacher model and extracts prior knowledge of feature distribution; The teacher model parameters are compressed to 15%-20% of their original size using knowledge distillation technology. The satellite in orbit receives false detection sample data transmitted back from other nodes in the constellation and updates the student model using a sliding window method; In step S6, the knowledge distillation technique selects the KL divergence of the intermediate feature map as the knowledge transfer target. The compression ratio is dynamically adjusted according to the real-time memory usage of the onboard processor, and the adjustment step does not exceed 5% of the original parameters.

8. A method for monitoring space debris as described in claim 7, characterized in that, The transmission method for the false detection sample data is as follows: The false detection image region is segmented into 32×32 pixel blocks and then compressed using JPEG2000; Transmit compressed data blocks during idle periods of the inter-satellite laser communication link; The receiving satellite performs a weighted average fusion of the data blocks and then stores them in its local training cache.

9. A method for monitoring space debris as described in claim 8, characterized in that, In step S6, after the onboard node detects a false target, the corresponding pixel region is divided into blocks of fixed size and compressed efficiently. The specific process includes: Define the false detection image region The pixel size is The side length of the block is set as: Divided into rows and columns respectively A complete block; No. line, number The pixel submatrix of a column block is defined as: ， in, Indicates the first line, number The pixel block matrix of the column, This represents the pixel matrix of the false detection image region. , , This indicates the block's side length is 32 pixels. This indicates the floor function. and These represent the height and width of the region, and the number of pixels, respectively. For each block Perform the following operations in sequence: 1) Wavelet Transform: ， in, This represents the output of a two-dimensional discrete wavelet transform operator. The wavelet coefficient matrix, 2) Quantization: Select the quantization step size Perform scalar quantization on the coefficient matrix: ， in, This represents the quantized coefficient matrix. Represents the coordinates within the coefficient matrix. This indicates the quantization step size, which is dynamically adjusted based on the real-time memory usage of the onboard processor. For floor operations, 3) Entropy encoding, for the quantization coefficient matrix Perform EBCOT entropy encoding according to the JPEG2000 standard to generate a bitstream: ， in, This represents the embedded block coding operator in JPEG2000. Indicates the first The compressed bitstream output of the block; An overall compression ratio assessment was performed, with a total number of bits transmitted: ， Actual compression ratio: ， in, Indicates bitstream length and number of bits. This represents the average number of bits per pixel and is a feedback indicator for quality control.

10. A method for monitoring space debris as described in claim 9, characterized in that, The following optimization measures are adopted when deploying space debris monitoring methods: Configure the onboard processor with dual buffered storage areas to alternately execute model inference and parameter updates; The number of training iterations for the generative adversarial network is dynamically adjusted based on the satellite's remaining battery power. When the attitude control system is detected to have entered maneuver mode, it automatically switches to high-sensitivity monitoring mode.

Images