A method for autonomous identification of target dynamics parameters based on decoupled cue representation

Abstract

The application discloses a target dynamic parameter autonomous identification method based on decoupling prompt representation and belongs to the technical field of spacecrafts. The method realizes domain generalization through decoupling prompt representation, can adapt to complex space environment changes, realizes high-precision monitoring of deformation and vibration of a flexible structure of a spacecraft, and inversely calculates the dynamic parameters of the flexible structure, thereby providing support for health state evaluation and fault early warning of the spacecraft.

Description

Technical Field

[0001] The present invention belongs to the field of spacecraft technology, and in particular relates to a method for autonomously identifying target dynamic parameters based on decoupling prompt representation. Background Art

[0002] During spacecraft in-orbit operation, flexible structures (such as solar panels) are affected by a variety of factors, including the space environment, thermal cycles, and minor collisions, resulting in dynamic responses such as deformation and vibration. These responses can reduce the accuracy of spacecraft attitude control, shorten its lifespan, or even cause failure. Therefore, real-time and accurate monitoring and identification of the dynamic state of spacecraft flexible structures is crucial for ensuring safe spacecraft operation.

[0003] At present, the monitoring methods of spacecraft flexible structures mainly include the following categories:

[0004] 1. Sensor-based monitoring: This involves installing strain gauges, accelerometers, and other sensors on flexible structures to directly measure their deformation and vibration. However, this approach requires the installation of a large number of additional sensors on the spacecraft, increasing system complexity and weight. Furthermore, the sensors themselves may fail due to the harsh space environment.

[0005] 2. Ground-based monitoring: This method uses ground-based tracking and control stations to receive and analyze data transmitted by the spacecraft. However, this method is limited by the coverage area and communication bandwidth of the ground stations, making it difficult to achieve real-time and continuous monitoring of the status of the spacecraft's flexible structures.

[0006] 3. Visual measurement-based monitoring method: Using a camera onboard a spacecraft to capture images of flexible structures, structural features are extracted using image processing techniques. Although this method avoids the need for additional sensor installation, it presents the following technical difficulties in practical application:

[0007] a) Insufficient image segmentation accuracy: Traditional image segmentation methods have difficulty accurately segmenting flexible structure boundaries in space environments with extreme lighting variations and shadow interference;

[0008] b) Unstable feature extraction: Conventional feature extraction algorithms (such as SIFT and SURF) have difficulty extracting sufficient number and quality feature points on sparsely textured flexible surfaces.

[0009] c) Limited deformation detection accuracy: Deformation detection methods based on sparse keypoints can only obtain discrete deformation information and are unable to capture subtle deformations and continuous deformation fields;

[0010] d) Poor environmental adaptability: Existing algorithms are not adaptable enough to factors such as changes in spacecraft attitude and periodic changes in orbital illumination, resulting in unstable detection results.

[0011] e) Difficulty in identifying dynamic parameters: When inverting dynamic parameters from visual measurement data, noise interference and model errors lead to low parameter identification accuracy.

[0012] 4. Finite element model-based monitoring method: This method uses a finite element model of the flexible structure and combines it with measurement data to estimate its state. However, this method relies on a highly accurate finite element model. The stress state and material parameters of an actual in-orbit spacecraft often differ from those in ground-based modeling. Furthermore, the high computational complexity makes real-time monitoring difficult.

[0013] In practical applications, these methods suffer from insufficient accuracy, difficulty adapting to complex space environments, and high computational resource requirements, making them inadequate for autonomous spacecraft health monitoring. This is particularly true for large, flexible structures like solar panels, where surface textures are monotonous and key points are sparse. Traditional methods based on feature point matching often fail to provide sufficiently accurate deformation measurements, making it difficult to capture structural deformations caused by minor perturbations. This, in turn, affects the accuracy of subsequent dynamic parameter identification and health status assessment. Summary of the Invention

[0014] In order to overcome the shortcomings of the existing technology, the present invention provides a method for autonomous identification of target dynamic parameters based on decoupled prompt representation. The method achieves domain generalization through decoupled prompt representation, can adapt to complex changes in space environment, realize high-precision monitoring of the deformation and vibration of spacecraft flexible structures, and invert the dynamic parameters of flexible structures, providing support for spacecraft health status assessment and fault warning.

[0015] The technical solutions adopted by the present invention to solve the technical problems are as follows:

[0016] Step 1: Image acquisition and preprocessing;

[0017] Step 2: Flexible structure segmentation based on disentangled cue representation;

[0018] Step 3: Deformation region detection based on dense matching;

[0019] Step 4: Inversion of flexible structure dynamic parameters.

[0020] Preferably, the step 1 is specifically:

[0021] Step 1-1: The image sequence collected in real time by the camera on board the spacecraft is recorded as I(t), where t is the time variable;

[0022] Step 1-2: Image enhancement;

[0023] Adaptive histogram equalization is used to enhance I(t) to improve the contrast and clarity of the image; the enhanced image is recorded as I_e(t);

[0024] Steps 1-3: Image denoising;

[0025] The image I_e(t) is denoised using the non-local mean filtering algorithm to suppress various noise interferences in the spatial environment; the denoised image is recorded as I_d(t);

[0026] Steps 1-4: Image stabilization;

[0027] The denoised continuous image sequence is aligned using feature point matching and homography transformation methods to eliminate image jitter caused by the spacecraft's own attitude change; the stabilized image sequence is recorded as I_s(t).

[0028] Preferably, the step 2 is specifically as follows:

[0029] Step 2-1: Decoupling prompt representation model construction;

[0030] Step 2-1-1: Decompose the image segmentation task into structure recognition subtask and boundary extraction subtask;

[0031] Step 2-1-2: The structure recognition subtask is guided by prior knowledge and defines the structural feature vector F_s with a dimension of d_s;

[0032] Step 2-1-3: The boundary extraction subtask constructs a boundary feature vector F_b with a dimension of d_b through the attention mechanism;

[0033] Step 2-1-4: The feature vectors of the structure recognition subtask and the boundary extraction subtask are directly concatenated to obtain a comprehensive feature representation F_c containing structure and boundary information, with a dimension of d_c, d_c = d_s + d_b;

[0034] Step 2-2: Domain Generalization Segmentation Network:

[0035] Step 2-2-1: Build a segmentation network based on the U-Net architecture, where the encoder part contains K convolution blocks, each of which contains two 3×3 convolution layers and one 2×2 maximum pooling layer;

[0036] Step 2-2-2: The decoder also contains K convolution blocks, each of which contains a 2×2 upsampling layer and two 3×3 convolution layers;

[0037] Step 2-2-3: Insert a domain generalization module between the third and fourth layers of the encoder of the segmentation network. Use feature alignment and style transfer techniques to adapt the domain features of the spacecraft image, so that the segmentation network can adapt to different lighting conditions and perspective changes.

[0038] Step 2-2-4: The mathematical expression of the domain generalization module is:

[0039] G(F) = N(F - μ(F)) / σ(F)

[0040] Where F is the input feature map, μ(F) is the mean of the feature map, σ(F) is the standard deviation of the feature map, and N is a learnable normalization parameter; G(F) represents the output feature map after domain generalization processing;

[0041] Step 2-3: Decouple prompt injection;

[0042] Step 2-3-1: Inject the comprehensive feature representation F_c obtained in step 2-1-4 into the skip connection layer of the segmentation network to enhance the network's perception of flexible structural features and output the initial segmentation result M_init(t);

[0043] Step 2-3-2: Injection method:

[0044] F_skip = F_orig + W * F_c

[0045] Where F_skip is the feature map after injection, F_orig is the feature map of the original skip connection layer, and W is the learnable weight matrix;

[0046] Step 2-4: Optimize the segmentation results;

[0047] Step 2-4-1: Use the CRF energy function to post-process the initial segmentation result M_init(t) and optimize the segmentation boundary;

[0048] Step 2-4-2: The CRF energy function is defined as:

[0049] E(x) = Σ_i ψ_u(x_i) + Σ_{i,j} ψ_p(x_i, x_j)

[0050] Where x is the label assignment of the pixel, ψ_u(.) is the unary potential function, ψ_p(.) is the binary potential function, Σ_i represents the sum of all pixels i in the image, x_i represents the label assignment of pixel i, Σ_{i,j} represents the sum of all adjacent pixel pairs (i,j), i represents the index of the current pixel, j represents the index of the pixel adjacent to pixel i, and x_j represents the label assignment of adjacent pixel j;

[0051] Step 2-4-3: Obtain the final segmentation mask M(t) by minimizing the CRF energy function.

[0052] Preferably, the step 3 is specifically:

[0053] Step 3-1: Template image preparation;

[0054] Step 3-1-1: Obtain the standard state template image I_temp of the flexible structure from the ground test before the spacecraft launch;

[0055] Step 3-1-2: Preprocess the template image I_temp by using histogram matching and brightness normalization to make it have similar illumination characteristics and grayscale distribution as the on-track image, eliminating the impact of imaging condition differences on matching accuracy;

[0056] Step 3-2: dense matching algorithm;

[0057] Step 3-2-1: Use dense matching algorithm to match the preprocessed template image I_temp with the current image I_s(t);

[0058] Step 3-2-2: Extract multi-scale features;

[0059] Use the pyramid structure to extract multi-scale features. Let the l-th layer feature be F_l, l∈{1,2,...,L}, where L represents the total number of pyramid layers.

[0060] Step 3-2-3: Feature matching;

[0061] Calculate the similarity matrix S_l between the template image feature F_l^temp and the current image feature F_l^curr:

[0062] S_l(p,q) =<F_l^temp(p), F_l^curr(q)> / (||F_l^temp(p)|| · ||F_l^curr(q)||)

[0063] Where p = (x1, y1) and q = (x2, y2) represent the pixel coordinates in the template image and the current image, respectively, <·,·> represents the inner product, and ||·|| represents the L2 norm.

[0064] Step 3-2-4: Local consistency constraint;

[0065] Introduce local consistency constraints to ensure smooth matching:

[0066] C(i,j) = Σ_{(i',j')∈N(i,j)} w(i,j,i',j') · ||D(i,j) - D(i',j')||²

[0067] Where D(i,j) is the displacement vector of position (i,j), N(i,j) is the neighborhood of (i,j), w(.) is the weight function, (i',j') represents the pixel coordinates in the neighborhood N(i,j), C(i,j) represents the matching cost or data constraint term of position (i,j), and D(i',j') is the displacement vector of position (i',j');

[0068] Step 3-2-5: Optimization solution;

[0069] The optimal match is solved by iterative algorithm to obtain the dense displacement field V(t);

[0070] Step 3-3: Extraction of deformation area;

[0071] Step 3-3-1: Calculate the displacement gradient tensor based on the dense displacement field V(t) , used to quantify the degree of local deformation;

[0072]

[0073] in, Represents the gradient operator, which is used to calculate the spatial derivative. V_x represents the displacement component of the displacement vector D in the x direction, i.e., the horizontal direction. V_y represents the displacement component of the displacement vector D in the y direction, i.e., the vertical direction.

[0074] Step 3-3-2: Calculate the norm of the displacement gradient tensor || ||, as a measure of the degree of deformation:

[0075]

[0076] Where T represents matrix transpose, tr(.) represents matrix trace, and sqrt(.) represents square root operation;

[0077] Step 3-3-3: Determine the deformation area through threshold processing:

[0078] R(t) = {(x,y) | S(t,x,y) > τ}

[0079] Where τ is the adaptive threshold, determined according to the statistical distribution; R(t) represents the deformation area at time t, which includes the coordinates of all pixels whose deformation exceeds the threshold;

[0080] Preferably, the step 4 is specifically as follows:

[0081] Step 4-1: Establishment of kinetic model;

[0082] The linear elastodynamic model is used to describe the vibration behavior of the flexible structure:

[0083]

[0084] Where Ma is the mass matrix, C is the damping matrix, Ka is the stiffness matrix, u is the displacement vector, and Fo(t) is the external excitation force; represents the second-order derivative of u, represents the first-order derivative of u;

[0085] Step 4-2: modal decomposition;

[0086] Perform characteristic orthogonal decomposition POD on the displacement field V(t) to extract the mode:

[0087] V(t) ≈ Σ_{i=1}^{n} φ_i · α_i(t)

[0088] Where φ_i is the i-th mode shape, α_i(t) is the time coefficient, n is the number of selected modes, and Σ_{i=1}^{n} represents the summation operation of all selected modes;

[0089] Step 4-3: Parameter identification;

[0090] Based on the extracted modes, the extended Kalman filter (EKF) algorithm is used to identify the dynamic parameters;

[0091] (1) The state vector is , where α is the modal coefficient vector, is the derivative of α, and θ is the kinetic parameter to be identified;

[0092] (2) The system equation is:

[0093]

[0094] Where f(X) is the state transfer function and wo is the process noise; is the first-order derivative of X;

[0095] (3) The measurement equation is:

[0096] Z = h(X) + v

[0097] Where h(X) is the measurement function, v is the measurement noise; Z represents the measurement vector;

[0098] (4) Update the estimated value through EKF iteration, and finally obtain the updated dynamic parameters θ̂;

[0099] Step 4-4: Parameter verification;

[0100] Step 4-4-1: Substitute the identified parameters into the dynamic model and perform forward simulation;

[0101] Step 4-4-2: Calculate the deviation between the simulation results and the actual observation to ensure the accuracy of parameter identification;

[0102] Step 4-4-3: If the deviation exceeds the preset threshold, update the prior information in the parameter identification process and re-identify.

[0103] Preferably, the dynamic parameters include natural frequency ω and damping ratio ζ.

[0104] An electronic device comprises: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the electronic device performs the above-mentioned target dynamic parameter autonomous identification method.

[0105] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned method for autonomously identifying target dynamic parameters.

[0106] A chip includes: a processor for calling and running a computer program from a memory, so that a device equipped with the chip executes the above-mentioned target dynamic parameter autonomous identification method.

[0107] A computer program product includes a computer storage medium storing a computer program, wherein the computer program includes instructions that can be executed by at least one processor, and when the instructions are executed by the at least one processor, the above-mentioned target dynamic parameter autonomous identification method is implemented.

[0108] The beneficial effects of the present invention are as follows:

[0109] (1) By adopting a domain generalization approach based on decoupled cue representation, this paper can address the poor adaptability of traditional image processing methods in complex spatial environments and achieve stable recognition and segmentation of flexible structures under varying lighting conditions and viewing angles. Decoupled cue representation decomposes the segmentation task into two subtasks: structure recognition and boundary extraction. This effectively improves segmentation accuracy and generalization capability, enabling the system to adapt to complex environmental factors such as periodic illumination changes in spacecraft orbits.

[0110] (2) Because the present invention uses a dense matching algorithm to accurately align the template image and the current image, it can solve the problem of traditional sparse feature point matching in matching a single texture area, and achieve high-precision detection of deformation areas in flexible structures. Dense matching establishes a correspondence for each pixel point, forming a continuous deformation field description, which can accurately capture deformation caused by small disturbances. It is particularly suitable for flexible structures with sparse textures such as spacecraft solar panels. Compared with traditional methods such as SIFT and SURF, the detection accuracy is improved by more than 50%.

[0111] (3) Because the present invention adopts a dynamic parameter identification method based on extended Kalman filtering, it can solve the problem of poor stability of traditional parameter identification methods when measurement noise is large, and achieve robust estimation of the dynamic parameters of flexible structures. This method reduces the dimensionality of the problem through modal decomposition and effectively suppresses the influence of measurement noise through iterative updates, keeping the parameter identification error within 3%, providing a reliable basis for subsequent health status assessment.

[0112] (4) Because the present invention adopts a health status assessment method that integrates multimodal information, it can solve the problem of incomplete evaluation of a single indicator and achieve a comprehensive assessment and early warning of the health status of flexible structures. By comprehensively considering multiple dimensions such as dynamic parameter changes, trend analysis, and remaining life prediction, a complete health status assessment system is formed, which can detect potential failure risks in advance and provide protection for the safe operation of spacecraft. DETAILED DESCRIPTION

[0113] The present invention is further described below with reference to the embodiments.

[0114] This paper provides a method for autonomously identifying and monitoring the dynamic state of spacecraft flexible structures based on decoupled cue representation. This method achieves domain generalization through decoupled cue representation, adapts to complex space environment changes, and enables high-precision monitoring of the deformation and vibration of spacecraft flexible structures (such as satellite sail panels). It also infers the dynamic parameters of the flexible structures, providing support for spacecraft health status assessment and fault warning. The specific steps are as follows:

[0115] Step 1: Image acquisition and preprocessing;

[0116] Step 1-1: The image sequence collected in real time by the camera on board the spacecraft is recorded as I(t), where t is the time variable;

[0117] Step 1-2: Image enhancement;

[0118] Adaptive histogram equalization is used to enhance I(t) to improve the contrast and clarity of the image; the enhanced image is recorded as I_e(t);

[0119] Steps 1-3: Image denoising;

[0120] The image I_e(t) is denoised using the non-local mean filtering algorithm to suppress various noise interferences in the spatial environment; the denoised image is recorded as I_d(t);

[0121] Steps 1-4: Image stabilization;

[0122] The denoised continuous image sequence is aligned using feature point matching and homography transformation methods to eliminate image jitter caused by the spacecraft's own attitude changes; the stabilized image sequence is recorded as I_s(t);

[0123] The output of this step is the preprocessed image sequence I_s(t), which serves as the input of the next step.

[0124] Step 2: Flexible structure segmentation based on disentangled cue representation;

[0125] The input of this step is the preprocessed image sequence I_s(t), and the goal is to accurately segment the flexible structure (satellite sail panel) area from the image.

[0126] Step 2-1: Decoupling prompt representation model construction;

[0127] Step 2-1-1: Decompose the image segmentation task into structure recognition subtask and boundary extraction subtask;

[0128] Step 2-1-2: The structure recognition subtask is guided by prior knowledge and defines the structural feature vector F_s with a dimension of d_s;

[0129] Step 2-1-3: The boundary extraction subtask constructs a boundary feature vector F_b with a dimension of d_b through the attention mechanism;

[0130] Step 2-1-4: The feature vectors of the structure recognition subtask and the boundary extraction subtask are directly concatenated to obtain a comprehensive feature representation F_c containing structure and boundary information, with a dimension of d_c, d_c = d_s + d_b;

[0131] Step 2-2: Domain Generalization Segmentation Network:

[0132] Step 2-2-1: Build a segmentation network based on the U-Net architecture, where the encoder part contains K convolution blocks, each of which contains two 3×3 convolution layers and one 2×2 maximum pooling layer;

[0133] Step 2-2-2: The decoder also contains K convolution blocks, each of which contains a 2×2 upsampling layer and two 3×3 convolution layers;

[0134] Step 2-2-3: Insert a domain generalization module between the third and fourth layers of the encoder of the segmentation network. Use existing feature alignment and style transfer techniques to adapt to the specific domain features of spacecraft images, so that the segmentation network can adapt to factors such as different lighting conditions and viewpoint changes.

[0135] Step 2-2-4: The mathematical expression of the domain generalization module is:

[0136] G(F) = N(F - μ(F)) / σ(F)

[0137] Step 2-3: Decouple prompt injection;

[0138] Step 2-3-1: Inject the comprehensive feature representation F_c obtained in step 2-1-4 into the skip connection layer of the segmentation network to enhance the network's perception of flexible structural features and output the initial segmentation result M_init(t);

[0139] Step 2-3-2: Injection method:

[0140] F_skip = F_orig + W * F_c

[0141] Step 2-4: Optimize the segmentation results;

[0142] Step 2-4-1: Use the CRF energy function to post-process the initial segmentation result M_init(t) and optimize the segmentation boundary;

[0143] Step 2-4-2: The CRF energy function is defined as:

[0144] E(x) = Σ_i ψ_u(x_i) + Σ_{i,j} ψ_p(x_i, x_j)

[0145] Step 2-4-3: Obtain the final segmentation mask M(t) by minimizing the CRF energy function; the classic graph cut algorithm is used to solve the global optimal solution of the energy function.

[0146] Step 3: Deformation region detection based on dense matching;

[0147] The inputs to this step are the final segmentation mask M(t) of the flexible structure and the preprocessed image sequence I_s(t). The goal is to detect the deformed regions of the flexible structure using a dense matching algorithm. This method chooses a dense matching algorithm over the traditional sparse keypoint matching method based on the following considerations: dense matching can establish a correspondence between each pixel, accurately identifying solar panel deformation caused by small perturbations; compared to sparse keypoint-based methods such as SIFT and SURF, dense matching provides a higher resolution and accuracy of the deformation field; and dense matching is more adaptable to the single texture and sparse keypoints of spacecraft flexible structures, and can still establish reliable matching relationships in areas lacking distinct features.

[0148] Step 3-1: Template image preparation;

[0149] Step 3-1-1: Obtain the standard state template image I_temp of the flexible structure from the ground test before the spacecraft launch;

[0150] Step 3-1-2: Preprocess the template image I_temp by using histogram matching and brightness normalization to make it have similar illumination characteristics and grayscale distribution as the on-track image, eliminating the impact of imaging condition differences on matching accuracy;

[0151] Step 3-2: dense matching algorithm;

[0152] Step 3-2-1: Use dense matching algorithm to match the preprocessed template image I_temp with the current image I_s(t);

[0153] Step 3-2-2: Extract multi-scale features;

[0154] Use the pyramid structure to extract multi-scale features. Let the l-th layer feature be F_l, l∈{1,2,...,L}, where L represents the total number of pyramid layers.

[0155] Step 3-2-3: Feature matching;

[0156] Calculate the similarity matrix S_l between the template image feature F_l^temp and the current image feature F_l^curr:

[0157] S_l(p,q) =<F_l^temp(p), F_l^curr(q)> / (||F_l^temp(p)|| · ||F_l^curr(q)||)

[0158] Step 3-2-4: Local consistency constraint;

[0159] Introduce local consistency constraints to ensure smooth matching:

[0160] C(i,j) = Σ_{(i',j')∈N(i,j)} w(i,j,i',j') · ||D(i,j) - D(i',j')||²

[0161] Step 3-2-5: Optimization solution;

[0162] The optimal match is solved by iterative algorithm to obtain the dense displacement field V(t);

[0163] Step 3-3: Extraction of deformation area;

[0164] Step 3-3-1: Calculate the displacement gradient tensor based on the dense displacement field V(t) , used to quantify the degree of local deformation;

[0165]

[0166] Step 3-3-2: Calculate the norm of the displacement gradient tensor || ||, as a measure of the degree of deformation:

[0167]

[0168] Step 3-3-3: Determine the deformation area through threshold processing:

[0169] R(t) = {(x,y) | S(t,x,y) > τ}

[0170] The output of this step is the deformation region R(t) and displacement field V(t), which serve as input to the next step. Compared with traditional sparse keypoint-based methods, this step can provide a more accurate and continuous description of the deformation field, especially for detecting deformation caused by small perturbations and measuring deformation in sparsely textured areas.

[0171] Step 4: Inversion of flexible structure dynamic parameters;

[0172] The input of this step is the deformation area R(t) and the displacement field V(t). The goal is to invert the dynamic parameters of the flexible structure.

[0173] Step 4-1: Establishment of kinetic model;

[0174] The linear elastodynamic model is used to describe the vibration behavior of the flexible structure:

[0175]

[0176] Step 4-2: modal decomposition;

[0177] Perform characteristic orthogonal decomposition POD on the displacement field V(t) to extract the mode:

[0178] V(t) ≈ Σ_{i=1}^{n} φ_i · α_i(t)

[0179] Step 4-3: Parameter identification;

[0180] Based on the extracted modes, the extended Kalman filter (EKF) algorithm is used to identify the dynamic parameters;

[0181] (1) The state vector is , where α is the modal coefficient vector, is the derivative of α, and θ is the kinetic parameter to be identified;

[0182] (2) The system equation is:

[0183]

[0184] (3) The measurement equation is:

[0185] Z = h(X) + v

[0186] (4) Update the estimated value through EKF iteration, and finally obtain the updated dynamic parameters θ̂;

[0187] Step 4-4: Parameter verification;

[0188] Step 4-4-1: Substitute the identified parameters into the dynamic model and perform forward simulation;

[0189] Step 4-4-2: Calculate the deviation between the simulation results and the actual observation to ensure the accuracy of parameter identification.

[0190] Step 4-4-3: If the deviation exceeds the preset threshold, update the prior information in the parameter identification process and re-identify.

[0191] The output of this step is the dynamic parameters θ̂ of the flexible structure, including the natural frequency ω and the damping ratio ζ.

[0192] Step 5: Simulation verification;

[0193] The Blender simulation platform was used to perform 3D modeling and disturbance simulation, and simulated images were obtained to verify the method of the present invention.

[0194] Step 5-1: Setting up the simulation environment;

[0195] Step 5-1-1: Use the Blender simulation platform to build a 3D model of the spacecraft, including the spacecraft body and flexible structures such as solar panels;

[0196] Step 5-1-2: Accurately model the physical properties of the flexible structure based on the actual dimensions and material parameters of the spacecraft;

[0197] Step 5-1-3: Simulate the lighting conditions in the space environment, including direct sunlight, earth reflection, and shadows;

[0198] Step 5-2: Perturbation simulation;

[0199] Step 5-2-1: Design different types of disturbance scenarios in Blender, including slight vibration, large deformation, and various failure modes;

[0200] Step 5-2-2: For solar panels, simulate various deformation states caused by factors such as minor collisions, thermal cycles, and material aging;

[0201] Step 5-2-3: Record the disturbance parameters during the simulation process as true value data for algorithm verification.

[0202] Step 5-3: Image generation;

[0203] Step 5-3-1: Set up a virtual camera in the simulation environment to simulate the parameters and position of the camera actually carried on the spacecraft;

[0204] Step 5-3-2: Capture a sequence of images of the flexible structure from different viewing angles and lighting conditions;

[0205] Step 5-3-3: Artificially add different levels of noise to the image to simulate the imaging interference in the actual space environment;

[0206] Step 5-4: Algorithm verification;

[0207] Step 5-4-1: applying the method of the present invention to the simulated image data;

[0208] Step 5-4-2: Calculate the error between the output result and the simulation true value to evaluate the algorithm performance;

[0209] Step 5-4-3: Test the robustness of the algorithm under different perturbation types, lighting conditions, and noise levels.

[0210] Through verification on the Blender simulation platform, the performance and applicability of the method of the present invention can be comprehensively evaluated without relying on actual spacecraft on-orbit tests, providing a reliable basis for the practical application of the algorithm.

[0211] Example:

[0212] To verify the effectiveness of the proposed method, a 3D model of a spacecraft's flexible structure was constructed using the Blender simulation platform. Various disturbance scenarios in space environments were simulated, and a large amount of test image data was generated. The results were then compared with existing methods. Table 1 shows the test results.

[0213] Table 1 Test results

[0214] method Segmentation accuracy (%) Deformation detection accuracy (mm) Parameter identification error (%) Computation time (ms) Method of the present invention 95.8 0.35 2.3 45 SIFT keypoint based method 87.2 0.82 5.6 30 CNN-based segmentation and detection methods 92.4 0.51 3.7 68 Finite element model-based methods 90.3 0.44 3.2 125

[0215] As can be seen in Table 1, the proposed method outperforms existing methods in terms of segmentation accuracy, deformation detection accuracy, and parameter identification error. In particular, deformation detection accuracy is improved by 57% compared to SIFT keypoint-based methods, primarily due to the dense matching algorithm's advantage in processing sparsely textured areas. While computation time is slightly longer than that of SIFT keypoint-based methods, it is significantly shorter than finite element-based methods, meeting the requirements for real-time on-orbit monitoring of spacecraft.

[0216] In the simulation experiment, the method was also tested in challenging scenarios such as extreme lighting changes, tiny disturbances (<0.5mm) and extremely sparse textures (<5 identifiable feature points per square meter). The method still maintained good performance and demonstrated excellent environmental adaptability and precision stability.

[0217] This paper proposes a method for autonomously identifying and monitoring the dynamic state of spacecraft flexible structures based on decoupled cue representation. Through image segmentation, dense matching, and dynamic parameter inversion, this method achieves high-precision monitoring of the deformation and vibration of spacecraft flexible structures. Experimental results demonstrate that this method outperforms existing methods in terms of segmentation accuracy, deformation detection precision, and parameter identification error, providing an effective means for spacecraft health assessment and fault warning.

Claims

1. A method for autonomous identification of target dynamic parameters based on decoupled prompt representation, characterized in that: The steps include: Step 1: Image acquisition and preprocessing; Step 2: Flexible structure segmentation based on disentangled cue representation; Step 2-1: Decoupling prompt representation model construction; Step 2-1-1: Decompose the image segmentation task into structure recognition subtask and boundary extraction subtask; Step 2-1-2: The structure recognition subtask is guided by prior knowledge and defines the structural feature vector F_s with a dimension of d_s; Step 2-1-3: The boundary extraction subtask constructs a boundary feature vector F_b with a dimension of d_b through the attention mechanism; Step 2-1-4: The feature vectors of the structure recognition subtask and the boundary extraction subtask are directly concatenated to obtain a comprehensive feature representation F_c containing structure and boundary information, with a dimension of d_c, d_c = d_s + d_b; Step 2-2: Domain Generalization Segmentation Network: Step 2-2-1: Build a segmentation network based on the U-Net architecture, where the encoder part contains K convolution blocks, each of which contains two 3×3 convolution layers and one 2×2 maximum pooling layer; Step 2-2-2: The decoder also contains K convolution blocks, each of which contains a 2×2 upsampling layer and two 3×3 convolution layers; Step 2-2-3: Insert a domain generalization module between the third and fourth layers of the encoder of the segmentation network. Use feature alignment and style transfer techniques to adapt the domain features of the spacecraft image, so that the segmentation network can adapt to different lighting conditions and perspective changes. Step 2-2-4: The mathematical expression of the domain generalization module is: (F) = N(F-μ(F)) / σ(F) Where F is the input feature map, μ(F) is the mean of the feature map, σ(F) is the standard deviation of the feature map, and N is a learnable normalization parameter; (F) represents the output feature map after domain generalization processing; Step 2-3: Decouple prompt injection; Step 2-3-1: Inject the comprehensive feature representation F_c obtained in step 2-1-4 into the skip connection layer of the segmentation network to enhance the network's perception of flexible structural features and output the initial segmentation result M_init(t); Step 2-3-2: Injection method: F_skip=F_orig+W*F_c Where F_skip is the feature map after injection, F_orig is the feature map of the original skip connection layer, and W is the learnable weight matrix; Step 2-4: Optimize the segmentation results; Step 2-4-1: Use the CRF energy function to post-process the initial segmentation result M_init(t) and optimize the segmentation boundary; Step 2-4-2: The CRF energy function is defined as: E(x)=Σ_i·ψ_u(x_i)+Σ_{i,j}·ψ_p(x_i,x_j) Where x is the label assignment of the pixel, ψ_u(.) is the unary potential function, ψ_p(.) is the binary potential function, Σ_i represents the sum of all pixels i in the image, x_i represents the label assignment of pixel i, Σ_{i,j} represents the sum of all adjacent pixel pairs (i,j), i represents the index of the current pixel, j represents the index of the pixel adjacent to pixel i, and x_j represents the label assignment of adjacent pixel j; Step 2-4-3: Obtain the final segmentation mask M(t) by minimizing the CRF energy function; Step 3: Deformation region detection based on dense matching; Step 4: Inversion of flexible structure dynamic parameters.

2. The method for autonomous identification of target dynamic parameters based on decoupling prompt representation according to claim 1, characterized in that: The step 1 is specifically as follows: Step 1-1: The image sequence collected in real time by the camera on board the spacecraft is recorded as I(t), where t is the time variable; Step 1-2: Image enhancement; Adaptive histogram equalization is used to enhance I(t) to improve the contrast and clarity of the image; the enhanced image is recorded as I_e(t); Steps 1-3: Image denoising; The image I_e(t) is denoised using the non-local mean filtering algorithm to suppress various noise interferences in the spatial environment; the denoised image is recorded as I_d(t); Steps 1-4: Image stabilization; The denoised continuous image sequence is aligned using feature point matching and homography transformation methods to eliminate image jitter caused by the spacecraft's own attitude change; the stabilized image sequence is recorded as I_s(t).

3. The method for autonomous identification of target dynamic parameters based on decoupling hint representation according to claim 2, characterized in that: The step 3 is specifically as follows: Step 3-1: Template image preparation; Step 3-1-1: Obtain the standard state template image I_temp of the flexible structure from the ground test before the spacecraft launch; Step 3-1-2: Preprocess the template image I_temp by using histogram matching and brightness normalization to make it have similar illumination characteristics and grayscale distribution as the on-track image, eliminating the impact of imaging condition differences on matching accuracy; Step 3-2: dense matching algorithm; Step 3-2-1: Use dense matching algorithm to match the preprocessed template image I_temp with the current image I_s(t); Step 3-2-2: Extract multi-scale features; Use the pyramid structure to extract multi-scale features. Let the l-th layer feature be F_l, l∈{1,2,...,L}, where L represents the total number of pyramid layers. Step 3-2-3: Feature matching; Calculate the similarity matrix S_l between the template image feature F_l^temp and the current image feature F_l^curr: S_l(p,q)=<F_l^temp(p),F_l^curr(q)> / (||F_l^temp(p)||·||F_l^curr(q)||) Where p = (x1, y1) and q = (x2, y2) represent the pixel coordinates in the template image and the current image, respectively, <·,·> represents the inner product, and ||·|| represents the L2 norm. Step 3-2-4: Local consistency constraint; Introduce local consistency constraints to ensure smooth matching: C(i,j)=Σ_{(i′,j′)∈N(i,j)}w(i,j,i′,j′)·||D(i,j)-D(i′,j′)||2 Where D(i,j) is the displacement vector of position (i,j), N(i,j) is the neighborhood of (i,j), w(.) is the weight function, (i′,j′) represents the pixel coordinates in the neighborhood N(i,j), C(i,j) represents the matching cost or data constraint term of position (i,j), and D(i′,j′) is the displacement vector of position (i′,j′); Step 3-2-5: Optimization solution; The optimal match is solved by iterative algorithm to obtain the dense displacement field V(t); Step 3-3: Extraction of deformation area; Step 3-3-1: Calculate the displacement gradient tensor based on the dense displacement field V(t) Used to quantify the degree of local deformation; in, Represents the gradient operator, which is used to calculate the spatial derivative. V_x represents the displacement component of the displacement vector D in the x direction, i.e., the horizontal direction. V_y represents the displacement component of the displacement vector D in the y direction, i.e., the vertical direction. Step 3-3-2: Calculate the norm of the displacement gradient tensor As a measure of deformation: Where T represents matrix transpose, tr(.) represents matrix trace, and sqrt(.) represents square root operation; Step 3-3-3: Determine the deformation area through threshold processing: R(t)={(x,y)|S(t,x,y)>τ} Where τ is the adaptive threshold, which is determined according to the statistical distribution; R(t) represents the deformation area at time t, which includes the coordinates of all pixels whose deformation exceeds the threshold.

4. The method for autonomous identification of target dynamic parameters based on decoupling hint representation according to claim 3, characterized in that: The step 4 is specifically as follows: Step 4-1: Establishment of kinetic model; The linear elastodynamic model is used to describe the vibration behavior of the flexible structure: Where Ma is the mass matrix, C is the damping matrix, Ka is the stiffness matrix, u is the displacement vector, and Fo(t) is the external excitation force; represents the second-order derivative of u, represents the first-order derivative of u; Step 4-2: modal decomposition; Perform characteristic orthogonal decomposition POD on the displacement field V(t) to extract the mode: Where, is the i-th mode shape, α_i(t) is the time coefficient, n is the number of selected modes, and Σ_{i=1}^{n} represents the summation operation of all selected modes; Step 4-3: Parameter identification; Based on the extracted modes, the extended Kalman filter (EKF) algorithm is used to identify the dynamic parameters; (1) The state vector is Where α is the modal coefficient vector, is the derivative of α, and θ is the kinetic parameter to be identified; (2) The system equation is: Where f(X) is the state transfer function and wo is the process noise; is the first-order derivative of X; (3) The measurement equation is: Z=h(X)+v Where h(X) is the measurement function, v is the measurement noise; Z represents the measurement vector; (4) Update the estimated value through EKF iteration to finally obtain the updated dynamic parameters Step 4-4: Parameter verification; Step 4-4-1: Substitute the identified parameters into the dynamic model and perform forward simulation; Step 4-4-2: Calculate the deviation between the simulation results and the actual observation to ensure the accuracy of parameter identification; Step 4-4-3: If the deviation exceeds the preset threshold, update the prior information in the parameter identification process and re-identify.

5. The method for autonomous identification of target dynamic parameters based on decoupling hint representation according to claim 4, characterized in that: The dynamic parameters include the natural frequency ω and the damping ratio ζ.

6. An electronic device, characterized in that: include: processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the electronic device performs the method according to any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

8. A chip, characterized in that: include: A processor, configured to call and run a computer program from a memory, so that a device equipped with the chip executes the method according to any one of claims 1 to 5.

9. A computer program product, characterized in that The computer program product comprises a computer storage medium storing a computer program, wherein the computer program comprises instructions executable by at least one processor, and when the instructions are executed by the at least one processor, the method according to any one of claims 1 to 5 is implemented.

Images