A semi-parametric compensation method for unmodeled pointing errors in optical communication with a moving platform
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INST OF OPTICS & ELECTRONICS CHINESE ACAD OF SCI
- Filing Date
- 2026-03-09
- Publication Date
- 2026-06-09
Smart Images
Figure CN122179040A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of space laser communication and optomechanical control technology, specifically relating to a semi-parametric compensation method for unmodeled pointing error in optical communication of a moving platform, applicable to compensating for pointing error caused by unmodeled error in optical communication systems of moving platforms. Background Technology
[0002] In space laser communication systems, accurate pointing of the moving optical communication terminal is crucial for achieving stable and efficient communication. Currently, mainstream methods primarily rely on mechanistic modeling and parameter estimation, establishing geometric models and employing algorithms such as least squares to identify and compensate for installation errors, structural errors, and constant attitude deviations. These methods have clear physical meaning and can effectively improve the system's basic pointing accuracy. However, even after mechanistic model compensation, complex unmodeled error sources and nonlinear residuals still exist in the system, such as dynamic deformation, micro-vibrations, thermal effects, and attitude estimation errors. These errors typically exhibit strong spatial correlation, local abrupt changes, and irregular global distribution, making them difficult to accurately describe using simple linear or polynomial models.
[0003] Existing research on compensating for the aforementioned unmodeled errors mainly falls into two categories: First, nonparametric modeling methods, such as multinomial regression, spline interpolation, or neural networks, directly utilize data to learn error mappings, but often lack physical interpretability and exhibit poor out-of-sample generalization ability. Second, semiparametric modeling methods, which superimpose data-driven compensation terms onto the mechanistic model, such as combining biased least squares with local interpolation, or introducing kernel functions and semi-supervised learning to correct residuals. These methods, to a certain extent, integrate the advantages of both mechanistic and data-driven approaches, improving compensation accuracy.
[0004] Despite this, existing methods still have significant limitations: most techniques rely on traditional interpolation or shallow learning models, failing to systematically and effectively model the complex spatial correlations and dynamic dependencies of errors. This leads to a significant decrease in the model's prediction accuracy and robustness in areas with complex motion trajectories, varying operating conditions, or sparse samples. With the development of graph neural network technology, using graph structures to model spatial correlations has become a new research direction. Therefore, it is necessary to introduce inductive graph attention networks, which can explicitly characterize spatial correlations and possess strong nonlinear learning capabilities, based on the existing semiparametric framework. This would enable more accurate, robust, and adaptable online modeling and compensation of unmodeled errors in dynamic environments, further improving the pointing performance and system reliability of motion optical communication terminals under complex tasks. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention provides a semi-parametric compensation method for unmodeled pointing errors in optical communication of motion platforms. It utilizes an inductive graph attention network to model spatially correlated residuals online and additively fuses them with a mechanistic model to achieve high-precision, robust, and extrapolable online pointing error compensation under complex dynamic conditions.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A semi-parametric compensation method for unmodeled pointing error in optical communication of a motion platform includes:
[0008] Step 1: Construct a spatial association graph: Use multiple measurement points and prediction points as graph nodes. Calculate the angular distance between nodes based on the unit direction vector of the nodes in the platform coordinate system. Calculate the adjacency weight matrix between nodes based on the angular distance. Characterize the spatial association graph structure based on the adjacency weight matrix.
[0009] Step 2: Define node features and training objectives: Extract sine and cosine codes of the platform attitude angle and encoder angle guidance value for each graph node, and combine them to form a node feature vector; calculate the difference between the total pointing error at the measurement point and the output of the mechanism model, and use it as the training objective for the corresponding graph node;
[0010] Step 3: Train the spatial residual prediction model: Based on the graph structure, the node feature vectors, and the training objective, an inductive multi-head graph attention network is used for training. During the training process, at least one strategy including residual standardization, cross-validation, regularization, and early stopping is adopted to obtain a spatial residual prediction model that can predict unmodeled residuals.
[0011] Step 4: Perform semi-parametric fusion compensation: For the real-time prediction point to be compensated, the physical error estimate calculated by the mechanism model is additively fused with the residual prediction value obtained after inputting the real-time prediction point to be compensated as a new node into the spatial residual prediction model, and the final pointing error compensation value is output to complete the online compensation.
[0012] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned semi-parametric compensation method for unmodeled pointing errors in optical communication of a motion platform.
[0013] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned semi-parametric compensation method for unmodeled pointing errors in optical communication of a motion platform.
[0014] The beneficial effects of this invention are as follows:
[0015] Strong spatial correlation modeling capability: By introducing a graph structure based on angular distance and a Gaussian kernel weighted adjacency matrix, this invention can explicitly and quantitatively characterize the spatial geometric correlation between measurement points and prediction points, effectively capture the local similarity and global distribution law of unmodeled errors on the sphere, and overcome the shortcomings of traditional methods in modeling spatial correlation.
[0016] Dynamic prediction and strong generalization ability: By adopting an inductive graph attention network, the model can not only learn residual patterns on the training set, but also perform dynamic reasoning on new and unseen prediction points through the attention mechanism. It can quickly output by only constructing its features and adjacency relationships, achieving good out-of-sample generalization and spatial extrapolation ability.
[0017] The semi-parametric framework offers significant advantages: This method innovatively integrates a data-driven inductive graph attention network model with a physical mechanism model, forming a semi-parametric compensation structure. This framework maintains the physical interpretability and stability of the mechanistic model while incorporating the powerful nonlinear fitting and adaptive capabilities of the data model, thus balancing accuracy and reliability.
[0018] High robustness and practicality: By comprehensively utilizing training strategies such as residual standardization, K-fold cross-validation, regularization, and early stopping, the model's training stability and generalization performance under conditions of small samples and non-uniform data are significantly enhanced. The computational complexity of the online compensation process is controllable, meeting real-time requirements and possessing good engineering practical value. Attached Figure Description
[0019] Figure 1 This is a flowchart of a semi-parametric compensation method for unmodeled pointing error in optical communication of a motion platform according to the present invention.
[0020] Figure 2 This is a schematic diagram of the platform coordinate system and navigation coordinate system. Detailed Implementation
[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0022] like Figure 1 As shown, this invention provides a semi-parametric compensation method for unmodeled pointing errors in optical communication of a motion platform, the method comprising:
[0023] Step 1: Construct a spatial association graph: Use multiple measurement points and prediction points as graph nodes. Calculate the angular distance between nodes based on the unit direction vector of the nodes in the platform coordinate system. Calculate the adjacency weight matrix between nodes based on the angular distance. Characterize the spatial association graph structure based on the adjacency weight matrix.
[0024] Step 2: Define node features and training objectives: Extract sine and cosine codes of the platform attitude angle and encoder angle guidance value for each graph node, and combine them to form a node feature vector; calculate the difference between the total pointing error at the measurement point and the output of the mechanism model, and use it as the training objective for the corresponding graph node;
[0025] Step 3: Train the spatial residual prediction model: Based on the graph structure, the node feature vectors, and the training objective, an inductive multi-head graph attention network is used for training; during the training process, various model generalization strategies are employed to obtain a spatial residual prediction model that can predict unmodeled residuals.
[0026] Step 4: Perform semi-parametric fusion compensation: For the real-time prediction point to be compensated, the physical error estimate calculated by the mechanism model is additively fused with the residual prediction value obtained after inputting the real-time prediction point to be compensated as a new node into the spatial residual prediction model, and the final pointing error compensation value is output to complete the online compensation.
[0027] Specifically, in step 1, the measurement point is a stellar point observable in different directions from the moving optical communication terminal, and the prediction point is a target point in different directions from the moving optical communication terminal where the pointing error needs to be predicted. The direction vector of the node is the unit direction vector of the node relative to the moving optical communication terminal in the platform coordinate system. Calculated using the following formula:
[0028]
[0029]
[0030] in, , , The attitude transformation matrix is composed of the heading, horizontal, and roll angles measured by the inertial navigation equipment mounted on the motion optical communication terminal, respectively:
[0031]
[0032] in, , , These are the heading, pitch, and roll angles measured by the inertial navigation system mounted on the motion optical communication terminal, respectively. The unit direction vector in the navigation coordinate system calculated for the node based on information such as ephemeris, platform latitude and longitude, and time; and The azimuth and pitch guidance values of the node in the platform coordinate system are given without considering the influence of errors; the formula for calculating the angular distance and the adjacency weight matrix between the nodes is as follows:
[0033]
[0034] in, , Let be the unit direction vectors of nodes i and j relative to the platform coordinate system. The width of the Gaussian kernel. Let be the angular distance between node i and node j. Let represent the weights between node i and node j. After obtaining the adjacency matrix, kNN is used for truncation to construct the neighbor set of each node. The calculation formula is as follows:
[0035]
[0036] in, For nodes The set of neighbors of node i, where formula (5) represents the k nodes closest to node i. The graph structure is as follows: The platform coordinate system and navigation coordinate system are shown as follows. Figure 2 As shown, Figure 2 middle Using the Earth-centered Earth-fixed coordinate system, For the navigation coordinate system, For the platform coordinate system, For encoder zero-position coordinate system, For the azimuth coordinate system, Using the pitch axis coordinate system, This is the line-of-sight coordinate system.
[0037] Specifically, in step 2, the node feature vector is:
[0038]
[0039] in, , , The attitude data obtained by the motion optical communication terminal carrying inertial navigation measurement when measuring node i. , This is the encoder angle guide value when measuring node i. Additionally, the residual used for network training is:
[0040]
[0041] in, The pointing error obtained by measurement, The pointing error estimated by the mechanistic model. The residuals are estimated for the mechanistic model. The training objective of the model is to minimize the residual estimation error.
[0042]
[0043] in, The target value for training; The number of measurement nodes; The measurement point set is defined as follows: The measurement pointing error is the difference between the azimuth and pitch encoder readings and the azimuth and pitch guidance values when the line of sight of the motion optical communication terminal is aligned with the target. The mechanism model is a pointing correction model established based on the geometry of the motion optical communication terminal, which includes inertial navigation system installation errors, axis non-orthogonality errors, and aiming errors.
[0044] Specifically, in step 3, the structure of the inductive multi-head graph attention network is an L-layer stack with M attention heads per layer, and the hidden dimension and number of heads are selected through cross-validation. The node update formula for the multi-head attention mechanism of the inductive multi-head graph attention network is:
[0045]
[0046] in, For the first The node, the Features of the layer; For the first The node, the Layer characteristics; when , Represents a node initial features It is the first The first point of attention, the first Layer feature mapping matrix; It is an activation function. It is a node With nodes The attention weights are learned through a mechanism as follows:
[0047]
[0048] in, The attention vector is the first... A trainable parameter vector for a layer. For the first Layer adjacent nodes The relevance score, Here, is a ReLU activation function with a negative slope, and || denotes vector concatenation. The output layer averages the multi-head results to achieve steady-state processing.
[0049]
[0050] The inductive multi-head graph attention network uses the residual after compensation by a mechanistic model. As the label, the objective function is shown in Equation (8). During the network training phase: first, the unit direction vector of each observed star is calculated using Equation (1); then, the weight matrix A between nodes is calculated using Equation (4); and finally, the node neighbor set is constructed using Equation (5). Next, the initial feature vectors of each observed star are constructed using formula (6). Then, the initial feature vectors are input into the network as shown in formula (9). The weight learning mechanism of the network is shown in formula (10), and the output layer uses the average result as shown in formula (11). Finally, the entire network uses the Adam optimizer, and the learning rate and batch size are determined by K-fold cross-validation. The above is the training process of the inductive multi-head graph attention network.
[0051] The inductive multi-head graph attention network introduces weight decay (L2 regularization) and sets early stopping (stopping if there is no improvement after several rounds of validation set loss). The hyperparameter selection and generalization strategy of the inductive multi-head graph attention network are as follows: The number of layers L, number of heads M, hidden dimension d, learning rate, weight decay, batch size, and early stopping threshold are selected using K-fold cross-validation; hierarchical cross-validation is performed for different sky regions / pose segments to ensure the model has extrapolation capabilities for unseen regions; when observations are abnormal or samples are sparse, the model is maintained... To avoid jitter, the weights of data-driven terms are gradually increased, prioritizing them. While maintaining a uniform spatial distribution, the samples are divided into training, validation, and test sets.
[0052] Specifically, in step 4, the additive fusion method is: online calculation to obtain... , and the output of the mechanism model Additive fusion yields Weighted fusion can be used to adapt to different working conditions.
[0053]
[0054] Specifically, in step 4, during the network prediction stage: first, the unit direction vector of the prediction point is calculated using formula (1); then, the new weight matrix A between nodes is calculated using formula (4); next, the set of the node neighbors of the prediction point is constructed using formula (5); then, the initial feature vector of the prediction point is constructed using formula (6); then, the initial feature vector of the prediction node is input into the trained network, and the model output result is obtained. Finally, using formula (12), the pointing error compensation value of the predicted point is obtained. The above describes the prediction process of the inductive multi-head graph attention network.
[0055] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned semi-parametric compensation method for unmodeled pointing errors in optical communication of a motion platform.
[0056] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned semi-parametric compensation method for unmodeled pointing errors in optical communication of a motion platform.
[0057] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A semi-parametric compensation method for unmodeled pointing error in optical communication of a motion platform, characterized in that, include: Step 1: Construct a spatial association graph: Use multiple measurement points and prediction points as graph nodes. Calculate the angular distance between nodes based on the unit direction vector of the nodes in the platform coordinate system. Calculate the adjacency weight matrix between nodes based on the angular distance. Characterize the spatial association graph structure based on the adjacency weight matrix. Step 2: Define node features and training objectives: Extract sine and cosine codes of the platform attitude angle and encoder angle guidance value for each graph node, and combine them to form a node feature vector; calculate the difference between the total pointing error at the measurement point and the output of the mechanism model, and use it as the training objective for the corresponding graph node; Step 3: Train the spatial residual prediction model: Based on the graph structure, the node feature vectors, and the training objective, train the model using an inductive multi-head graph attention network. During training, at least one strategy, including residual standardization, cross-validation, regularization, and early stopping, is employed to obtain a spatial residual prediction model capable of predicting unmodeled residuals. Step 4: Perform semi-parametric fusion compensation: For the real-time prediction point to be compensated, the physical error estimate calculated by the mechanism model is additively fused with the residual prediction value obtained after inputting the real-time prediction point to be compensated as a new node into the spatial residual prediction model, and the final pointing error compensation value is output to complete the online compensation.
2. The semi-parametric compensation method for unmodeled pointing error in optical communication of a motion platform according to claim 1, characterized in that, In step 1, the measurement point is a star point observed in different directions of the moving optical communication terminal, and the prediction point is the target point where the pointing error needs to be predicted.
3. The semi-parametric compensation method for unmodeled pointing error in optical communication of a motion platform according to claim 2, characterized in that, In step 1, the unit direction vector of a node is calculated by the attitude transformation matrix and the unit direction vector in the navigation coordinate system. The angular distance is calculated by the inverse cosine of the dot product of the unit direction vectors of the node. The adjacency weight is obtained by weighting the angular distance using the Gaussian kernel function, and the set of the k nearest neighbors of each node is constructed based on the weight truncation.
4. The semi-parametric compensation method for unmodeled pointing error in optical communication of a motion platform according to claim 1, characterized in that, In step 2, the node feature vector consists of the sine and cosine values of the heading, pitch, and roll angles measured by the inertial navigation system, as well as the sine and cosine values of the encoder's azimuth and pitch guidance values; the training objective is the residual between the total pointing error and the estimation error of the mechanism model, and the objective function during the training process is to minimize the residual estimation error.
5. The semi-parametric compensation method for unmodeled pointing error in optical communication of a motion platform according to claim 1, characterized in that, In step 3, the inductive multi-head graph attention network adopts a multi-layer multi-head attention mechanism, with each layer containing multiple attention heads. When the node features are updated, the features of neighboring nodes are weighted and aggregated through attention weights. The output layer averages the results of the multi-head attention. During training, the Adam optimizer is used, and the learning rate and batch size are determined through cross-validation.
6. A semi-parametric compensation method for unmodeled pointing error in optical communication of a motion platform according to claim 5, characterized in that, During training, weight decay regularization and early stopping strategies are introduced. Model hyperparameters, including Gaussian kernel width, number of network layers, number of attention heads, and hidden dimension, are selected through K-fold cross-validation. Layered cross-validation is performed for different sky regions or pose segments to enhance extrapolation capabilities.
7. A semi-parametric compensation method for unmodeled pointing error in optical communication of a motion platform according to claim 1, characterized in that, In step 4, the additive fusion method involves directly adding the physical error estimate output by the mechanism model to the predicted residual value to obtain the final pointing error compensation value.
8. A semi-parametric compensation method for unmodeled pointing error in optical communication of a motion platform according to claim 7, characterized in that, The mechanism model is a pointing correction model based on the geometric structure of the motion optical communication terminal, which includes inertial navigation installation error, axis non-orthogonality error and aiming error.
9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the semi-parametric compensation method for unmodeled pointing error in optical communication of a motion platform as described in any one of claims 1-8.
10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement a semi-parametric compensation method for unmodeled pointing errors in optical communication of a motion platform, as described in any one of claims 1-8.