Method for correcting dynamic model of space photoelectric tracking turntable

Abstract

The invention discloses a method for correcting a dynamic model of a space photoelectric tracking turntable, which comprises the following steps of: obtaining main parameters influencing model correction by adopting a sensitivity analysis method, and reducing the calculated amount of correction; carrying out secondary development on finite element software, generating a plurality of random samples containing locking mechanism and bearing parameters, calculating frequency response functions in batches, and completing the construction of a data set; aiming at the data set generated by the finite element and the data of the mechanical test, reducing the order of a large amount of data by adopting a mathematical method to obtain a characteristic value of a frequency response function; constructing a neural network, and learning the relationship between the frequency response function eigenvalue and the finite element parameter; using data generated by the finite element after order reduction as a training set to train a neural network; and with an error between a mechanical test frequency response function eigenvalue and a neural network output eigenvalue as a target, an optimization algorithm is adopted to invert finite element model parameters. By adopting the method provided by the invention, the precision of the dynamical model can be greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of optoelectronic tracking and turning stage technology, and relates to a method for correcting parameter errors in the dynamic model of a space optoelectronic tracking and turning stage, and more particularly to a method for correcting the dynamic model of a space optoelectronic tracking and turning stage. Background Technology

[0002] The dynamic model of the space tracking turntable serves as the foundation for evaluating its ability to withstand complex mechanical environments and prevent structural failure during launch. It determines the overall design layout of the space tracking turntable and is crucial to the tracking control system. Improving the modeling accuracy of the space tracking turntable's dynamic model is of great significance for accurately assessing its mechanical resistance and designing the control system.

[0003] Current photoelectric tracking and aiming turntable dynamics modeling is primarily based on finite element method (FEM) parametric modeling. While the modeling approach is clear and each parameter has physical meaning, to constrain the degrees of freedom of the space tracking and aiming turntable during launch, locking mechanisms are typically designed on each axis. These locking mechanisms rely on bolt connections, exhibiting complex nonlinear dynamic characteristics at the connection interfaces. Furthermore, nonlinear dynamic characteristics are introduced between the rolling elements and inner and outer raceways of the bearings. Under the influence of these factors, the accuracy and precision of FEM-based dynamics parametric modeling are generally low, resulting in significant discrepancies with mechanical experiments. To further improve the accuracy of the dynamic model, it is necessary to revise the model.

[0004] In existing work on revising the dynamic models of space tracking and turning platforms, the usual approach is to rely on the results of mechanical experiments, empirically determine the sources of error, and then modify the dynamic model. The target for correction is typically the natural frequencies, which contain limited physical information and lose crucial details such as mode shapes and acceleration amplification. The dynamic frequency response function, on the other hand, includes natural frequencies, damping, mode shapes, and acceleration amplification. Using the frequency response function as the correction target would make the dynamic correction more accurate, but it is computationally intensive and inefficient. Therefore, a more efficient and accurate method using the frequency response function as the correction target is needed for revising the dynamic models of space tracking and turning platforms. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to efficiently and accurately correct the dynamic model of the space tracking turntable by using the frequency response function as the correction target.

[0006] To address the aforementioned technical problems, this invention proposes a method for correcting the dynamic model of a space optoelectronic tracking turntable, which includes the following steps: Step 1: Finite element modeling, which includes: S1. Select element type: Commonly used three-dimensional elements are usually tetrahedral and hexahedral elements. This invention uses tetrahedral elements for geometric objects. S2. For loads with complex structures such as optical lenses, a 0-dimensional mass point is used for equivalent simplification. The mass point includes mass and inertia. S3. The locking mechanism typically includes pyrotechnic explosives / shape memory alloys, connecting bolts, etc. During launch, the simplified modeling method of the dynamic characteristics of the locking mechanism of the space optoelectronic tracking turntable is as follows: The bolts are simplified as beam elements and a preload of 1000N is applied; between the upper and lower connection interfaces, an anisotropic virtual material layer is set, with the anisotropic elastic modulus and shear modulus as input parameters; S4. In photoelectric tracking turntables, angular contact bearings and deep groove ball bearings are usually used. Based on the dynamic characteristics of the bearings, they are simplified into a 6-direction spring unit, which includes linear stiffness in 3 directions and angular stiffness in 3 directions, with each direction stiffness as the input parameter. S5. After meshing the model, add loads and load them into the solver. The solution order is as follows: first solve the statics under a preload of 1000N; use the statics analysis results as the prestress and load them into the modal solver; load the modal solution results into the sinusoidal vibration solver, set the damping to 0.02~0.08, and set the damping as the input parameter. S6. Use the frequency response function of the sinusoidal vibration output as the training set.

[0007] Step 2: Sensitivity analysis is used to obtain the main parameters affecting model correction, reducing the computational burden. Key parameters that significantly affect the eigenvalues ​​of the frequency response function are selected, including: the anisotropic elastic modulus (Ex, Ey, Ez), shear modulus (Gxy, Gyz, Gzx), and Poisson's ratio (νxy, νyz, νzx) of the virtual material layer; the linear stiffness (kx, ky, kz) and angular stiffness (krx, kry, krz) of the bearing; and the structural damping coefficient. Local sensitivity analysis is used to calculate the partial derivatives of the parameters with respect to the characteristic frequencies within the target frequency band (5Hz~200Hz) using the perturbation method, quantifying the sensitivity index of each parameter. For coupled parameters, a global sensitivity analysis is performed, and the Sobol exponent method is used to evaluate the contribution of individual and interactive effects of the parameters to the output response. Finally, based on the sensitivity ranking results, the core parameter set with sensitivity higher than the threshold (e.g., relative sensitivity > 5%) is selected as the optimization variables for subsequent model correction, reducing the dimensionality of the inversion problem.

[0008] Step 3: Perform secondary development on the finite element software to generate multiple random samples containing locking mechanism and bearing parameters, calculate the frequency response function in batches, and complete the construction of the dataset.

[0009] Step 4: Using mathematical methods to reduce the order of the large amount of data from the finite element method dataset and the mechanical test data, the eigenvalues ​​of the frequency response function are obtained. The dataset generated by finite element method and the original frequency response function data obtained from mechanical experiments are collected. This dataset contains high-dimensional, large-sample frequency domain response information. To reduce computational complexity and extract core features, mathematical reduction methods such as Principal Component Analysis (PCA) are used. Specific steps include: standardizing the frequency response function matrix to eliminate dimensional influences; calculating the eigenvalues ​​and eigenvectors of the covariance matrix using singular value decomposition (SVD), retaining principal components with a cumulative contribution exceeding 95%; and projecting the original data into a low-dimensional subspace to obtain the reduced-order frequency response function eigenvectors and their corresponding eigenvalues. This process effectively compresses the data size while preserving key dynamic characteristics, providing efficient input for subsequent neural network training. The reduced-order eigenvalue set will serve as a representative index of the frequency response function, used to quantify the difference between the model and the experiment.

[0010] Step 5: Training a neural network model based on reduced-order data of the frequency response function This invention employs various neural networks, such as BP neural networks (Backpropagation Neural Networks), radial basis function neural networks (RBFNNs), or deep neural networks (DNNs). The network structure includes an input layer, several hidden layers, and an output layer. The number of nodes in the input layer corresponds to the dimension of the eigenvalues ​​of the frequency response function after order reduction, which is determined by the number of principal components retained by principal component analysis (PCA). The number of nodes in the output layer is the number of core dynamic parameters selected through sensitivity analysis, including elastic parameters of the virtual material layer, bearing stiffness parameters, and damping coefficients. The hidden layer structure needs to be flexibly configured according to the data complexity, and fully connected layers can be used. The activation function uses ReLU (Rectified Linear Unit) or its variants to alleviate the gradient vanishing problem and improve training efficiency. Before network training, the input (order-reduced eigenvalues) and output (sensitive parameters) need to be normalized to map the data to the [0,1] or [-1,1] interval to avoid the difference in dimensions affecting model convergence. The training process employs a supervised learning strategy, using the reduced-order frequency response function feature values ​​generated by the finite element model as input samples, and the corresponding model parameters (such as virtual materials Ex, Ey, Ez, Gxy, etc.) as labels. The loss function is either mean squared error (MSE) or mean absolute error (MAE), and the network weights and biases are iteratively updated through the backpropagation algorithm. During training, a validation set (e.g., accounting for 20% of the total samples) needs to be partitioned to monitor overfitting, and an early stopping strategy or L2 regularization is adopted to improve generalization ability. The trained neural network can establish a nonlinear mapping relationship from the frequency response function feature space to the physical parameter space, providing an efficient surrogate model for subsequent parameter inversion.

[0011] Step 6: Input parameter inversion to minimize the error between the eigenvalues ​​of the frequency response function obtained from the mechanical experiment and the output eigenvalues ​​of the neural network; define the objective error function as the mean square error (MSE) between the eigenvalues ​​of the reduced-order frequency response function obtained from the mechanical experiment and the eigenvalues ​​predicted by the neural network. This function quantifies the deviation between the model parameters and the true parameters; the objective function can be expressed as: ,in The vector of core parameters to be inverted (including virtual material elastic modulus, bearing stiffness, and damping coefficient, etc.) These are the eigenvalues ​​of the frequency response function obtained after order reduction from mechanical test data. These are the eigenvalues ​​of the frequency response function predicted by the neural network based on the current parameter θ. The number of feature value samples; Global optimization algorithms (such as genetic algorithms or particle swarm optimization) are used for parameter inversion to overcome local minima and improve convergence efficiency. During initialization, parameters are selected within the range defined by sensitivity analysis (e.g., elastic modulus Ex is between 50 GPa and 70 GPa, bearing stiffness kx is between 10 GPa and 10 GPa). 6 N / mm~10 8 A random parameter population (N / mm) is generated. During iteration, the objective function value of each individual in the population is evaluated in each round: the predicted feature value is calculated through forward propagation of a neural network and compared with the experimental feature value to obtain the error; based on the error value, selection, crossover, and mutation operations are performed on the population to update the parameter vector; the convergence condition is set to a relative error change rate below a threshold (e.g., 10). -4 (or reach the maximum number of iterations (e.g., 1000 times); After the inversion is complete, the optimal parameter set will be output. The finite element model was directly imported for updating. To verify the correction effect, the frequency response function of the corrected model was recalculated and compared with the original experimental data, such as calculating the fitting degree of the frequency response function curve or the characteristic frequency error. Finally, this method achieved efficient and high-precision dynamic model correction, significantly improving the accuracy of predicting the anti-mechanical characteristics of the space optoelectronic tracking and aiming turntable in the launch environment.

[0012] Compared with the prior art, the beneficial effects of the present invention are: (i) The method of the present invention has a more comprehensive correction target and significantly improved accuracy: the "frequency response function (FRF)" containing complete dynamic information is used as the correction target, so that the corrected model can more realistically and comprehensively reflect the actual dynamic characteristics of the turntable, and greatly improve the model accuracy.

[0013] (ii) Effectively solves the modeling problem of nonlinear dynamic characteristics: By introducing a "virtual material layer" to simulate the anisotropic characteristics of the bolt connection surface and simplifying the bearing into a "6-axis spring unit", the dynamic behavior of these key components is more accurately characterized at the modeling level, reducing model errors from the source.

[0014] (III) Achieving a balance between high efficiency and high accuracy: Sensitivity analysis was used to screen key parameters, significantly reducing the dimensionality of the optimization problem; Principal Component Analysis (PCA) was used to reduce the order of the frequency response function data, greatly compressing the data volume while retaining core information; a neural network surrogate model was constructed to replace time-consuming finite element calculations, realizing a fast nonlinear mapping from eigenvalues ​​to physical parameters; and global optimization algorithms (such as genetic algorithms) were combined for parameter inversion, avoiding getting trapped in local optima. This combined strategy of "dimensionality reduction-order reduction-surrogate model-global optimization" significantly improves the computational efficiency of model correction while ensuring high accuracy, making it suitable for engineering practice.

[0015] (iv) It provides a systematic and automated correction process: From parametric modeling, sensitivity analysis, data reduction, neural network training to parameter inversion, this invention forms a complete, systematic, and automatically executable dynamic model correction process. This reduces reliance on human experience and improves the repeatability and reliability of the correction process. Attached Figure Description

[0016] Figure 1 This is a flowchart of a method for correcting the dynamic model of a space optoelectronic tracking and aiming turntable according to the present invention; Figure 2 This is a schematic diagram of the present invention simplifying bolts into virtual materials for beam units; Figure 3 This is a schematic diagram of the photoelectric tracking turntable bearing of the present invention simplified to a 6-direction spring unit. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, a method for correcting the dynamic model of a space optoelectronic tracking turntable includes the following steps: Step 1: Finite element modeling, which includes: S1. Select element type: Commonly used three-dimensional elements are usually tetrahedral and hexahedral elements. This invention uses tetrahedral elements for geometric objects. S2. For loads with complex structures such as optical lenses, a 0-dimensional mass point is used for equivalent simplification. The mass point includes mass and inertia. S3, such as Figure 2 As shown, the locking mechanism includes components such as pyrotechnic explosives / shape memory alloys and connecting bolts. When the space optoelectronic tracking turntable arrives at the designated orbit, the upper and lower connection interfaces of the connecting bolts are separated by igniting the pyrotechnic explosives / triggering the shape memory alloys. During launch, the simplified modeling of the locking mechanism dynamics of the space optoelectronic tracking turntable is as follows: the bolts are simplified as beam elements, and a preload of 1000N is applied. An anisotropic virtual material layer is set between the upper and lower connection interfaces. The constitutive model of the virtual material is as follows: ; The anisotropic elastic modulus and shear modulus are used as input parameters; where: ε is normal strain, γ is shear strain, σ is normal stress, τ is shear stress, E is elastic modulus, and the subscripts x, y, z indicate the direction; G is shear modulus, and the subscript indicates the plane of action; ν is Poisson's ratio, and the subscript νxy indicates the transverse strain coefficient in the y direction when subjected to force in the x direction; S4, such as Figure 3 As shown, in the space optoelectronic tracking turntable, angular contact bearings and deep groove ball bearings are used. Considering the dynamic characteristics of the bearings, they are simplified into a 6-axis spring element, including linear stiffness in 3 directions and angular stiffness in 3 directions. Its stiffness matrix is ​​as follows: The stiffness in each direction is used as the input parameter; where: kx, ky, kz: are the linear stiffness (N / mm) of the bearing in the three translational directions (X, Y, Z), krx, kry, krz: are the angular stiffness or rotational stiffness (N·mm / rad) of the bearing about the three coordinate axes (X, Y, Z); for angular contact bearings, krz is usually not included; for deep groove ball bearings, only ky and kz are usually included; S5. After meshing the model, add loads and load them into the solver. The solution order is as follows: first solve the statics under a preload of 1000N; use the statics analysis results as prestress and load them into the modal solver; load the modal solution results into the sinusoidal vibration solver, set the damping to 0.02~0.08, and set the damping as the input parameter. S6. Organize the frequency response function of the sinusoidal vibration output into a training set.

[0019] Step 2: Sensitivity analysis is used to obtain the main parameters affecting model correction, reducing the computational burden. First, key parameters significantly affecting the eigenvalues ​​of the frequency response function are selected, including: the anisotropic elastic modulus (Ex, Ey, Ez), shear modulus (Gxy, Gyz, Gzx), and Poisson's ratio (νxy, νyz, νzx) of the virtual material layer; the linear stiffness (kx, ky, kz) and angular stiffness (krx, kry, krz) of the bearing; and the structural damping coefficient. Local sensitivity analysis is used to calculate the partial derivatives of the parameters with respect to the characteristic frequencies within the target frequency band (5Hz~200Hz) using the perturbation method, quantifying the sensitivity of each parameter. For coupled parameters, a global sensitivity analysis is performed, using the Sobol exponent method to evaluate the contribution of individual and interactive effects of the parameters to the output response. Finally, based on the sensitivity ranking results, a set of core parameters with sensitivity higher than a threshold (e.g., relative sensitivity > 5%) is selected as optimization variables for subsequent model correction, significantly reducing the dimensionality of the inversion problem.

[0020] Step 3: Perform secondary development on the finite element software to generate multiple random samples containing locking mechanism and bearing parameters, calculate the frequency response function in batches, and complete the construction of the dataset.

[0021] Step 4: Using mathematical methods to reduce the order of the large amount of data from the finite element method dataset and the mechanical test data, the eigenvalues ​​of the frequency response function are obtained. The dataset generated by finite element method and the original frequency response function data obtained from mechanical experiments are collected. This dataset contains high-dimensional, large-sample frequency domain response information. To reduce computational complexity and extract core features, mathematical reduction methods such as principal component analysis (PCA) are used for processing. The specific steps include: standardizing the frequency response function matrix to eliminate the influence of dimensions; calculating the eigenvalues ​​and eigenvectors of the covariance matrix through singular value decomposition (SVD), retaining the principal components with a cumulative contribution rate of over 95%; projecting the original data into a low-dimensional subspace to obtain the reduced frequency response function eigenvectors and corresponding eigenvalues. The above process can effectively compress the data scale while retaining key dynamic characteristics, providing efficient input for subsequent neural network training. The reduced eigenvalue set will serve as a representative index of the frequency response function to quantify the difference between the model and the experiment.

[0022] Step 5: Training a neural network model based on reduced-order data of the frequency response function This invention employs various neural networks, such as BP neural networks (Backpropagation Neural Networks), radial basis function neural networks (RBFNNs), or deep neural networks (DNNs). The network structure typically includes an input layer, several hidden layers, and an output layer. The number of nodes in the input layer corresponds to the dimension of the frequency response function's eigenvalues ​​after order reduction, which is determined by the number of principal components retained by principal component analysis (PCA). The number of nodes in the output layer represents the number of core dynamic parameters selected through sensitivity analysis, including elastic parameters of the virtual material layer, bearing stiffness parameters, and damping coefficients. The hidden layer structure needs to be flexibly configured according to data complexity, and fully connected layers can be used. The activation function uses ReLU (Rectified Linear Unit) or its variants to alleviate the gradient vanishing problem and improve training efficiency. Before network training, the input (order-reduced eigenvalues) and output (sensitivity parameters) need to be normalized, mapping the data to the [0,1] or [-1,1] interval to avoid dimensional differences affecting model convergence. The training process employs a supervised learning strategy, using the reduced-order frequency response function feature values ​​generated by the finite element model as input samples, and the corresponding model parameters (such as virtual materials Ex, Ey, Ez, Gxy, etc.) as labels. The loss function is either mean squared error (MSE) or mean absolute error (MAE), and the network weights and biases are iteratively updated through the backpropagation algorithm. During training, a validation set (e.g., accounting for 20% of the total samples) needs to be partitioned to monitor overfitting, and an early stopping strategy or L2 regularization is adopted to improve generalization ability. The trained neural network can establish a nonlinear mapping relationship from the frequency response function feature space to the physical parameter space, providing an efficient surrogate model for subsequent parameter inversion.

[0023] Step 6: Input parameter inversion to minimize the error between the eigenvalues ​​of the frequency response function obtained from the mechanical experiment and the output eigenvalues ​​of the neural network; define the objective error function as the mean square error (MSE) between the eigenvalues ​​of the reduced-order frequency response function obtained from the mechanical experiment and the eigenvalues ​​predicted by the neural network. This function quantifies the deviation between the model parameters and the true parameters; the objective error function can be expressed as: ,in The vector of core parameters to be inverted includes the virtual material elastic modulus, bearing stiffness, and damping coefficient. These are the eigenvalues ​​of the frequency response function obtained after order reduction from mechanical test data. These are the eigenvalues ​​of the frequency response function predicted by the neural network based on the current parameter θ. The number of feature value samples; Global optimization algorithms (such as genetic algorithms or particle swarm optimization) are used for parameter inversion to overcome local minima and improve convergence efficiency. During initialization, parameters are selected within the range defined by sensitivity analysis (e.g., elastic modulus Ex is between 50 GPa and 70 GPa, bearing stiffness kx is within 10 GPa). 6 N / mm~10 8 A random parameter population (N / mm) is generated. During iteration, the objective function value of each individual in the population is evaluated in each round: the predicted feature value is calculated through forward propagation of a neural network and compared with the experimental feature value to obtain the error; based on the error value, selection, crossover, and mutation operations are performed on the population to update the parameter vector; the convergence condition is set to a relative error change rate below a threshold (e.g., 10). -4 (or reach the maximum number of iterations (e.g., 1000 times); After the inversion is complete, the optimal parameter set will be output. The finite element model was directly imported for updating. To verify the correction effect, the frequency response function of the corrected model was recalculated and compared with the original experimental data, such as calculating the fitting degree of the frequency response function curve or the characteristic frequency error. Finally, this method achieved efficient and high-precision dynamic model correction, significantly improving the accuracy of predicting the anti-mechanical characteristics of the space optoelectronic tracking and aiming turntable in the launch environment.

Claims

1. A method for correcting the dynamic model of a space optoelectronic tracking turntable, characterized in that: It includes the following steps: Step 1: Create a finite element model of the geometry of the space optoelectronic tracking and turning platform; Step 2: After the model is meshed, loads are added and loaded into the solver. The output frequency response function is then organized as the training set. Step 3: Perform secondary development on the finite element software to generate multiple random samples containing locking mechanism and bearing parameters, calculate the frequency response function in batches, and complete the construction of the dataset; Step 4: Using mathematical methods to reduce the order of the large amount of data from the finite element method dataset and the mechanical test data, the eigenvalues ​​of the frequency response function are obtained. We collect datasets generated by finite element analysis and raw frequency response function data obtained from mechanical experiments. These datasets contain high-dimensional, large-sample frequency domain response information. We then process the data using principal component analysis, a mathematical method for order reduction. The reduced eigenvalue set serves as a representative index of the frequency response function, used to quantify the difference between the model and the experiment. Step 5: Training a neural network model based on reduced-order data of the frequency response function; Step 6: Invert the input parameters by minimizing the error between the eigenvalues ​​of the frequency response function of the mechanical test and the output eigenvalues ​​of the neural network; The objective error function is defined as the mean square error between the eigenvalues ​​of the reduced-order frequency response function obtained from the mechanical experiment and the eigenvalues ​​predicted by the neural network. This function quantifies the deviation between the model parameters and the true parameters; the objective error function is expressed as: ,in The vector of core parameters to be inverted includes the virtual material elastic modulus, bearing stiffness, and damping coefficient. These are the eigenvalues ​​of the frequency response function obtained after order reduction from mechanical test data. These are the eigenvalues ​​of the frequency response function predicted by the neural network based on the current parameter θ. The number of feature value samples; After the inversion is complete, the optimal parameter set will be output. The finite element model was directly imported for updating. To verify the correction effect, the frequency response function of the corrected model was recalculated, that is, the goodness of fit of the frequency response function curve or the characteristic frequency error was calculated, and compared with the original experimental data.

2. The method for correcting the dynamic model of a space optoelectronic tracking turntable according to claim 1, characterized in that: Step 1 specifically includes: S1. The geometry of the space optoelectronic tracking and aiming stage adopts a tetrahedral unit type; S2. The optical lens structure has complex loads, which are simplified by using a 0-dimensional mass point, which includes mass and inertia. S3. The locking mechanism includes pyrotechnic explosives / shape memory alloys and connecting bolt components. During launch, the simplified modeling method for the dynamics of the locking mechanism of the space optoelectronic tracking turntable is as follows: the bolts are simplified as beam elements, and a preload is applied; an anisotropic virtual material layer is set between the upper and lower connection interfaces, and the constitutive model of the virtual material is as follows: ； The elastic modulus and shear modulus of the anisotropic model are used as input parameters; In the formula: ε is normal strain, γ is shear strain, σ is normal stress, τ is shear stress, E is elastic modulus, and the subscripts x, y, z indicate the direction; G is shear modulus, and the subscript indicates the plane of action; ν is Poisson's ratio, and the subscript νxy indicates the transverse strain coefficient in the y direction when subjected to force in the x direction; S4. Considering the dynamic characteristics of the angular contact bearings and deep groove ball bearings of the space optoelectronic tracking turntable, it is simplified into a 6-axis spring element, including linear stiffness in 3 directions and angular stiffness in 3 directions. The stiffness matrix is: ； Using anisotropic stiffness as the input parameter; In the formula: kx, ky, kz: linear stiffness of the bearing in the three translational directions X, Y, Z, unit: N / mm; krx, kry, krz: angular stiffness or rotational stiffness of the bearing about the three coordinate axes X, Y, Z, unit: N·mm / rad; For angular contact bearings, krz is not included; for deep groove ball bearings, only ky and kz are included.

3. The method for correcting the dynamic model of a space optoelectronic tracking turntable according to claim 1, characterized in that: The solution sequence for step 2 is as follows: First, solve the statics under preload; The static analysis results are used as prestress and applied to the modal solver; The modal solution results are loaded into the sinusoidal vibration solver, the damping is set to 0.02~0.08, and the damping is set as the input parameter; the frequency response function of the sinusoidal vibration output is organized as the training set.

4. The method for correcting the dynamic model of a space optoelectronic tracking turntable according to claim 1, characterized in that: It also includes step 3, which precedes step 4: To reduce the computational burden of correction, sensitivity analysis was employed to obtain the main parameters affecting model correction. First, key parameters significantly influencing the eigenvalues ​​of the frequency response function were identified, including: the anisotropic elastic modulus Ex, Ey, and Ez of the virtual material layer; the shear modulus Gxy, Gyz, and Gzx; the Poisson's ratio νxy, νyz, and νzx; the linear stiffness kx, ky, and kz of the bearing; the angular stiffness krx, kry, and krz; and the structural damping coefficient. Local sensitivity analysis was used to calculate the partial derivatives of the parameters with respect to the characteristic frequencies within the target frequency band of 5Hz to 200Hz using the perturbation method, quantifying the sensitivity of each parameter. For coupled parameters, a global sensitivity analysis was conducted, employing the Sobol exponent method to evaluate the contribution of individual and interactive effects of the parameters to the output response. Finally, based on the sensitivity ranking results, a set of core parameters with a sensitivity higher than the threshold (i.e., relative sensitivity > 5%) was selected as optimization variables for subsequent model correction.

5. The method for correcting the dynamic model of a space optoelectronic tracking turntable according to claim 1, characterized in that: The specific steps for processing using the principal component analysis mathematical order reduction method include: The frequency response function matrix is ​​standardized and preprocessed to eliminate the influence of dimensions; The eigenvalues ​​and eigenvectors of the covariance matrix are calculated by singular value decomposition, and principal components with a cumulative contribution rate of over 95% are retained. The original data is projected onto a low-dimensional subspace to obtain the reduced frequency response function eigenvectors and corresponding eigenvalues, providing efficient input for subsequent neural network training.

6. The method for correcting the dynamic model of a space optoelectronic tracking turntable according to claim 1, characterized in that: The neural network model adopts a BP neural network, a radial basis function neural network, or a deep neural network; the network structure includes an input layer, several hidden layers, and an output layer; the number of nodes in the input layer corresponds to the dimension of the eigenvalues ​​of the frequency response function after order reduction; the number of nodes in the output layer is the number of core dynamic parameters selected through sensitivity analysis, including the elastic parameters of the virtual material layer, bearing stiffness parameters, and damping coefficients; the hidden layer structure is configured according to the data complexity and adopts a fully connected layer. The activation function uses ReLU or its variants to alleviate the gradient vanishing problem and improve training efficiency; before network training, the input of reduced-order feature values ​​and the output of sensitive parameters are normalized to map the data to the interval [0,1] or [-1,1] to avoid the difference in scale affecting the model convergence; The training process employs a supervised learning strategy, using the reduced-order frequency response function eigenvalues ​​generated by the finite element model as input samples, and the corresponding virtual material Ex, Ey, Ez, Gxy model parameters as labels. The loss function is either mean squared error or mean absolute error, and the network weights and biases are iteratively updated through the backpropagation algorithm. During training, a validation set, comprising 20% ​​of the total samples, is defined to monitor overfitting. Early stopping strategies or L2 regularization are employed to enhance generalization ability. The trained neural network can establish a nonlinear mapping relationship from the frequency response function feature space to the physical parameter space, providing an efficient surrogate model for subsequent parameter inversion.

7. The method for correcting the dynamic model of a space optoelectronic tracking turntable according to claim 1, characterized in that: The inversion employs a genetic algorithm or a global optimization algorithm based on particle swarm optimization to perform parameter inversion, thereby overcoming the local minima problem and improving convergence efficiency. During initialization, the parameters are selected within the range defined by sensitivity analysis, i.e., the elastic modulus Ex is between 50 GPa and 70 GPa, and the bearing stiffness kx is between 10 GPa and 10 GPa. 6 N / mm~10 8 N / mm, generate a random parameter population; during iteration, evaluate the objective function value of each individual in the population in each round: calculate the predicted feature value through forward propagation of the neural network and compare it with the experimental feature value to obtain the error; based on the error value, perform selection, crossover and mutation operations on the population to update the parameter vector; set the convergence condition as the relative error change rate being less than a threshold of 10. -4 Or it can reach 1000 iterations.

8. The method for correcting the dynamic model of a space optoelectronic tracking turntable according to claim 3, characterized in that: The preload force is 1000N.

Images