In-situ fast measurement and calibration method for large aperture optical system

Abstract

The present application relates to a kind of measurement and calibration method of optical system, specifically relates to a kind of in-situ fast measurement and calibration method of large aperture optical system, for solving the existing large aperture optical system adjustment technology relies on hardware sensing resource, and calculation is time-consuming, and the existing deep learning method is sensitive to platform microvibration, and image translation is misjudged as the problem of optical aberration.This method includes constructing defocus star image generation model, generates misadjustment sample using multidimensional layered random sampling algorithm, to generate defocus star image construction dataset;VGG convolutional neural network model is constructed;Using translation invariance loss function trains network, forces it to output consistent prediction value for original and translated image;Real-time multi-view defocus star image is obtained, input into trained special network model, output normalized misadjustment quantity prediction value, after inverse normalization is solved, control instruction is generated, and multi-degree-of-freedom adjustment mechanism is driven to carry out reverse compensation.

Description

Technical Field

[0001] This invention relates to a method for measuring and calibrating an optical system, specifically a method for in-situ rapid measurement and calibration of a large-aperture optical system. Background Technology

[0002] With the rapid development of space optical remote sensing technology, large-aperture optical systems (such as space telescopes and high-resolution reconnaissance satellites) have become core payloads for Earth observation and deep space exploration. The imaging resolution of such systems is theoretically proportional to the aperture size; therefore, as the aperture of optical systems continues to increase, the requirements for image quality become increasingly stringent. However, large-aperture optical systems face severe challenges throughout their entire lifecycle:

[0003] First, after large-aperture optical systems are assembled and adjusted on the ground, they must endure the severe vibrations and overload impacts during rocket launch. Once in orbit, they face stress reconstruction due to gravity release and structural deformation caused by extreme alternating thermal environments. All of these factors can lead to minute rigid body displacements in optical components (especially secondary mirrors), including misalignments such as lateral eccentricity, axial defocusing, and tilting. These misalignments severely damage the wavefront quality of large-aperture optical systems, introduce asymmetric aberrations, and result in blurred images and a significant decrease in modulation transfer function (MTF). Therefore, high-precision, high-response-speed on-orbit active assembly and calibration technology is a core technology for ensuring the on-orbit performance of large-aperture space optical systems.

[0004] Currently, wavefront sensing and assembly techniques for large-aperture optical systems are mainly divided into two categories: hardware sensing methods and image inversion methods. However, both existing technologies have significant limitations in practical applications.

[0005] Hardware-based wavefront sensing technologies, such as Shaker-Hartmann sensors and shearing interferometers, can achieve high measurement accuracy, but their application requires the additional integration of complex beam-splitting wavefront sensing optical paths into large-aperture optical systems. During ground assembly and adjustment, this increases the complexity and time cost of optical path setup; in space applications, it significantly consumes valuable weight, volume, and power resources of the satellite platform. More critically, these methods generally suffer from non-common optical path error, meaning the detection optical path and the main imaging optical path are inconsistent, resulting in measurement results that cannot fully represent the actual image quality. The complexity of the system structure also reduces overall reliability.

[0006] Traditional image inversion techniques, represented by the Phase Diversity (PD) method, are currently the mainstream wavefront-free sensing methods. This method acquires multiple frames of images at both the focal and defocus planes and uses a nonlinear optimization algorithm to iteratively calculate the wavefront phase error. However, the PD algorithm is inherently a computationally intensive iterative process with slow convergence and high computational complexity, making it difficult to meet the timeliness requirements of real-time or near-real-time assembly and adjustment in a space-on-orbit environment. Furthermore, this method requires the simultaneous acquisition of multiple images with different phase states, imposing additional operational requirements on the focusing mechanism of the imaging system and increasing the risk of dynamic operation.

[0007] In recent years, with the rapid development of deep learning technology, end-to-end wavefront sensing methods based on neural networks have attracted widespread attention due to their fast inference capabilities. However, directly applying general convolutional neural networks (CNNs) to the on-orbit assembly and adjustment of large-aperture optical systems in space faces a core technical challenge: the interference of micro-vibrations on the imaging platform. During on-orbit operation, the imaging platform inevitably experiences slight jitter due to factors such as the rotation of the flywheel in the satellite attitude control mechanism and the movement of the robotic arm; similarly, environmental vibrations and air disturbances exist in the ground laboratory environment. These factors can cause random spatial translations of star images on the detector target surface. Although the shape and position of star points are physically decoupled according to optical aberration theory, traditional convolutional neural networks are highly sensitive to the spatial position of the input image. In the absence of sufficient training data or a network structure that is not specifically designed, the network is prone to confusing "positional features" with "morphological features," incorrectly interpreting the image translation caused by platform jitter as asymmetric aberrations of the optical system (such as coma), thus outputting false misalignment predictions and leading to a significant decrease in measurement accuracy. This problem severely restricts the practical application of deep learning methods in the field of high-precision on-orbit assembly and adjustment. Summary of the Invention

[0008] The purpose of this invention is to solve the technical problem in existing large-aperture optical system wavefront sensing and assembly technology, where end-to-end wavefront sensing methods based on neural networks are sensitive to platform micro-vibrations and easily misjudge image translation as optical aberrations. The invention provides an in-situ rapid measurement and calibration method for large-aperture optical systems.

[0009] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0010] A rapid in-situ measurement and calibration method for a large-aperture optical system is disclosed, used for the rapid measurement and calibration of the large-aperture optical system, which includes optical elements and a multi-degree-of-freedom adjustment mechanism for driving the movement of the optical elements; its key feature is that it includes the following steps:

[0011] S1. Constructing the dataset

[0012] S1.1. Establish a defocused star point image generation model using optical design software; generate multiple sets of misaligned samples in the misalignment parameter space using a multidimensional hierarchical random sampling algorithm; the misalignment parameter space is a multidimensional space spanned by the variables of each dimension in the multiple sets of misaligned samples.

[0013] S1.2 Input the misaligned samples into the off-focus star image generation model to generate corresponding off-focus star images and construct a dataset containing the mapping relationship between the misaligned samples and the off-focus star images; the off-focus star images include at least off-focus star images of the center field of view, the 0.7 field of view and the edge field of view;

[0014] S2. Constructing the VGG convolutional neural network model

[0015] A VGG convolutional neural network model for multi-view feature fusion is constructed. The VGG convolutional neural network model includes multiple convolutional layers for extracting deep features of out-of-focus star images, a global average pooling layer at the end of the convolutional layers, and a fully connected layer at the end of the global average pooling layer for outputting the offset prediction value. The global average pooling layer is used to perform spatial averaging on each feature channel to eliminate the spatial position information of the out-of-focus star images, so as to extract wavefront features with translation invariance, thereby suppressing the interference of imaging platform jitter on feature extraction.

[0016] S3. Training the VGG convolutional neural network model

[0017] The dataset generated in step S1.2 is preprocessed; the preprocessed dataset is divided into a training set and a validation set, and the training set and validation set are sequentially input into the VGG convolutional neural network model constructed in step S2 to train and validate the VGG convolutional neural network model, thereby obtaining a dedicated network model suitable for the current environment.

[0018] S4. Acquire real-time off-focus star images and perform image preprocessing.

[0019] Acquire real-time defocused star point images output by the large-aperture optical system and perform preprocessing operations;

[0020] S5, Online Inference and Closed-Loop Assembly Control

[0021] The real-time defocused star image preprocessed in step S4 is input into the dedicated network model obtained in step S3, and the normalized offset prediction value is output. The normalized offset prediction value is subjected to inverse normalization calculation to obtain the actual offset in physical space and generate the corresponding control command. The control command is sent to the controller of the multi-degree-of-freedom adjustment mechanism, and the optical element is driven by the multi-degree-of-freedom adjustment mechanism to perform inverse compensation motion, thereby completing the in-situ rapid measurement and calibration of the large-aperture optical system.

[0022] Furthermore, in step S1.1, the defocused star point image generation model is a parameterized model of a large-aperture optical system, with its object side at infinity and its field of view selected according to the star target.

[0023] Alternatively, the defocused star point image generation model is a joint simulation model composed of a collimator and a large-aperture optical system; in the joint simulation model, the point light source at the focal plane of the collimator is the object point, and after collimation by the collimator, it simulates an infinity target;

[0024] When using the co-simulation model, the residual aberration of the collimator wavefront is introduced as an error term of the large-aperture optical system during the generation of misaligned samples, enabling the co-simulation model to adapt to the non-ideal characteristics of ground-based detection equipment.

[0025] Furthermore, in step S1.1, the multidimensional stratified random sampling algorithm is the Latin hypercube sampling algorithm.

[0026] Furthermore, in step S1.1, the specific method for generating multiple sets of imbalanced samples in the imbalance parameter space using the Latin hypercube sampling algorithm is as follows:

[0027] The Latin hypercube sampling algorithm is used to determine the variable range [L] of each of the K dimensions in the imbalance parameter space. k U k The data is divided into N equally probable intervals, and one imbalanced sample is randomly sampled from each interval to obtain multiple sets of imbalanced samples formed by combining the K dimensions and the imbalanced samples sampled from the corresponding N intervals; where the value X of the i-th imbalanced sample in the k-th dimension is... ik The calculation formula is as follows:

[0028]

[0029] In the formula, i is the disordered sample number collected in the k-th dimension, i = 1, 2, ..., N, where N represents the total number of interval numbers and the total number of disordered samples collected in the k-th dimension, N ≥ 2; k is the dimension number, k = 1, 2, ..., K, where K is the total number of dimensions, K ≥ 1; L k with U k These represent the lower and upper bounds of the range of the k-th dimension variable, respectively; π k (i) -1 indicates that the interval numbering is changed from 1~N to 0~N-1; π k (i) represents the interval number randomly assigned to the i-th imbalanced sample in the k-th dimension; π k This represents a random permutation sequence from 1 to N, used to ensure that only one out-of-balance sample is selected in each interval of each dimension; It represents a random number that follows a uniform distribution in [0, 1) and is used to achieve local random perturbation within a small interval.

[0030] Further, in step S1.1, each group of misaligned samples is an misalignment vector V, whose expression is:

[0031] V = [dx, dy, dz, tx, ty];

[0032] In the formula, dx is the lateral eccentricity of the optical element along the X-axis in a plane perpendicular to the optical axis; dy is the lateral eccentricity of the optical element along the Y-axis in a plane perpendicular to the optical axis; dz is the axial defocusing of the optical element along the optical axis; tx is the tilt angle of the optical element rotating around the X-axis; and ty is the tilt angle of the optical element rotating around the Y-axis.

[0033] Furthermore, in step S3, the training and validation of the VGG convolutional neural network model to obtain a dedicated network model suitable for the current environment is specifically as follows:

[0034] During training, a translation-invariant loss function is introduced to force the VGG convolutional neural network model to output consistent offset prediction values ​​for both the original defocused star images acquired in the training set and the defocused star images after random translation. The parameters of the VGG convolutional neural network model are optimized through backpropagation, and the training state of the VGG convolutional neural network model is verified through the validation set to prevent overfitting, until the VGG convolutional neural network model converges on the validation set, resulting in a dedicated network model suitable for on-orbit environments.

[0035] And / or, during training, data augmentation is performed using off-focus star images from the training set to simulate star jitter and deformation caused by air turbulence. Global average pooling layers are used to extract spot morphology features that are insensitive to airflow jitter. The enhanced off-focus star image data is used to optimize the parameters of the VGG convolutional neural network model through backpropagation. The training state of the VGG convolutional neural network model is verified using the validation set to prevent overfitting, until the VGG convolutional neural network model converges on the validation set, resulting in a dedicated network model suitable for ground environments.

[0036] Furthermore, in step S3, the specific steps for preprocessing the dataset generated in step S1.2 are as follows:

[0037] Square root operations are performed on the defocused star images in the central field of view, 0.7 field of view, and edge field of view to compress the dynamic range, and then normalized. Then, the normalized defocused star images of the three fields of view are stacked in the channel dimension to construct a three-channel tensor of size 3×H×W, which is input into the VGG convolutional neural network model. Here, H represents the height of the defocused star image, and W represents the width of the defocused star image.

[0038] Furthermore, in step S3, the formulas for the square root operation and normalization are as follows:

[0039] ;

[0040] In the formula, I temp (u, v) represents the pixel value of the intermediate image obtained by taking the square root of each pixel of the original acquired defocused star point image, where (u, v) are the image pixel coordinates, I raw (u, v) represents the pixel values ​​of the original acquired out-of-focus star image. To prevent extremely small positive numbers from being introduced abnormally in numerical calculations, take... ;

[0041]

[0042] In the formula, Iinput(u, v) represents the normalized pixel value input to the VGG convolutional neural network model, with a value range of [0, 1], and max(I temp ) represents the maximum value of the intermediate image obtained by taking the square root of each pixel of the original defocused star point image, min(T) temp () represents the minimum value of the intermediate image obtained by taking the square root of each pixel of the original defocused star point image.

[0043] Furthermore, in step S3, the translation invariance loss function L... total The calculation formula is:

[0044]

[0045] In the formula, L MSE L represents the mean square error term; Consist λ is the consistency constraint term; λ is the weighting coefficient that balances prediction accuracy and jitter resistance.

[0046] The mean square error term L MSE The calculation formula is:

[0047]

[0048] In the formula, j is the batch number of the VGG convolutional neural network model training, j=1, 2, ..., B, and B is the total number of batches trained by the VGG convolutional neural network model; x m Let y be the tensor of the m-th input defocused star image, where m is the number of the input defocused star image in the j-th batch, m = 1, 2, ..., M, and M is the total number of input defocused star images in the j-th batch; m Let f be the true imbalance label vector corresponding to the m-th imbalance sample, and f(·) be the forward inference mapping function of the VGG convolutional neural network model. The square of the L2 norm;

[0049] The consistency constraint term L Consist The calculation formula is:

[0050]

[0051] Where T(x) m , Δu, Δv) are spatial affine transformation operators, which represent shifting the tensor of the input m-th defocused star image by Δu and Δv pixels in the horizontal and vertical directions, respectively.

[0052] Furthermore, in step S4, the specific method for obtaining the real-time defocused star map output by the large-aperture optical system is as follows:

[0053] Adjust the satellite attitude so that the large-aperture optical system is aligned with a star of moderate brightness and with a clean background. Control the optical elements to acquire images of defocused star points and capture the corresponding multi-field regions to obtain real-time defocused star point images output by the large-aperture optical system.

[0054] Alternatively, the collimator can be turned on to project the artificial star points it generates onto the large-aperture optical system, and the exposure time of the large-aperture optical system can be adjusted to obtain a real-time defocused star point image output by the large-aperture optical system.

[0055] Compared with the prior art, the present invention has the following beneficial technical effects:

[0056] 1. This invention discloses an in-situ rapid measurement and calibration method for large-aperture optical systems. It constructs a complete technical chain from simulation dataset generation, neural network construction and training to online inference and closed-loop control, achieving end-to-end rapid measurement of misalignment in large-aperture optical systems using a single frame image. By introducing multi-field-of-view defocused star point images as input, it fully utilizes the sensitivity differences in misalignment across different fields of view, significantly improving the information redundancy and solution accuracy of wavefront inversion. Furthermore, by extracting translation-invariant features through a global average pooling layer, it effectively suppresses the interference of imaging platform jitter on the measurement results, laying a solid foundation for subsequent high-precision closed-loop assembly and calibration.

[0057] 2. This invention provides an in-situ rapid measurement and calibration method for large-aperture optical systems. By setting two selectable defocused star point image generation models, this method can be adapted to both ground-based laboratory assembly and on-orbit calibration in space. The parametric model uses stars as targets and is suitable for on-orbit environments; the co-simulation model uses collimators as light sources and is suitable for ground testing. This modular design enhances the method's versatility and engineering adaptability, avoiding the technical redundancy of repeatedly developing algorithm models for different scenarios.

[0058] 3. This invention provides an in-situ rapid measurement and calibration method for large-aperture optical systems. By actively introducing the residual aberration of the collimator as an error term into the co-simulation model for training, the neural network can learn in advance and automatically compensate for the non-ideal characteristics of the ground detection equipment. This effectively eliminates the common-path error introduced by the detection equipment in traditional methods, significantly improving the accuracy and reliability of transferring ground assembly and calibration results to on-orbit performance. Simultaneously, a Latin hypercube sampling algorithm is used to generate misalignment samples, ensuring that every dimension and interval of the misalignment parameter space is uniformly covered with a limited number of samples. Compared to random sampling, this method greatly increases the information entropy of the misalignment sample space, avoiding misalignment sample clustering or gaps, thereby supporting higher-precision neural network training with a smaller dataset and improving data generation efficiency.

[0059] 4. This invention provides an in-situ rapid measurement and calibration method for large-aperture optical systems. By introducing a random perturbation factor and a random permutation sequence within an interval, the Latin hypercube sampling algorithm maintains both overall uniformity and local randomness, ensuring that the generated misaligned samples are both representative and diverse. This "globally uniform + locally random" sampling strategy effectively prevents the neural network from overfitting to a specific sampling pattern and enhances the model's generalization ability.

[0060] 5. This invention provides an in-situ rapid measurement and calibration method for large-aperture optical systems. It precisely quantifies the misalignment state into a five-degree-of-freedom vector V, completely describing the rigid body displacement of optical elements (especially secondary mirrors). These five physical quantities directly correspond to the control axes of the multi-degree-of-freedom adjustment mechanism, allowing the predicted values ​​output by the neural network to directly drive the actuator without complex transformations. This achieves a seamless transition from "measurement" to "adjustment," simplifying control system design.

[0061] 6. This invention provides an in-situ rapid measurement and calibration method for large-aperture optical systems. By stacking defocused star point images from the center field of view, 0.7 field of view, and edge field of view into a three-channel input, it simulates the channel structure of a color image, enabling the VGG network to simultaneously extract the associated features of the three fields of view. Compared to single-field-of-view input, multi-field-of-view fusion significantly improves the network's sensitivity and recognition ability to asymmetric aberrations (such as coma and astigmatism), achieving efficient feature extraction of "seeing three fields at a glance." Simultaneously, the square root operation effectively compresses the high dynamic range of the point spread function, allowing the details of faint diffraction rings to be preserved and participate in network training; normalization eliminates the absolute light intensity differences under different acquisition conditions, enabling the network to focus on the morphological features of the light spot rather than absolute brightness. These two preprocessing steps jointly improve the model's robustness to different lighting conditions and exposure parameters, enhancing its cross-scene generalization ability.

[0062] 7. This invention discloses an in-situ rapid measurement and calibration method for large-aperture optical systems. By innovatively introducing a translation-invariant loss function, it forces the network to output consistent predicted values ​​for the original image and the translated image under the same wavefront state. This design fundamentally solves the technical problem of traditional convolutional neural networks being sensitive to image position and easily misinterpreting platform jitter as optical aberrations. Combined with a mean square error term, it achieves the dual optimization goals of "maintaining accuracy and resisting interference," representing a core technological breakthrough for on-orbit applications.

[0063] 8. This invention provides an in-situ rapid measurement and calibration method for large-aperture optical systems, offering clear real-time image acquisition schemes for two typical scenarios: ground-based and in-orbit. The ground-based mode utilizes collimators to provide stable and controllable artificial star points, facilitating repeated adjustments; the in-orbit mode uses stars as a natural light source, requiring no additional hardware and achieving wavefront measurement with "zero resource consumption." Both modes support a closed-loop process of "acquisition, inference, and adjustment," providing a complete solution for the full lifecycle health maintenance of large-aperture optical systems. Attached Figure Description

[0064] Figure 1 This is a flowchart illustrating an embodiment of the in-situ rapid measurement and calibration method for a large-aperture optical system according to the present invention.

[0065] Figure 2 This is a schematic diagram illustrating the principle of an embodiment of the in-situ rapid measurement and calibration method for a large-aperture optical system according to the present invention.

[0066] Figure 3 This is a schematic diagram of the VGG convolutional neural network model in step S3 of an embodiment of the in-situ rapid measurement and calibration method for a large-aperture optical system of the present invention. Detailed Implementation

[0067] To make the objectives, advantages, and features of the present invention clearer, the following detailed description of an in-situ rapid measurement and calibration method for a large-aperture optical system, in conjunction with the accompanying drawings and specific embodiments, provides further insight into this invention. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0068] Example 1

[0069] This invention discloses an in-situ rapid measurement and calibration method for large-aperture optical systems, applicable to scenarios where the imaging quality of a large-aperture optical system deteriorates after satellite launch and orbit insertion due to gravity release, changes in thermal environment, and satellite attitude control jitter. The core meaning of "in-situ" is to perform measurements while maintaining the large-aperture optical system in its original state and working position, without disassembling optical components, moving the large-aperture optical system to a dedicated testing platform, or inserting complex interferometers or Shaker-Hartmann wavefront sensors into the main optical path. The large-aperture optical system includes optical components and multi-degree-of-freedom adjustment mechanisms that drive the movement of these components; such as... Figure 1 , Figure 2 As shown, it includes the following steps:

[0070] S1. Constructing the dataset

[0071] S1.1. Establish a defocused star point image generation model (such as a parametric model of a large-aperture optical system) using optical design software (such as Zemax OpticStudio). The object side of the parametric model of the large-aperture optical system is set to infinity, and the field of view is set according to the imaging position of the star target on the detector. Typically, three typical fields of view are selected: the central field of view, the 0.7 field of view, and the edge field of view. A multidimensional hierarchical random sampling algorithm is used to generate multiple sets of misaligned samples in the misalignment parameter space. The misalignment parameter space is a multidimensional space spanned by the variables of each dimension in the multiple sets of misaligned samples.

[0072] Each set of misalignment samples is represented by an misalignment vector V, which is expressed as: V = [dx, dy, dz, tx, ty]; where dx is the lateral eccentricity of the optical element along the X-axis in a plane perpendicular to the optical axis; dy is the lateral eccentricity of the optical element along the Y-axis in a plane perpendicular to the optical axis; dz is the axial defocus of the optical element along the optical axis; tx is the tilt angle of the optical element rotating around the X-axis; and ty is the tilt angle of the optical element rotating around the Y-axis.

[0073] Based on the physical stroke of the secondary mirror adjustment mechanism, the value ranges of each dimension are set. For example, the value ranges of dx and dy are [-0.1mm, 0.1mm], the value range of dz is [-0.05mm, 0.05mm], and the value ranges of tx and ty are [-0.02°, 0.02°].

[0074] The multidimensional stratified random sampling algorithm, which is the Latin hypercube sampling algorithm (LHS), generates N sets of imbalance vectors in the five-dimensional imbalance parameter space (N=5000 in this embodiment). The LHS algorithm divides the range of variable parameters in each dimension into N equally probable non-overlapping intervals and randomly samples within each interval to ensure that the projection of the imbalance samples in any dimension is uniformly distributed.

[0075] The specific method for generating multiple sets of imbalanced samples in the imbalance parameter space using the Latin hypercube sampling algorithm is as follows: The Latin hypercube sampling algorithm (LHS) is used to divide the parameter range [L] of each of the K dimensions in the imbalance parameter space. k U k The system is divided into N equally probable non-overlapping intervals, and one imbalance sample is randomly sampled from each interval to obtain multiple sets of imbalance samples formed by combining the imbalance samples from the K dimensions and their corresponding N intervals; where the value X of the i-th imbalance sample in the k-th dimension is... ik The calculation formula is as follows:

[0076]

[0077] In the formula, i is the disordered sample number collected in the k-th dimension, i = 1, 2, ..., N, where N represents the total number of interval numbers and the total number of disordered samples collected in the k-th dimension, N ≥ 2; k is the dimension number, k = 1, 2, ..., K, where K is the total number of dimensions, K ≥ 1; L k with U k These represent the lower and upper bounds of the range of the k-th dimension variable, respectively; π k (i) -1 indicates that the interval numbering is changed from 1~N to 0~N-1; π k (i) represents the interval number randomly assigned to the i-th imbalanced sample in the k-th dimension; π k This represents a random permutation sequence from 1 to N, used to ensure that only one out-of-balance sample is selected in each interval of each dimension; It represents a random number that follows a uniform distribution in [0, 1) and is used to achieve local random perturbation within a small interval.

[0078] S1.2 Input the misaligned samples into the defocused star image generation model, drive the defocused star image generation model to generate the corresponding defocused star images, and construct a dataset containing the mapping relationship between the misaligned samples and the defocused star images; the defocused star images include at least the defocused star images of the central field of view, the 0.7 field of view and the edge field of view.

[0079] For example, a Python script can be written to automatically write each set of misalignment vectors into the optical design software via the Dynamic Data Exchange (DDE) interface or the ZOS-API interface, driving the parametric model to update the secondary mirror position. The Huygens PSF calculation engine is then invoked to generate defocused star images in the center field of view, 0.7 field of view, and edge field of view under the current misalignment state. During the generation process, measured secondary mirror surface shape error data (represented by Zernike polynomial coefficients) and simulated stray light background and detector readout noise under on-orbit conditions can be injected to improve the realism of the simulation data. Finally, a dataset containing 5000 sets of mapping relationships between the misalignment vector V and the defocused star image is constructed.

[0080] S2. Constructing the VGG convolutional neural network model

[0081] Build as Figure 3 The VGG convolutional neural network model shown is a multi-view feature fusion model. The VGG model includes multiple convolutional layers for extracting deep features from out-of-focus star images, a global average pooling (GAP) layer at the end of the convolutional layers, and a fully connected layer at the end of the GAP layer for outputting the offset prediction value. The GAP layer performs spatial averaging on each feature channel to eliminate spatial position information of the out-of-focus star image, thereby extracting translation-invariant wavefront features and suppressing the interference of imaging platform jitter on feature extraction. The GAP layer calculates the spatial average value for each feature channel, compressing the H×W×C feature map into a 1×1×C feature vector, thus eliminating spatial position information of the feature map. This makes the wavefront features extracted by the network sensitive only to the spot morphology and insensitive to the absolute position of the spot on the detector target surface. Finally, a fully connected layer is connected as a regression head to output a 5-dimensional offset prediction value.

[0082] In this embodiment, the VGG convolutional neural network model can be built based on the VGG16 network architecture. Its original fully connected layers are removed, while all convolutional and pooling layers are retained for feature extraction. A global average pooling layer is then added at the end, followed by a fully connected layer with 5 neurons as the output layer, used for regression prediction of the five-dimensional imbalance vector. The input image dimensions H and W can be set to 64×64 pixels.

[0083] S3. Training the VGG convolutional neural network model

[0084] The dataset generated in step S1.2 is preprocessed; the preprocessed dataset is divided into a training set and a validation set, and the training set and validation set are sequentially input into the VGG convolutional neural network model constructed in step S2 to train and validate the VGG convolutional neural network model, resulting in a dedicated network model suitable for the current environment. During training, a translation-invariant loss function is introduced to force the VGG convolutional neural network model to output consistent offset prediction values ​​for both the original defocused star images and the randomly translated defocused star images in the training set. The parameters of the VGG convolutional neural network model are optimized through backpropagation, and the training state of the VGG convolutional neural network model is verified using the validation set to prevent overfitting, until the VGG convolutional neural network model converges on the validation set, resulting in a dedicated network model suitable for the on-orbit environment (corresponding to the on-orbit application mode in subsequent steps S4 and S5).

[0085] The specific steps for preprocessing the dataset generated in step S1.2 are as follows: Square root operations are performed on the out-of-focus star images in the central field of view, 0.7 field of view, and edge field of view to compress the dynamic range, and then normalization is performed. Then, the normalized out-of-focus star images of the three fields of view are stacked in the channel dimension to construct a three-channel tensor of size 3×H×W, which is input into the VGG convolutional neural network model. Here, H represents the height of the out-of-focus star image, and W represents the width of the out-of-focus star image.

[0086] The formula for square root operation is:

[0087] ;

[0088] In the formula, I temp (u, v) represents the pixel value of the intermediate image obtained by taking the square root of each pixel of the original acquired defocused star point image, where (u, v) are the image pixel coordinates, I raw (u, v) represents the pixel values ​​of the original acquired out-of-focus star image. To prevent extremely small positive numbers from being introduced abnormally in numerical calculations, take... .

[0089] The formula for normalization is:

[0090]

[0091] In the formula, Iinput(u, v) represents the normalized pixel value input to the VGG convolutional neural network model, with a value range of [0, 1], and max(I temp ) represents the maximum value of the intermediate image obtained by taking the square root of each pixel of the original defocused star point image, min(T) temp () represents the minimum value of the intermediate image obtained by taking the square root of each pixel of the original defocused star point image.

[0092] Translation invariance loss function L total The calculation formula is:

[0093]

[0094] In the formula, L MSE L represents the mean square error term; Consist For consistency constraints; Weighting coefficients to balance prediction accuracy and jitter resistance performance;

[0095] Mean square error term L MSE The calculation formula is:

[0096]

[0097] In the formula, j is the batch number of the VGG convolutional neural network model training, j=1, 2, ..., B, and B is the total number of batches trained by the VGG convolutional neural network model; x m The tensor m for the m-th input defocused star image is the number of the input defocused star image in the j-th batch, where m = 1, 2, ..., M, and M is the total number of input defocused star images in the j-th batch; y m Let f be the true imbalance label vector corresponding to the m-th imbalance sample; f(·) is the forward inference mapping function of the VGG convolutional neural network model; The square of the L2 norm;

[0098] Consistency constraint L Consist The calculation formula is:

[0099]

[0100] Where T(x) m , Δu, Δv) are spatial affine transformation operators, which represent shifting the tensor of the input m-th defocused star image by Δu and Δv pixels in the horizontal and vertical directions, respectively.

[0101] S4. Acquire real-time off-focus star images and perform image preprocessing.

[0102] To obtain the real-time defocused star image output by the large-aperture optical system, a preprocessing operation is performed. The specific method for obtaining the real-time defocused star image output by the large-aperture optical system is as follows: adjust the satellite attitude so that the large-aperture optical system is aligned with a star of moderate brightness (to avoid detector saturation) and a clean background. Control the optical elements to acquire the defocused star image and crop the corresponding multi-field of view (center field of view, 0.7 field of view and edge field of view) region to obtain the real-time defocused star image output by the large-aperture optical system.

[0103] S5, Online Inference and Closed-Loop Assembly Control

[0104] The real-time defocused star image preprocessed in step S4 is input into the dedicated network model obtained in step S3, and the normalized offset prediction value is output. The normalized offset prediction value is subjected to inverse normalization calculation to obtain the actual offset in physical space and generate the corresponding control command. The control command is sent to the controller of the multi-degree-of-freedom adjustment mechanism, and the optical element is driven to perform inverse compensation motion through the multi-degree-of-freedom adjustment mechanism to complete the rapid measurement and calibration of the large-aperture optical system.

[0105] For example, by performing inverse normalization on the normalized value, the offset vector V[dx, dy, dz, tx, ty] in physical space is obtained. This offset vector V is then converted into control commands for the secondary mirror adjustment mechanism and sent to the six-degree-of-freedom parallel mechanism (Hexapod) controller connected to the secondary mirror, driving the secondary mirror to perform inverse compensation motion. After calibration, star images can be acquired again to calculate the modulation transfer function (MTF) or circumferential energy to verify the calibration effect. If the results are not as expected, steps S4-S5 can be repeated for iterative fine-tuning.

[0106] Example 2

[0107] The difference between this embodiment and Embodiment 1 is that this embodiment is applied to the ground integration test (AIT) phase before the launch of a large-aperture optical system, in a scenario affected by ground airflow disturbances (Seeing) and platform micro-vibrations.

[0108] S1. Constructing the dataset

[0109] The difference between this step and Example 1 is that the defocused star point image generation model is a joint simulation model of a collimator and a large-aperture optical system; in the joint simulation model, the point light source at the focal plane of the collimator is the object point, which is collimated by the collimator to simulate an infinitely distant target; when using the joint simulation model, the residual aberration of the wavefront emitted by the collimator is introduced as an error term of the large-aperture optical system during the generation of misaligned samples to participate in the training, so that the joint simulation model can adapt to the non-ideal characteristics of the ground detection equipment.

[0110] S2. The steps for constructing the VGG convolutional neural network model are the same as in Example 1.

[0111] S3. The difference between this step and Example 1 is as follows:

[0112] During training, data augmentation is performed using off-focus star images from the training set to simulate star jitter and deformation caused by air turbulence (e.g., adding random noise, Gaussian blur, etc.). Global average pooling layers are used to extract spot morphology features that are insensitive to airflow jitter. Using the augmented off-focus star image data, the parameters of the VGG convolutional neural network model are optimized through backpropagation. The training state of the VGG convolutional neural network model is verified using the validation set to prevent overfitting until the VGG convolutional neural network model converges on the validation set, resulting in a dedicated network model suitable for ground environments (corresponding to the ground application mode in subsequent steps S4 and S5).

[0113] S4. The difference between this step and Embodiment 1 is that the specific method for obtaining the real-time defocused star point image output by the large-aperture optical system is as follows: turn on the collimator, project the artificial star points it generates onto the large-aperture optical system, and adjust the exposure time of the large-aperture optical system to obtain the real-time defocused star point image output by the large-aperture optical system.

[0114] S5. The difference between this step and Example 1 is that the preprocessed artificial star point image is input into the processing module deployed on the ground workstation. The model quickly outputs the misalignment of the secondary mirror. Personnel or automated robotic arms adjust the position of the secondary mirror based on the output. Because this method does not require complex interferometer optical path construction, it can achieve efficient convergence of large-aperture systems from "coarse adjustment" to "fine adjustment".

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A method for in-situ rapid measurement and calibration of a large-aperture optical system, used for rapid measurement and calibration of a large-aperture optical system, wherein the large-aperture optical system includes optical elements and a multi-degree-of-freedom adjustment mechanism for driving the movement of the optical elements; characterized in that, Includes the following steps: S1. Construct the dataset; S1.

1. Establish a defocused star point image generation model using optical design software; generate multiple sets of misaligned samples in the misalignment parameter space using a multidimensional hierarchical random sampling algorithm; the misalignment parameter space is a multidimensional space spanned by the variables of each dimension in the multiple sets of misaligned samples. S1.2 Input the misaligned samples into the off-focus star image generation model to generate corresponding off-focus star images and construct a dataset containing the mapping relationship between the misaligned samples and the off-focus star images; the off-focus star images include at least off-focus star images of the center field of view, the 0.7 field of view and the edge field of view; S2. Construct the VGG convolutional neural network model; A VGG convolutional neural network model for multi-view feature fusion is constructed. The VGG convolutional neural network model includes multiple convolutional layers for extracting deep features of defocused star point images, a global average pooling layer at the end of the convolutional layers, and a fully connected layer at the end of the global average pooling layer for outputting the offset prediction value. The global average pooling layer is used to perform spatial averaging on each feature channel to eliminate the spatial position information of the out-of-focus star image, so as to extract wavefront features with translation invariance and thus suppress the interference of imaging platform jitter on feature extraction. S3. Train the VGG convolutional neural network model; The dataset generated in step S1.2 is preprocessed; the preprocessed dataset is divided into a training set and a validation set, and the training set and validation set are sequentially input into the VGG convolutional neural network model constructed in step S2 to train and validate the VGG convolutional neural network model, thereby obtaining a dedicated network model suitable for the current environment. S4. Acquire real-time defocused star point images and perform image preprocessing; Acquire the real-time defocused star image output by the large-aperture optical system and perform preprocessing; S5, Online Inference and Closed-Loop Assembly Control; The real-time defocused star image preprocessed in step S4 is input into the dedicated network model obtained in step S3, and the normalized misalignment prediction value is output. The normalized offset prediction value is denormalized to obtain the actual offset in the physical space and generate the corresponding control command. The control command is sent to the controller of the multi-degree-of-freedom adjustment mechanism, which drives the optical element to perform reverse compensation motion, thereby completing the in-situ rapid measurement and calibration of the large-aperture optical system.

2. The in-situ rapid measurement and calibration method for a large-aperture optical system according to claim 1, characterized in that: In step S1.1, the defocused star point image generation model is a parameterized model of a large-aperture optical system, with its object side at infinity and its field of view selected according to the star target. Alternatively, the defocused star point image generation model is a joint simulation model composed of a collimator and a large-aperture optical system; in the joint simulation model, the point light source at the focal plane of the collimator is the object point, and after collimation by the collimator, it simulates an infinity target.

3. The in-situ rapid measurement and calibration method for a large-aperture optical system according to claim 2, characterized in that: In step S1.1, the multidimensional stratified random sampling algorithm is the Latin hypercube sampling algorithm.

4. The in-situ rapid measurement and calibration method for a large-aperture optical system according to claim 3, characterized in that, In step S1.1, the specific method for generating multiple sets of imbalanced samples in the imbalance parameter space using the Latin hypercube sampling algorithm is as follows: The Latin hypercube sampling algorithm is used to determine the variable range [L] of each of the K dimensions in the imbalance parameter space. k U k The data is divided into N equally probable intervals, and one imbalanced sample is randomly sampled from each interval to obtain multiple sets of imbalanced samples formed by combining the K dimensions and the imbalanced samples sampled from the corresponding N intervals; where the value X of the i-th imbalanced sample in the k-th dimension is... ik The calculation formula is as follows: ； In the formula, i represents the interval number and the number of the disordered sample collected in the k-th dimension, i = 1, 2, ..., N, where N represents the total number of interval numbers and the total number of disordered samples collected in the k-th dimension, N ≥ 2; k represents the dimension number, k = 1, 2, ..., K, where K is the total number of dimensions, K ≥ 1; L k with U k These represent the lower and upper bounds of the range of the k-th dimension variable, respectively; π k (i) -1 indicates that the interval numbering is changed from 1~N to 0~N-1; π k (i) represents the interval number randomly assigned to the i-th imbalanced sample in the k-th dimension; π k This represents a random permutation sequence from 1 to N, used to ensure that only one out-of-balance sample is selected in each interval of each dimension; It represents a random number that follows a uniform distribution in [0, 1) and is used to achieve local random perturbation within a small interval.

5. The in-situ rapid measurement and calibration method for a large-aperture optical system according to claim 1, characterized in that: In step S1.1, each group of misaligned samples is an misalignment vector V, whose expression is: V = [dx, dy, dz, tx, ty]; In the formula, dx is the lateral eccentricity of the optical element along the X-axis in a plane perpendicular to the optical axis; dy is the lateral eccentricity of the optical element along the Y-axis in a plane perpendicular to the optical axis; dz is the axial defocusing of the optical element along the optical axis; tx is the tilt angle of the optical element rotating around the X-axis; and ty is the tilt angle of the optical element rotating around the Y-axis.

6. The in-situ rapid measurement and calibration method for a large-aperture optical system according to claim 1, characterized in that, In step S3, the VGG convolutional neural network model is trained and validated to obtain a dedicated network model suitable for the current environment. The specific training process is as follows: During training, a translation-invariant loss function is introduced to force the VGG convolutional neural network model to output consistent offset prediction values ​​for both the original defocused star images acquired in the training set and the defocused star images after random translation. The parameters of the VGG convolutional neural network model are optimized through backpropagation, and the training state of the VGG convolutional neural network model is verified through the validation set to prevent overfitting, until the VGG convolutional neural network model converges on the validation set, resulting in a dedicated network model suitable for on-orbit environments. And / or, during training, data augmentation is performed using off-focus star images from the training set to simulate star jitter and deformation caused by air turbulence. Global average pooling layers are used to extract spot morphology features that are insensitive to airflow jitter. The enhanced off-focus star image data is used to optimize the parameters of the VGG convolutional neural network model through backpropagation. The training state of the VGG convolutional neural network model is verified using the validation set to prevent overfitting, until the VGG convolutional neural network model converges on the validation set, resulting in a dedicated network model suitable for ground environments.

7. The in-situ rapid measurement and calibration method for a large-aperture optical system according to claim 6, characterized in that, In step S3, the specific steps for preprocessing the dataset generated in step S1.2 are as follows: Square root operations are performed on the defocused star images in the central field of view, 0.7 field of view, and edge field of view to compress the dynamic range, and then normalized. Then, the normalized defocused star images of the three fields of view are stacked in the channel dimension to construct a three-channel tensor of size 3×H×W, which is input into the VGG convolutional neural network model. Here, H represents the height of the defocused star image, and W represents the width of the defocused star image.

8. The in-situ rapid measurement and calibration method for a large-aperture optical system according to claim 7, characterized in that, In step S3, the formulas for the square root operation and normalization are as follows: ； In the formula, I temp (u, v) represents the pixel value of the intermediate image obtained by taking the square root of each pixel of the original acquired defocused star point image, where (u, v) are the image pixel coordinates, I raw (u, v) represents the pixel values ​​of the original acquired out-of-focus star image. To prevent extremely small positive numbers from being introduced abnormally in numerical calculations, take... ; ； In the formula, I input (u, v) are the normalized pixel values ​​input to the VGG convolutional neural network model, ranging from [0, 1], max(I temp ) represents the maximum value of the intermediate image obtained by taking the square root of each pixel of the original defocused star point image, min(T) temp () represents the minimum value of the intermediate image obtained by taking the square root of each pixel of the original defocused star point image.

9. The in-situ rapid measurement and calibration method for a large-aperture optical system according to claim 6, characterized in that, In step S3, the translation invariance loss function L total The calculation formula is: ； In the formula, L MSE L represents the mean square error term; Consist λ is the consistency constraint term; λ is the weighting coefficient that balances prediction accuracy and jitter resistance. The mean square error term L MSE The calculation formula is: ； In the formula, j is the batch number of the VGG convolutional neural network model training, j=1, 2, ..., B, and B is the total number of batches trained by the VGG convolutional neural network model; x m Let y be the tensor of the m-th input defocused star image, where m is the number of the input defocused star image in the j-th batch, m = 1, 2, ..., M, and M is the total number of input defocused star images in the j-th batch; m Let f be the true imbalance label vector corresponding to the m-th imbalance sample; f(·) is the forward inference mapping function of the VGG convolutional neural network model; The square of the L2 norm; The consistency constraint term L Consist The calculation formula is: ； Where T(x) m , Δu, Δv) are spatial affine transformation operators, which represent shifting the tensor of the input m-th defocused star image by Δu and Δv pixels in the horizontal and vertical directions, respectively.

10. The in-situ rapid measurement and calibration method for a large-aperture optical system according to claim 1, characterized in that, In step S4, the specific method for obtaining the real-time defocused star map output by the large-aperture optical system is as follows: Adjust the satellite attitude so that the large-aperture optical system is aligned with a star of moderate brightness and with a clean background. Control the optical elements to acquire images of defocused star points and capture the corresponding multi-field regions to obtain real-time defocused star point images output by the large-aperture optical system. Alternatively, the collimator can be turned on to project the artificial star points it generates onto the large-aperture optical system, and the exposure time of the large-aperture optical system can be adjusted to obtain a real-time defocused star point image output by the large-aperture optical system.

Images