Optical system adjustment method based on double-branch neural network model

Abstract

The invention relates to an optical system installation and adjustment method based on a double-branch neural network model, relates to the technical field of crossing of optical installation and adjustment and artificial intelligence, and solves the technical problems that in the prior art, a model is limited in misalignment prediction precision, low in generalization ability and difficult to apply to an actual installation and adjustment scene. The method comprises the following steps: analyzing element imbalance sensitivity; analyzing actual installation and adjustment scene noise; generating a neural network model training data set; constructing and training a neural network model; and the neural network model is utilized to guide actual installation and adjustment. According to the optical system installation and adjustment method based on the double-branch neural network model, the noise influence in an actual installation and adjustment scene is considered in the data set of model training, the anti-noise capability and robustness of the model are enhanced, and the applicability of applying the neural network model to real optical installation and adjustment is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intersection of optical alignment and artificial intelligence, and in particular to an optical system alignment method based on a dual-branch neural network model. Background Art

[0002] With the advancement of space optics, deep space exploration, and gravitational wave detection, the application environments of optical systems are becoming increasingly demanding, and the demands for imaging performance are becoming increasingly extreme. This poses significant challenges to the development of optical instruments, and new breakthroughs are urgently needed in optical system integrated assembly technology. Traditional optical assembly relies primarily on the experience of assemblers and conventional assembly equipment, resulting in low efficiency and difficulty ensuring accuracy and consistency. With the continuous advancement of computer simulation, optical design, optical processing, and testing technologies, computer-aided alignment (CAA) has emerged. This technology uses mathematical models to establish the relationship between optical system wavefront aberrations and component misalignment. It then uses measured wavefront data to determine component position errors, guiding assembly. Compared to traditional assembly methods, it can significantly improve optical assembly efficiency and system performance. Existing methods for calculating misalignment include the sensitivity matrix method, inverse optimization, vector aberration method, and artificial neural network method.

[0003] The sensitivity matrix method has the widest application range due to its simple model, mature research, and ease of use. The currently commonly used optical alignment approach combines manual alignment based on experience with computer-assisted alignment using the sensitivity matrix method. However, the sensitivity matrix method has the following drawbacks: 1) it only considers the low-order effects of component position on system imaging quality while ignoring higher-order effects. Consequently, a single iteration cannot yield an exact solution, requiring multiple iterations to approximate the solution. 2) it fails to consider the coupling between component degrees of freedom and the inter-coupling relationships between different wavefront Zernike coefficients, failing to reflect the fundamental laws governing the effects of alignment errors on system wavefront aberrations. 3) it fails to account for the impact of actual alignment noise, resulting in low precision in practical applications.

[0004] Artificial neural network methods, with their nonlinear mapping capabilities, autonomous feature learning mechanisms, and dynamic adaptive characteristics, have demonstrated significant advantages in the field of optical system assembly. By deeply exploring the complex nonlinear relationship between optical system wave aberrations and assembly errors, they can efficiently construct mathematical models of multivariable coupling relationships. Their application has gradually extended from basic research to engineering practice and will become an important technical means for precision optical assembly. The significant advantages of using artificial intelligence technology for optical assembly are reflected in the following aspects: 1) In terms of assembly accuracy, by adaptively learning the error transmission laws of optical systems, the aberration compensation accuracy and system stability are significantly improved; 2) In terms of efficiency improvement, by combining intelligent optimization algorithms to achieve rapid iteration of assembly parameters, promoting the intelligent upgrade of automated assembly processes; 3) In terms of cost control, by using data-driven virtual assembly previews, the frequency of physical experiments is reduced, effectively reducing resource loss. These technological breakthroughs will lay a key technical foundation for the evolution of optical assembly towards intelligence and integration.

[0005] Current research on the application of artificial neural networks to optical alignment primarily focuses on using optical simulation software to simulate misaligned optical systems, generating large datasets of misalignment and wavefront aberrations for training. Model training focuses on employing various artificial intelligence algorithms to construct more accurate nonlinear mappings between misalignment and wavefront Zernike coefficients. Existing methods are limited to using a single optical performance evaluation parameter as model input and lack the full utilization of multidimensional optical system information. This results in limited accuracy in misalignment prediction and poor generalization. Furthermore, the training process lacks consideration of noise factors such as optical component surface errors, mechanical actuator adjustment errors, and system wavefront measurement errors in actual alignment scenarios. This results in low robustness of high-precision models trained using simulation data, making them difficult to apply in real-world alignment scenarios. Summary of the Invention

[0006] The present invention aims to solve the technical problems in the prior art of limited accuracy of misalignment prediction by models, low model generalization ability, and difficulty in application to actual alignment scenarios, and provides an optical system alignment method based on a dual-branch neural network model.

[0007] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0008] An optical system assembly method based on a dual-branch neural network model includes the following steps:

[0009] Step 1: Component imbalance sensitivity analysis;

[0010] Construct a theoretical optical system model and calculate the misalignment sensitivity matrix of the optical system;

[0011] Step 2: Noise analysis of actual installation and adjustment scene;

[0012] Analyze and determine the noise and its distribution characteristics during the optical system installation process based on the actual installation environment;

[0013] Step 3: Generate neural network model training data set;

[0014] Generate a large number of data sets for training neural network assembly agent models;

[0015] Step 4: Neural network model construction and training;

[0016] Construct a dual-branch neural network assembly agent model and conduct neural network assembly agent model training;

[0017] Step 5: Use the neural network model to guide the actual installation and adjustment;

[0018] A neural network assembly agent model is used to guide the actual optical system assembly.

[0019] In the above technical solution, step 1 specifically includes:

[0020] Step 11: Establish a theoretical optical system model in optical analysis software;

[0021] Step 12: Build an integrated simulation analysis link to calculate the curve of the wavefront aberration evaluation index corresponding to each optical element and single position degree of freedom in the optical system as the misalignment changes, and obtain the misalignment sensitivity matrix of the optical system through curve fitting.

[0022] In the above technical solution, step 2 specifically includes:

[0023] Step 21: machining error and measurement error of optical element surface shape;

[0024] According to the statistical distribution law of errors under actual processing and measurement conditions, the specific distribution characteristics and ranges of processing errors and measurement errors are obtained and converted into the form of wavefront Zernike coefficients. Then, the processing errors and measurement errors are superimposed to comprehensively represent the surface error of the optical element after actual processing.

[0025] Step 22: wavefront detection error of the optical system;

[0026] Through empirical analysis and measured data analysis, the error statistical distribution law of the wavefront measurement instrument is obtained and converted into the form of wavefront Zernike coefficients;

[0027] Step 23: Mechanical actuator error;

[0028] Through empirical judgment and measured data analysis, the distribution law and range of the system cumulative error and random noise of the actuator are determined, and converted into the expression of the six degrees of freedom of the component posture.

[0029] In the above technical solution, step 3 specifically includes:

[0030] Step 31: Determine the input and output of the neural network assembly agent model training;

[0031] Step 32: Based on the offset range of each degree of freedom of the component determined in step 1 and the distribution pattern and range of various noises determined in step 2, batch generate a set of data pairs of the offset of each degree of freedom of the component for neural network model training and taking into account the influence of various actual noises.

[0032] In the above technical solution, the specific method of adding various noises in step 32 is as follows:

[0033] First, the surface error of the optical element determined in step 21 is added to the theoretical optical design model, and the model is used as the ideal model in the subsequent data generation and iterative regression process of adjustment;

[0034] Next, when generating each set of randomly uniformly distributed misalignment combinations, the noise of the adjustment mechanical actuator is added to each adjustable component position degree of freedom according to the noise distribution pattern and range of the adjustment mechanical actuator determined in step 23. Then, the optical system wavefront diagram and its Zernike coefficient corresponding to the random misalignment combination after adding the noise are calculated.

[0035] Finally, wavefront detection noise is added to the calculated wavefront aberration data of the optical system;

[0036] The wavefront detection noise of the optical system determined in step 22 and expressed in the form of Zernike coefficients is directly added to the wavefront Zernike coefficients calculated in the previous step, and the noise distribution is converted into an image form and superimposed on the wavefront diagram.

[0037] In the above technical solution,

[0038] Step 4 specifically includes:

[0039] Step 41: data preprocessing;

[0040] A neural network assembly proxy model is proposed to realize information acquisition and feature fusion of two modal data. By normalizing and standardizing the input optical system wavefront image and its fitted Zernike coefficients and the output optical system component posture degree of freedom misalignment, the input and output data meet the requirements of the neural network assembly proxy model for numerical scale and distribution consistency, thereby improving the stability and accuracy of the neural network assembly proxy model training.

[0041] Step 42: Model structure design;

[0042] The neural network assembly agent model structure includes: a dual-branch feature extraction module, a feature fusion module, and a component misalignment prediction module;

[0043] The dual-branch feature extraction module extracts features from the wavefront image branch and the wavefront Zernike coefficient branch, and then fuses the extracted features through the feature fusion module. Finally, the component misalignment prediction module achieves high-precision prediction of component misalignment.

[0044] Step 43: Model training;

[0045] By using large-scale labeled data, the parameters of the dual-branch feature extraction module, feature fusion module, and component misalignment prediction module are optimized to achieve high-precision prediction of optical component misalignment. Based on the input data and the training set of corresponding multi-degree-of-freedom misalignment labels, the mean squared error loss function is used to optimize the parameters of the neural network alignment proxy model. The parameters are iteratively updated using a mini-batch gradient descent algorithm, combined with regularization and early stopping mechanisms to effectively prevent overfitting, ultimately achieving accurate estimation of the optical component misalignment state.

[0046] Step 44: Model verification;

[0047] First, from the perspective of prediction accuracy, the mean square error and mean absolute error indicators are used to evaluate the difference between the misalignment output by the neural network tuning agent model in the test set and the true value:

[0048] ,

[0049] ,

[0050] in, is the model's predicted misalignment, is the actual offset, is the total number of test samples, is the test sample number, represents the mean square error, represents the mean absolute error;

[0051] Next, to further evaluate the application effect of the neural network adjustment agent model in the actual optical system adjustment, the wavefront Zernike coefficients after the optical system is adjusted using the predicted misalignment are used to calculate the wavefront RMS value, and the differences in the wavefront RMS values ​​of the optical system before and after adjustment are compared. The formula for calculating the RMS value using the Zernike coefficients is:

[0052] ,

[0053] in, For the wavefront Zernike coefficients, is the polynomial order used, Indicates the Zernike coefficient number of the Zernike wavefront.

[0054] In the above technical solution,

[0055] Step 41 specifically includes:

[0056] Step 411: pre-processing of the wavefront image;

[0057] The pixel values ​​of the wavefront image are normalized and the channels are normalized to ensure the stability of the training process. The formulas for pixel value normalization and channel normalization are as follows:

[0058] ,

[0059] ,

[0060] in, Represents the original wavefront image The pixel value after channel normalization, and Represents the pixel coordinates, Represents the original wavefront image in coordinates The first Channel pixel value, Channels representing the three colors red, green, and blue, is the channel index of the image, Represents the wavefront diagram The pixel value after channel normalization, and The first The mean and standard deviation of each channel;

[0061] Step 412: pre-processing of wavefront Zernike coefficients;

[0062] Based on the results of sensitivity analysis, the Zernike polynomial components that have a significant impact on the imaging quality change of the wavefront system are selected as network inputs, thereby reducing feature dimensions and redundant information. All wavefront Zernike coefficients are transformed into data with a mean of 0 and a standard deviation of 1 through the Z-score normalization operation, eliminating the dimensional differences between coefficients of different orders, making all features have similar scales and distributions, providing a more stable gradient path for the neural network optimizer, and reducing the learning difficulty of the neural network adjustment agent model. For each Zernike term coefficient , the standardization method formula used is as follows:

[0063] ,

[0064] in, is the sample mean of the coefficient, is the standard deviation, Representative coefficient To standardize the process, Indicates the item coefficient number.

[0065] In the above technical solution,

[0066] Step 42 specifically includes:

[0067] Step 421, dual-branch feature extraction module;

[0068] Wavefront image branch: focuses on the texture and color distribution of the image space, capturing the local disturbance and distortion of the wavefront; the wavefront image branch is based on the wavefront color image As input, it reflects the spatial distribution and phase change of the optical system wavefront, contains rich local texture and color information, and reveals the slight disturbance of the wavefront; the wavefront branch network structure uses a multi-layer convolutional neural network to extract spatial local features, and each layer of convolution operation is performed by a filter and bias Complete, use the ReLU activation function to ensure the capture of nonlinear information. The specific formula is:

[0069] ,

[0070] in, Indicates the The feature map of the layer, Indicates the Layer, The feature map of each channel, is the convolution kernel of the corresponding channel, Indicates the layer index of the network, Indicates the channel number. Represents the total number of layers of the convolutional neural network;

[0071] The output of the last layer of the convolutional neural network is a three-dimensional feature map , Indicates that the output is a three-dimensional real tensor, is the height of the feature map, corresponding to the number of rows of output, is the width of the feature map and the number of columns of the corresponding output, is the number of channels. It is converted into a one-dimensional vector after flattening operation. , represents the flattening operation, Represents a vector space, dimension Equal to the total number of all elements of the feature map, this one-dimensional vector contains the high-order spatial characteristics of the wavefront image, reflecting the complex local and global optical structure information; Wavefront Zernike coefficient branch: The optical system wavefront is fitted with the Zernike polynomial coefficient vector is the input, the number of coefficients Typically covers low to mid-order major aberration terms, represents transpose, express dimensional real vector space; the wavefront Zernike coefficient branch network structure uses a multi-layer fully connected network for nonlinear feature mapping, and each layer is connected by a weight matrix and the bias vector Implement affine transformation and combine it with ReLU activation function to implement nonlinear operation. The specific formula is:

[0072] ,

[0073] ,

[0074] Finally, the deep feature vector is output As a high-level representation of this branch;

[0075] in, represents the output vector of the first hidden layer, represents the weight matrix of the first layer, represents the input vector of the current layer, represents the bias vector of the first layer, Indicates the The output vector of the hidden layer, Indicates the The output vector of the hidden layer, Indicates the hidden layer index, the value range is , represents the total number of layers in the fully connected network, Indicates the dimension The real number space of ;

[0076] Step 422, feature fusion module;

[0077] The high-order eigenvectors from the wavefront map branch and the wavefront Zernike coefficient branch are and Effectively integrate and give full play to the complementary advantages of the two modal data to achieve collaborative expression of information;

[0078] In order to achieve effective fusion, the eigenvectors of the wavefront image branch and the wavefront Zernike coefficient branch are first and Mapped to the same dimensional space :

[0079] ,

[0080] ,

[0081] in, and are the mapping matrices of the wavefront Zernike coefficient branch and the wavefront graph branch, are bias vectors; Represents the dimensional space as The real number space, Represents the wavefront branch mapped to the dimensional space The eigenvector of Represents the wavefront Zernike coefficient branch mapping to the dimensional space The eigenvector of

[0082] Next, the weighted summation method is used to achieve feature fusion of the wavefront branch and the Zernike coefficient branch:

[0083] ,

[0084] Among them, the weight coefficient To balance the contribution of the wavefront graph branch and the wavefront Zernike coefficient branch, Represents the eigenvector after the wavefront image branch and the wavefront Zernike coefficient branch are fused;

[0085] Step 423, component imbalance prediction module;

[0086] Based on the fused feature vector , a fully connected network is used to regress and predict the misalignment of multiple adjustable degrees of freedom of the optical system; through multi-layer nonlinear mapping, the complex correspondence between features and misalignment parameters is extracted to achieve high-precision misalignment estimation. The specific formula is as follows:

[0087] ,

[0088] in, The optical element degree of freedom misalignment parameter predicted by the element misalignment prediction module, is the mapping function of the multilayer perceptron network.

[0089] In the above technical solution,

[0090] Step 5 specifically includes:

[0091] Step 51, calibrating the optical axis reference;

[0092] Install the laser interferometer, theodolite, plane mirror, and six-degree-of-freedom adjustment frame, and place the main support structure of the optical system on the air-floating platform; use the six-degree-of-freedom adjustment frame and theodolite to continuously adjust the plane mirror so that the plane mirror and the optical axis reference plane of the main support structure are perpendicular to each other; adjust the laser interferometer to align with the optical axis reference plane to achieve optical axis alignment between the laser interferometer and the main support structure;

[0093] Step 52, rough assembly and adjustment;

[0094] First, the primary mirror is adjusted: using a laser interferometer, theodolite, three-dimensional coordinate measuring arm, altimeter, and six-degree-of-freedom adjustment frame, the angle between the primary mirror reference plane and the plane reflector is adjusted to the ideal angle. Then, the laser interferometer is used to monitor the primary mirror shape in real time, and the primary mirror position is continuously adjusted to ensure that the primary mirror shape error is consistent with the measured result.

[0095] Secondly, perform coarse adjustment of the secondary and tertiary mirrors: Use a six-degree-of-freedom adjustment frame to preliminarily place the secondary and tertiary mirrors in their designed positions. During the coarse adjustment process, use a theodolite and micrometer to monitor the adjustment of each degree of freedom of the secondary and tertiary mirrors in real time. The adjustment of each degree of freedom is compared with the designed value to ensure the relative spatial position accuracy between the plane reflectors in the optical system.

[0096] Step 53, hardcover adjustment;

[0097] The optical system's wavefront is detected using an autocollimation method, with wavefront detection at multiple viewpoints achieved through deflection of a plane mirror and a laser interferometer. A trained neural network adjustment proxy model is used to calculate the predicted adjustment for each degree of freedom of the optical element based on the wavefront detection results of multiple viewpoints measured by the system, thereby guiding the adjustment of the secondary and tertiary mirrors.

[0098] During the fine assembly and adjustment process, after each measurement, the wavefront diagrams of multiple fields of view and the fitted Zernike coefficients are input into the neural network assembly and adjustment agent model to calculate the recommended adjustment amount of each component degree of freedom; then, the component position is adjusted according to the recommended adjustment amount through the six-degree-of-freedom adjustment frame; after the adjustment, the optical system wavefront is measured again; through multiple rounds of iterations of the above process, the wavefront aberration of the actual optical system is made close to the wavefront aberration of the optical system after considering various errors, ensuring that all optical performance indicators are within an acceptable range, and the assembly and adjustment is completed.

[0099] The present invention has the following beneficial effects:

[0100] The optical system assembly method based on the dual-branch neural network model of the present invention considers the noise influence in the actual assembly scene in the data set of the neural network model training, enhances the noise resistance and robustness of the neural network model, and improves the applicability of the neural network model in real optical assembly.

[0101] The optical system assembly and adjustment method based on the dual-branch neural network model of the present invention adopts the dual-branch neural network model in model training, fully mines and utilizes optical system information through feature fusion of two modal data of system wavefront image and Zernike coefficient, and improves the accuracy of component misalignment solution and the stability of the assembly and adjustment process.

[0102] The optical system assembly method based on the dual-branch neural network model of the present invention adopts an artificial neural network model to establish a nonlinear mapping relationship between component misalignment and optical system performance, which can fully explore the essential laws of the influence of assembly errors. While improving the assembly accuracy, it effectively reduces the dependence of traditional assembly methods on manual assembly experience and improves the efficiency of optical assembly.

[0103] The optical system assembly method based on the dual-branch neural network model of the present invention can also reduce the frequency of physical experiments through data-driven virtual assembly preview, effectively reducing resource loss and cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0105] Figure 1 It is a flow chart of the optical system assembly method based on the dual-branch neural network model of the present invention.

[0106] Figure 2 Schematic diagram of the off-axis three-mirror optical system structure, where (a) is a plane diagram and (b) is a three-dimensional diagram.

[0107] Figure 3 Generate a flow diagram for the training dataset.

[0108] Figure 4 Schematic diagram of the dual-branch neural network model architecture.

[0109] Figure 5 Schematic diagram of the off-axis three-mirror optical system assembly scene.

[0110] The reference numerals in the figures indicate:

[0111] M1-primary mirror; M2-secondary mirror; M3-tertiary mirror. DETAILED DESCRIPTION

[0112] The present invention will be described in detail below with reference to the accompanying drawings.

[0113] Figure 1This is a flow chart of the optical system assembly method based on a dual-branch neural network model of the present invention, which shows the overall flow chart of optical assembly using an artificial neural network method, mainly including: constructing a theoretical optical system model, component misalignment sensitivity analysis, determining the component misalignment range and participating in neural network model training parameters, actual assembly scene noise analysis, neural network model training data set generation, neural network model construction and training, and using the trained neural network model to guide actual optical system assembly (the neural network model of the present invention is a dual-branch neural network model, and the neural network assembly agent model involved in the following content description refers to the dual-branch neural network model).

[0114] The optical system assembly method based on the dual-branch neural network model of the present invention accurately calculates the component misalignment by fusing the features of two modal data, namely the optical system wavefront image and the wavefront Zernike coefficient, and adds the optical component surface error, the assembly mechanical actuator adjustment error and the optical system wavefront detection (measurement) error to the training data to enhance the robustness of the neural network model in practical applications. The optical system sensitivity matrix is ​​added to the loss function of the neural network model training as physical information guidance, and the optical system wavefront RMS is additionally used as a physical evaluation indicator in the evaluation indicators of the neural network model verification to further improve the accuracy and robustness of the neural network model and enhance the practical application capability.

[0115] The optical system assembly and adjustment method based on the dual-branch neural network model of the present invention comprises the following steps:

[0116] Step 1: Component imbalance sensitivity analysis;

[0117] Construct a theoretical optical system model and calculate the misalignment sensitivity matrix of the optical system;

[0118] Based on the actual alignment conditions, the misalignment range of the six degrees of freedom (DOF) of each optical system component is determined. Optical software is then used to calculate the corresponding wavefront evaluation indicators (such as wavefront Zernike coefficients, RMS values, MTF, distortion, and other parameters) for individual components and individual degrees of freedom within the misalignment range. By analyzing the response of these evaluation indicators to changes in misalignment, the impact of each degree of freedom on the imaging quality of the optical system is evaluated. This provides a physical foundation and prior knowledge for the subsequent establishment of a neural network alignment proxy model, guiding the selection of network input parameters and the setting of loss functions, and improving the efficiency and physical rationality of the neural network alignment proxy model training.

[0119] Step 2: Noise analysis of actual installation and adjustment scene;

[0120] The installation and adjustment of high-precision optical systems place extremely stringent environmental demands, requiring multi-dimensional control to ensure the stability of environmental parameters such as temperature, vibration, cleanliness, and humidity. However, in addition to environmental disturbances, several types of noise remain difficult to eliminate through control. These include machining and measurement errors in the optical component's surface shape, wavefront detection errors in the optical system, and adjustment errors in the mechanical actuators used for installation and adjustment. This requires combining full-link error decomposition of the measurement equipment with experimental data analysis to determine the noise distribution pattern. This noise distribution should then be added to the training dataset to develop a misalignment prediction model that is unaffected by the noise in the actual installation and adjustment environment.

[0121] Step 3: Generate neural network model training data set;

[0122] First, the machining errors and measurement errors of the optical component surface shape are added to the theoretical optical system model, and the optical system model affected by the actual component surface shape errors is used as the reference model for the assembly regression. Then, optical simulation software and self-written programs are used to batch generate a large number of random misalignment data pairs of components within the determined misalignment range, and the corresponding optical system wavefront distribution diagram and its fitted wavefront Zernike coefficient are calculated. Finally, the adjustment noise (i.e., adjustment error) of the assembly mechanical actuator obtained by analysis is added to the random misalignment combination data according to its distribution law, and the wavefront detection error of the optical system is added to the simulated wavefront data according to its distribution law, thereby obtaining a "misalignment combination and system wavefront aberration" data pair set that takes into account the influence of noise and is more in line with the actual assembly environment. It is used for training the neural network assembly agent model, see. Figure 3 .

[0123] Step 4: Build and train the neural network model (i.e., neural network assembly agent model);

[0124] A dual-branch neural network model is proposed for information acquisition and feature fusion of two modal data, training a high-precision component misalignment state estimation model with noise immunity. First, the dual-branch feature extraction module processes the input data of the two modalities. A multi-layer fully connected network (MLP) is used to extract deep semantic information from the wavefront Zernike coefficients, while a multi-layer convolutional neural network (CNN) is used to extract local spatial features from the wavefront image. To improve the physical validity of the input features, the input wavefront Zernike coefficients are screened based on sensitivity analysis results, retaining only those coefficients that significantly impact the imaging quality of the optical system and suppressing redundant features dominated by noise.

[0125] Then, a feature fusion module effectively integrates high-dimensional feature vectors from the wavefront image and wavefront Zernike coefficient branches, achieving collaborative information expression and enhancing feature discrimination capabilities, providing more comprehensive and robust input for downstream component misalignment estimation. In the misalignment state prediction module, a fully connected neural network (FCN) utilizes the fused feature vectors to regress and predict the misalignment state of multiple degrees of freedom in the optical system. During training, physical prior knowledge based on the sensitivity matrix is ​​introduced to construct a weighted error term for misalignment, enhancing the model's accuracy in predicting key degrees of freedom and ensuring that the prediction results are more consistent with the actual system response characteristics. Ultimately, a neural network assembly proxy model with physical interpretation capabilities is constructed, which can be used to guide the efficient assembly and error compensation of complex optical systems.

[0126] Step 5: Use the neural network model to guide the actual installation and adjustment;

[0127] The actual optical alignment process requires the following steps: optical axis datum calibration, rough alignment, and fine alignment. The rough alignment process requires placing the optical system's various components as close to their ideal design state as possible, ensuring that the wavefront aberrations of the actual optical system can be measured, and keeping the misalignment of each component's degree of freedom within the applicable range of the neural network alignment proxy model, thus providing a good iterative starting point for the fine alignment process. The fine alignment process involves inputting the measured wavefront data into a trained neural network alignment proxy model (misalignment state estimation model) to calculate the predicted misalignment of each degree of freedom. The alignment mechanical actuator is then used to adjust the positional degrees of freedom of the corresponding component according to the predicted misalignment. This process is iterated repeatedly until the measured performance indicators of the optical system meet the design requirements.

[0128] The following will take the assembly and adjustment process of an off-axis three-mirror optical system as an example and further describe the technical solution process of the present invention in conjunction with the accompanying drawings.

[0129] Step 1: Component imbalance sensitivity analysis;

[0130] Construct a theoretical optical system model and calculate the misalignment sensitivity matrix of the optical system. Specifically including:

[0131] Step 11: Create a theoretical optical system model (such as Figure 2 As shown, Figure 2 The optical path diagrams of the multi-path light transmission paths between the various reflectors in an off-axis three-mirror optical system are shown, where (a) is a plan view diagram and (b) is a three-dimensional diagram. The secondary mirror M2 in the figure is the aperture, and the image plane is the image plane. The misalignment range of each degree of freedom of the component posture is determined based on actual assembly experience.

[0132] Step 12. Use MATLAB + CODEV software (or other programming languages ​​+ optical analysis software) to build an integrated simulation analysis link to calculate the curve of the wavefront aberration evaluation indicators (such as MTF, RMS value, wavefront Zernike coefficient, etc.) corresponding to each optical element and single position degree of freedom in the optical system as the misalignment changes, and obtain the misalignment sensitivity matrix of the optical system through curve fitting.

[0133] Step 2: Noise analysis of actual installation and adjustment scene;

[0134] According to the actual installation environment, the noise and its distribution characteristics during the installation of the optical system are analyzed and determined.

[0135] High-precision optical system assembly and adjustment has extremely stringent environmental requirements. In addition to environmental disturbance factors, there are still the following types of noise that are difficult to eliminate through control. It is necessary to analyze and determine the distribution characteristics of the noise and add it to the training data of the neural network assembly and adjustment agent model to reduce the impact of noise on assembly and adjustment accuracy. Specifically, they include:

[0136] Step 21: machining error and measurement error of optical element surface shape;

[0137] Machining error is the deviation between the finished optical component's surface shape and the ideal designed shape caused by factors such as machining equipment (such as grinders and polishers), tool accuracy, material properties, process parameter control, and environmental factors (such as temperature fluctuations and clamping deformation) during the optical component manufacturing process. Measurement error is the error between the actual component's surface shape and the measured result during the actual surface measurement of an optical component due to factors such as the accuracy of the measurement equipment (interferometers, profilometers, etc.), environmental interference, the operation and installation of the measurement fixture, and post-measurement data processing and algorithms. Based on the statistical distribution of errors under actual machining and measurement conditions, the specific distribution characteristics and ranges of these two errors are derived and converted into wavefront Zernike coefficients (or other forms). The two errors are then superimposed to comprehensively represent the actual surface error of the optical component after machining.

[0138] Step 22: wavefront detection error of the optical system;

[0139] When using a laser interferometer to measure the wavefront of an optical system, there is a deviation between the optical system wavefront detection result and the actual wavefront due to factors such as environmental disturbances, laser light source fluctuations, electronic signal processing, and wavefront fitting deviations. Through empirical rules and analysis of measured data, the error statistical distribution law of the wavefront measurement instrument is obtained and converted into the form of wavefront Zernike coefficients.

[0140] Step 23: Adjust the adjustment error of the mechanical actuator;

[0141] During the actual assembly and adjustment of optical systems, mechanical actuators are used to control and adjust the position and posture of optical components. Due to factors such as environmental disturbances, mechanical manufacturing and assembly defects, material deformation, motor drive and control errors, there is a deviation between the nominal adjustment amount of the component and the actual movement amount. Through empirical judgment and analysis of measured data, the system cumulative error and random noise distribution law and range of the assembly and adjustment mechanical actuator are determined, and converted into an expression of the component's six degrees of freedom.

[0142] Step 3: Generate neural network model training data set;

[0143] Generate a large number of data sets for training neural network assembly agent models. Specifically including:

[0144] Step 31. Determine that the input for training the neural network assembly agent model is: the wavefront distribution diagram of the optical system and the 4th to 16th items (Z4-Z16) of its fitted Zernike coefficients; the output is: the misalignment of the component position freedom that can be adjusted during the actual assembly process of the optical system (for the off-axis three-mirror optical system, this is: the eccentricity of the secondary mirror M2 and the tertiary mirror M3 in the X, Y, and Z directions, and the rotational freedom in the X and Y directions).

[0145] Step 32: Using MATLAB + CODE V software (or other programming language + optical analysis software), based on the misalignment range of each degree of freedom of the component determined in step 1 and the distribution pattern and range of various noises determined in step 2, batch generate a set of data pairs of "optical system wavefront image + Zernike coefficient ~ component misalignment of each degree of freedom" for neural network assembly agent model training and taking into account the influence of various actual noises. The training data set generation process is as follows: Figure 3 As shown ( Figure 3 This paper demonstrates the process of generating a training dataset for a neural network proxy model for optical assembly. To develop a noise-resistant neural network proxy model suitable for practical optical assembly scenarios, the training dataset considers component surface errors (i.e., machining and measurement errors of component surface shapes), adjustment errors of the mechanical actuators, and wavefront detection errors of the optical system. The training data pairs are: (combinations of random component misalignments and system wavefront aberrations). The specific methods for adding various types of noise are as follows:

[0146] First, the component surface error expressed in the form of Zernike coefficients (or other forms) determined in step 21 is added to the theoretical optical design model, and this optical model is used as the "ideal model", i.e., the reference model, in the subsequent data generation and iterative regression process of assembly.

[0147] Next, when generating each set of randomly uniformly distributed misalignment combinations, the noise of the adjustment mechanical actuator is added to each adjustable component position degree of freedom according to the noise distribution pattern and range determined in step 23. Then, the optical system wavefront diagram and its Zernike coefficient corresponding to the random misalignment combination after adding this noise are calculated.

[0148] Finally, the wavefront measurement noise of the optical system is added to the calculated wavefront aberration data of the optical system (i.e., the wavefront diagram of the optical system and its Zernike coefficients). Specifically, the wavefront measurement noise of the optical system (i.e., the detection error) determined in step 22 and expressed in the form of Zernike coefficients is directly added to the wavefront Zernike coefficients calculated in the previous step, and the noise distribution is converted into an image form and superimposed on the wavefront diagram.

[0149] Step 4: Neural network model construction and training;

[0150] Construct a neural network assembly agent model and conduct neural network assembly agent model training. Specifically include:

[0151] Step 41: data preprocessing;

[0152] A dual-branch neural network model is proposed to achieve information acquisition and feature fusion of two modal data. By normalizing and standardizing the input wavefront image and its fitted Zernike coefficients and the output component posture misalignment, the model meets the requirements of the neural network for numerical scale and distribution consistency, thereby improving the stability and accuracy of the neural network assembly agent model training. Specifically, it includes:

[0153] Step 411: pre-processing of the wavefront image;

[0154] A wavefront image is typically a color image that maps two-dimensional wavefront phase or optical path difference data to a color image. Color represents the phase value and gradient, encoding the spatial distribution and structural information of the wavefront and intuitively reflecting the spatial changes of the wavefront. This paper proposes to normalize the pixel values ​​of the wavefront image and normalize the channels to ensure the stability of the training process. The formulas for pixel value normalization and channel normalization are as follows:

[0155] ,

[0156] ,

[0157] in, Represents the original wavefront image The pixel value after channel normalization, and Represents the pixel coordinates, Represents the original wavefront image in coordinates The first Channel pixel value, Channels representing the three colors red, green, and blue, is the channel index of the image, Represents the wavefront diagram The pixel value after channel normalization, and The first The mean and standard deviation of each channel.

[0158] Step 412: pre-processing of wavefront Zernike coefficients;

[0159] Zernike polynomials are a set of ordered orthogonal polynomials that can be used to express wavefronts, and their coefficients are often used to describe the size of each aberration (such as spherical aberration, coma, astigmatism, etc.). Based on the results of sensitivity analysis, the present invention preferentially selects Zernike polynomial components that have a significant impact on changes in the imaging quality of the optical system (such as wavefront RMS, Zernike coefficients, etc.) as network inputs, thereby reducing feature dimensions and reducing redundant information. In addition, since the physical meanings represented by each coefficient are different, their numerical distribution and magnitude are also inconsistent for a specific optical system. Through the Z-score normalization operation, all wavefront Zernike coefficients are transformed into data with a mean of 0 and a standard deviation of 1, eliminating the dimensional differences between coefficients of different orders, so that all features have similar scales and distributions, providing a more stable gradient path for the neural network optimizer and reducing the learning difficulty of the dual-branch neural network model. For each Zernike coefficient term , the standardization method formula used is as follows:

[0160] ,

[0161] in, is the term coefficient The sample mean of is the standard deviation, Represents each Zernike coefficient To standardize the process, Representative coefficient number.

[0162] Step 42: Model structure design;

[0163] The present invention designs a dual-branch neural network model, namely a neural network adjustment agent model, which is used to fuse the wavefront map and the wavefront Zernike coefficient, two key data used to describe the wavefront aberration of the optical system, and jointly predict the multi-degree-of-freedom misalignment of the optical system. The dual-branch neural network model consists of two parallel branches, which process different modal data respectively and output accurate misalignment through feature fusion. The dual-branch neural network model architecture is divided into three parts, including: a dual-branch feature extraction module, a feature fusion module and a component misalignment prediction module, such as Figure 4 As shown, Figure 4 The dual-branch neural network model architecture diagram is shown, which mainly includes: dual-branch feature extraction module, feature fusion module and component misalignment prediction module. The dual-branch feature extraction module extracts features from the wavefront image branch and the wavefront Zernike coefficient branch, and then fuses the extracted features through the feature fusion module to fully explore and utilize the optical system information. Finally, the component misalignment prediction module achieves high-precision prediction of component misalignment. Specifically including:

[0164] Step 421, dual-branch feature extraction module;

[0165] This module is designed with two independent branches to process the input data of the wavefront image and the wavefront Zernike coefficient respectively, realizing the multi-level and multi-angle feature extraction of the wavefront information of the optical system. By identifying the physical meaning of different optical performance indicators and their inherent deep correlation, this design can effectively mine the key features in multi-source data, thereby improving the expression ability of the model. The two branches output high-order feature vectors respectively. and , representing the deep semantic information and local spatial features of the input data of the two branches respectively.

[0166] Wavefront image branch: focuses on the texture and color distribution of the image space, captures the local disturbance and distortion of the wavefront, and makes up for the inadequacy of the wavefront Zernike coefficient branch in expressing local information. As input, it reflects the spatial distribution and phase change of the optical system wavefront, contains rich local texture and color information, and reveals the slight perturbation of the wavefront. The wavefront branch network structure uses a multi-layer convolutional neural network (CNN) to extract spatial local features. Each layer of convolution operation is performed by a filter. and bias Complete, use the ReLU activation function to ensure the capture of nonlinear information. The specific formula is:

[0167] ,

[0168] in, Indicates the The feature map of the layer, Indicates the Layer, feature maps of channels is the convolution kernel of the corresponding channel. Indicates the layer index of the network, Indicates the channel number. Represents the total number of layers in the convolutional neural network.

[0169] The output of the last layer of the convolutional neural network is a three-dimensional feature map , Indicates that the output is a three-dimensional real tensor, is the height of the feature map, corresponding to the number of output rows, is the width of the feature map, corresponding to the number of output columns, is the number of channels, which is converted into a one-dimensional vector after flattening , represents the flattening operation, Represents a vector space, dimension Equal to the total number of all elements of the feature map, this one-dimensional vector contains the high-order spatial features of the wavefront image, reflecting the complex local and global optical structure information.

[0170] Wavefront Zernike coefficient branch: Fitting the Zernike polynomial coefficient vector to the optical system wavefront is the input, the number of coefficients Typically covers low to mid-order major aberrations (select items Z4 to Z16), represents transpose, is a A column vector of coefficients, express dimensional real vector space. The wavefront Zernike coefficient branch network structure uses a multi-layer fully connected network (MLP) for nonlinear feature mapping, and each layer is calculated by the weight matrix and the bias vector Implement affine transformation and combine it with ReLU activation function to implement nonlinear operation. The specific formula is:

[0171] ,

[0172] ,

[0173] Finally, the deep feature vector is output As a high-order representation of this branch, it can comprehensively express the global characteristics of the wavefront and contains rich physical meanings. represents the output vector of the first hidden layer, represents the weight matrix of the first layer, represents the input vector of the current layer, represents the bias vector of the first layer, Indicates the The output vector of the hidden layer, Indicates the The output vector of the hidden layer, Indicates the hidden layer index, the value range is , represents the total number of layers in the fully connected network, Indicates the dimension The real number space of ;

[0174] Step 422, feature fusion module;

[0175] The feature fusion module combines the high-dimensional feature vectors from the wavefront image branch and the wavefront Zernike coefficient branch and Effective integration leverages the complementary strengths of the two modal data to achieve collaborative information expression. The fusion process not only enhances the discriminative power of features but also provides a more comprehensive and robust input for downstream component degree of freedom misalignment prediction.

[0176] In order to achieve effective fusion, the high-order feature vectors of the two branches are first and Mapped to the same dimensional space :

[0177] ,

[0178] ,

[0179] in, and are the mapping matrices of the wavefront Zernike coefficient branch and the wavefront graph branch, is the bias vector, Indicates the dimension The real number space, Represents the wavefront branch mapped to the dimensional space The eigenvector of Represents the wavefront Zernike coefficient branch mapping to the dimensional space Then, the weighted summation method is used to achieve the feature fusion of the two branches:

[0180] ,

[0181] Among them, the weight coefficient To balance the contributions of the two branches, It represents the eigenvector after the wavefront image branch and the wavefront Zernike coefficient branch are fused.

[0182] Step 423, component imbalance prediction module;

[0183] Based on the fused feature vector , using a multi-layer fully connected network (MLP, Multi-LayerPerceptron) to regress and predict the misalignment of multiple adjustable degrees of freedom of the optical system. This module extracts the complex correspondence between features and misalignment parameters through multi-layer nonlinear mapping, achieving high-precision misalignment estimation. The specific formula is as follows:

[0184] ,

[0185] in, The optical element degree of freedom misalignment parameter predicted by the element misalignment prediction module, is the mapping function of the multilayer perceptron network.

[0186] Step 43: Model training;

[0187] By using large-scale labeled data, the parameters of the three modules—the dual-branch feature extraction module, the feature fusion module, and the component misalignment prediction module—are optimized to achieve high-precision prediction of optical component misalignment. The specific training process is based on a training set containing input data (wavefront images and wavefront Zernike coefficients) and corresponding multi-degree-of-freedom misalignment labels. The model parameters are optimized using a mean square error loss function. The parameters are iteratively updated using a mini-batch gradient descent algorithm (such as the Adam optimizer). Regularization and early stopping mechanisms are combined to effectively prevent overfitting, ultimately achieving accurate estimation of the optical component's misalignment state. To guide the network to focus more on physically sensitive degree-of-freedom directions, the loss function incorporates weighting factors based on the sensitivity matrix. The training objective can be expressed as:

[0188] ,

[0189] in, is the diagonal weighted matrix obtained based on sensitivity analysis, which is used to emphasize the misalignment degrees of freedom that have a more significant impact on the performance of the optical system and improve the prediction accuracy of key directions. are model parameters, and Respectively The actual imbalance and predicted imbalance of samples, represents the model training target, Indicates the sample number, Indicates the total number of samples.

[0190] Step 44: Model verification;

[0191] To comprehensively evaluate the effectiveness of the designed dual-branch neural network model in optical system alignment tasks, multiple performance metrics were used to verify the model's performance on a test set. This validation process not only included conventional regression evaluation metrics but also introduced the RMS value, a physical evaluation metric used to measure the wavefront quality of optical systems, to physically verify the model's actual adjustment effects.

[0192] First, from the perspective of prediction accuracy, the mean square error (MSE) and mean absolute error (MAE) indicators are used to evaluate the difference between the offset output by the model in the test set and the true value:

[0193] ,

[0194] ,

[0195] in, is the model's predicted misalignment, is the actual offset, is the total number of test samples, Number the test sample.

[0196] Next, to further evaluate the application effect of the dual-branch neural network model in the actual optical system alignment, the wavefront Zernike coefficients after the system is adjusted using the predicted misalignment are used to calculate the wavefront RMS value, and the difference between the system wavefront RMS values ​​before and after alignment is compared. The formula for calculating the RMS value using the Zernike coefficients is:

[0197] ,

[0198] in, For the wavefront Zernike coefficients, is the polynomial order used (take ), Indicates the wavefront Zernike coefficient number.

[0199] Step 5: Use the neural network model to guide the actual installation and adjustment;

[0200] A neural network assembly agent model is used to guide the actual optical system assembly.

[0201] Taking the off-axis three-mirror optical system as an example, the process of using the neural network adjustment agent model to guide the actual adjustment is described as follows: Figure 5 As shown ( Figure 5A schematic diagram of the off-axis three-mirror optical system assembly scenario is shown. During assembly, the primary mirror M1 is used as a reference, and the positions of the secondary mirror M2 and the tertiary mirror M3 are continuously adjusted through a high-precision six-degree-of-freedom translation stage and a neural network assembly agent model. A laser interferometer is used to monitor the system wavefront in real time, ultimately ensuring that the optical system wavefront meets the usage requirements). Figure 5 Where XDE, YDE and ZDE represent the translation of the optical system along the X, Y and Z directions respectively, and ADE, BDE and CDE represent the rotation of the optical system along the X, Y and Z directions respectively. Specifically including:

[0202] Step 51, calibrating the optical axis reference;

[0203] Install the laser interferometer, theodolite, plane mirror, and six-degree-of-freedom adjustment mount (also known as a six-degree-of-freedom translation stage), and place the main optical support structure on the air-bearing platform. Use the six-degree-of-freedom adjustment mount and theodolite to continuously adjust the plane mirror until it is perpendicular to the optical axis reference plane of the main support structure. Align the laser interferometer to this optical axis reference plane to achieve optical axis alignment between the laser interferometer and the main support structure.

[0204] Step 52, rough assembly and adjustment;

[0205] First, the primary mirror M1 is installed and adjusted: using tools such as laser interferometer, theodolite, three-coordinate measuring arm, altimeter, and six-degree-of-freedom adjustment frame, the angle between the reference plane of the primary mirror M1 and the plane reflector is adjusted to the ideal angle; then, the laser interferometer is used to monitor the surface shape of the primary mirror M1 in real time, and the position of the primary mirror M1 is continuously adjusted to make its surface shape error the same as the measured result.

[0206] Secondly, perform coarse adjustment of the secondary mirror M2 and the tertiary mirror M3: use a six-degree-of-freedom adjustment frame to preliminarily place the secondary mirror M2 and the tertiary mirror M3 to the designed position, and use a theodolite and micrometer to monitor the adjustment amounts of each degree of freedom of the secondary mirror M2 and the tertiary mirror M3 in real time during the coarse adjustment process, and compare them with the design values ​​to ensure the relative spatial position accuracy between the plane mirrors in the optical system (it is required that the errors of each adjustment dimension of the components after coarse adjustment are within the applicable range of the neural network adjustment agent model).

[0207] Step 53, hardcover adjustment;

[0208] The optical system's wavefront is detected using an autocollimation method, with wavefront detection at multiple fields of view achieved through deflection mirrors and laser interferometers. A trained neural network alignment proxy model is used to calculate the predicted adjustment for each degree of freedom of the optical components based on the wavefront detection results from multiple fields of view of the optical system, thus guiding the alignment of the secondary mirror M2 and the tertiary mirror M3.

[0209] During the fine-tuning process, after each measurement, the wavefront images and fitted Zernike coefficients for multiple fields of view are input into the neural network alignment agent model to calculate the recommended adjustment values ​​for each component's degree of freedom. The component's position is then adjusted according to the recommended adjustment values ​​using a six-degree-of-freedom alignment mount. After these adjustments, the system wavefront is measured again. Through multiple iterations of this process, the wavefront aberration of the actual optical system is brought as close as possible to that of the optical system after various errors have been accounted for, ensuring that all optical performance indicators are within acceptable ranges. This completes the alignment process.

[0210] The optical system assembly method based on the dual-branch neural network model of the present invention considers the noise influence in the actual assembly scene in the data set of the neural network model training, enhances the noise resistance and robustness of the neural network model, and improves the applicability of the neural network model in real optical assembly.

[0211] The optical system assembly and adjustment method based on the dual-branch neural network model of the present invention adopts the dual-branch neural network model in model training, fully mines and utilizes optical system information through feature fusion of two modal data of system wavefront image and Zernike coefficient, and improves the accuracy of component misalignment solution and the stability of the assembly and adjustment process.

[0212] The optical system assembly method based on the dual-branch neural network model of the present invention adopts an artificial neural network model to establish a nonlinear mapping relationship between component misalignment and optical system performance, which can fully explore the essential laws of the influence of assembly errors. While improving the assembly accuracy, it effectively reduces the dependence of traditional assembly methods on manual assembly experience and improves the efficiency of optical assembly.

[0213] The optical system assembly method based on the dual-branch neural network model of the present invention can also reduce the frequency of physical experiments through data-driven virtual assembly preview, effectively reducing resource loss and cost.

[0214] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will readily appreciate that other variations or modifications based on the above descriptions are possible. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. An optical system assembly method based on a dual-branch neural network model, characterized in that: The following steps are involved: Step 1: Component imbalance sensitivity analysis; Construct a theoretical optical system model and calculate the misalignment sensitivity matrix of the optical system; Step 2: Noise analysis of actual installation and adjustment scene; Analyze and determine the noise and its distribution characteristics during the optical system installation process based on the actual installation environment; Step 3: Generate neural network model training data set; Generate a large number of data sets for training neural network assembly agent models; Step 4: Neural network model construction and training; Construct a dual-branch neural network assembly agent model and conduct neural network assembly agent model training; Step 5: Use the neural network model to guide the actual installation and adjustment; A neural network assembly agent model is used to guide the actual optical system assembly.

2. The optical system assembly method based on the dual-branch neural network model according to claim 1, characterized in that: Step 1 specifically includes: Step 11: Establish a theoretical optical system model in optical analysis software; Step 12: Build an integrated simulation analysis link to calculate the curve of the wavefront aberration evaluation index corresponding to each optical element and single position degree of freedom in the optical system as the misalignment changes, and obtain the misalignment sensitivity matrix of the optical system through curve fitting.

3. The optical system assembly and adjustment method based on the dual-branch neural network model according to claim 2, characterized in that: Step 2 specifically includes: Step 21: machining error and measurement error of optical element surface shape; According to the statistical distribution law of errors under actual processing and measurement conditions, the specific distribution characteristics and ranges of processing errors and measurement errors are obtained and converted into the form of wavefront Zernike coefficients. Then, the processing errors and measurement errors are superimposed to comprehensively represent the surface error of the optical element after actual processing. Step 22: wavefront detection error of the optical system; Through empirical analysis and measured data analysis, the error statistical distribution law of the wavefront measurement instrument is obtained and converted into the form of wavefront Zernike coefficients; Step 23: Adjust the adjustment error of the mechanical actuator; Through empirical judgment and analysis of measured data, the system cumulative error and random noise distribution law and range of the mechanical actuator are determined, and converted into the expression of six degrees of freedom of the component posture.

4. The optical system assembly method based on the dual-branch neural network model according to claim 3 is characterized in that: Step 3 specifically includes: Step 31: Determine the input and output of the neural network assembly agent model training; Step 32: Based on the misalignment range of each degree of freedom of the component determined in step 1 and the distribution pattern and range of various noises determined in step 2, batch generate a set of data pairs of the misalignment of each degree of freedom of the component for training the neural network assembly agent model and taking into account the influence of various actual noises.

5. The optical system assembly method based on the dual-branch neural network model according to claim 4 is characterized in that: The specific method of adding various noises in step 32 is as follows: First, the surface error of the optical element determined in step 21 is added to the theoretical optical design model, and the model is used as the ideal model in the subsequent data generation and iterative regression process of adjustment; Next, when generating each set of randomly uniformly distributed misalignment combinations, the noise of the adjustment mechanical actuator is added to each adjustable component position degree of freedom according to the noise distribution pattern and range of the adjustment mechanical actuator determined in step 23. Then, the optical system wavefront diagram and its Zernike coefficient corresponding to the random misalignment combination after adding the noise are calculated. Finally, wavefront detection noise is added to the calculated wavefront aberration data of the optical system; The wavefront detection noise of the optical system determined in step 22 and expressed in the form of Zernike coefficients is directly added to the wavefront Zernike coefficients calculated in the previous step, and the noise distribution is converted into an image form and superimposed on the wavefront diagram.

6. The optical system assembly and adjustment method based on the dual-branch neural network model according to claim 4, characterized in that: Step 4 specifically includes: Step 41: data preprocessing; A neural network assembly proxy model is proposed to realize information acquisition and feature fusion of two modal data. By normalizing and standardizing the input optical system wavefront image and its fitted Zernike coefficients and the output optical system component posture degree of freedom misalignment, the input and output data meet the requirements of the neural network assembly proxy model for numerical scale and distribution consistency, thereby improving the stability and accuracy of the neural network assembly proxy model training. Step 42: Model structure design; The neural network assembly agent model structure includes: a dual-branch feature extraction module, a feature fusion module, and a component misalignment prediction module; The dual-branch feature extraction module extracts features from the wavefront image branch and the wavefront Zernike coefficient branch, and then fuses the extracted features through the feature fusion module. Finally, the component misalignment prediction module achieves high-precision prediction of component misalignment. Step 43: Model training; By using large-scale labeled data, the parameters of the dual-branch feature extraction module, feature fusion module, and component misalignment prediction module are optimized to achieve high-precision prediction of optical component misalignment. Based on the input data and the training set of corresponding multi-degree-of-freedom misalignment labels, the mean squared error loss function is used to optimize the parameters of the neural network alignment proxy model. The parameters are iteratively updated using a mini-batch gradient descent algorithm, combined with regularization and early stopping mechanisms to effectively prevent overfitting, ultimately achieving accurate estimation of the optical component misalignment state. Step 44: Model verification; First, from the perspective of prediction accuracy, the mean square error and mean absolute error indicators are used to evaluate the difference between the misalignment output by the neural network tuning agent model in the test set and the true value: ， ， in, is the model's predicted misalignment, is the actual offset, is the total number of test samples, is the test sample number, represents the mean square error, represents the mean absolute error; Next, to further evaluate the application effect of the neural network adjustment agent model in the actual optical system adjustment, the wavefront Zernike coefficients after the optical system is adjusted using the predicted misalignment are used to calculate the wavefront RMS value, and the differences in the wavefront RMS values ​​of the optical system before and after adjustment are compared. The formula for calculating the RMS value using the Zernike coefficients is: ， in, For the wavefront Zernike coefficients, is the polynomial order used, Indicates the wavefront Zernike coefficient number.

7. The optical system assembly and adjustment method based on the dual-branch neural network model according to claim 6, characterized in that: Step 41 specifically includes: Step 411: pre-processing of the wavefront image; The pixel values ​​of the wavefront image are normalized and the channels are normalized to ensure the stability of the training process. The formulas for pixel value normalization and channel normalization are as follows: ， ， in, Represents the original wavefront image The pixel value after channel normalization, and Represents the pixel coordinates, Represents the original wavefront image in coordinates The first Channel pixel value, Channels representing the three colors red, green, and blue, is the channel index of the image, Represents the wavefront diagram The pixel value after channel normalization, and The first The mean and standard deviation of each channel; Step 412: pre-processing of wavefront Zernike coefficients; Based on the sensitivity analysis results, the Zernike polynomial components that have a significant impact on the imaging quality change of the optical system are selected as network inputs, thereby reducing feature dimensions and redundant information. All wavefront Zernike coefficients are transformed into data with a mean of 0 and a standard deviation of 1 through the Z-score normalization operation, eliminating the dimensional differences between coefficients of different orders, making all features have similar scales and distributions, providing a more stable gradient path for the neural network optimizer, and reducing the learning difficulty of the neural network adjustment agent model. For each Zernike term coefficient , the standardization method formula used is as follows: ， in, is the sample mean of the coefficient, is the standard deviation, Representative coefficient To standardize the process, Indicates the item coefficient number.

8. The optical system assembly and adjustment method based on a dual-branch neural network model according to claim 6, characterized in that: Step 42 specifically includes: Step 421, dual-branch feature extraction module; Wavefront image branch: focuses on the texture and color distribution of the image space, capturing the local disturbance and distortion of the wavefront; the wavefront image branch is based on the wavefront color image As input, it reflects the spatial distribution and phase change of the optical system wavefront, contains rich local texture and color information, and reveals the slight disturbance of the wavefront; the wavefront branch network structure uses a multi-layer convolutional neural network to extract spatial local features, and each layer of convolution operation is performed by a filter and bias Complete, use the ReLU activation function to ensure the capture of nonlinear information. The specific formula is: ， in, Indicates the The feature map of the layer, Indicates the Layer, The feature map of each channel, is the convolution kernel of the corresponding channel, Indicates the layer index of the network, Indicates the channel number. Represents the total number of layers of the convolutional neural network; The output of the last layer of the convolutional neural network is a three-dimensional feature map , Indicates that the output is a three-dimensional real tensor, is the height of the feature map, corresponding to the number of rows of output, is the width of the feature map and the number of columns of the corresponding output, is the number of channels; converted to a one-dimensional vector after flattening , represents the flattening operation, Represents a vector space, dimension Equal to the total number of all elements of the feature map, this one-dimensional vector contains the high-order spatial characteristics of the wavefront image, reflecting the complex local and global optical structure information; Wavefront Zernike coefficient branch: The optical system wavefront is fitted with the Zernike polynomial coefficient vector is the input, the number of coefficients Typically covers low to mid-order major aberration terms, represents transpose, express dimensional real vector space; the wavefront Zernike coefficient branch network structure uses a multi-layer fully connected network for nonlinear feature mapping, and each layer is connected by a weight matrix and the bias vector Implement affine transformation and combine it with ReLU activation function to implement nonlinear operation. The specific formula is: ， ， Finally, the deep feature vector is output As a high-level representation of this branch; in, represents the output vector of the first hidden layer, represents the weight matrix of the first layer, represents the input vector of the current layer, represents the bias vector of the first layer, Indicates the The output vector of the hidden layer, Indicates the The output vector of the hidden layer, Indicates the hidden layer index, the value range is , represents the total number of layers in the fully connected network, Indicates the dimension The real number space of ; Step 422, feature fusion module; The high-order eigenvectors from the wavefront map branch and the wavefront Zernike coefficient branch are and Effectively integrate and give full play to the complementary advantages of the two modal data to achieve collaborative expression of information; In order to achieve effective fusion, the eigenvectors of the wavefront image branch and the wavefront Zernike coefficient branch are first and Mapped to the same dimensional space : ， ， in, and are the mapping matrices of the wavefront Zernike coefficient branch and the wavefront graph branch, are bias vectors, Represents the dimensional space as The real number space, Represents the wavefront branch mapped to the dimensional space The eigenvector of Represents the wavefront Zernike coefficient branch mapping to the dimensional space The eigenvector of Next, the weighted summation method is used to achieve feature fusion of the wavefront image branch and the wavefront Zernike coefficient branch: ， Among them, the weight coefficient , used to balance the contribution of the wavefront graph branch and the wavefront Zernike coefficient branch, Represents the eigenvector after the wavefront image branch and the wavefront Zernike coefficient branch are fused; Step 423, component imbalance prediction module; Based on the fused feature vector , a fully connected network is used to regress and predict the misalignment of multiple adjustable degrees of freedom of the optical system; through multi-layer nonlinear mapping, the complex correspondence between features and misalignment parameters is extracted to achieve high-precision misalignment estimation. The specific formula is as follows: ， in, The optical element degree of freedom misalignment parameter predicted by the element misalignment prediction module, is the mapping function of the multilayer perceptron network.

9. The optical system assembly and adjustment method based on a dual-branch neural network model according to claim 1, characterized in that: Step 5 specifically includes: Step 51, calibrating the optical axis reference; Install the laser interferometer, theodolite, plane mirror, and six-degree-of-freedom adjustment frame, and place the main support structure of the optical system on the air-floating platform; use the six-degree-of-freedom adjustment frame and theodolite to continuously adjust the plane mirror so that the plane mirror and the optical axis reference plane of the main support structure are perpendicular to each other; adjust the laser interferometer to align with the optical axis reference plane to achieve optical axis alignment between the laser interferometer and the main support structure; Step 52, rough assembly and adjustment; First, the primary mirror (M1) is installed and adjusted: using a laser interferometer, theodolite, three-dimensional coordinate measuring arm, altimeter, and six-degree-of-freedom adjustment frame, the angle between the primary mirror (M1) reference plane and the plane reflector is adjusted to the ideal angle; then, the laser interferometer is used to monitor the surface shape of the primary mirror (M1) in real time, and the position of the primary mirror (M1) is continuously adjusted to ensure that the surface shape error of the primary mirror (M1) is consistent with the measured result; Secondly, the secondary mirror (M2) and the tertiary mirror (M3) are coarsely adjusted: the secondary mirror (M2) and the tertiary mirror (M3) are initially placed in the designed positions using a six-degree-of-freedom adjustment frame. During the coarse adjustment process, the theodolite and micrometer are used to monitor the adjustment amount of each degree of freedom of the secondary mirror (M2) and the tertiary mirror (M3) in real time. The adjustment amount of each degree of freedom is compared with the design value to ensure the relative spatial position accuracy between the plane reflectors in the optical system. Step 53, hardcover adjustment; The self-collimation method is used to detect the wavefront of the optical system, and the wavefront detection of multiple field positions is achieved by deflecting plane mirrors and laser interferometers; the trained neural network adjustment agent model is used to solve the predicted adjustment amount of each degree of freedom of the optical component through the wavefront detection results of multiple fields of view measured by the system, thereby guiding the adjustment of the secondary mirror (M2) and the tertiary mirror (M3). In the fine adjustment process, after each measurement, the wavefront diagrams of multiple fields of view and the fitted Zernike coefficients are input into the neural network adjustment agent model to solve the recommended adjustment amount of each degree of freedom of each component; then, the component position is adjusted according to the recommended adjustment amount through the six-degree-of-freedom adjustment frame; after the adjustment, the optical system wavefront is measured again; through multiple rounds of iteration of the above process, the wave aberration of the actual optical system is close to the wave aberration of the optical system after considering various errors, ensuring that all optical performance indicators are within an acceptable range, and the adjustment is completed.

Images