A Joint Prediction Method for Offset and Distortion in Optical Systems Based on Multi-Field Wavefront Information

By using a joint prediction method for misalignment and distortion in a multi-field wavefront information optical system, the problem of independently handling misalignment and distortion in optical assembly and adjustment is solved. This enables rapid feedback and synchronous optimization of multiple performance indicators in the optical system, thereby improving assembly and adjustment efficiency and automation level.

CN123069084APending Publication Date: 2026-11-10CHANGCHUN INST OF OPTICS FINE MECHANICS & PHYSICS CHINESE ACAD OF SCI
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611544599.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-08
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In existing optical assembly and adjustment technologies, misalignment prediction and distortion assessment are independent of each other, and the processing procedures are cumbersome. This makes it difficult to meet the comprehensive constraints of high-end optical systems on wavefront quality and geometric performance, and existing technologies lack a unified framework for the entire assembly and adjustment process.

Method used

A joint prediction method for offset and distortion in a multi-field wavefront information optical system is adopted. Through a multi-field wavefront data acquisition and characterization module, a joint prediction model construction module, and an assembly and adjustment closed-loop optimization module, a nonlinear mapping relationship between offset and distortion is established, and a dynamic closed-loop adjustment mechanism is constructed to achieve synchronous prediction and optimization of offset and distortion.

Benefits of technology

It simplifies the distortion assessment process, improves the automation level of the assembly and adjustment process and the convenience of engineering implementation, and realizes the simultaneous prediction of misalignment and distortion, meeting the rapid feedback requirements of high-end optical systems for multiple performance indicators.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069084A_ABST
    Figure CN123069084A_ABST
Patent Text Reader

Abstract

This invention relates to a joint prediction method for offset and distortion in optical systems based on multi-field wavefront information. It falls within the interdisciplinary field of optical assembly and artificial intelligence, addressing the technical problems of independent offset inversion and distortion assessment during optical assembly, the complexity of distortion detection using wavefronts, and the difficulty in simultaneously verifying multiple performance indicators. The method includes the following steps: acquiring multi-field wavefront data during optical design simulation or actual assembly; establishing a joint prediction model combining shared feature extraction and dual-task output using the multi-field wavefront data as input; obtaining the current system's offset prediction and distortion assessment results, and using these to guide system adjustments. This invention establishes a mapping relationship between multi-field wavefronts and the offset and distortion of system optical components, simultaneously outputting component offset and distortion information under unified wavefront input conditions, thus improving the integration of state recognition and performance evaluation during optical assembly.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of optical assembly and artificial intelligence, and in particular to a method for jointly predicting the misalignment and distortion of an optical system based on multi-field wavefront information. Background Technology

[0002] As high-end optical equipment continues to develop towards higher resolution, larger field of view, wider spectral band, and larger aperture, optical system structures are becoming increasingly complex and system performance indicators are becoming increasingly stringent. Examples include high-resolution remote sensing cameras, precision interferometry systems, and high-end spectrometers. The core optical systems in these devices are characterized by a large number of components, strong parameter coupling, and high error sensitivity, placing higher demands on the precision, stability, and consistency of optical assembly and adjustment. Meanwhile, for special optical systems such as lithography projection lenses, performance requirements are no longer limited to single image quality indicators such as wavefront error; geometric performance parameters such as distortion must also be considered. That is, the design and assembly and adjustment process must ensure that both wavefront and distortion indicators simultaneously meet predetermined constraints. However, traditional optical assembly and adjustment methods often employ "local linear approximation + manual experience judgment + repeated measurement and adjustment," which suffers from long assembly and adjustment cycles, low efficiency, strong reliance on operator experience, insufficient global optimization capabilities, and poor repeatability and consistency. These methods are gradually becoming inadequate for meeting the assembly and adjustment needs of complex, multi-degree-of-freedom, high-precision optical systems.

[0003] In recent years, artificial intelligence technology has shown significant advantages in the field of optical assembly and adjustment. Currently, there are various technical solutions based on artificial neural network algorithms that use wavefront information to predict the misalignment of optical components. For example, the technical solution disclosed in the patent with publication number CN117537711A, "A Deep Learning-Based Method for Identifying Optical Lens Assembly Pose Errors," and the journal article "Improving neural-network-based prediction models for misalignment in off-axis three-mirror anastigmat telescopes" (Source: Applied Optics), is based on the idea of ​​constructing a set of data pairs between the deviation of the optical component from the ideal pose and the wavefront Zernike coefficients of the system. Then, a mapping model from the wavefront Zernike coefficients to the component assembly pose deviation is established using deep learning methods. In actual assembly and adjustment, this model is used to predict the component misalignment to guide the assembly and adjustment process. These technologies can significantly improve the level of automation in assembly and adjustment, reduce reliance on manual operation experience, and have efficient misalignment state recognition capabilities and strong adaptability for multi-degree-of-freedom, strongly coupled optical systems. However, existing research mainly focuses on inverting the misalignment of optical elements from wavefront information, while paying relatively little attention to the joint characterization and prediction of geometric performance indicators such as system distortion.

[0004] In existing engineering technologies, wavefront information can usually be conveniently measured using equipment such as Shack-Hartmann wavefront sensors and interferometers. However, distortion detection typically employs target imaging, exposure measurement, aerial image measurement, or wavefront measurement methods, combined with processes such as full-field sampling, coordinate mapping, feature point extraction, and model fitting, making the measurement process more cumbersome. Among these, wavefront measurement demonstrates an internal correlation between wavefront information and distortion. Existing research also includes various technical solutions for recovering distortion parameters using wavefront information, such as the patent with publication number CN106768882A, "A Distortion Measurement Method for Optical Systems Based on Shack-Hartmann Wavefront Sensors," and the journal article "Distortion measurement of a lithography projection lens based on multichannel grating lateral shearing interferometry" (Source: Applied Optics). These approaches typically utilize Shack-Hartmann wavefront sensors, shearing interferometers, or other wavefront detection methods to first obtain the wavefront slope, wavefront reconstruction results, or wavefront Zernike expansion coefficients, and then further calculate image point offsets or fit the distortion field distribution. This type of technique can indirectly obtain distortion results using relatively easily measurable wavefront information, reducing the complexity of direct distortion measurement. However, its primary objective remains distortion measurement or simultaneous wavefront / distortion detection, and it has not yet been integrated with the identification of component misalignment states during optical assembly.

[0005] Although existing technologies have proposed schemes for predicting optical element misalignment based on wavefront information and recovering optical system distortion based on wavefront measurement, the following shortcomings still exist from the perspective of the actual needs of optical assembly and adjustment processes:

[0006] First, existing technologies typically treat "misalignment prediction" and "distortion assessment" as two separate problems. The former serves the adjustment control during the assembly and adjustment process, while the latter serves the geometric performance testing of the optical system. The two are separated in terms of input information, analysis process, and result application, and a unified technical framework serving the entire assembly and adjustment process has not yet been formed.

[0007] Secondly, although existing wavefront-based distortion recovery techniques can avoid some of the direct distortion measurement process, they still require multiple steps such as wavefront reconstruction, wavefront slope calculation, image point offset calculation, and full-field fitting. The processing chain is long and the calculation process is cumbersome, making it difficult to meet the needs of rapid feedback and efficient evaluation in the optical assembly process.

[0008] Third, existing intelligent assembly and adjustment technologies based on wavefront prediction of misalignment mainly focus on component pose error inversion, lacking synchronous prediction and verification of geometric performance indicators such as optical system distortion during the adjustment process. This makes it difficult to support the comprehensive constraints on wavefront quality and geometric performance of high-end optical systems during assembly and adjustment. Summary of the Invention

[0009] This invention aims to address the technical problems in the prior art where misalignment inversion and distortion assessment are independent during optical assembly and adjustment, the distortion detection process using wavefronts is complex, and it is difficult to achieve simultaneous verification of multiple performance indicators. It provides a method for joint prediction of misalignment and distortion in optical systems based on multi-field wavefront information.

[0010] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0011] A method for jointly predicting the misalignment and distortion of an optical system based on multi-field wavefront information is applicable to systems that include interconnected modules: a multi-field wavefront data acquisition and characterization module, a joint prediction model construction module, and an assembly and adjustment closed-loop optimization module.

[0012] The multi-field wavefront data acquisition and characterization module is used to map wavefront information from different fields of view to a unified feature space.

[0013] The joint prediction model building module is used to establish the nonlinear mapping relationship between the multi-field wavefront input features and the offset and distortion of the optical system, respectively.

[0014] The assembly and adjustment closed-loop optimization module is used to construct a dynamic adaptive closed-loop adjustment mechanism based on the misalignment prediction value and distortion prediction value output by the joint prediction model, gradually correct the misalignment state of the optical system, reduce system distortion, and bring the image quality index of the optical system to near the preset range.

[0015] The method includes the following steps:

[0016] Step 1: The multi-field wavefront data acquisition and characterization module acquires multi-field wavefront data during optical design simulation or actual assembly and adjustment, and constructs corresponding misalignment and distortion labels;

[0017] Step 2: The joint prediction model construction module takes multi-field wavefront data as input and establishes a joint prediction model that combines shared feature extraction and dual-task output. One branch of the joint prediction model outputs the predicted values ​​of the misalignment of the optical components of the optical system, and the other branch outputs the predicted values ​​of the distortion of the entire field of view.

[0018] Step 3: After the joint prediction model is trained by joint loss constraints, the assembly and adjustment closed-loop optimization module inputs the measured multi-field wavefront data into the joint prediction model, and simultaneously obtains the current optical system's misalignment prediction results and distortion assessment results, and guides the adjustment of the optical system accordingly; after each optical system adjustment, the multi-field wavefront data is measured again and input into the joint prediction model to form a closed-loop iterative optimization assembly and adjustment process.

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

[0020] For the For each field of view, its wavefront function is defined as:

[0021]

[0022] in, For the first Continuous wavefront function under each field of view; This represents the total number of fields of view of the optical system. For the field of view index; the wavefront function is represented in normalized polar coordinates, i.e.:

[0023]

[0024] in, The wavefront function in normalized polar coordinates. The normalized polar radius has a range of values. ; The polar angle has a range of values ​​of 100. ;

[0025] The wavefront function in normalized polar coordinates is parametrically expressed using the Zernike coefficient form of wavefront expansion, specifically as follows:

[0026]

[0027] in, For the first The first field wavefront function Zernike polynomial coefficients of order 1; For the first Zernike polynomial of order Z; This represents the total order of the Zernike polynomials used. Number the Zernike polynomial order;

[0028] Define a single-field-of-view wavefront eigenvector to integrate all Zernike polynomial coefficients for a single field of view, in the following form:

[0029]

[0030] in, For the first Wavefront feature vectors of each field of view, with dimension ; superscript This represents the transpose operation of a vector; This indicates that the feature vector is 3D real vector; For the first The first field wavefront function Zernike polynomial coefficients of order 1; For the first The first field wavefront function Zernike polynomial coefficients of order 1; For the first The first field wavefront function Zernike polynomial coefficients of order 1;

[0031] Next, the feature vectors of all single-field wavefronts across the entire field of view are integrated to construct a multi-field wavefront input matrix, namely:

[0032]

[0033] in, The multi-field wavefront input matrix has a dimension of . ; This matrix is ​​represented as OK A real matrix of columns; For the first Wavefront eigenvectors of each field of view For the first Wavefront eigenvectors of each field of view For the first Wavefront eigenvectors of each field of view;

[0034] The above multi-field wavefront input matrix is ​​expanded into a one-dimensional eigenvector, as follows:

[0035]

[0036] in, The expanded one-dimensional feature vector has a dimension of . ; A matrix column vectorization operator, used to convert... dimensional matrix Concatenate the columns in order to form a one-dimensional vector; This indicates that the one-dimensional eigenvector is 3D real vector;

[0037] The index information for each field of view is vectorized and specifically defined as follows:

[0038]

[0039] in, For the first Encoding vectors for each field of view, with dimension ; For field-of-view coding function; The dimension of the field-of-view encoding vector;

[0040] The wavefront feature vector of a single field of view is concatenated with the field of view encoding vector to construct an enhanced single field of view wavefront feature vector, thereby achieving the fusion of wavefront features and field of view related information. The specific form is as follows:

[0041]

[0042] in, For the enhanced version Each field-of-view wavefront feature vector has a dimension of _ ... ;symbol This represents the vertical concatenation operation of vectors;

[0043] Based on the enhanced single-view wavefront eigenvectors of the entire field of view, the final enhanced multi-view input matrix is ​​constructed, specifically in the following form:

[0044]

[0045] in, The enhanced multi-view input matrix has dimensions of . ;

[0046] The mean and standard deviation of the enhanced multi-view input matrix are normalized to ensure that the processed feature data follows a standard normal distribution. The specific formula is as follows:

[0047]

[0048] in, This is the standardized enhanced multi-view input matrix; The original input matrix The mean of all elements in the dataset; The original input matrix The standard deviation of all elements in the dataset.

[0049] In the above technical solution, step 2 specifically includes:

[0050] The overall output relationship of the joint prediction model is defined as follows:

[0051]

[0052] in, The joint output of the joint prediction model corresponds to the predicted values ​​of optical system misalignment and distortion, respectively. The joint prediction model is a nonlinear mapping function composed of multiple neural networks; This is the standardized enhanced multi-view input matrix; The set of all learnable parameters for the joint prediction model;

[0053] The input features are subjected to multi-level nonlinear transformations to extract high-level shared features that can simultaneously support offset and distortion prediction, thereby achieving efficient feature reuse. The specific implementation process is as follows:

[0054] First, the enhanced multi-view input matrix is ​​used as the initial input to the shared encoder, and the initial input features are defined as follows:

[0055]

[0056] in, The initial input features of the shared encoder; if If it is in matrix form, then If they are matrices of the same dimension; If it is in tensor form, then Tensors of the same dimension;

[0057] Next, the shared encoder extracts high-level shared features step by step through multi-layer nonlinear transformations. The feature update formula for each layer is as follows:

[0058]

[0059] in, For shared encoders The output features of the layer; For shared encoders The input features of the layer; For the first Layer weight matrix; For the first Layer bias vector; It is a non-linear activation function; This is the layer index of the shared encoder, with a value range of [value range missing]. , This represents the total number of layers in the shared encoder.

[0060] go through After the layer nonlinear transformation, the final output of the shared encoder is obtained, which is the high-level shared feature, specifically in the following form:

[0061]

[0062] in, For high-level shared feature vectors, the dimension is ; For shared encoders The output features of the layer share feature vectors with the higher layers. Consistent dimensions; This indicates that the high-level shared feature vector is 3D real vector To share feature dimensions;

[0063] High-level shared features based on shared encoder output An independent misalignment prediction branch is constructed to achieve accurate prediction of the misalignment of each degree of freedom of the optical system, specifically expressed as follows:

[0064]

[0065] in, Let be the prediction vector of the system offset, with dimension . ; The weight matrix for the imbalance prediction branch has dimensions of . ; The bias vector for the offset prediction branch has dimension . ;

[0066] The distortion predictions in vector field form are discretized into three-dimensional tensor form, as follows:

[0067]

[0068] in, The discretized distortion prediction tensor has a dimension of . ; The height of the discrete sampling of the image plane; The width of the discrete sampling of the image plane; dimension 2 corresponds to... direction and Predicted value of directional offset; This indicates that the distortion prediction tensor is a three-dimensional real tensor;

[0069] The distortion prediction tensor is fed back to the imbalance prediction branch. After mapping the output of the distortion prediction branch, it is fused with the shared features of the higher layers to optimize the input features of the imbalance prediction branch and improve the accuracy of imbalance prediction. The specific formula is as follows:

[0070]

[0071] in, The optimized offset prediction branch inputs features, and the dimensions share features with the higher layers. Consistent; High-level shared features output by the shared encoder; , which is the coupling coefficient, used to adjust the feedback weight of the distortion prediction results; This is a feature mapping function used to map the distortion prediction tensor. Mapping to shared features with higher levels Feature vectors with consistent dimensions enable feature fusion.

[0072] In the above technical solution, step 3 specifically includes:

[0073] Step 31, Initial State Detection:

[0074] Initial multi-field wavefront detection is performed on the optical system. Full-field wavefront data is acquired through the multi-field wavefront data acquisition and characterization module to obtain the multi-field wavefront input features. , Let be the number of iterations, the initial iteration. ;

[0075] Step 32, Model Prediction Output:

[0076] Input features from the initial multi-field wavefront Input the joint prediction model after training convergence to obtain the prediction vector of the system offset in the initial state. With distortion prediction tensor ;

[0077] Step 33, Adjustment of assembly parameters:

[0078] Based on the Prediction vector of system offset in round iteration An adaptive step size adjustment strategy is adopted to adjust the parameters of the optical system's assembly and adjustment mechanism, thereby obtaining the adjusted system misalignment. ;

[0079] Step 34, Status Retesting and Evaluation:

[0080] Multi-field wavefront retesting was performed on the adjusted optical system to obtain new multi-field wavefront input features. And a new prediction vector for the system imbalance is obtained through a joint prediction model. and distortion prediction tensor Evaluate whether the current state of the optical system satisfies the convergence criterion;

[0081] Step 35, Closed-loop iteration judgment:

[0082] If the state of the optical system satisfies the convergence criterion, then stop the iteration and output the optimal setup parameters; otherwise, let Repeat the above steps to enter the next round of closed-loop iteration until the convergence requirement is met, thus forming a closed-loop iterative optimization and commissioning process.

[0083] In the above technical solution, the iterative adjustment strategy for the closed-loop iterative optimization assembly and adjustment process in step 3 is as follows:

[0084] The offset state evolution equation of the optical system is used to describe the update law of the offset in each closed-loop iteration. The specific formula is as follows:

[0085]

[0086] in, For the first The actual system misalignment after rounds of iteration; For the first The actual system misalignment after rounds of iteration; For the first The step size for each iteration of the assembly and adjustment; For the first The prediction vector of the system offset in each iteration is the predicted value of the offset; the negative sign indicates that the adjustment direction is opposite to the direction of the predicted offset value, that is, the adjustment is made in the direction of reducing the offset.

[0087] An adaptive step size mechanism is adopted to dynamically adjust the step size according to the current misalignment state of the optical system. The specific implementation method is as follows:

[0088]

[0089] in, The initial step size is determined based on the adjustment accuracy of the assembly and adjustment mechanism; This is the step size decay coefficient; This represents the number of iterations.

[0090] In the above technical solution, step 3, the construction of the multi-objective joint optimization objective function for the closed-loop iterative optimization assembly and adjustment process, specifically involves:

[0091] To achieve synergistic optimization of the misalignment and distortion of the optical system, a multi-objective joint optimization objective function is constructed to measure the adjustment quality of the optical system after each iteration and guide the direction of the closed-loop iteration. The specific formula is as follows:

[0092]

[0093] in, For the first The system's multi-objective joint optimization objective function value is obtained through round-by-round iteration. , These are the weighting coefficients for wavefront error and distortion error, respectively, satisfying... ; For the first Wavefront error of round iteration, For the first The distortion error of each iteration is defined as follows:

[0094] wavefront error The root mean square error of the full field of view wavefront is used to measure the quality of the full field of view wavefront. The formula is as follows:

[0095]

[0096] in, This represents the total number of fields of view of the optical system. For the first Round of iteration Continuous wavefront function for each field of view; For the first Ideal wavefront function for each field of view; For the first Wavefront sampling area of ​​each field of view; For field of view index; 、 These are the two-dimensional spatial coordinate parameters of the wavefront function;

[0097] Distortion error The vector field form of the distortion is expressed as the mean absolute error of the full-plane distortion migration, and the formula is:

[0098]

[0099] in, , These represent the height and width of the discrete sampling of the image plane, respectively. , The first Iterating through image plane coordinates place direction and The actual distortion offset in the direction; , They are respectively the corresponding positions direction and Ideal design distortion offset in direction; 、 These are the two-dimensional coordinate parameters of the image plane of the optical system.

[0100] In the above technical solution, the convergence criterion in step 3 is specifically as follows:

[0101] A dual convergence criterion is established, simultaneously constraining both misalignment and distortion, as follows:

[0102] 1) Criterion for convergence of imbalance: The first The 2-norm of the prediction vector of the system offset in the round of iterations is less than a preset threshold. ,Right now:

[0103]

[0104] in, Let be the 2-norm of the prediction vector of the system offset; This is the convergence threshold for the imbalance.

[0105] 2) Distortion convergence criterion: The first Distortion error of round iteration Less than the preset threshold ,Right now:

[0106]

[0107] in, This is the distortion convergence threshold;

[0108] 3) Additional conditions for iteration termination: If the number of iterations reaches the preset maximum number of iterations. Even if the above dual convergence criteria are not met, the iteration stops, the current optimal assembly and adjustment parameters are output, and an assembly and adjustment accuracy evaluation report is given. When both the misalignment convergence criterion and the distortion convergence criterion are met, the assembly and adjustment closed-loop optimization is determined to be completed, and the optical system reaches the optimal assembly and adjustment state. At this time, the final assembly and adjustment parameters are output as the final assembly and adjustment basis of the optical system.

[0109] In the above technical solution, the process of training the joint prediction model in steps 2 and 3 through joint loss constraints is as follows:

[0110] First, construct the training sample dataset and preprocess it;

[0111] Secondly, define a joint loss function for multiple tasks;

[0112] Then, an adaptive optimization algorithm is used to iteratively update the model parameters;

[0113] Finally, the training process is monitored by validating the dataset, and regularization and early stopping strategies are introduced to prevent overfitting until the model converges and reaches the preset performance metrics.

[0114] In the above technical solution, the construction and preprocessing of the training sample dataset specifically involves:

[0115] 1) Sample set composition: The training sample dataset adopts... It means that among them The total number of samples in the training sample dataset; For the first Multi-field wavefront input features of each sample; For the first The actual misalignment vector of each sample, with dimension [missing information]. ; For the first The actual distortion value of each sample; For sample index;

[0116] 2) Sample preprocessing:

[0117] Outlier removal: The 3σ criterion is used to remove outliers from the training dataset. Outliers are removed when the multi-view wavefront input features, misalignment, or distortion value of a sample exceeds the mean of the corresponding features. When the sample falls within the specified range, it is identified as an abnormal sample and removed. This represents the sample mean for the corresponding feature. The sample standard deviation is the corresponding feature.

[0118] Sample splitting: The preprocessed training sample dataset is divided into training set, validation set and test set in a ratio of 7:2:1. The training set is used for iterative updates of model parameters, the validation set is used to monitor the model training process and adjust hyperparameters, and the test set is used to finally evaluate the model's generalization performance.

[0119] 3) Label normalization: for the first... The actual misalignment vector of each sample With the The actual distortion value of each sample Perform normalization to map it to The interval, specifically:

[0120]

[0121] in, , These are the normalized th The actual misalignment vector of the sample and the first The actual distortion value of each sample; , These are the minimum and maximum values ​​of the actual misalignment vector for all training samples, respectively. , These are the minimum and maximum values ​​of the actual distortion values ​​for all training samples, respectively.

[0122] The multi-task joint loss function integrates the losses of two tasks using a weighted summation method. The specific formula is as follows:

[0123]

[0124] in, This is the model's overall joint loss function, used to measure the overall deviation between the model's predicted values ​​and the actual labels; These are the weighting coefficients for the offset prediction loss and the distortion prediction loss, respectively. They are adaptively adjusted based on the importance and prediction difficulty of the two tasks. If higher offset prediction accuracy is required, then the weighting coefficients are increased. Conversely, it increases. And satisfy ; The loss function for predicting the offset is... The distortion prediction loss function is defined as follows:

[0125] Offset Prediction Loss Function The mean squared error loss is used, and the calculation formula is as follows:

[0126]

[0127] in, For the first The first sample Predicted values ​​of misalignment for each degree of freedom of assembly; For the first The first sample The actual misalignment amount label for each degree of freedom of assembly adjustment; This represents the total number of degrees of freedom in assembly and adjustment.

[0128] Distortion prediction loss function The mean absolute error loss is used, and the calculation formula is as follows:

[0129]

[0130] in, , The first Image coordinates of each sample Place , Predicted value of orientation distortion offset; , These are the actual distortion offsets corresponding to the position; , These represent the height and width of the discrete sampling on the image plane; dimension 2 corresponds to... , Offset in both directions; 、 These are the two-dimensional coordinate parameters of the image plane of the optical system.

[0131] The adaptive optimization algorithm is used to iteratively update the model parameters as follows:

[0132] Adaptive momentum optimization algorithm is used to optimize the learnable parameters of the model. The iterative update process is as follows:

[0133] 1) Initialize parameters: Initialize all learnable parameters of the model. The Adam algorithm is initialized using a random normal distribution; the hyperparameters for initialization include the learning rate. Momentum parameters , and numerical stability parameters ;

[0134] 2) Gradient calculation: For each training batch, calculate the total joint loss function. For all learnable parameters gradient The backpropagation algorithm is used to complete the backpropagation and calculation of the gradient;

[0135] 3) Adaptive Parameter Update: Based on the update rule of the Adam algorithm, an exponential moving average is applied to the gradient to update the first and second moment estimates of the parameters. The learning rate for each parameter is then adjusted accordingly, ultimately completing the parameter update. The specific formula is as follows:

[0136]

[0137] in, This represents the number of iterations. , These are the first-order moment estimates and second-order moment estimates of the gradient, respectively. , These are the deviation correction values ​​for the first-order moment estimate and the second-order moment estimate, respectively; , The first Wheel, First Model parameters after one iteration; , The first First-order and second-order moment estimates of the gradient in each iteration; For gradient operators;

[0138] The model converges and reaches the preset performance index as follows:

[0139] 1) Training convergence criterion: The model is considered to have converged when any of the following conditions are met: (a) the number of iterations reaches the preset maximum number of iterations; (b) in a consecutive preset number of iterations, the changes in both the training set loss and the validation set loss are less than a preset threshold, i.e.:

[0140]

[0141] in, , The first Wheel, First The training set loss in each iteration; , The first Wheel, First The validation set loss in each iteration; The threshold for loss variation;

[0142] 2) Model performance evaluation metrics: The performance of the converged model is evaluated using a test set, with evaluation metrics designed for offset prediction and distortion prediction respectively:

[0143] (a) Imbalance prediction and evaluation indicators: using mean absolute error and coefficient of determination. The formula is:

[0144]

[0145]

[0146] in, The total number of samples in the test set. For the test set sample index, It represents the total number of degrees of freedom in assembly and adjustment; For the assembly and adjustment of degrees of freedom index, For the first In the test sample, the first The average value of the true label for the degree of freedom mismatch of each assembly. For the first In the test sample, the first The true value of the label for the degree of freedom mismatch of an assembly. For the first In the test sample, the first The predicted value of the misalignment label for each degree of freedom of assembly. The mean absolute deviation between the predicted and actual values ​​of the imbalance quantity;

[0147] (b) Distortion prediction evaluation index: using the average endpoint error The distance between the endpoints of the predicted offset vector and the actual offset vector is measured by the following formula:

[0148]

[0149] in, , The first Image coordinates of each sample Place , Predicted value of orientation distortion offset; , These are the labels for the actual distortion offset at the corresponding position.

[0150] The method for joint prediction of misalignment and distortion in optical systems based on multi-field wavefront information of the present invention has the following beneficial effects:

[0151] (1) To achieve simultaneous prediction of misalignment and distortion, and improve the integration of assembly and adjustment analysis.

[0152] This invention establishes a mapping relationship between multi-field wavefronts and system optical element offset and system distortion, simultaneously outputting element offset and distortion information under a unified wavefront input condition. Compared to existing technologies that perform offset inversion and distortion detection separately, this reduces the processes of independent modeling, independent measurement, and step-by-step processing, thereby improving the integration of state recognition and performance evaluation during optical assembly.

[0153] (2) Simplify the distortion assessment process and improve the efficiency of status analysis during assembly and adjustment.

[0154] This invention directly predicts the system distortion field distribution using multi-field wavefront information, eliminating the need for a separate, complex distortion measurement process and avoiding the tedious solution process of existing wavefront distortion recovery methods. This technical solution allows for simultaneous distortion assessment results after acquiring wavefront information, thereby shortening the state analysis chain during assembly and adjustment, improving distortion assessment efficiency, and meeting the need for rapid feedback in the assembly and adjustment of complex optical systems.

[0155] (3) Supports multi-objective optimization and improves the overall performance of the system.

[0156] Existing assembly and adjustment methods based on wavefront prediction of misalignment mainly focus on component pose error identification and typically use optimal wavefront quality as the sole optimization objective, making it difficult to consider imaging geometric performance. This invention introduces distortion information into the assembly and adjustment quality evaluation system by simultaneously predicting system distortion information. This allows the assembly and adjustment process to obtain "how to adjust" and "what the performance will be after adjustment" simultaneously, while also considering wavefront quality and system geometric performance. Through multi-objective joint optimization, it improves the overall system performance.

[0157] (4) Improve the level of automation in assembly and adjustment and the convenience of project implementation.

[0158] This invention, based on a unified wavefront input, completes offset prediction and distortion assessment, reducing the reliance on manual experience and multiple iterative adjustments in traditional assembly and adjustment processes. Through this technical solution, assembly and adjustment personnel can more quickly obtain information on the system's offset state and distortion field distribution, providing a unified basis for state determination, adjustment control, and comprehensive performance evaluation during the assembly and adjustment process, thereby improving the automation level and ease of engineering implementation in the optical assembly and adjustment process. Attached Figure Description

[0159] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0160] Figure 1 This is a schematic diagram of the overall algorithm architecture and reasoning process of the method for joint prediction of misalignment and distortion in optical systems based on multi-field wavefront information of the present invention.

[0161] Figure 2 This is a schematic diagram showing the adjustable degrees of freedom of an optical system component.

[0162] Figure 3 This is the wavefront distribution diagram of the optical system and the Zernike polynomial and its coefficients for the fitted partial fringes.

[0163] Figure 4 This is a diagram showing the distortion vector distribution of the optical system. Detailed Implementation

[0164] The inventive concept of this invention is as follows:

[0165] The present invention provides a method for joint prediction of misalignment and distortion in optical systems based on multi-field wavefront information. This method uses simulated or measured multi-field wavefront data as a unified input to construct a joint prediction model for the misalignment and distortion of optical components. This enables rapid inversion and multi-dimensional characterization of the optical system state during assembly and adjustment. Furthermore, a closed-loop optimization mechanism for the assembly and adjustment process is established based on this method, thereby improving assembly and adjustment efficiency and comprehensive performance evaluation capabilities.

[0166] The present invention provides a method for joint prediction of misalignment and distortion in optical systems based on multi-field wavefront information. This method simultaneously predicts misalignment and assesses distortion based on a unified wavefront information input, reducing the cumbersome measurement and solution process in existing methods that use wavefronts to recover distortion. It improves the efficiency of misalignment state identification and optical performance verification during optical assembly and provides unified and efficient technical support for optical system assembly and adjustment.

[0167] The present invention will now be described in detail with reference to the accompanying drawings.

[0168] This invention discloses a method for jointly predicting the misalignment and distortion of an optical system based on multi-field wavefront information. The applicable system includes interconnected modules: a multi-field wavefront data acquisition and characterization module, a joint prediction model construction module, and an assembly / adjustment closed-loop optimization module. The multi-field wavefront data acquisition and characterization module maps wavefront information from different fields of view to a unified feature space. The joint prediction model construction module establishes nonlinear mapping relationships between the multi-field wavefront input features and the misalignment and distortion of the optical system. The assembly / adjustment closed-loop optimization module constructs a dynamically adaptive closed-loop adjustment mechanism based on the predicted misalignment and distortion values ​​output by the joint prediction model, gradually correcting the misalignment state of the optical system, reducing system distortion, and bringing the image quality indicators of the optical system to near a preset range. The method includes the following steps:

[0169] Step 1: The multi-field wavefront data acquisition and characterization module acquires multi-field wavefront data during optical design simulation or actual assembly and adjustment, and constructs corresponding misalignment and distortion labels;

[0170] Step 2: The joint prediction model construction module takes multi-field wavefront data as input and establishes a joint prediction model that combines shared feature extraction and dual-task output. One branch of the joint prediction model outputs the predicted values ​​of the misalignment of the optical components of the optical system, and the other branch outputs the predicted values ​​of the distortion of the entire field of view.

[0171] Step 3: After the joint prediction model is trained by joint loss constraints, the assembly and adjustment closed-loop optimization module inputs the measured multi-field wavefront data into the joint prediction model, and simultaneously obtains the current optical system's misalignment prediction results and distortion assessment results, and guides the adjustment of the optical system accordingly; after each optical system adjustment, the multi-field wavefront data is measured again and input into the joint prediction model to form a closed-loop iterative optimization assembly and adjustment process.

[0172] The overall algorithmic framework of the method of this invention revolves around dual-task modeling of multi-field wavefront driving and closed-loop optimization of optical assembly and adjustment, with the core modules as described above. The method of this invention will be described in detail below through its overall algorithmic architecture and the composition of its modules:

[0173] (1) Overall methodology:

[0174] The overall algorithm architecture and reasoning process of the proposed method for joint prediction of misalignment and distortion in optical systems based on multi-field wavefront information are as follows: Figure 1 As shown.

[0175] Multi-field wavefront information, as a state observation of an optical system, can be represented as a field index. The corresponding wavefront function is expressed as To construct the input tensor, a unified model is used for wavefronts from different fields of view.

[0176]

[0177] in, It serves as a unified input tensor for multi-field wavefronts, used to integrate wavefront information from all fields of view to comprehensively characterize the aberration distribution of the optical system. For the first Wavefront function of each field of view, subscript This is the field index, and its value range is... , This represents the total number of fields of view of the optical system. For the first Wavefront function of each field of view For the first Wavefront function of each field of view For the first Wavefront function of each field of view 、 These are the two-dimensional spatial coordinate parameters of the wavefront function, used to describe the spatial distribution of the wavefront.

[0178] The misalignment of the degrees of freedom of the optical element is expressed in vector form:

[0179]

[0180] in, This is the component offset vector, used to integrate offset parameters from all adjustable degrees of freedom; For the first The misalignment of a single degree of freedom of assembly, subscript This is the assembly and adjustment degree of freedom index (specifically characterizing eccentricity, tilt, and component spacing error), and its value range is... , This represents the total number of degrees of freedom for the assembly and adjustment of the optical system.

[0181] Optical system distortion can be expressed as a function (discrete vector field):

[0182]

[0183] in, This is the system distortion function, used to describe the degree of offset between the actual image point and the ideal image point of the optical system. Its input is the image plane coordinates. The output is an offset vector; 、 These are the two-dimensional coordinate parameters of the image plane of the optical system; Image plane coordinates Image point at The offset in direction; Image plane coordinates Image point at The offset in direction.

[0184] Construct a joint prediction mapping relationship from multi-field wavefronts to system offset and distortion:

[0185]

[0186] in, The mapping function between the multi-field wavefront and the system offset and distortion is the core of the joint prediction model of this invention. This is the input to the mapping function, i.e., the aforementioned multi-field wavefront input tensor; One of the outputs of the mapping function is the aforementioned system multi-degree-of-freedom misalignment vector; The second output of the mapping function, namely the aforementioned system distortion function; symbol The mapping symbol represents the input. To output The correspondence.

[0187] (2) Multi-field wavefront data acquisition and characterization module:

[0188] This module is used to construct a unified observation representation of optical system information across the entire field of view. Its core function is to map wavefront information from different fields of view to a unified feature space, providing highly consistent input data for subsequent joint prediction models and ensuring the accuracy and generalization ability of the mapping relationship. Specifically, this module includes: multi-field wavefront acquisition, wavefront parameterization representation, multi-field data stitching, field information encoding, data normalization processing, and a unified multi-field wavefront output process.

[0189] Specifically, step 1 of the method of the present invention includes the following steps:

[0190] For the For each field of view, its wavefront function is defined as:

[0191]

[0192] in, For the first Continuous wavefront function under each field of view; 、 These are the two-dimensional spatial coordinate parameters of the wavefront function; The total number of fields of view in the optical system is denoted by , and its value depends on the design requirements or measurement needs of the optical system. Meanwhile, to enhance the numerical stability and physical interpretability of the wavefront data, it is preferable to represent the wavefront function in normalized polar coordinates, i.e.:

[0193]

[0194] in, This is the wavefront function in normalized polar coordinates; The normalized polar radius has a range of values. , used to characterize the normalized distance from the measurement point to the optical axis; The polar angle has a range of values ​​of 100. , is used to characterize the azimuth angle of the measurement point relative to the optical axis.

[0195] Parameterized representation of wavefront Zernike coefficients:

[0196] The wavefront function in normalized polar coordinates is parametrically expressed using the Zernike coefficient form of wavefront expansion, specifically as follows:

[0197]

[0198] in, For the first The first field wavefront function The Zernike polynomial coefficients of order Z are used to represent the magnitude of the aberration of that order. For the first The Zernike polynomial of order 1 is an orthogonal basis function characterizing the wavefront distribution and has a clear physical meaning (corresponding to different types of aberrations such as astigmatism, spherical aberration, and coma). This represents the total order of the Zernike polynomials used. Number the Zernike polynomial order;

[0199] Based on this, a single-field-of-view wavefront eigenvector is defined to integrate all Zernike polynomial coefficients of a single field of view, specifically in the form of:

[0200]

[0201] in, For the first Wavefront feature vectors of each field of view, with dimension The total order of the Zernike polynomial Maintain consistency; superscript This indicates the transpose operation of a vector (for easier subsequent matrix operations). This indicates that the feature vector is A real vector, encompassing Zernike coefficient information of all orders under this field of view; For the first The first field wavefront function Zernike polynomial coefficients of order 1; For the first The first field wavefront function Zernike polynomial coefficients of order 1; For the first The first field wavefront function Zernike polynomial coefficients of order 1;

[0202] Next, the feature vectors of all single-field wavefronts across the entire field of view are integrated to construct a multi-field wavefront input matrix, namely:

[0203]

[0204] in, The multi-field wavefront input matrix has a dimension of . Each row of the matrix corresponds to a wavefront eigenvector of a field of view, thus achieving the orderly integration of all wavefront information of the field of view. This matrix is ​​represented as OK The real matrix of columns, and the total number of fields of view and Zernike polynomial order To ensure compatibility with the input format of the subsequent joint prediction model, the multi-field wavefront input matrix is ​​expanded into a one-dimensional feature vector, as follows:

[0205]

[0206] in, The expanded one-dimensional feature vector has a dimension of . ; A matrix column vectorization operator, used to convert... dimensional matrix Concatenate the columns in order to form a one-dimensional vector, preserving all wavefront feature information in the matrix; This indicates that the one-dimensional eigenvector is A 3D real vector that fully encompasses the wavefront Zernike coefficient features across the entire field of view.

[0207] Enhanced Field-of-View Association Modeling:

[0208] To uncover and enhance the coupling correlation information between different fields of view in an optical system, and to improve the ability of the joint prediction model to capture wavefront correlation features across the entire field of view, a field-of-view encoding mechanism is introduced. This mechanism vectorizes the index information of each field of view, specifically defined as follows:

[0209]

[0210] in, For the first Encoding vectors for each field of view, with dimension This is used to characterize the positional association information of the field of view within the full field of view sequence; For field-of-view coding function; The dimension of the field-of-view encoding vector, whose value depends on the total number of fields of view. Adaptive adjustments are made to the complexity of the field of view association.

[0211] The wavefront feature vector of a single field of view is concatenated with the field of view encoding vector to construct an enhanced single field of view wavefront feature vector, thereby achieving the fusion of wavefront features and field of view related information. The specific form is as follows:

[0212]

[0213] in, For the enhanced version Each field-of-view wavefront feature vector has a dimension of _ ... ;symbol This represents the vertical concatenation operation of vectors, which combines wavefront feature vectors. With field of view encoding vector By stitching together the data along the dimensions, the complete features of both types of information are preserved.

[0214] Based on the enhanced single-view feature vector of the entire field of view, the final enhanced multi-view input matrix is ​​constructed, specifically in the following form:

[0215]

[0216] in, The enhanced multi-view input matrix has dimensions of . Each row of the matrix corresponds to an enhanced single-view feature vector; the matrix simultaneously covers wavefront feature information and field-of-view correlation information, which can provide more comprehensive input features for subsequent joint prediction models and improve the prediction accuracy of the model.

[0217] Normalization and standardization processing:

[0218] To eliminate the impact of dimensional differences between different feature dimensions and avoid problems such as unstable model training, slow convergence, and decreased prediction accuracy caused by excessive differences in feature numerical scales, this module normalizes the mean and standard deviation of the enhanced multi-view input matrix, ensuring that the processed feature data follows a standard normal distribution. The specific formula is as follows:

[0219]

[0220] in, This is the standardized enhanced multi-view input matrix; The original input matrix The mean of all elements in the dataset is used to characterize the central tendency of the feature data; The original input matrix The standard deviation of all elements in the dataset is used to characterize the dispersion of the feature data.

[0221] (3) Joint prediction model construction module:

[0222] The core function of this module is to establish a nonlinear mapping relationship between the multi-field wavefront input features and the offset and distortion of the optical system. By designing a shared feature extraction mechanism and dual-task dedicated prediction branches, it achieves synchronous, efficient, and accurate prediction of offset and distortion, providing reliable prediction output for the subsequent assembly and adjustment closed-loop optimization module. Specifically, this module includes: input feature construction, shared feature extraction, offset prediction branch outputting the component offset, and distortion prediction branch outputting the system distortion field and coupling process.

[0223] Specifically, step 2 of the method of this invention includes the following steps: The overall output relationship of the joint prediction model is defined as:

[0224]

[0225] in, The joint output of the joint prediction model corresponds to the predicted values ​​of optical system misalignment and system distortion, respectively. The joint prediction model is a nonlinear mapping function composed of multiple neural networks; This is the standardized enhanced multi-view input matrix; This is the set of all learnable parameters for the joint prediction model, including the shared encoder weight matrix, bias vector, weights and biases of each task branch, etc.

[0226] Shared feature extraction mechanism:

[0227] To fully exploit the common information in multi-field wavefront input features and avoid feature redundancy and computational resource waste caused by single-task prediction, a shared encoder is constructed. This encoder performs multi-level nonlinear transformations on the input features to extract high-level shared features that can simultaneously support offset and distortion prediction, thus achieving efficient feature reuse. The specific implementation process is as follows:

[0228] First, the multi-field wavefront input features (i.e., the enhanced multi-field input matrix) are used as the initial input to the shared encoder, and the initial input features are defined as follows:

[0229]

[0230] in, The initial input features of the shared encoder, and the multi-field wavefront input features The dimensions are completely consistent, preserving the original information of the input features; if If it is in matrix form, then If they are matrices of the same dimension; If it is in tensor form (such as a spatial sampling input tensor), then It is a tensor of the same dimension.

[0231] Next, the shared encoder extracts high-level shared features step by step through multi-layer nonlinear transformations. The feature update formula for each layer is as follows:

[0232]

[0233] in, For shared encoders The output features of a layer are high-level features after nonlinear transformation; For shared encoders The input features of the layer can be the initial input features. Or the output features of the previous layer; For the first The layer's weight matrix, with dimensions matching the input and output feature dimensions, is used to achieve linear transformation of the features; For the first The layer's bias vector is used to compensate for the offset of the linear transformation and improve the model's feature fitting ability. It is a non-linear activation function; This is the layer index of the shared encoder, with a value range of [value range missing]. , This represents the total number of layers in the shared encoder.

[0234] go through After the layer nonlinear transformation, the final output of the shared encoder is obtained, which is the high-level shared feature, specifically in the following form:

[0235]

[0236] in, For high-level shared feature vectors, the dimension is ; For shared encoders The output features of the layer share feature vectors with the higher layers. Consistent dimensions; This indicates that the high-level shared feature vector is 3D real vector To share feature dimensions, the dimensions are determined based on the complexity of the input features and the model's prediction requirements, ensuring that the core common information of multi-field wavefronts can be fully represented while avoiding the introduction of redundant features.

[0237] Imbalance prediction branch:

[0238] High-level shared features based on shared encoder output An independent misalignment prediction branch is constructed to achieve accurate prediction of the misalignment of each degree of freedom of the optical system, specifically expressed as follows:

[0239]

[0240] in, Let be the prediction vector of the system offset, with dimension . (The total number of adjustable degrees of freedom of the optical system), and the actual misalignment vector. Consistent dimensions; The weight matrix for the imbalance prediction branch has dimensions of . This is used to achieve a linear transformation from high-level shared features to offset quantities; The bias vector for the offset prediction branch has dimension . It is used to improve the fitting accuracy of the model and compensate for the systematic errors of the linear transformation.

[0241] Distortion prediction branch:

[0242] To meet the numerical computation and training requirements of the joint prediction model, the distortion prediction values ​​in vector field form are discretized into three-dimensional tensor form, as follows:

[0243]

[0244] in, The discretized distortion prediction tensor has a dimension of . ; The height of the discrete sampling of the image plane, i.e., the number of rows of sampled pixels; Dimension 2 represents the width of the discrete sampling of the image plane, i.e., the number of columns of sampled pixels; Dimension 2 corresponds to... direction and Predicted value of directional offset; This indicates that the tensor is a three-dimensional real tensor, which matches the discrete sampling dimension of the image plane and can completely preserve the distortion offset information of all sampling points on the image plane.

[0245] Task coupling enhancement mechanism:

[0246] To uncover the intrinsic relationship between offset and distortion, and to achieve information exchange and collaborative optimization between the two prediction tasks, addressing the issue of insufficient prediction accuracy caused by ignoring inter-task correlations in single-task prediction, this module introduces a task coupling enhancement mechanism. Distortion prediction is fed back to the offset prediction branch. After mapping, the output of the distortion prediction branch is fused with shared features from higher levels, optimizing the input features of the offset prediction branch and improving the accuracy of offset prediction. The specific formula is as follows:

[0247]

[0248] in, The optimized offset prediction branch inputs features, and the dimensions share features with the higher layers. Consistent; High-level shared features output by the shared encoder; , which is the coupling coefficient, used to adjust the feedback weight of the distortion prediction results; This is the feature mapping function, used to map the distorted predicted values ​​(i.e., the distorted prediction tensor). Mapping to shared features with higher levels Feature vectors with consistent dimensions enable feature fusion.

[0249] (4) Assembly and adjustment closed-loop optimization module:

[0250] The core function of this module is to predict the offset value based on the output of the joint prediction model. With distortion prediction value A dynamic adaptive closed-loop adjustment mechanism is constructed. Through a cyclical iterative process of "wavefront detection → model prediction → assembly parameter adjustment → wavefront retest (system state evaluation and iterative update)," the misalignment state of the optical system is gradually corrected, the system distortion is reduced, and the image quality index of the optical system converges to near the preset range.

[0251] The core logic of optical assembly and adjustment closed-loop optimization is: based on multi-field wavefront information and the prediction results of the joint prediction model, the system assembly and adjustment parameters are dynamically updated through adaptive step size control and multi-objective optimization strategy until the system misalignment and distortion meet the preset accuracy requirements, thereby forming an assembly and adjustment closed-loop optimization link of "measurement-prediction-adjustment-remeasurement".

[0252] Specifically, step 3 of the method of the present invention includes the following steps:

[0253] Assembly and adjustment closed-loop process:

[0254] The assembly and adjustment closed-loop optimization process adopts an iterative loop structure. Each iteration includes four core steps: wavefront detection (i.e., initial state detection), model prediction output, assembly and adjustment parameter adjustment, and state retesting and evaluation. It also includes a closed-loop iteration judgment step. The specific process is as follows:

[0255] 1) Initial State Detection: Perform initial multi-field wavefront detection on the optical system to acquire full-field wavefront data. Then, obtain the multi-field wavefront input features through the multi-field wavefront data acquisition and characterization module. ( Let be the number of iterations, the initial iteration. );

[0256] 2) Model prediction output: Input features from the initial multi-field wavefront Input the joint prediction model after training convergence to obtain the predicted value of the offset in the initial state (the predicted vector of the system offset). With distortion prediction values ​​(distortion prediction tensor) ;

[0257] 3) Adjustment of assembly parameters: based on the predicted value of the misalignment (i.e., the first... (Prediction vector of system offset in round iteration) An adaptive step size adjustment strategy is adopted to adjust the parameters of the optical system's assembly and adjustment mechanism, thereby obtaining the adjusted system misalignment. ;

[0258] 4) State Retesting and Evaluation: Perform multi-field wavefront retesting on the adjusted optical system to obtain new multi-field wavefront input characteristics. And new predicted values ​​are obtained through a joint prediction model. , (i.e., the prediction vector of the system offset and the distortion prediction tensor) are used to evaluate whether the current state of the system satisfies the convergence criterion.

[0259] 5) Closed-loop iterative judgment: If the system state satisfies the convergence criterion, stop the iteration and output the optimal setup parameters; if not, let... Repeat the above steps to enter the next round of closed-loop iteration until the convergence requirement is met, thus forming a closed-loop iterative optimization and commissioning process.

[0260] The above closed-loop iterative optimization and adjustment process can be simplified to the following iterative relationship:

[0261]

[0262] in, For the first Multi-field wavefront input features in round-iteration iteration; , For the first The model predictions in each iteration (i.e., the prediction vector of system offset and the distortion prediction tensor). For the first The system offset during iteration; during the iteration process, the system offset gradually approaches the optimal value, the distortion gradually decreases, and the wavefront quality gradually improves.

[0263] Iterative adjustment strategy:

[0264] The offset state evolution equation of the optical system describes the update law of the offset in each closed-loop iteration. The specific formula is as follows:

[0265]

[0266] in, For the first The actual system misalignment after rounds of iteration; For the first The actual system misalignment after rounds of iteration; For the first The step size for each iteration of the assembly and adjustment; For the first The offset prediction value (i.e., the prediction vector of the system offset) of the round of iterations; the negative sign indicates that the adjustment direction is opposite to the direction of the offset prediction value, that is, the adjustment is made in the direction of reducing the offset.

[0267] To ensure the stability and accuracy of the adjustment, an adaptive step size mechanism is adopted. This mechanism dynamically adjusts the step size based on the current system misalignment state, avoiding system oscillations caused by excessively large step sizes and inefficient iterations caused by excessively small step sizes. This problem is solved by using an exponentially decaying adaptive step size, as specifically implemented below:

[0268]

[0269] in, The initial step size is determined based on the adjustment accuracy of the assembly and adjustment mechanism; This is the step size decay coefficient, used to control the step size decay rate; The step size is the number of iterations. As the number of iterations increases, the step size gradually decreases to achieve the transition from "coarse adjustment" to "fine adjustment", balancing iteration efficiency and adjustment accuracy.

[0270] Construction of multi-objective optimization indicators:

[0271] To achieve synergistic optimization of system offset and distortion, a multi-objective joint optimization objective function is constructed to measure the adjustment quality of the optical system after each iteration and guide the direction of closed-loop iteration. The specific formula is as follows:

[0272]

[0273] in, For the first The smaller the objective function value of the multi-objective joint optimization of the system in each iteration, the better the system assembly and adjustment quality. , These are the weighting coefficients for wavefront error and distortion error, respectively, both being positive real numbers, satisfying... It can be adjusted according to actual assembly and adjustment needs. If priority is given to ensuring wavefront quality, then increase the [specific value]. If distortion is suppressed preferentially, then the effect will increase. ; For the first Wavefront error of round iteration, For the first The distortion error of each iteration is defined as follows:

[0274] wavefront error The root mean square error (RMS) of the full field of view wavefront is used to measure the quality of the full field of view wavefront. The formula is as follows:

[0275]

[0276] in, This represents the total number of fields of view in the system. For the first Round of iteration Continuous wavefront function for each field of view; For the first Ideal wavefront function for each field of view; For field of view index; 、 These are the two-dimensional spatial coordinate parameters of the wavefront function; For the first The wavefront sampling area of ​​each field of view; this index directly reflects the aberration level of the system, and the smaller the value, the better the wavefront quality.

[0277] Distortion error The vector field form of the distortion is expressed as the mean absolute error of the full-plane distortion migration, and the formula is:

[0278]

[0279] in, , These represent the height and width of the discrete sampling of the image plane, respectively. , For the first Iterating through image plane coordinates place direction and The actual distortion offset in the direction; , For the corresponding position direction and Ideal design distortion offset in direction; 、 These are the two-dimensional coordinate parameters of the image plane of the optical system.

[0280] Closed-loop convergence criterion:

[0281] To ensure that the assembly and adjustment closed-loop optimization can converge stably to the optimal state and avoid excessive or insufficient iteration, a dual convergence criterion is set, which simultaneously constrains the misalignment and distortion, as follows:

[0282] 1) Criterion for convergence of imbalance: The first The 2-norm of the prediction vector of the system offset in the round of iterations is less than a preset threshold. ,Right now:

[0283]

[0284] in, Let be the 2-norm of the prediction vector of the system offset, used to characterize the overall magnitude of the predicted offset; The threshold for misalignment convergence is determined based on the assembly and adjustment accuracy requirements of the optical system. The smaller the threshold, the higher the assembly and adjustment accuracy.

[0285] 2) Distortion convergence criterion: The first Distortion error of round iteration Less than the preset threshold ,Right now:

[0286]

[0287] in, The distortion convergence threshold is determined based on the design requirements of the optical system. The smaller the threshold, the better the distortion suppression effect.

[0288] 3) Additional conditions for iteration termination: If the number of iterations reaches the preset maximum number of iterations. Even if the above dual convergence criteria are not met, the iteration stops, the current optimal assembly and adjustment parameters are output, and an assembly and adjustment accuracy evaluation report is given to avoid inefficiency caused by infinite iteration. When both the misalignment convergence criterion and the distortion convergence criterion are met, the assembly and adjustment closed-loop optimization is considered complete, and the system has reached the optimal assembly and adjustment state. At this time, the final assembly and adjustment parameters are output as the final assembly and adjustment basis of the optical system.

[0289] (5) Joint prediction model training and optimization:

[0290] This process iteratively optimizes the learnable parameters of the aforementioned joint prediction model. By constructing a reasonable loss function and selecting suitable optimization algorithms and training strategies, the model can accurately learn the mapping relationship between multi-field wavefront input features and optical system offset and distortion. Simultaneously, it improves the model's generalization ability, convergence speed, and prediction stability, ensuring that the model output meets the accuracy requirements of practical engineering applications. The overall training and optimization process is as follows: First, a training sample dataset is constructed and preprocessed. Second, a multi-task joint loss function is defined. Then, an adaptive optimization algorithm is used to iteratively update the model parameters. Finally, the training process is monitored using a validation dataset, and regularization and early stopping strategies are introduced to prevent overfitting until the model converges and reaches the preset performance indicators. The following describes each process in more detail.

[0291] The construction and preprocessing of the training sample dataset specifically includes: The quality of the training sample data set directly determines the model training effect. It is necessary to construct a sample dataset covering different working states and different field-of-view settings of the optical system to ensure the diversity and representativeness of the samples. The specific construction and preprocessing process is as follows:

[0292] 1) Sample set composition: The training sample dataset adopts... It means that among them The total number of samples in the training sample dataset; For the first Multi-field wavefront input features of each sample; For the first The actual misalignment vector of each sample, with dimension [missing information]. ; For the first The actual distortion value of each sample can be expressed as a parameterized vector or a discrete tensor of a vector field, consistent with the form of the model's predicted output. For sample index;

[0293] 2) Sample preprocessing:

[0294] Outlier removal: The 3σ criterion is used to remove outliers from the training dataset. Outliers are removed when the multi-view wavefront input features, misalignment, or distortion value of a sample exceeds the mean of the corresponding features. When the range is exceeded, samples are identified as outliers and removed to prevent them from affecting the model's training accuracy. This represents the sample mean for the corresponding feature. This represents the sample standard deviation for the corresponding feature.

[0295] Sample splitting: The preprocessed training sample dataset is divided into a training set, a validation set, and a test set in a 7:2:1 ratio. The training set is used for iterative updates of model parameters, the validation set is used to monitor the model training process and adjust hyperparameters, and the test set is used to finally evaluate the model's generalization performance.

[0296] 3) Label normalization: for the first... The actual misalignment vector of each sample With the The actual distortion value of each sample Perform normalization to map it to The interval eliminates the influence of dimensional differences on the training process. The specific normalization formula is consistent with the range normalization of the multi-field wavefront input features mentioned above:

[0297]

[0298]

[0299] in, , These are the normalized th The actual misalignment vector of the sample and the first The actual distortion value of each sample; , These are the minimum and maximum values ​​of the actual misalignment vector for all training samples, respectively. , These are the minimum and maximum values ​​of the actual distortion values ​​for all training samples, respectively.

[0300] Design of a joint loss function for multiple tasks:

[0301] To achieve synergistic optimization of the two tasks of offset prediction and distortion prediction, while considering the prediction accuracy of both tasks, a multi-task joint loss function is designed. This function integrates the losses from both tasks using a weighted summation method. The specific formula is as follows:

[0302]

[0303] in, This is the model's overall joint loss function, used to measure the overall deviation between the model's predicted values ​​and the actual labels; These are the weighting coefficients for the offset prediction loss and the distortion prediction loss, respectively. They are adaptively adjusted based on the importance and prediction difficulty of the two tasks. If higher offset prediction accuracy is required, then the weighting coefficients are increased. Conversely, it increases. And satisfy ; The loss function for predicting the offset is... The distortion prediction loss function is defined as follows:

[0304] Offset Prediction Loss Function The mean squared error loss (MSE) is used, and the calculation formula is as follows:

[0305]

[0306] in, For the first The first sample Predicted values ​​of misalignment for each degree of freedom of assembly; For the first The first sample The actual misalignment amount label for each degree of freedom of assembly adjustment; This represents the total number of degrees of freedom in assembly and adjustment.

[0307] This loss function can converge quickly and is sensitive to small biases, making it easy to accurately optimize the prediction accuracy of the offset.

[0308] Distortion prediction loss function The mean absolute error loss (MAE) is used, and the calculation formula is as follows:

[0309]

[0310] in, , The first Image coordinates of each sample Place , Predicted value of orientation distortion offset; , These are the actual distortion offsets corresponding to the position; , These represent the height and width of the discrete sampling on the image plane; dimension 2 corresponds to... , Offset in both directions; 、 These are the two-dimensional coordinate parameters of the image plane of the optical system.

[0311] Model optimization algorithm selection and parameter update:

[0312] Adaptive momentum optimization algorithm is used to optimize the learnable parameters of the model. The algorithm combines the advantages of momentum gradient descent and adaptive learning rate adjustment to perform iterative updates, which can accelerate model convergence, avoid local optima, and improve the stability of the training process. The specific parameter update process is as follows:

[0313] 1) Initialize parameters: Initialize all learnable parameters of the model. The Adam algorithm is initialized using a random normal distribution (mean 0, standard deviation 0.01). The hyperparameters for initializing the Adam algorithm include: learning rate. Momentum parameters , and numerical stability parameters .

[0314] 2) Gradient calculation: For each training batch, calculate the total joint loss function. For all learnable parameters gradient The backpropagation algorithm is used to complete the backpropagation and calculation of the gradient.

[0315] 3) Adaptive Parameter Update: Based on the update rule of the Adam algorithm, an exponential moving average is applied to the gradient to update the first and second moment estimates of the parameters. The learning rate for each parameter is then adjusted accordingly, ultimately completing the parameter update. The specific formula is as follows:

[0316]

[0317] in, This represents the number of iterations. , These are the first-order moment estimates and second-order moment estimates of the gradient, respectively. , These are the deviation correction values ​​for the first-order moment estimate and the second-order moment estimate, respectively; , The first Wheel, First Model parameters after one iteration; The learning rate; This is a numerical stability parameter used to avoid cases where the denominator is 0; , The first First-order and second-order moment estimates of the gradient in each iteration; This is the gradient operator.

[0318] Training convergence criteria and performance evaluation metrics:

[0319] 1) Training convergence criterion: The model is considered to have converged when any of the following conditions are met: (a) the number of iterations reaches the preset maximum number of iterations; (b) in a consecutive preset number of iterations, the changes in both the training set loss and the validation set loss are less than a preset threshold, i.e.:

[0320]

[0321] in, , The first Wheel, First The training set loss in each iteration; , The first Wheel, First The validation set loss in each iteration; This is the threshold for the change in loss.

[0322] 2) Model performance evaluation metrics: The performance of the converged model is evaluated using a test set, with evaluation metrics designed for offset prediction and distortion prediction respectively:

[0323] (a) Indicators for predicting and evaluating imbalance: Mean Absolute Error (MAE) and Coefficient of Determination. The formulas are as follows:

[0324]

[0325]

[0326] in, The total number of samples in the test set. For the test set sample index, It represents the total number of degrees of freedom in assembly and adjustment; For the assembly and adjustment of degrees of freedom index, For the first In the test sample, the first The average value of the true label for the degree of freedom mismatch of each assembly. For the first In the test sample, the first The true value of the label for the degree of freedom mismatch of an assembly. For the first In the test sample, the first The predicted value of the misalignment label for each degree of freedom of assembly. The mean absolute deviation between the predicted and actual values ​​of the imbalance quantity. The evaluation model's overall explanatory power for the patterns of imbalance changes, and The smaller The closer it is to 1, the better the performance in predicting the imbalance.

[0327] (b) Distortion prediction evaluation index: using the average endpoint error The distance between the endpoints of the predicted offset vector and the actual offset vector is measured by the following formula:

[0328]

[0329] in, , The first Image coordinates of each sample Place , Predicted value of orientation distortion offset; , These are the labels for the actual distortion offset at the corresponding position; , These represent the height and width of the discrete sampling on the image plane; dimension 2 corresponds to... , Offset in both directions The smaller the value, the higher the accuracy of distortion prediction.

[0330] Figure 2 The adjustable degrees of freedom of the optical elements within the optical system are demonstrated. These adjustable optical elements possess translational degrees of freedom along the X, Y, and Z directions and rotational degrees of freedom about the X, Y, and Z axes, such as... Figure 2 In the coordinate system, XDE / YDE / ZDE represent the elements' orbital distances around the local coordinate system. x / y / z Axis eccentricity, ADE / BDE / CDE respectively represent: the element's eccentricity around the local coordinate system x / y / z The axis is tilted.

[0331] Figure 3This paper presents the wavefront distribution plot of an optical system and its fitted partial fringe Zernike polynomials and their coefficients. The wavefront distribution plot is used to characterize the error state of the actual wavefront relative to the ideal reference wavefront, while the fringe Zernike coefficients are used to quantitatively characterize the various components of the wavefront error, such as... Represents the piston term. and Indicates tilt in the X and Y directions. This indicates out of focus. and represent and Like scattered, and Represents coma in the X and Y directions. The first term represents the basic ball difference, and the remaining terms represent the advanced ball difference. Figure 3 middle (Distinguished from its subscript); p is the normalized radial coordinate in the pupil plane, used to characterize the radial position of any point within the pupil relative to the pupil center. Its value range is 0≤p≤1, where p=0 represents the pupil center and p=1 represents the pupil edge; A is the azimuth angle in the pupil plane, starting from the +X axis of the local coordinate system and increasing counterclockwise. p describes the spatial distribution of wavefront aberrations in the radial direction of the pupil, and A describes the spatial distribution of wavefront aberrations in the circumferential direction of the pupil. Combinations of radial and angular functions of different orders constitute different types of Zernike wavefront aberration modes.

[0332] Figure 4 The distortion vector distribution diagram of the optical system is shown. The horizontal and vertical axes represent the field of view angles in the X and Y directions of the object side, respectively. The arrows in the diagram indicate the relative magnitude of the offset direction of the actual image point relative to the ideal image point at each field of view. The distortion distribution is used to characterize the offset law of the image point in the field of view under examination and to analyze the imaging performance and misalignment characteristics of the optical system.

[0333] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for joint prediction of misalignment and distortion in an optical system based on multi-field wavefront information, the applicable system comprising: a multi-field wavefront data acquisition and characterization module, a joint prediction model construction module, and an assembly and adjustment closed-loop optimization module that are interconnected; The multi-field wavefront data acquisition and characterization module is used to map wavefront information from different fields of view to a unified feature space. The joint prediction model construction module is used to establish the nonlinear mapping relationship between the multi-field wavefront input features and the misalignment and distortion of the optical system, respectively. The assembly and adjustment closed-loop optimization module is used to construct a dynamic adaptive closed-loop adjustment mechanism based on the misalignment prediction value and distortion prediction value output by the joint prediction model, gradually correct the misalignment state of the optical system, reduce system distortion, and bring the image quality index of the optical system to near the preset range. Its features are, The method includes the following steps: Step 1: The multi-field wavefront data acquisition and characterization module acquires multi-field wavefront data during optical design simulation or actual assembly and adjustment, and constructs corresponding misalignment and distortion labels; Step 2: The joint prediction model construction module takes multi-field wavefront data as input and establishes a joint prediction model that combines shared feature extraction and dual-task output. One branch of the joint prediction model outputs the predicted values ​​of the misalignment of the optical components of the optical system, and the other branch outputs the predicted values ​​of the distortion of the entire field of view. Step 3: After the joint prediction model is trained by joint loss constraints, the assembly and adjustment closed-loop optimization module inputs the measured multi-field wavefront data into the joint prediction model, and simultaneously obtains the current optical system's misalignment prediction results and distortion assessment results, and guides the adjustment of the optical system accordingly; after each optical system adjustment, the multi-field wavefront data is measured again and input into the joint prediction model to form a closed-loop iterative optimization assembly and adjustment process.

2. The method for joint prediction of misalignment and distortion in an optical system based on multi-field wavefront information as described in claim 1, characterized in that, Step 1 is as follows: For the For each field of view, its wavefront function is defined as: in, For the first Continuous wavefront function under each field of view; 、 These are the two-dimensional spatial coordinate parameters of the wavefront function; This represents the total number of fields of view in the optical system. For the field of view index; the wavefront function is represented in normalized polar coordinates, i.e.: in, The wavefront function in normalized polar coordinates. The normalized polar radius has a range of values. ; The polar angle has a range of values ​​of 100. ; The wavefront function in normalized polar coordinates is parametrically expressed using the Zernike coefficient form of wavefront expansion, specifically as follows: in, For the first The first field wavefront function Zernike polynomial coefficients of order 1; For the first Zernike polynomial of order Z; This represents the total order of the Zernike polynomials used. Number the Zernike polynomial order; Define a single-field-of-view wavefront eigenvector to integrate all Zernike polynomial coefficients for a single field of view, in the following form: in, For the first Wavefront feature vectors of fields of view, with dimension ; superscript This represents the transpose operation of a vector; This indicates that the feature vector is 3D real vector; For the first The first field wavefront function Zernike polynomial coefficients of order 1; For the first The first field wavefront function Zernike polynomial coefficients of order 1; For the first The first field wavefront function Zernike polynomial coefficients of order 1; Next, the feature vectors of all single-field wavefronts across the entire field of view are integrated to construct a multi-field wavefront input matrix, namely: in, The multi-field wavefront input matrix has a dimension of . ; This matrix is ​​represented as OK A matrix of real numbers in columns; For the first Wavefront eigenvectors of each field of view For the first Wavefront eigenvectors of each field of view For the first Wavefront eigenvectors of each field of view; The above multi-field wavefront input matrix is ​​expanded into a one-dimensional eigenvector, as follows: in, The expanded one-dimensional feature vector has a dimension of . ; A matrix column vectorization operator, used to convert... dimensional matrix Concatenate the columns in order to form a one-dimensional vector; This indicates that the one-dimensional eigenvector is 3D real vector; The index information for each field of view is vectorized and specifically defined as follows: in, For the first Encoding vectors for each field of view, with dimension ; For field-of-view coding function; The dimension of the field-of-view encoding vector; The wavefront feature vector of a single field of view is concatenated with the field of view encoding vector to construct an enhanced single field of view wavefront feature vector, thereby achieving the fusion of wavefront features and field of view related information. The specific form is as follows: in, For the enhanced version Each field-of-view wavefront feature vector has a dimension of _ ... ;symbol This represents the vertical concatenation operation of vectors; Based on the enhanced single-view wavefront eigenvectors of the entire field of view, the final enhanced multi-view input matrix is ​​constructed, specifically in the following form: in, The enhanced multi-view input matrix has dimensions of . ; The mean and standard deviation of the enhanced multi-view input matrix are normalized to ensure that the processed feature data follows a standard normal distribution. The specific formula is as follows: in, This is the standardized enhanced multi-view input matrix; The original input matrix The mean of all elements in the dataset; The original input matrix The standard deviation of all elements in the matrix.

3. The method for joint prediction of misalignment and distortion in an optical system based on multi-field wavefront information as described in claim 1, characterized in that, Step 2 is as follows: The overall output relationship of the joint prediction model is defined as follows: in, The joint output of the joint prediction model corresponds to the predicted values ​​of optical system misalignment and distortion, respectively. The joint prediction model is a nonlinear mapping function composed of multiple neural networks; This is the standardized enhanced multi-view input matrix; The set of all learnable parameters for the joint prediction model; The input features are subjected to multi-level nonlinear transformations to extract high-level shared features that can simultaneously support offset and distortion prediction, thereby achieving efficient feature reuse. The specific implementation process is as follows: First, the enhanced multi-view input matrix is ​​used as the initial input to the shared encoder, and the initial input features are defined as follows: in, The initial input features of the shared encoder; if If it is in matrix form, then If they are matrices of the same dimension; If it is in tensor form, then Tensors of the same dimension; Next, the shared encoder extracts high-level shared features step by step through multi-layer nonlinear transformations. The feature update formula for each layer is as follows: in, For shared encoders The output features of the layer; For shared encoders The input features of the layer; For the first Layer weight matrix; For the first Layer bias vector; It is a non-linear activation function; This is the layer index of the shared encoder, with a value range of [value range missing]. , This represents the total number of layers in the shared encoder. go through After the layer nonlinear transformation, the final output of the shared encoder is obtained, which is the high-level shared feature, specifically in the following form: in, For high-level shared feature vectors, the dimension is ; For shared encoders The output features of the layer share feature vectors with the higher layers. Consistent dimensions; This indicates that the high-level shared feature vector is 3D real vector To share feature dimensions; High-level shared features based on shared encoder output An independent misalignment prediction branch is constructed to achieve accurate prediction of the misalignment of each degree of freedom of the optical system, specifically expressed as follows: in, Let be the prediction vector of the system offset, with dimension . ; The weight matrix for the imbalance prediction branch has dimensions of . ; The bias vector for the offset prediction branch has dimension . ; The distortion predictions in vector field form are discretized into three-dimensional tensor form, as follows: in, The discretized distortion prediction tensor has a dimension of . ; The height of the discrete sampling of the image plane; The width of the discrete sampling of the image plane; dimension 2 corresponds to... direction and Predicted value of directional offset; This indicates that the distortion prediction tensor is a three-dimensional real tensor; The distortion prediction tensor is fed back to the imbalance prediction branch. After mapping the output of the distortion prediction branch, it is fused with the shared features of the higher layers to optimize the input features of the imbalance prediction branch and improve the accuracy of imbalance prediction. The specific formula is as follows: in, The optimized offset prediction branch inputs features, and the dimensions share features with the higher layers. Consistent; High-level shared features output by the shared encoder; , which is the coupling coefficient, used to adjust the feedback weight of the distortion prediction results; This is a feature mapping function used to map the distortion prediction tensor. Mapping to shared features with higher levels Feature vectors with consistent dimensions enable feature fusion.

4. The method for joint prediction of misalignment and distortion in an optical system based on multi-field wavefront information as described in claim 1, characterized in that, Step 3 specifically includes: Step 31, Initial State Detection: Initial multi-field wavefront detection is performed on the optical system. Full-field wavefront data is acquired through the multi-field wavefront data acquisition and characterization module to obtain the multi-field wavefront input features. , Let be the number of iterations, the initial iteration. ; Step 32, Model Prediction Output: Input features from the initial multi-field wavefront Input the joint prediction model after training convergence to obtain the prediction vector of the system offset in the initial state. With distortion prediction tensor ; Step 33, Adjustment of assembly parameters: Based on the Prediction vector of system offset in round iteration An adaptive step size adjustment strategy is adopted to adjust the parameters of the optical system's assembly and adjustment mechanism, thereby obtaining the adjusted system misalignment. ; Step 34, Status Retesting and Evaluation: Multi-field wavefront remeasurement was performed on the adjusted optical system to obtain new multi-field wavefront input features. And a new prediction vector for the system imbalance is obtained through a joint prediction model. and distortion prediction tensor Evaluate whether the current state of the optical system satisfies the convergence criterion; Step 35, Closed-loop iteration judgment: If the state of the optical system satisfies the convergence criterion, then stop the iteration and output the optimal setup parameters; otherwise, let Repeat the above steps to enter the next round of closed-loop iteration until the convergence requirement is met, thus forming a closed-loop iterative optimization and commissioning process.

5. The method for joint prediction of misalignment and distortion in an optical system based on multi-field wavefront information according to claim 4, characterized in that, In step 3, the iterative adjustment strategy for the closed-loop iterative optimization process is as follows: The offset state evolution equation of the optical system is used to describe the update law of the offset in each closed-loop iteration. The specific formula is as follows: in, For the first The actual system misalignment after rounds of iteration; For the first The actual system misalignment after rounds of iteration; For the first The step size for each iteration of the assembly and adjustment; For the first The prediction vector of the system offset in each iteration is the predicted value of the offset; the negative sign indicates that the adjustment direction is opposite to the direction of the predicted offset value, that is, the adjustment is made in the direction of reducing the offset. An adaptive step size mechanism is adopted to dynamically adjust the step size according to the current misalignment state of the optical system. The specific implementation method is as follows: in, The initial step size is determined based on the adjustment accuracy of the assembly and adjustment mechanism; This is the step size decay coefficient; This represents the number of iterations.

6. The method for joint prediction of misalignment and distortion in an optical system based on multi-field wavefront information according to claim 4, characterized in that, In step 3, the construction of the multi-objective joint optimization objective function for the closed-loop iterative optimization process is as follows: To achieve synergistic optimization of the misalignment and distortion of the optical system, a multi-objective joint optimization objective function is constructed to measure the adjustment quality of the optical system after each iteration and guide the direction of the closed-loop iteration. The specific formula is as follows: in, For the first The system's multi-objective joint optimization objective function value is obtained through round-by-round iteration. , These are the weighting coefficients for wavefront error and distortion error, respectively, satisfying... ; For the first Wavefront error of round iteration, For the first The distortion error of each iteration is defined as follows: wavefront error The root mean square error of the full field of view wavefront is used to measure the quality of the full field of view wavefront. The formula is as follows: in, This represents the total number of fields of view for the optical system. For the first Round of iteration Continuous wavefront function for each field of view; For the first Ideal wavefront function for each field of view; For the first Wavefront sampling area of ​​each field of view; For field of view index; 、 These are the two-dimensional spatial coordinate parameters of the wavefront function; Distortion error The vector field form of the distortion is expressed as the mean absolute error of the full-plane distortion migration, and the formula is: in, , These represent the height and width of the discrete sampling of the image plane, respectively. , The first Iterating through image plane coordinates place direction and The actual distortion offset in the direction; , They are respectively the corresponding positions direction and Ideal design distortion offset in direction; 、 These are the two-dimensional coordinate parameters of the image plane of the optical system.

7. The method for joint prediction of misalignment and distortion in an optical system based on multi-field wavefront information according to claim 4, characterized in that, In step 3, the convergence criterion is as follows: A dual convergence criterion is established, simultaneously constraining both misalignment and distortion, as follows: 1) Criterion for convergence of imbalance: The first The 2-norm of the prediction vector of the system offset in the round of iterations is less than a preset threshold. ,Right now: in, Let be the 2-norm of the prediction vector of the system offset; This is the convergence threshold for the imbalance. 2) Distortion convergence criterion: The first Distortion error of round iteration Less than the preset threshold ,Right now: in, This is the distortion convergence threshold; 3) Additional conditions for iteration termination: If the number of iterations reaches the preset maximum number of iterations. Even if the above dual convergence criteria are not met, the iteration stops, the current optimal assembly and adjustment parameters are output, and an assembly and adjustment accuracy evaluation report is given. When both the misalignment convergence criterion and the distortion convergence criterion are met, the assembly and adjustment closed-loop optimization is determined to be completed, and the optical system reaches the optimal assembly and adjustment state. At this time, the final assembly and adjustment parameters are output as the final assembly and adjustment basis of the optical system.

8. The method for joint prediction of misalignment and distortion in an optical system based on multi-field wavefront information as described in claim 1, characterized in that, The process of training the joint prediction model in steps 2 and 3 using joint loss constraints is as follows: First, construct the training sample dataset and preprocess it; Secondly, define a joint loss function for multiple tasks; Then, an adaptive optimization algorithm is used to iteratively update the model parameters; Finally, the training process is monitored by validating the dataset, and regularization and early stopping strategies are introduced to prevent overfitting until the model converges and reaches the preset performance metrics.

9. The method for joint prediction of misalignment and distortion in an optical system based on multi-field wavefront information as described in claim 8, characterized in that, The specific steps for constructing and preprocessing the training sample dataset are as follows: 1) Sample set composition: The training sample dataset adopts... It means that among them The total number of samples in the training sample dataset; For the first Multi-field wavefront input features of each sample; For the first The actual misalignment vector of each sample, with dimension [missing information]. ; For the first The actual distortion value of each sample; For sample index; 2) Sample preprocessing: Outlier removal: The 3σ criterion is used to remove outliers from the training dataset. Outliers are removed when the multi-view wavefront input features, misalignment, or distortion value of a sample exceeds the mean of the corresponding features. When the sample falls within the specified range, it is identified as an abnormal sample and removed. This represents the sample mean for the corresponding feature. The sample standard deviation is the corresponding feature. Sample splitting: The preprocessed training sample dataset is divided into training set, validation set and test set in a ratio of 7:2:

1. The training set is used for iterative updates of model parameters, the validation set is used to monitor the model training process and adjust hyperparameters, and the test set is used to finally evaluate the model's generalization performance. 3) Label normalization: for the first... The actual misalignment vector of each sample With the The actual distortion value of each sample Perform normalization to map it to The interval, specifically: in, , These are the normalized th The actual misalignment vector of the sample and the first The actual distortion value of each sample; , These are the minimum and maximum values ​​of the actual misalignment vector for all training samples, respectively. , These are the minimum and maximum values ​​of the actual distortion values ​​for all training samples, respectively. The multi-task joint loss function integrates the losses of two tasks using a weighted summation method. The specific formula is as follows: in, This is the model's overall joint loss function, used to measure the overall deviation between the model's predicted values ​​and the actual labels; These are the weighting coefficients for the offset prediction loss and the distortion prediction loss, respectively. They are adaptively adjusted based on the importance and prediction difficulty of the two tasks. If higher offset prediction accuracy is required, then the weighting coefficients are increased. Conversely, it increases. And satisfy ; The loss function for predicting the offset is... The distortion prediction loss function is defined as follows: Offset Prediction Loss Function The mean squared error loss is used, and the calculation formula is as follows: in, For the first The first sample Predicted values ​​of misalignment for each degree of freedom of assembly; For the first The first sample The actual misalignment amount label for each degree of freedom of assembly; This represents the total number of degrees of freedom in assembly and adjustment. Distortion prediction loss function The mean absolute error loss is used, and the calculation formula is as follows: in, , The first Image coordinates of each sample Place , Predicted value of orientation distortion offset; , These are the actual distortion offsets corresponding to the position; , These represent the height and width of the discrete sampling on the image plane; dimension 2 corresponds to... , The offset in both directions, 、 These are the two-dimensional coordinate parameters of the image plane of the optical system; The adaptive optimization algorithm is used to iteratively update the model parameters as follows: Adaptive momentum optimization algorithm is used to optimize the learnable parameters of the model. The iterative update process is as follows: 1) Initialize parameters: Initialize all learnable parameters of the model. The Adam algorithm is initialized using a random normal distribution; the hyperparameters for initialization include the learning rate. Momentum parameters , and numerical stability parameters ; 2) Gradient calculation: For each training batch, calculate the total joint loss function. For all learnable parameters gradient The backpropagation algorithm is used to complete the backpropagation and calculation of the gradient; 3) Adaptive Parameter Update: Based on the update rule of the Adam algorithm, an exponential moving average is applied to the gradient to update the first and second moment estimates of the parameters. The learning rate for each parameter is then adjusted accordingly, ultimately completing the parameter update. The specific formula is as follows: in, This represents the number of iterations. , These are the first-order moment estimates and second-order moment estimates of the gradient, respectively. , These are the deviation correction values ​​for the first-order moment estimate and the second-order moment estimate, respectively; , The first Wheel, First Model parameters after one iteration; , The first First-order and second-order moment estimates of the gradient in each iteration; For gradient operators; The model converges and reaches the preset performance index as follows: 1) Training convergence criterion: The model is considered to have converged when any of the following conditions are met: (a) the number of iterations reaches the preset maximum number of iterations; (b) in a consecutive preset number of iterations, the changes in both the training set loss and the validation set loss are less than a preset threshold, i.e.: in, , The first Wheel, First The training set loss during each iteration; , The first Wheel, First The validation set loss in each iteration; The threshold for loss variation; 2) Model performance evaluation metrics: The performance of the converged model is evaluated using a test set, with evaluation metrics designed for offset prediction and distortion prediction respectively: (a) Imbalance prediction and evaluation indicators: using mean absolute error and coefficient of determination. The formula is: in, The total number of samples in the test set. For the test set sample index, It represents the total number of degrees of freedom in assembly and adjustment; For the assembly and adjustment of degrees of freedom index, For the first In the test sample, the first The average value of the true label for the degree of freedom mismatch of each assembly. For the first In the test sample, the first The true value of the label for the degree of freedom mismatch of an individual assembly. For the first In the test sample, the first The predicted value of the misalignment label for each degree of freedom of assembly. The mean absolute deviation between the predicted and actual values ​​of the imbalance quantity; (b) Distortion prediction evaluation index: using average endpoint error The distance between the endpoints of the predicted offset vector and the actual offset vector is measured by the following formula: in, , The first Image coordinates of each sample Place , Predicted value of orientation distortion offset; , These are the labels for the actual distortion offset at the corresponding position.

Citation Information

Patent Citations

  • Optical system distortion measuring method based on Shack-Hartmann wave-front sensor

    CN106768882A

  • Optical lens installation and adjustment pose error identification method based on deep learning

    CN117537711A