Optical system alignment method based on a double-branch neural network model
By adopting an optical system assembly and adjustment method based on a dual-branch neural network model, the problems of optical assembly and adjustment accuracy and consistency are solved, achieving efficient and stable optical system assembly and adjustment, reducing reliance on manual experience and the frequency of physical experiments.
Patent Information
- Application Number
- CN202511326764.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-17
- Publication Date
- 2025-11-25
- Estimated Expiration
- 2045-09-17
AI Technical Summary
Existing optical assembly and adjustment methods are insufficient to meet the requirements of harsh environments in terms of accuracy and consistency. Traditional methods are inefficient and lack model generalization ability, making them difficult to apply to real-world assembly and adjustment scenarios.
An optical system assembly and adjustment method based on a dual-branch neural network model is adopted. By constructing a theoretical optical system model, analyzing the noise in the actual assembly and adjustment scenario, generating a training dataset, and constructing a dual-branch neural network model, the method combines wavefront plots and Zernike coefficient features to perform high-precision prediction and adjustment of component misalignment.
It improves the accuracy and stability of optical assembly and adjustment, reduces reliance on manual experience, increases assembly and adjustment efficiency, reduces the frequency of physical experiments, and enhances the applicability of the model in actual assembly and adjustment.
Smart Images

Figure CN120822547B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of optical alignment and artificial intelligence, and particularly relates to an optical system alignment method based on a double-branch neural network model. BACKGROUND
[0002] With the development of space optical technology, deep space exploration technology, gravitational wave detection, etc., the application environment of optical systems is becoming more and more severe, and the imaging performance requirements are becoming more and more extreme, which brings great challenges to the development of optical instruments, and the integrated alignment technology of optical systems urgently needs new breakthroughs. Traditional optical alignment mainly relies on the experience of alignment personnel and conventional alignment process equipment, and the alignment efficiency is low, and the precision and consistency are difficult to guarantee. With the continuous development of computer simulation technology, optical design, optical processing and detection technology, etc., computer-aided alignment technology (Computer-Aided Alignment, CAA) emerges as the times require, which uses mathematical models to establish the relationship between optical system wave aberration and element misalignment, and solves the element position error according to the measured wavefront data to guide the alignment, which can greatly improve the optical alignment efficiency and system performance compared with the traditional alignment method. The existing misalignment solving methods are: sensitivity matrix method, inverse optimization method, vector aberration method and artificial neural network method, etc.
[0003] Because the sensitivity matrix method has simple model, mature research and convenient use, its application range is the most extensive, and the current commonly used optical alignment route is: manual alignment relying on experience + computer-aided alignment based on sensitivity matrix method. However, the sensitivity matrix method has the following problems: 1) only considers the low-order influence of element position on system imaging quality and ignores the high-order influence, and the precise solution cannot be obtained through single iteration, which needs to be iterated multiple times; 2) does not consider the mutual coupling relationship between element degrees of freedom and different wavefront Zernike coefficients, and cannot reflect the essential law of the influence of alignment error on system wave aberration; 3) cannot consider the influence of actual alignment noise, resulting in low precision in actual application.
[0004] Artificial neural network method has shown significant advantages in optical system alignment field due to its nonlinear mapping ability, autonomous feature learning mechanism and dynamic adaptive characteristics. It can efficiently construct a mathematical model of multivariate coupling relationship by deeply mining the complex nonlinear relationship between optical system wavefront aberration and alignment error. Its application has gradually extended from basic research to engineering practice, and will become an important technical means for precise optical alignment. The significant advantages of using artificial intelligence technology for optical alignment are as follows: 1) in terms of alignment accuracy, the error propagation law of optical system is learned through adaptive learning, which significantly improves the aberration compensation accuracy and system stability; 2) in terms of efficiency improvement, intelligent optimization algorithm is combined to realize rapid iteration of alignment parameters, which promotes the intelligent upgrading of automatic alignment process; 3) in terms of cost control, the frequency of physical experiments is reduced through data-driven virtual alignment rehearsal, which effectively reduces resource consumption. These technical breakthroughs will lay a key technical foundation for the evolution of optical alignment towards intelligence and integration.
[0005] At present, the application of artificial neural network method in optical alignment mainly focuses on using optical simulation software to simulate the large amount of simulation data set of misalignment optical system and wavefront aberration for training. The model training focuses on using different artificial intelligence algorithms to construct more accurate nonlinear mapping relationship between misalignment and wavefront Zernike coefficient. The existing methods are limited to using a single optical performance evaluation parameter as the model input, which lacks sufficient utilization of multi-dimensional optical system information, resulting in limited accuracy of misalignment prediction by the model and low model generalization ability. At the same time, the noise factors such as optical element surface error, mechanical actuator adjustment error and system wavefront measurement error in actual alignment scene are not considered in the training process of the model, resulting in low robustness of the high-precision model trained by simulation data and difficulty in applying it to actual alignment scene. SUMMARY
[0006] The present application solves the technical problems of limited accuracy of misalignment prediction by the model, low model generalization ability and difficulty in applying to actual alignment scene in the prior art, and provides an optical system alignment method based on a double-branch neural network model.
[0007] To solve the above technical problems, the technical scheme of the present application is as follows:
[0008] An optical system alignment method based on a double-branch neural network model, comprising the following steps:
[0009] Step 1, component misalignment sensitivity analysis;
[0010] Construct a theoretical optical system model and calculate the misalignment sensitivity matrix of the optical system;
[0011] Step 2, analysis of noise in actual alignment scene;
[0012] According to the actual installation environment, analyze and determine the noise in the optical system installation process and its distribution characteristics;
[0013] Step 3, neural network model training data set generation;
[0014] A large number of data sets for training the neural network installation agent model are generated;
[0015] Step 4, neural network model construction and training;
[0016] A double-branch neural network installation agent model is constructed, and the neural network installation agent model training is carried out;
[0017] Step 5, using the neural network model to guide the actual installation;
[0018] The neural network installation agent model is used to guide the actual optical system installation.
[0019] In the above technical solution, step 1 specifically includes:
[0020] Step 11, a theoretical optical system model is established in the optical analysis software;
[0021] Step 12, an integrated simulation analysis link is built, the curves of the wavefront aberration evaluation index corresponding to each optical element and single position degree of freedom in the optical system with respect to the misadjustment amount are calculated, and the misadjustment sensitivity matrix of the optical system is obtained through curve fitting.
[0022] In the above technical solution, step 2 specifically includes:
[0023] Step 21, optical element surface shape machining error and measurement error;
[0024] According to the statistical distribution law of the error under the real machining and measurement conditions, the specific distribution characteristics and range of the machining error and the measurement error are obtained, and are converted into the form of wavefront Zernike coefficient, and then the machining error and the measurement error are superimposed to comprehensively represent the optical element surface shape error after actual machining;
[0025] Step 22, wavefront detection error of the optical system;
[0026] Through experience and measured data analysis, the error statistical distribution law of the wavefront measurement instrument is obtained, and is converted into the form of wavefront Zernike coefficient;
[0027] Step 23, mechanical actuator error;
[0028] Through empirical judgment and measured data analysis, the system cumulative error and random noise distribution law and range of the actuator are determined, and are converted into the expression form of the element position and posture six degrees of freedom.
[0029] In the above technical solution, step 3 specifically includes:
[0030] Step 31: Determine the inputs and outputs for training the neural network assembly agent model;
[0031] Step 32: Based on the misalignment range of each pose degree of freedom of the component determined in Step 1 and the distribution law and range of various noises determined in Step 2, batch generate a set of data pairs of misalignment of each pose degree of freedom of the component for neural network model training, which takes into account the influence of various actual noises.
[0032] In the above technical solution, the specific methods of adding various noises in step 32 are as follows:
[0033] First, the optical element surface shape error determined in step 21 is added to the theoretical optical design model, and the model is used as the ideal model in the subsequent data generation and assembly adjustment iteration regression process.
[0034] Next, when generating each set of randomly uniformly distributed misalignment combinations, according to the noise distribution law and range of the assembly and adjustment mechanical actuator determined in step 23, the noise of the assembly and adjustment mechanical actuator is added to each adjustable element position degree of freedom, and then the wavefront diagram of the optical system and its Zernike coefficient corresponding to the random misalignment combination after adding the noise are calculated.
[0035] Finally, wavefront detection noise is added to the calculated optical system wavefront aberration data;
[0036] The wavefront detection noise of the optical system, determined in step 22 and expressed in Zernike coefficient form, is directly added to the wavefront Zernike coefficients calculated in the previous step, and the noise distribution is converted into an image and superimposed on the wavefront map.
[0037] In the above technical solutions,
[0038] Step 4 specifically includes:
[0039] Step 41: Data preprocessing;
[0040] A neural network assembly surrogate model is proposed to achieve information acquisition and feature fusion of two modal data. By normalizing and standardizing the wavefront diagram of the input optical system and its fitted Zernike coefficients and the component pose degree of freedom misalignment of the output optical system, the input and output data can meet the requirements of the neural network assembly surrogate model for numerical scale and distribution consistency, thereby improving the stability and accuracy of the training of the neural network assembly surrogate model.
[0041] Step 42, Model Structure Design;
[0042] The neural network assembly agent model structure includes: a two-branch feature extraction module, a feature fusion module, and a component misalignment prediction module;
[0043] Feature extraction is performed through the wavefront plot branch and wavefront Zernike coefficient branch of the dual-branch feature extraction module, and then the extracted features are fused through the feature fusion module. Finally, the component offset prediction module achieves high-precision prediction of component offset.
[0044] Step 43: Model training;
[0045] By using large-scale labeled data, the parameters of the dual-branch feature extraction module, feature fusion module, and component misalignment prediction module are optimized to achieve high-precision prediction of optical component misalignment. Based on the input data and the training set of corresponding multi-degree-of-freedom misalignment labels, the parameters of the neural network assembly surrogate model are optimized using the mean square error loss function. The parameters are iteratively updated using the mini-batch gradient descent algorithm, and overfitting is effectively prevented by combining regularization and early stopping mechanisms. Finally, accurate estimation of the misalignment state of optical components is achieved.
[0046] Step 44: Model Validation;
[0047] First, from the perspective of prediction accuracy, the mean squared error and mean absolute error metrics are used to evaluate the difference between the offset output of the neural network surrogate model and the true value on the test set:
[0048] ,
[0049] ,
[0050] in, To predict the misalignment in the model, For the true degree of disorder, The total number of test samples, The test sample number, Indicates mean square error. Indicates the mean absolute error;
[0051] Next, to further evaluate the application effect of the neural network assembly surrogate model in the actual assembly of optical systems, the wavefront Zernike coefficients after adjusting the optical system by predicting the misalignment were used to calculate the root mean square (RMS) value of the wavefront, and the difference between the wavefront RMS values before and after assembly was compared. The formula for calculating the RMS value using the Zernike coefficients is as follows:
[0052] ,
[0053] in, For the first Zernike coefficients of wavefronts, The order of the polynomial used. This indicates the Zernike coefficient number of the Zernike wavefront.
[0054] In the above technical solutions,
[0055] Step 41 specifically includes:
[0056] Step 411: Preprocessing of wavefront diagrams;
[0057] To ensure the stability of the training process, the pixel values of the wavefront image are normalized and the channels are standardized. The formulas for pixel value normalization and channel standardization are as follows:
[0058] ,
[0059] ,
[0060] in, This indicates the original wavefront diagram. Pixel values after channel normalization and Represents pixel coordinates. Indicates the original wavefront plot in coordinates The first Channel pixel value, Channels representing the three colors: red, green, and blue. For the channel index of the image, Indicates wavefront diagram Pixel values after channel normalization and The first in the dataset The mean and standard deviation of each channel;
[0061] Step 412: Preprocessing of wavefront Zernike coefficients;
[0062] Based on the sensitivity analysis results, Zernike polynomial components that significantly influence the imaging quality of the wavefront system are selected as network inputs to reduce feature dimensionality and redundant information. Z-score normalization transforms all wavefront Zernike coefficients into data with a mean of 0 and a standard deviation of 1, eliminating dimensional differences between coefficients of different orders and ensuring all features have similar scales and distributions. This provides a more stable gradient path for the neural network optimizer and reduces the learning difficulty of the neural network surrogate model. For each Zernike term coefficient... The standardized formula used is as follows:
[0063] ,
[0064] in, This is the sample mean of the coefficient. Standard deviation, Representative coefficient Standardization processing is required. Indicates the coefficient number of the term.
[0065] In the above technical solutions,
[0066] Step 42 specifically includes:
[0067] Step 421, Dual-branch feature extraction module;
[0068] Wavefront plotting branch: Focuses on the texture and color distribution in image space, capturing local perturbations and distortions of the wavefront; this branch uses wavefront color images. As input, the wavefront map reflects the spatial distribution and phase changes of the optical system's wavefront, containing rich local texture and color information, revealing subtle perturbations in the wavefront. The wavefront map branch network structure employs a multi-layer convolutional neural network to extract local spatial features, with each convolutional operation consisting of a filter. and bias Complete. The ReLU activation function is used to ensure the capture of nonlinear information. The specific formula is as follows:
[0069] ,
[0070] in, Indicates the first Feature map of the layer Indicates the first Layer, First Feature map of each channel It is the convolution kernel for the corresponding channel. Indicates the layer index of the network. Indicates the channel number. This represents the total number of layers in a convolutional neural network;
[0071] The output of the last layer of a convolutional neural network is a three-dimensional feature map. , This indicates that the output is a three-dimensional real tensor. This represents the height of the feature map and the corresponding number of rows in the output. This represents the width of the feature map and the corresponding number of columns in the output. This represents the number of channels. After flattening, it is converted into a one-dimensional vector. , Indicates the flattening operation. Represents a vector space, dimension This is equal to the total number of elements in the feature map. This one-dimensional vector contains high-order spatial features of the wavefront image, reflecting complex local and global optical structure information; Wavefront Zernike coefficient branch: Fitting the Zernike polynomial coefficient vector to the wavefront of the optical system. For input, the number of coefficients It typically covers the main aberration terms from low to medium order. Indicates transpose. express A real vector space; the wavefront Zernike coefficient branch network structure employs a multi-layer fully connected network for nonlinear feature mapping, with each layer using a weight matrix. and bias vector To implement affine transformation and achieve nonlinear operation using the ReLU activation function, the specific formula is as follows:
[0072] ,
[0073] ,
[0074] Finally, the deep feature vector is output. As a higher-order representation of this branch;
[0075] in, This represents the output vector of the first hidden layer. This represents the weight matrix of the first layer. This represents the input vector of the current layer. This represents the bias vector of the first layer. Indicates the first The output vector of the hidden layer. Indicates the first The output vector of the hidden layer. This represents the index of the hidden layer, with a value range of 1. , This represents the total number of layers in a fully connected network. The dimension is The real number space;
[0076] Step 422, Feature Fusion Module;
[0077] High-order eigenvectors from the wavefront plot branch and the wavefront Zernike coefficient branch and Effective integration should be carried out to fully leverage the complementary advantages of the two modalities and achieve collaborative information expression.
[0078] To achieve effective fusion, the eigenvectors of the wavefront plot branch and the wavefront Zernike coefficient branch are first... and Mapped to the same dimensional space :
[0079] ,
[0080] ,
[0081] in, and These are the mapping matrices for the Zernike coefficient branch and the wavefront diagram branch, respectively. All are bias vectors; The dimension space is The real space, This indicates that the wavefront diagram branch is mapped to the dimensional space. eigenvectors, This indicates that the wavefront Zernike coefficient branch is mapped to the dimensional space. eigenvectors;
[0082] Next, a weighted summation method is used to achieve feature fusion between the wavefront plot branch and the Zernike coefficient branch:
[0083] ,
[0084] Among them, the weighting coefficient Used to balance the contributions of the wavefront plot branch and the wavefront Zernike coefficient branch. This represents the eigenvector after fusing the wavefront plot branch and the wavefront Zernike coefficient branch;
[0085] Step 423, Component Offset Prediction Module;
[0086] Based on the fused feature vector This paper utilizes a fully connected network to perform regression prediction on the offset of multiple adjustable degrees of freedom of an optical system. Through multi-layer nonlinear mapping, it extracts the complex correspondence between features and offset parameters to achieve high-precision offset estimation. The specific formula is as follows:
[0087] ,
[0088] in, The component misalignment prediction module predicts the degree-of-freedom misalignment parameters of the optical component. This is the mapping function for a multilayer perceptron network.
[0089] In the above technical solutions,
[0090] Step 5 specifically includes:
[0091] Step 51: Calibrate the optical axis reference;
[0092] Install a laser interferometer, theodolite, plane mirror, and six-degree-of-freedom adjustment frame, and place the main support structure of the optical system on an air-floating platform; continuously adjust the plane mirror using the six-degree-of-freedom adjustment frame and theodolite until the plane mirror is perpendicular to the optical axis reference plane of the main support structure; adjust the laser interferometer to align with this optical axis reference plane to achieve optical axis alignment between the laser interferometer and the main support structure;
[0093] Step 52, rough adjustment;
[0094] First, the primary mirror is assembled and adjusted: using a laser interferometer, theodolite, coordinate measuring arm, altimeter, and six-degree-of-freedom adjustment frame, the angle between the primary mirror reference plane and the plane mirror is adjusted to the ideal angle; then, the primary mirror surface shape is monitored in real time using a laser interferometer, and the primary mirror position is continuously adjusted to make the primary mirror surface shape error the same as the measured result;
[0095] Secondly, the secondary and tertiary mirrors are coarsely adjusted: the secondary and tertiary mirrors are initially placed in the design position using a six-degree-of-freedom adjustment frame. During the coarse adjustment process, the adjustment amount of each degree of freedom of the secondary and tertiary mirrors is monitored in real time using a theodolite and micrometer. The adjustment amount of each degree of freedom is compared with the design value to ensure the relative spatial position accuracy between the plane mirrors in the optical system.
[0096] Step 53, Adjust the fine packaging;
[0097] The wavefront of the optical system is detected by self-collimation method. Wavefront detection at multiple field positions is achieved by deflecting plane mirrors and laser interferometers. A trained neural network assembly and adjustment surrogate model is used to calculate the predicted adjustment amount of each degree of freedom of the optical element by using the wavefront detection results of multiple fields of view measured by the system, thereby guiding the assembly and adjustment of the secondary and tertiary mirrors.
[0098] During the fine-tuning process, after each measurement, the wavefront maps of multiple fields of view and the fitted Zernike coefficients are input into the neural network surrogate model to calculate the suggested adjustment amount for each component's degree of freedom. Then, the component pose is adjusted according to the suggested adjustment amount using a six-degree-of-freedom adjustment frame. After adjustment, the wavefront of the optical system is measured again. Through multiple iterations of the above process, the wavefront aberration of the actual optical system is made close to that of the optical system after considering various errors, ensuring that all optical performance indicators are within acceptable range, thus completing the fine-tuning process.
[0099] The present invention has the following beneficial effects:
[0100] The optical system assembly and adjustment method based on a dual-branch neural network model of the present invention considers the noise influence in the actual assembly and adjustment scenario in the training dataset of the neural network model, thereby enhancing the noise resistance and robustness of the neural network model and improving the applicability of applying the neural network model to real optical assembly and adjustment.
[0101] The optical system assembly and adjustment method based on a dual-branch neural network model of the present invention adopts a dual-branch neural network model in model training. By fusing the features of two modal data, namely the system wavefront diagram and Zernike coefficients, it fully mines and utilizes the information of the optical system, thereby improving the accuracy of component misalignment calculation and the stability of the assembly and adjustment process.
[0102] The optical system assembly and adjustment method based on the dual-branch neural network model of the present invention establishes a nonlinear mapping relationship between component misalignment and optical system performance using an artificial neural network model. This method can fully explore the essential laws of the influence of assembly errors, effectively reduce the dependence on manual assembly experience in traditional assembly and adjustment methods, and improve the efficiency of optical assembly and adjustment while improving assembly and adjustment accuracy.
[0103] The optical system assembly and adjustment method based on the dual-branch neural network model of the present invention can reduce the frequency of physical experiments through data-driven virtual assembly and adjustment pre-simulation, effectively reducing resource consumption and costs. Attached Figure Description
[0104] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0105] Figure 1 This is a flowchart illustrating the optical system assembly and adjustment method based on a dual-branch neural network model according to the present invention.
[0106] Figure 2 The diagram shows the structure of an off-axis three-mirror optical system, where (a) is a planar diagram and (b) is a three-dimensional diagram.
[0107] Figure 3 A schematic diagram illustrating the process of generating the training dataset.
[0108] Figure 4 This is a schematic diagram of a two-branch neural network model architecture.
[0109] Figure 5 This is a schematic diagram of the assembly and adjustment of an off-axis three-mirror optical system.
[0110] The reference numerals in the figure are:
[0111] M1 - Primary lens; M2 - Secondary lens; M3 - Third lens. Detailed Implementation
[0112] The present invention will now be described in detail with reference to the accompanying drawings.
[0113] Figure 1This is a flowchart illustrating the optical system assembly and adjustment method based on a two-branch neural network model of the present invention. It shows the overall flowchart of optical assembly and adjustment using an artificial neural network method, which mainly includes: constructing a theoretical optical system model, analyzing component misalignment sensitivity, determining the component misalignment range and parameters for neural network model training, analyzing noise in the actual assembly and adjustment scenario, generating a neural network model training dataset, constructing and training the neural network model, and using the trained neural network model to guide the actual optical system assembly and adjustment (the neural network model of the present invention is a two-branch neural network model, and the neural network assembly and adjustment proxy model mentioned in the following description refers to the two-branch neural network model).
[0114] The optical system assembly and adjustment method based on a dual-branch neural network model of the present invention accurately calculates the component misalignment by fusing features from two modal data: the wavefront diagram and the wavefront Zernike coefficients of the optical system. Furthermore, it adds optical component surface shape errors, assembly and adjustment mechanism adjustment errors, and optical system wavefront detection (measurement) errors to the training data to enhance the robustness of the neural network model in practical applications. The optical system sensitivity matrix is added to the loss function of the neural network model training as physical information guidance, and the optical system wavefront RMS is additionally used as a physical evaluation index in the evaluation index for neural network model validation, further improving the accuracy and robustness of the neural network model and enhancing its practical application capabilities.
[0115] The optical system assembly and adjustment method based on a dual-branch neural network model of the present invention includes the following steps:
[0116] Step 1: Component misalignment sensitivity analysis;
[0117] Construct a theoretical optical system model and calculate the offset sensitivity matrix of the optical system.
[0118] Based on the actual assembly and adjustment conditions, the misalignment range of the six pose degrees of freedom for each component of the optical system is determined. Optical software is then used to calculate the wavefront evaluation indices (e.g., wavefront Zernike coefficient, RMS value, MTF, distortion, etc.) corresponding to changes in individual components and individual degrees of freedom within the misalignment range. By analyzing the response of these evaluation indices to changes in misalignment, the impact of each degree of freedom on the imaging quality of the optical system is assessed. This provides a physical basis and prior knowledge for the subsequent establishment of a neural network assembly and adjustment surrogate model, guiding the selection of network input parameters and the setting of the loss function, thereby improving the efficiency and physical rationality of the neural network assembly and adjustment surrogate model training.
[0119] Step 2: Noise analysis of the actual assembly and adjustment scenario;
[0120] The assembly and adjustment of high-precision optical systems have extremely stringent environmental requirements, necessitating multi-dimensional control to ensure the stability of environmental parameters such as temperature, vibration, cleanliness, and humidity. However, in addition to environmental disturbances, several types of noise remain that are difficult to eliminate through control. These mainly include: machining and measurement errors of optical component surfaces, wavefront detection errors of the optical system, and adjustment errors of the assembly and adjustment mechanical actuators. It is necessary to combine methods such as end-to-end error decomposition of measurement equipment and experimental data analysis to obtain the distribution patterns of noise, and add the noise distribution to the training dataset to train an offset prediction model that is unaffected by noise in the actual assembly and adjustment environment.
[0121] Step 3: Generate the neural network model training dataset;
[0122] First, the processing and measurement errors of the optical element surface shape are added to the theoretical optical system model, and the optical system model affected by the actual element surface shape errors is used as the reference model for assembly and adjustment regression. Then, using optical simulation software and a self-developed program, a large set of random misalignment data pairs for the elements is generated in batches within the determined misalignment range, and the corresponding wavefront distribution map of the optical system and its fitted wavefront Zernike coefficients are calculated. Finally, the adjustment noise (i.e., adjustment error) of the assembly and adjustment mechanical actuator obtained from the analysis is added to the random misalignment combination data according to its distribution law, and the wavefront detection error of the optical system is added to the simulated wavefront data according to its distribution law. This results in a set of "misalignment combination and system wavefront aberration" data pairs that considers the noise effect and better reflects the actual assembly and adjustment environment, which is used for training the neural network assembly and adjustment surrogate model. See [link to relevant documentation]. Figure 3 .
[0123] Step 4: Construction and training of the neural network model (i.e., the neural network assembly agent model);
[0124] A dual-branch neural network model is proposed to acquire information and fuse features from two modalities, training a high-precision component misalignment state estimation model with noise resistance. First, the input data from both modalities is processed through a dual-branch feature extraction module. A multilayer fully connected network (MLP) is used to extract deep semantic information from the wavefront Zernike coefficients, and a multilayer convolutional neural network (CNN) is used to extract local spatial features from the wavefront image. To improve the physical effectiveness of the input features, the input wavefront Zernike coefficients are filtered based on sensitivity analysis results, retaining only coefficients that significantly affect the imaging quality of the optical system and suppressing noise-dominated redundant features.
[0125] Then, a feature fusion module effectively integrates high-dimensional feature vectors from wavefront diagrams and wavefront Zernike coefficient branches to achieve collaborative information expression and enhance feature discrimination capabilities, providing a more comprehensive and robust input for downstream component misalignment estimation. In the misalignment state prediction module, a fully connected neural network (FCN) is used to regress and predict the misalignment states of multiple pose degrees of freedom in the optical system using the fused feature vectors. During training, physical prior knowledge based on the sensitivity matrix is introduced, and a weighted error term for misalignment is constructed to enhance the model's accuracy in predicting key degrees of freedom, ensuring that the prediction results better match the actual system response characteristics. Finally, a neural network assembly proxy model with physical interpretation capabilities is constructed, which can be used to guide efficient assembly and error compensation in complex optical systems.
[0126] Step 5: Use neural network models to guide actual assembly and adjustment;
[0127] The actual optical assembly and adjustment process requires three steps: optical axis reference calibration, coarse assembly and adjustment, and fine assembly and adjustment. The coarse assembly and adjustment process aims to place each optical component of the optical system into its ideal design state, ensuring that the wavefront aberration of the actual optical system can be measured. This ensures that the misalignment of each degree of freedom of the components is controlled within the applicable range of the neural network assembly and adjustment surrogate model, providing a good iterative starting point for the fine assembly and adjustment process. The fine assembly and adjustment process involves inputting measured wavefront data into a trained neural network assembly and adjustment surrogate model (misalignment state estimation model) to calculate the predicted misalignment of each degree of freedom. Then, the assembly and adjustment mechanical actuator adjusts the pose degrees of freedom of the corresponding components according to the predicted misalignment. This process is iterated repeatedly until the measured performance indicators of the optical system meet the design requirements.
[0128] The following will take the assembly and adjustment process of an off-axis three-mirror optical system as an example, and further describe the technical solution flow of the present invention in conjunction with the accompanying drawings.
[0129] Step 1: Component misalignment sensitivity analysis;
[0130] Construct a theoretical optical system model and calculate the offset sensitivity matrix of the optical system. Specifically, this includes:
[0131] Step 11: Establish the designed theoretical optical system model in optical analysis software (e.g., Figure 2 As shown, Figure 2 The optical path diagram of the transmission path of multiple light rays in an off-axis three-mirror optical system between various mirrors is shown. (a) is a planar schematic diagram and (b) is a three-dimensional schematic diagram. In the figure, the secondary mirror M2 is the aperture stop and the image plane is the image plane. The misalignment range of each orientation degree of freedom of the component is determined based on actual assembly and adjustment experience.
[0132] Step 12: Use MATLAB+CODEV software (or other programming language+optical analysis software) to build an integrated simulation analysis link, calculate the curves of wave aberration evaluation indicators (such as MTF, RMS value, wavefront Zernike coefficient, etc.) corresponding to each optical element and single position degree of freedom in the optical system as a function of the misalignment, and obtain the misalignment sensitivity matrix of the optical system by curve fitting.
[0133] Step 2: Noise analysis of the actual assembly and adjustment scenario;
[0134] Based on the actual assembly and adjustment environment, analyze and determine the noise and its distribution characteristics during the assembly and adjustment process of the optical system.
[0135] The assembly and adjustment of high-precision optical systems have extremely stringent environmental requirements. Besides environmental disturbances, several types of noise remain that are difficult to eliminate through control. It is necessary to analyze and determine the distribution characteristics of this noise and add it to the training data of the neural network assembly and adjustment proxy model to mitigate its impact on assembly and adjustment accuracy. Specifically, these include:
[0136] Step 21: Machining and measurement errors of the optical element surface shape;
[0137] Manufacturing errors occur during the fabrication of optical components due to factors such as the precision of the processing equipment (e.g., grinding machines, polishing machines), tool accuracy, material properties, process parameter control, and environmental factors (e.g., temperature fluctuations, clamping deformation), resulting in deviations between the finished optical component's surface shape and the designed ideal surface shape. Measurement errors occur during the actual surface shape inspection of optical components due to factors such as the precision of the inspection equipment (interferometers, profilometers, etc.), environmental interference, the operation and installation of the measuring fixture, and post-inspection data processing and algorithms, resulting in errors between the actual surface shape of the component and the inspection results. Based on the statistical distribution patterns of errors under actual manufacturing and measurement conditions, the specific distribution characteristics and ranges of the aforementioned two types of errors are obtained and converted into expressions using wavefront Zernike coefficients (or other forms). Then, the two types of errors are superimposed to comprehensively represent the surface shape error of the optical component after actual manufacturing.
[0138] Step 22: Wavefront detection error of the optical system;
[0139] When using a laser interferometer to measure the wavefront of an optical system, factors such as environmental disturbances, laser source fluctuations, electronic signal processing, and wavefront fitting deviations can cause discrepancies between the wavefront detection results and the actual wavefront. Through empirical rules and analysis of measured data, the statistical distribution law of the error of the wavefront measuring instrument is obtained and converted into the form of wavefront Zernike coefficients.
[0140] Step 23: Adjust the mechanical actuator to correct any errors;
[0141] In the actual assembly and adjustment of optical systems, mechanical actuators are used to control and adjust the pose of optical components. Due to factors such as environmental disturbances, mechanical manufacturing and assembly defects, material deformation, and motor drive and control errors, there is a deviation between the nominal adjustment amount and the actual movement amount of the components. Through empirical judgment and analysis of measured data, the system cumulative error and random noise distribution law and range of the assembly and adjustment mechanical actuators are determined and transformed into a six-degree-of-freedom expression of the component pose.
[0142] Step 3: Generate the neural network model training dataset;
[0143] Generate a large dataset for training neural network agent models. Specifically, this includes:
[0144] Step 31: Determine the input for training the neural network assembly agent model as: the wavefront distribution diagram of the optical system and the 4th to 16th terms (Z4-Z16) of its fitted Zernike coefficients; the output is: the degree of freedom misalignment of the component pose that can be adjusted during the actual assembly process of the optical system (for an off-axis three-mirror optical system, this is: the eccentric degrees of freedom in the X, Y and Z directions of the secondary mirror M2 and the third mirror M3, as well as the rotational degrees of freedom in the X and Y directions).
[0145] Step 32: Using MATLAB + CODE V software (or other programming languages + optical analysis software), based on the misalignment ranges of each pose degree of freedom of the components determined in Step 1 and the distribution patterns and ranges of various noises determined in Step 2, batch generate a set of data pairs consisting of "optical system wavefront diagram + Zernike coefficient ~ misalignment of each pose degree of freedom of the components" for training the neural network assembly surrogate model, taking into account the influence of various actual noises. The training dataset generation process is as follows: Figure 3 As shown ( Figure 3 This paper demonstrates the process of generating the training dataset for a neural network assembly proxy model used in optical assembly. To obtain a noise-resistant neural network assembly proxy model suitable for practical optical assembly scenarios, the training dataset considers component surface errors (i.e., manufacturing and measurement errors of component surface shapes), adjustment errors of assembly mechanical actuators, and wavefront detection errors of the optical system. The data pairs used for training are: (a combination of random misalignment of components, system wavefront aberrations)). The specific methods for adding various types of noise are as follows:
[0146] First, the component surface error, expressed in Zernike coefficient form (or other form) as determined in step 21, is added to the theoretical optical design model, and this optical model is used as the "ideal model" or reference model in the subsequent data generation and assembly iteration regression process.
[0147] Next, when generating each set of randomly uniformly distributed misalignment combinations, according to the noise distribution pattern and range of the assembly and adjustment mechanical actuator determined in step 23, the noise of the assembly and adjustment mechanical actuator is added to each adjustable element position degree of freedom, and then the wavefront diagram of the optical system and its Zernike coefficient corresponding to the random misalignment combination after adding the noise are calculated.
[0148] Finally, the wavefront measurement noise of the optical system is added to the calculated wavefront aberration data of the optical system (i.e., the wavefront map of the optical system and its Zernike coefficients). Specifically, the wavefront measurement noise (i.e. detection error) of the optical system determined in step 22 and expressed in the form of Zernike coefficients is directly added to the wavefront Zernike coefficients calculated in the previous step, and the noise distribution is converted into an image and superimposed on the wavefront map.
[0149] Step 4: Neural network model construction and training;
[0150] Construct a neural network assembly agent model and train it. Specifically, this includes:
[0151] Step 41: Data preprocessing;
[0152] A dual-branch neural network model is proposed to achieve information acquisition and feature fusion of two modalities. By normalizing and standardizing the input wavefront image, its fitted Zernike coefficients, and the output component pose misalignment, the model meets the neural network's requirements for numerical scale and distribution consistency, thereby improving the stability and accuracy of the neural network assembly surrogate model training. Specifically, this includes:
[0153] Step 411: Preprocessing of wavefront diagrams;
[0154] Wavefront images are typically color images that map two-dimensional wavefront phase or optical path difference data using color mapping. Colors represent phase values and gradients, encoding the spatial distribution and structural information of the wavefront, and visually reflecting its spatial variations. This invention aims to normalize the pixel values and standardize the channels of the wavefront image to ensure the stability of the training process. The formulas for pixel value normalization and channel standardization are as follows:
[0155] ,
[0156] ,
[0157] in, This indicates the original wavefront diagram. Pixel values after channel normalization and Represents pixel coordinates. Indicates the original wavefront plot in coordinates The first Channel pixel value, Channels representing the three colors: red, green, and blue. For the channel index of the image, Indicates wavefront diagram Pixel values after channel normalization and The first in the dataset The mean and standard deviation of each channel.
[0158] Step 412: Preprocessing of wavefront Zernike coefficients;
[0159] Zernike polynomials are a set of ordered orthogonal polynomials that can be used to represent wavefronts. Their coefficients are commonly used to describe the magnitude of various aberrations (such as spherical aberration, coma, astigmatism, etc.). Based on sensitivity analysis results, this invention prioritizes Zernike polynomial components that significantly influence the imaging quality of optical systems (such as wavefront RMS, Zernike coefficients, etc.) as network inputs, thereby reducing feature dimensionality and redundant information. Furthermore, since the physical meanings represented by each coefficient are different, their numerical distributions and magnitudes are inconsistent for a specific optical system. Z-score normalization transforms all wavefront Zernike coefficients into data with a mean of 0 and a standard deviation of 1, eliminating dimensional differences between coefficients of different orders. This ensures that all features have similar scales and distributions, providing a more stable gradient path for the neural network optimizer and reducing the learning difficulty of the bi-branch neural network model. For each Zernike coefficient term... The standardized formula used is as follows:
[0160] ,
[0161] in, Coefficient of the term The sample mean, Standard deviation, Represents the Zernike coefficient for each term. Standardization processing is required. Representative coefficient number.
[0162] Step 42, Model Structure Design;
[0163] This invention designs a dual-branch neural network model, also known as a neural network assembly proxy model, to fuse wavefront maps and wavefront Zernike coefficients—two key data points describing wavefront aberrations in optical systems—to jointly predict the multi-degree-of-freedom misalignment of the optical system. This dual-branch neural network model consists of two parallel branches, each processing different modal data, and outputs accurate misalignment values through feature fusion. The dual-branch neural network model architecture is divided into three parts: a dual-branch feature extraction module, a feature fusion module, and a component misalignment prediction module, as follows: Figure 4 As shown, Figure 4 The diagram illustrates the architecture of a two-branch neural network model, primarily comprising: a two-branch feature extraction module, a feature fusion module, and a component offset prediction module. The feature extraction module extracts features from both the wavefront image branch and the wavefront Zernike coefficient branch. The feature fusion module then fuses these extracted features to fully leverage optical system information. Finally, the component offset prediction module achieves high-precision prediction of component offset. Specifically, it includes:
[0164] Step 421, Dual-branch feature extraction module;
[0165] This module is designed with two independent branches, processing wavefront diagrams and wavefront Zernike coefficients, respectively, to extract multi-level and multi-angle features from the wavefront information of the optical system. This design effectively mines key features from multi-source data by identifying the physical meaning and deep correlations of different optical performance indicators, thereby improving the model's expressive power. The two branches output high-order feature vectors. and , representing the deep semantic information and local spatial features of the two branch input data, respectively.
[0166] Wavefront plot branch: Focuses on the texture and color distribution in image space, capturing local perturbations and distortions of the wavefront, compensating for the insufficient representation of local information by the wavefront Zernike coefficient branch. The wavefront plot branch uses wavefront color images... As input, the wavefront map reflects the spatial distribution and phase changes of the optical system's wavefront, containing rich local texture and color information, revealing subtle perturbations in the wavefront. The wavefront map branch network structure employs a multi-layer convolutional neural network (CNN) to extract spatial local features, with each convolutional operation consisting of filters. and bias Complete. The ReLU activation function is used to ensure the capture of nonlinear information. The specific formula is as follows:
[0167] ,
[0168] in, Indicates the first Feature map of the layer Indicates the first Layer, First Feature map of each channel It is the convolution kernel for the corresponding channel. Indicates the layer index of the network. Indicates the channel number. This represents the total number of layers in the convolutional neural network.
[0169] The output of the last layer of a convolutional neural network is a three-dimensional feature map. , This indicates that the output is a three-dimensional real tensor. This represents the height of the feature map, corresponding to the number of rows in the output. The width of the feature map corresponds to the number of columns in the output. The number of channels is converted into a one-dimensional vector after flattening. , Indicates the flattening operation. Represents a vector space, dimension It equals the total number of all elements in the feature map. This one-dimensional vector contains high-order spatial features of the wavefront image, reflecting complex local and global optical structure information.
[0170] Zernike coefficient branching of wavefront: fitting Zernike polynomial coefficient vectors to the wavefront of the optical system For input, the number of coefficients Typically covers the main aberration terms from low to medium order (selecting terms Z4 to Z16). Indicates transpose, indicates It is a by A column vector consisting of coefficients express A real vector space. The wavefront Zernike coefficient branch network structure uses a multi-layer fully connected network (MLP) for nonlinear feature mapping, with each layer using a weight matrix. and bias vector To implement affine transformation and achieve nonlinear operation using the ReLU activation function, the specific formula is as follows:
[0171] ,
[0172] ,
[0173] Finally, the deep feature vector is output. As a higher-order representation of this branch, it can comprehensively express the global characteristics of the wavefront and contains rich physical meaning. This represents the output vector of the first hidden layer. This represents the weight matrix of the first layer. This represents the input vector of the current layer. This represents the bias vector of the first layer. Indicates the first The output vector of the hidden layer. Indicates the first The output vector of the hidden layer. This represents the index of the hidden layer, with a value range of 1. , This represents the total number of layers in a fully connected network. The dimension is The real number space;
[0174] Step 422, Feature Fusion Module;
[0175] The feature fusion module combines high-dimensional feature vectors from the wavefront diagram branch and the wavefront Zernike coefficient branch. and Effective integration fully leverages the complementary advantages of the two modalities to achieve synergistic information expression. The fusion process not only enhances the discriminative power of features but also provides a more comprehensive and robust input for downstream component degree-of-freedom misalignment prediction.
[0176] To achieve effective fusion, the higher-order feature vectors of the two branches are first... and Mapped to the same dimensional space :
[0177] ,
[0178] ,
[0179] in, and These are the mapping matrices for the Zernike coefficient branch and the wavefront diagram branch, respectively. For bias vectors, The dimension is The real space, This indicates that the wavefront diagram branch is mapped to the dimensional space. eigenvectors, This indicates that the wavefront Zernike coefficient branch is mapped to the dimensional space. The feature vectors of the two branches are then fused using a weighted summation method.
[0180] ,
[0181] Among them, the weighting coefficient Used to balance the contributions of the two branches, This represents the eigenvector resulting from the fusion of the wavefront plot branch and the wavefront Zernike coefficient branch.
[0182] Step 423, Component Offset Prediction Module;
[0183] Based on the fused feature vector This module utilizes a multi-layer perceptron (MLP) to regress and predict the offset of multiple adjustable degrees of freedom in an optical system. Through multi-layer nonlinear mapping, it extracts the complex correspondence between features and offset parameters, achieving high-precision offset estimation. The specific formula is as follows:
[0184] ,
[0185] in, The component misalignment prediction module predicts the degree-of-freedom misalignment parameters of the optical component. This is the mapping function for a multilayer perceptron network.
[0186] Step 43: Model training;
[0187] By leveraging large-scale labeled data, the parameters of three modules—the dual-branch feature extraction module, the feature fusion module, and the component misalignment prediction module—are optimized to achieve high-precision prediction of optical component misalignment. The specific training process is based on a training set containing input data (wavefront plots and wavefront Zernike coefficients) and corresponding multi-degree-of-freedom misalignment labels. The mean squared error loss function is used to optimize the model parameters, and the parameters are iteratively updated using a mini-batch gradient descent algorithm (such as the Adam optimizer). Regularization and early stopping mechanisms are combined to effectively prevent overfitting, ultimately achieving accurate estimation of the optical component misalignment state. To guide the network to focus more on physically sensitive degrees of freedom, the loss function incorporates weight factors based on the sensitivity matrix. The training objective can be expressed as:
[0188] ,
[0189] in, The diagonal weighted matrix, obtained from sensitivity analysis, is used to emphasize the misalignment degrees of freedom that have a more significant impact on the performance of the optical system, thereby improving the prediction accuracy in key directions. For model parameters, and The first The true imbalance and the predicted imbalance of a sample Indicates the model training objective. Indicates the sample number. This indicates the total number of samples.
[0190] Step 44: Model Validation;
[0191] To comprehensively evaluate the effectiveness of the designed dual-branch neural network model in optical system assembly and adjustment tasks, multiple performance metrics were used to validate the model's performance on the test set. The validation process included not only conventional regression evaluation metrics but also introduced the RMS value, used to measure the wavefront quality of the optical system, as a physical evaluation metric to verify the model's actual adjustment effect in a physical sense.
[0192] First, from the perspective of prediction accuracy, the mean squared error (MSE) and mean absolute error (MAE) metrics are used to evaluate the difference between the model's output offset and the true value on the test set:
[0193] ,
[0194] ,
[0195] in, To predict the misalignment in the model, For the true degree of disorder, The total number of test samples, This is the test sample number.
[0196] Next, to further evaluate the application effect of the dual-branch neural network model in the actual optical system assembly and adjustment, the wavefront Zernike coefficients after adjusting the system by predicting the misalignment were used to calculate the root mean square (RMS) value of the wavefront, and the difference between the wavefront RMS values before and after assembly and adjustment was compared. The formula for calculating the RMS value using the Zernike coefficients is as follows:
[0197] ,
[0198] in, For the first Zernike coefficients of wavefronts, The order of the polynomial used (take) ), This indicates the wavefront Zernike coefficient number.
[0199] Step 5: Use neural network models to guide actual assembly and adjustment;
[0200] A neural network assembly proxy model is used to guide the assembly and adjustment of actual optical systems.
[0201] Taking an off-axis three-mirror optical system as an example, this describes the process of using a neural network assembly and adjustment surrogate model to guide the actual assembly and adjustment. Figure 5 As shown ( Figure 5A schematic diagram of the assembly and adjustment of an off-axis three-mirror optical system is shown. During assembly and adjustment, the primary mirror M1 is used as a reference. The poses of the secondary mirror M2 and the third mirror M3 are continuously adjusted through a high-precision six-degree-of-freedom displacement stage and a neural network assembly and adjustment proxy model. A laser interferometer is used to monitor the wavefront of the system in real time, so that the wavefront of the optical system meets the usage requirements. Figure 5 In this context, XDE, YDE, and ZDE represent the translational magnitudes of the optical system along the X, Y, and Z directions, respectively, while ADE, BDE, and CDE represent the rotational magnitudes of the optical system along the X, Y, and Z directions, respectively. Specifically, this includes:
[0202] Step 51: Calibrate the optical axis reference;
[0203] Install the laser interferometer, theodolite, plane mirror, and six-degree-of-freedom adjustment frame (i.e., six-degree-of-freedom displacement stage), and place the main support structure of the optical system on the air-bearing platform. Continuously adjust the plane mirror using the six-degree-of-freedom adjustment frame and theodolite until it is perpendicular to the optical axis reference plane of the main support structure. Adjust the laser interferometer to align it with this optical axis reference plane, achieving optical axis alignment between the laser interferometer and the main support structure.
[0204] Step 52, rough adjustment;
[0205] First, the primary mirror M1 is assembled and adjusted: using tools such as a laser interferometer, theodolite, coordinate measuring arm, altimeter, and six-degree-of-freedom adjustment frame, the angle between the reference plane of the primary mirror M1 and the plane mirror is adjusted to the ideal angle; then, the surface shape of the primary mirror M1 is monitored in real time using a laser interferometer, and the position and pose of the primary mirror M1 are continuously adjusted to make its surface shape error the same as the measured result.
[0206] Secondly, coarse adjustment of secondary mirror M2 and tertiary mirror M3 is performed: using a six-degree-of-freedom adjustment frame, secondary mirror M2 and tertiary mirror M3 are initially placed in their designed positions. During the coarse adjustment process, a theodolite and micrometer are used to monitor the adjustment of each degree of freedom of secondary mirror M2 and tertiary mirror M3 in real time, and compare them with the design values to ensure the relative spatial position accuracy between the plane mirrors in the optical system (it is required that the error of each adjustment dimension of the components after coarse adjustment is within the applicable range of the neural network adjustment proxy model).
[0207] Step 53, Adjust the fine packaging;
[0208] The wavefront of the optical system is detected using a self-collimation method, and wavefront detection at multiple field-of-view positions is achieved through a deflecting plane mirror and a laser interferometer. A trained neural network assembly surrogate model is used to calculate the predicted adjustment amounts for each degree of freedom of the optical elements based on the measured wavefront detection results from multiple fields of view of the optical system, thereby guiding the assembly and adjustment of the secondary mirror M2 and the tertiary mirror M3.
[0209] During the fine-tuning process, after each measurement, wavefront maps of multiple fields of view and fitted Zernike coefficients are input into the neural network surrogate model to calculate the suggested adjustment amounts for each component's degrees of freedom. Then, the component poses are adjusted according to the suggested adjustment amounts using a six-DOF adjustment frame. After adjustment, the system wavefront is measured again. Through multiple iterations of the above process, the wavefront aberration of the actual optical system is made as close as possible to the wavefront aberration of the optical system after considering various errors, ensuring that all optical performance indicators are within acceptable ranges, at which point the fine-tuning is considered complete.
[0210] The optical system assembly and adjustment method based on a dual-branch neural network model of the present invention considers the noise influence in the actual assembly and adjustment scenario in the training dataset of the neural network model, thereby enhancing the noise resistance and robustness of the neural network model and improving the applicability of applying the neural network model to real optical assembly and adjustment.
[0211] The optical system assembly and adjustment method based on a dual-branch neural network model of the present invention adopts a dual-branch neural network model in model training. By fusing the features of two modal data, namely the system wavefront diagram and Zernike coefficients, it fully mines and utilizes the information of the optical system, thereby improving the accuracy of component misalignment calculation and the stability of the assembly and adjustment process.
[0212] The optical system assembly and adjustment method based on the dual-branch neural network model of the present invention establishes a nonlinear mapping relationship between component misalignment and optical system performance using an artificial neural network model. This method can fully explore the essential laws of the influence of assembly errors, effectively reduce the dependence on manual assembly experience in traditional assembly and adjustment methods, and improve the efficiency of optical assembly and adjustment while improving assembly and adjustment accuracy.
[0213] The optical system assembly and adjustment method based on the dual-branch neural network model of the present invention can reduce the frequency of physical experiments through data-driven virtual assembly and adjustment pre-simulation, effectively reducing resource consumption and costs.
[0214] 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. An optical system assembly and adjustment method based on a dual-branch neural network model, characterized in that, Includes the following steps: Step 1: Component misalignment sensitivity analysis; Construct a theoretical optical system model and calculate the offset sensitivity matrix of the optical system. Step 2: Noise analysis of the actual assembly and adjustment scenario; Based on the actual assembly and adjustment environment, analyze and determine the noise and its distribution characteristics during the assembly and adjustment process of the optical system; Step 3: Generate the neural network model training dataset; Generate a large dataset for training neural network assembly agent models; Step 4: Neural network model construction and training; Construct a two-branch neural network assembly agent model and train the neural network assembly agent model; Step 5: Use neural network models to guide actual assembly and adjustment; Using a neural network assembly proxy model to guide the assembly and adjustment of actual optical systems; Step 3 specifically includes: Step 31: Determine the inputs and outputs for training the neural network assembly agent model; Step 32: Based on the misalignment range of each pose degree of freedom of the component determined in Step 1 and the distribution law and range of various noises determined in Step 2, batch generate a set of data pairs of misalignment of each pose degree of freedom of the component for training the neural network assembly agent model and taking into account the influence of various actual noises. The specific methods for adding various noises in step 32 are as follows: First, after adding the surface shape error of optical components to the theoretical optical design model, the model is used as the ideal model in the subsequent data generation and assembly adjustment iteration regression process. Next, when generating each set of randomly uniformly distributed misalignment combinations, according to the noise distribution pattern and range of the assembly and adjustment mechanical actuator, the noise of the assembly and adjustment mechanical actuator is added to each adjustable element position degree of freedom, and then the wavefront diagram of the optical system and its Zernike coefficient corresponding to the random misalignment combination after adding the noise are calculated. Finally, wavefront detection noise is added to the calculated optical system wavefront aberration data; The wavefront detection noise of the optical system, expressed in Zernike coefficient form, is directly added to the wavefront Zernike coefficients calculated in the previous step, and the noise distribution is converted into an image and superimposed on the wavefront map.
2. The optical system assembly and adjustment method based on a dual-branch neural network model according to claim 1, characterized in that, Step 1 specifically includes: Step 11: Establish a theoretical optical system model in optical analysis software; Step 12: Build an integrated simulation analysis link, calculate the curves of wave aberration evaluation index corresponding to each optical element and single position degree of freedom in the optical system as a function of misalignment, and obtain the misalignment sensitivity matrix of the optical system through curve fitting.
3. The optical system assembly and adjustment method based on a dual-branch neural network model according to claim 2, characterized in that, Step 2 specifically includes: Step 21: Machining and measurement errors of the optical element surface shape; Based on the statistical distribution law of errors under actual processing and measurement conditions, the specific distribution characteristics and range of processing errors and measurement errors are obtained and converted into the form of wavefront Zernike coefficients. Then, the processing errors and measurement errors are superimposed to comprehensively represent the surface shape error of the optical element after actual processing. Step 22: Wavefront detection error of the optical system; Through empirical rules and analysis of measured data, the statistical distribution law of the error of the wavefront measuring instrument is obtained and converted into the form of wavefront Zernike coefficients; Step 23: Adjust the mechanical actuator to correct any errors; Through empirical judgment and analysis of measured data, the distribution law and range of the system cumulative error and random noise of the mechanical actuator are determined and transformed into a six-degree-of-freedom expression of the component's position and orientation.
4. The optical system assembly and adjustment method based on a dual-branch neural network model according to claim 1, characterized in that, Step 4 specifically includes: Step 41: Data preprocessing; A neural network assembly surrogate model is proposed to achieve information acquisition and feature fusion of two modal data. By normalizing and standardizing the wavefront diagram of the input optical system and its fitted Zernike coefficients and the component pose degree of freedom misalignment of the output optical system, the input and output data can meet the requirements of the neural network assembly surrogate model for numerical scale and distribution consistency, thereby improving the stability and accuracy of the training of the neural network assembly surrogate model. Step 42, Model Structure Design; The neural network assembly agent model structure includes: a two-branch feature extraction module, a feature fusion module, and a component misalignment prediction module; Feature extraction is performed through the wavefront plot branch and wavefront Zernike coefficient branch of the dual-branch feature extraction module, and then the extracted features are fused through the feature fusion module. Finally, the component offset prediction module achieves high-precision prediction of component offset. Step 43: Model training; By using large-scale labeled data, the parameters of the dual-branch feature extraction module, feature fusion module, and component misalignment prediction module are optimized to achieve high-precision prediction of optical component misalignment. Based on the input data and the training set of corresponding multi-degree-of-freedom misalignment labels, the parameters of the neural network assembly surrogate model are optimized using the mean square error loss function. The parameters are iteratively updated using the mini-batch gradient descent algorithm, and overfitting is effectively prevented by combining regularization and early stopping mechanisms. Finally, accurate estimation of the misalignment state of optical components is achieved. Step 44: Model Validation; First, from the perspective of prediction accuracy, the mean squared error and mean absolute error metrics are used to evaluate the difference between the offset output of the neural network surrogate model and the true value on the test set: , , in, To predict the misalignment in the model, For the true degree of disorder, The total number of test samples, The test sample number, Indicates mean square error. Indicates the mean absolute error; Next, to further evaluate the application effect of the neural network assembly surrogate model in the actual assembly of optical systems, the wavefront Zernike coefficients after adjusting the optical system by predicting the misalignment were used to calculate the root mean square (RMS) value of the wavefront, and the difference between the wavefront RMS values before and after assembly was compared. The formula for calculating the RMS value using the Zernike coefficients is as follows: , in, For the first Zernike coefficients of wavefronts, The order of the polynomial used. This indicates the wavefront Zernike coefficient number.
5. The optical system assembly and adjustment method based on a dual-branch neural network model according to claim 4, characterized in that, Step 41 specifically includes: Step 411: Preprocessing of wavefront diagrams; To ensure the stability of the training process, the pixel values of the wavefront image are normalized and the channels are standardized. The formulas for pixel value normalization and channel standardization are as follows: , , in, This indicates the original wavefront diagram. Pixel values after channel normalization and Represents pixel coordinates. Indicates the original wavefront plot in coordinates The first Channel pixel value, Channels representing the three colors: red, green, and blue. For the channel index of the image, Indicates wavefront diagram Pixel values after channel normalization and The first in the dataset The mean and standard deviation of each channel; Step 412: Preprocessing of wavefront Zernike coefficients; Based on the sensitivity analysis results, Zernike polynomial components that significantly affect the imaging quality of the optical system are selected as network inputs to reduce feature dimensionality and redundant information. Z-score normalization transforms all wavefront Zernike coefficients into data with a mean of 0 and a standard deviation of 1, eliminating dimensional differences between coefficients of different orders and ensuring all features have similar scales and distributions. This provides a more stable gradient path for the neural network optimizer and reduces the learning difficulty of the neural network surrogate model. For each Zernike term coefficient... The standardized formula used is as follows: , in, This is the sample mean of the coefficient. Standard deviation, Representative coefficient Standardization processing is required. Indicates the coefficient number of the term.
6. The optical system assembly and adjustment method based on a dual-branch neural network model according to claim 4, characterized in that, Step 42 specifically includes: Step 421, Dual-branch feature extraction module; Wavefront plotting branch: Focuses on the texture and color distribution in image space, capturing local perturbations and distortions of the wavefront; this branch uses wavefront color images. As input, the wavefront map reflects the spatial distribution and phase changes of the optical system's wavefront, containing rich local texture and color information, revealing subtle perturbations in the wavefront. The wavefront map branch network structure employs a multi-layer convolutional neural network to extract local spatial features, with each convolutional operation consisting of a filter. and bias Complete. The ReLU activation function is used to ensure the capture of nonlinear information. The specific formula is as follows: , in, Indicates the first Feature map of the layer Indicates the first Layer, First Feature map of each channel It is the convolution kernel for the corresponding channel. Indicates the layer index of the network. Indicates the channel number. This represents the total number of layers in a convolutional neural network; The output of the last layer of a convolutional neural network is a three-dimensional feature map. , This indicates that the output is a three-dimensional real tensor. This represents the height of the feature map and the corresponding number of rows in the output. This represents the width of the feature map and the corresponding number of columns in the output. The number of channels is used; after flattening, it is converted into a one-dimensional vector. , Indicates the flattening operation. Represents a vector space, dimension This is equal to the total number of elements in the feature map. This one-dimensional vector contains high-order spatial features of the wavefront image, reflecting complex local and global optical structure information; Wavefront Zernike coefficient branch: Fitting the Zernike polynomial coefficient vector to the wavefront of the optical system. For input, the number of coefficients It typically covers the main aberration terms from low to medium order. Indicates transpose. express A real vector space; the wavefront Zernike coefficient branch network structure employs a multi-layer fully connected network for nonlinear feature mapping, with each layer using a weight matrix. and bias vector To implement affine transformation and achieve nonlinear operation using the ReLU activation function, the specific formula is as follows: , , Finally, the deep feature vector is output. As a higher-order representation of this branch; in, This represents the output vector of the first hidden layer. This represents the weight matrix of the first layer. This represents the input vector of the current layer. This represents the bias vector of the first layer. Indicates the first The output vector of the hidden layer. Indicates the first The output vector of the hidden layer. This represents the index of the hidden layer, with a value range of 1. , This represents the total number of layers in a fully connected network. The dimension is The real number space; Step 422, Feature Fusion Module; High-order eigenvectors from the wavefront plot branch and the wavefront Zernike coefficient branch and Effective integration should be carried out to fully leverage the complementary advantages of the two modalities and achieve collaborative information expression. To achieve effective fusion, the eigenvectors of the wavefront plot branch and the wavefront Zernike coefficient branch are first... and Mapped to the same dimensional space : , , in, and These are the mapping matrices for the Zernike coefficient branch and the wavefront diagram branch, respectively. Both are bias vectors. The dimension space is The real space, This indicates that the wavefront diagram branch is mapped to the dimensional space. eigenvectors, This indicates that the wavefront Zernike coefficient branch is mapped to the dimensional space. eigenvectors; Next, a weighted summation method is used to achieve feature fusion of the wavefront plot branch and the wavefront Zernike coefficient branch: , Among them, the weighting coefficient This is used to balance the contributions of the wavefront plot branch and the wavefront Zernike coefficient branch. This represents the eigenvector after fusing the wavefront plot branch and the wavefront Zernike coefficient branch; Step 423, Component Offset Prediction Module; Based on the fused feature vector This paper utilizes a fully connected network to perform regression prediction on the offset of multiple adjustable degrees of freedom of an optical system. Through multi-layer nonlinear mapping, it extracts the complex correspondence between features and offset parameters to achieve high-precision offset estimation. The specific formula is as follows: , in, The component misalignment prediction module predicts the degree-of-freedom misalignment parameters of the optical component. This is the mapping function for a multilayer perceptron network.
7. The optical system assembly and adjustment method based on a dual-branch neural network model according to claim 1, characterized in that, Step 5 specifically includes: Step 51: Calibrate the optical axis reference; Install a laser interferometer, theodolite, plane mirror, and six-degree-of-freedom adjustment frame, and place the main support structure of the optical system on an air-floating platform; continuously adjust the plane mirror using the six-degree-of-freedom adjustment frame and theodolite until the plane mirror is perpendicular to the optical axis reference plane of the main support structure; adjust the laser interferometer to align with this optical axis reference plane to achieve optical axis alignment between the laser interferometer and the main support structure; Step 52, rough adjustment; First, the primary mirror (M1) is assembled and adjusted: using a laser interferometer, theodolite, coordinate measuring arm, altimeter, and six-degree-of-freedom adjustment frame, the angle between the primary mirror (M1) reference plane and the plane mirror is adjusted to the ideal angle; then, the surface shape of the primary mirror (M1) is monitored in real time using a laser interferometer, and the position and pose of the primary mirror (M1) are continuously adjusted to make the surface shape error of the primary mirror (M1) the same as the measured result; Secondly, coarse adjustment of the secondary mirror (M2) and the tertiary mirror (M3) is carried out: the secondary mirror (M2) and the tertiary mirror (M3) are initially placed in the design position using a six-degree-of-freedom adjustment frame, and the adjustment amount of each degree of freedom of the secondary mirror (M2) and the tertiary mirror (M3) is monitored in real time using a theodolite and a micrometer during the coarse adjustment process. The adjustment amount of each degree of freedom is compared with the design value to ensure the relative spatial position accuracy between the plane mirrors in the optical system. Step 53, Adjust the fine packaging; The wavefront of the optical system is detected using an autocollimation method, with wavefront detection at multiple field-of-view positions achieved through a deflecting plane mirror and a laser interferometer. A trained neural network assembly proxy model is used to calculate the predicted adjustment amounts for each degree of freedom of the optical components based on the wavefront detection results from multiple field-of-view measurements, thus guiding the assembly and adjustment of the secondary mirror (M2) and the tertiary mirror (M3). During fine-tuning, after each measurement, the wavefront maps of multiple field-of-views and the fitted Zernike coefficients are input into the neural network assembly proxy model to calculate the suggested adjustment amounts for each component's degree of freedom. Then, the component poses are adjusted according to the suggested adjustment amounts using a six-DOF adjustment frame. After adjustment, the wavefront of the optical system is measured again. Through multiple iterations of the above process, the wavefront aberration of the actual optical system is brought close to the wavefront aberration of the optical system after considering various errors, ensuring that all optical performance indicators are within acceptable ranges, at which point the assembly and adjustment is considered complete.
Citation Information
Patent Citations
Intelligent optical active adjustment method based on point spread function
CN113468802A
Neural network-based optical lens force-position collaborative adjustment method
CN120429906A