Autonomous decision-making and design method for high-precision detection of aspheric mirrors

Abstract

The present invention relates to the field of optical detection technology, and specifically provides an autonomous decision-making and design method for a high-precision detection scheme for an aspheric mirror. By performing standard normalization processing and principal component analysis on the characteristic parameters of the aspheric mirror itself, the category of historical detection schemes, and the historical detection design data, the effective characteristic parameters of the aspheric mirror that affect the category of detection schemes and the detection design data are obtained, and the parameters are used as the input of a multi-task learning neural network, and the category of detection schemes and the detection design data are used as outputs. A self-defined loss function is used, and the hyperparameters of the multi-task learning neural network are updated by using Bayesian search, and the training and testing of the multi-task learning neural network are completed. Then, the autonomous decision-making and design of the detection scheme category and the detection design data of the aspheric mirror to be tested are realized through the trained network model. While ensuring the high precision and accuracy of the aspheric mirror detection, the present invention greatly improves the detection efficiency of the aspheric mirror.

Description

Technical Field

[0001] The present invention relates to the field of optical detection technology, and specifically provides an autonomous decision-making and design method for a high-precision detection solution for an aspherical mirror. Background Art

[0002] In optical systems, aspheric optical elements have more design freedom and can meet more complex design requirements. The introduction of aspheric elements can expand the field of view and improve the resolution, thereby improving the imaging quality. Under the same performance index conditions, the use of aspheric surfaces can reduce the number of components and thus reduce the complexity of the system. Therefore, aspheric reflectors are increasingly used in space optics, astronomical optics, and high-tech civilian fields. For example, the primary mirror of the WM Keck Telescope (Keck) is composed of 36 small lenses with an aperture of 1.8 meters, and its aperture size reaches 10 meters; the aperture of the secondary mirror of the Large Synoptic Survey Telescope (LSST) under planning reaches 3.4 meters. With the rapid development of space optical technology, large-aperture aspheric reflectors are increasingly widely used. This trend puts higher requirements on the efficiency and accuracy of high-precision optical detection technology, and has become an important challenge in the current field of optical engineering.

[0003] At present, the most commonly used method for high-precision surface shape detection of aspheric mirrors is the interference detection method, which has nanometer-level detection accuracy and high detection efficiency. The interference detection method can be further divided into the aberration-free point method, the compensation mirror method, the computational holographic method, the sub-aperture stitching method and the hybrid compensation method. However, for the aberration-free point detection method, the required Hindle sphere usually needs to have an aperture twice or even larger than the aperture of the inspected mirror, and can only detect quadratic surfaces; for the compensation mirror method, it is necessary to design a group of compensators with spherical aberration, and the aperture of the required compensator group is slightly larger than the convex high-order aspheric surface. However, it is difficult to ensure the processing accuracy of such a large-aperture lens compensator; for the method of setting the CGH compensator, due to the size limitation of the existing micro-nano processing platform, it is difficult to ensure the processing accuracy of CGH with an aperture of more than 300mm; when the sub-aperture stitching method detects large-aperture aspheric mirrors, the increase in sub-apertures will increase the error transmission and affect the detection accuracy.

[0004] Since the existing detection methods for large-aperture aspheric reflectors have certain shortcomings and limitations, the selection and design of high-precision detection solutions for aspheric reflectors are extremely dependent on the experience and expertise of professionals. It usually takes a lot of time to simulate and design detection solutions, analyze the tolerances and detection accuracy of various detection solutions, and determine the final detection solution after comparison, which is extremely inefficient. At the same time, due to the scarcity of experienced professionals, faced with an increasing number of detection tasks, it is often impossible to provide detection solutions accurately and in a timely manner, resulting in extremely low detection efficiency for aspheric reflectors. Summary of the invention

[0005] In order to solve the above problems, the present invention provides an autonomous decision-making and design method for a high-precision detection scheme for aspheric mirrors. The effective characteristic parameters of the aspheric mirrors are screened according to the historical detection schemes and design data of the aspheric mirrors, and a neural network mapping model of the effective characteristic parameters of the aspheric mirrors and the detection schemes and design data is constructed by using a neural network. The model can be used to autonomously decide and output the detection scheme and design data of the aspheric mirror to be tested, which ensures the validity of the detection scheme and design data while greatly improving the detection efficiency of the aspheric mirror.

[0006] The autonomous decision-making and design method of the high-precision detection solution for aspheric mirrors provided by the present invention includes:

[0007] S1: Collect characteristic parameters of multiple aspherical mirror samples, as well as their corresponding detection scheme categories and detection design data, perform principal component analysis on the detection scheme category and detection design data of each aspherical mirror sample, and obtain effective characteristics of the aspherical mirror that determine the detection scheme category and effective characteristic parameters that affect the detection design data;

[0008] The effective feature parameters are used as input, and the detection scheme category and detection design data are used as output to form a training set;

[0009] S2: Establish a multi-task learning neural network, including: an input layer with effective feature parameters as input, a shared hidden layer, a regression task output layer that outputs detection design data prediction values, and a classification task output layer that outputs detection scheme category prediction values;

[0010] Customize the loss function of the multi-task learning neural network and select Bayesian search to update the hyperparameters of the multi-task learning neural network;

[0011] Use the training set to train the multi-task learning neural network;

[0012] S3: Acquire characteristic parameters of the aspheric mirror to be tested and input them into a trained multi-task learning neural network. The multi-task learning neural network automatically outputs a prediction value of the detection scheme category and a prediction value of the detection design data.

[0013] Preferably, in S1, each aspherical mirror sample is first subjected to standard normalization processing according to the following formula, and then principal component analysis is performed:

[0014]

[0015] in, represents the aspheric mirror sample after the i-th standard normalization process, represents the i-th aspherical mirror sample, Represents the average of all aspherical mirror samples, Represents the standard deviation of all aspherical mirror samples.

[0016] Optimized loss function for multi-task learning neural networks for:

[0017] ;

[0018] in, is the loss function of the output layer of the regression task, is the loss function of the output layer of the classification task, express In the loss function The weight ratio in express In the loss function The weight ratio in ;

[0019]

[0020] in, represents the number of aspherical mirror samples, represents the actual value of the detection design data in the i-th aspherical mirror sample, represents the predicted value of the detection design data corresponding to the i-th aspherical mirror sample;

[0021]

[0022] in, represents the actual value of the detection scheme category of the i-th aspherical mirror sample, represents the detection scheme category prediction value corresponding to the i-th aspherical mirror sample, j represents any detection scheme category, and C represents the total number of detection scheme categories.

[0023] Preferably, the aspherical mirror is an aspherical reflecting mirror.

[0024] Preferably, in S2, the loss function of the multi-task learning neural network is Update the hyperparameters when the loss function Convergence completes the training of the multi-task learning neural network.

[0025] Compared with the prior art, the present invention can achieve the following beneficial effects:

[0026] The present invention relies on historical data to establish a database of aspheric mirrors, and uses a multi-task learning neural network to construct a network mapping relationship between characteristic parameters of aspheric mirrors and detection scheme categories and specific detection design data. The multi-task learning neural network can realize autonomous decision-making and design of the detection scheme category and specific detection design data of the aspheric mirror to be tested. On the basis of ensuring that the aspheric mirror detection scheme is correct and the detection accuracy meets the design requirements, the present invention improves the automation level of the detection process and greatly improves the detection efficiency of optical components. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flow chart of an autonomous decision-making and design method for a high-precision detection solution for aspheric mirrors provided in an embodiment of the present invention;

[0028] Figure 2 is a specific implementation process diagram of the autonomous decision-making and design method of the aspheric mirror high-precision detection solution provided by an embodiment of the present invention;

[0029] Figure 3 is a screenshot of the interface of the autonomous decision-making and design software provided according to an embodiment of the present invention;

[0030] Figure 4 It is a wavefront residual image obtained by detecting an aspheric mirror through the method of the present invention according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the following description, the same modules are represented by the same reference numerals. In the case of the same reference numerals, their names and functions are also the same. Therefore, the detailed description thereof will not be repeated.

[0032] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.

[0033] See also Figure 1 and Figure 2 The embodiment of the present invention provides an autonomous decision-making and design method for a high-precision detection scheme for an aspheric mirror, which is mainly used to design a surface detection scheme for an aspheric mirror or a device with an aspheric surface, select a suitable detection scheme for detecting the aspheric surface, and provide specific detection scheme design data. The specific process is as follows:

[0034] S1: Collect characteristic parameters of multiple aspherical mirrors, as well as historical detection scheme categories and detection design data selected and designed by professionals based on experience, and form a sample data with the characteristic parameters, historical detection scheme categories and detection design data of each aspherical mirror. In this embodiment, an aspherical reflector is used as a target for training and detection, so the characteristic parameters of the aspherical reflector itself usually include: aperture, vertex curvature radius, quadratic coefficient, quartic coefficient, sextic coefficient, octadeconic coefficient, twelfth coefficient, fourteenth coefficient, sixteenth coefficient, detection accuracy, detection wavelength and reflector material, etc. The categories of historical detection schemes include but are not limited to aberration-free detection method, computational holographic detection method and compensator detection method. Accordingly, if the category of the historical detection scheme is aberration-free detection method, the corresponding detection design data shall at least include: the distance from the interferometer focus to the spherical shell, the distance from the spherical shell to the mirror to be tested, the caliber of the spherical shell, the radius of curvature, the thickness, the material, the quadratic coefficient and the higher-order coefficient, wherein the spherical shell refers to a circular mirror or lens with optical functions used in the aberration-free detection process, whose surface is in the shape of a partial sphere, used to generate interference patterns for surface detection. If the category of the historical detection scheme is computational holographic detection method, the corresponding detection design data shall at least include: the distance from the interferometer focus to the CGH, the distance from the CGH to the mirror to be tested, the caliber of the CGH, the thickness, the material and the first 37 Zernike coefficients characterizing the phase of the CGH, wherein CGH refers to the computer-generated hologram used in the computational holographic detection process. If the historical detection scheme category is the compensator detection method, the corresponding detection design data should at least include: the distance from the interferometer focus to the compensator, the distance from the compensator to the mirror to be tested, the aperture of the compensator, the radius of curvature, the thickness, material, quadratic coefficient, and higher-order coefficient.

[0035] In order to facilitate subsequent data processing and analysis, the sample data needs to be preprocessed first. Specifically, the characteristic parameters of the aspheric reflector itself, the historical detection scheme category and the detection design data are subjected to standard normalization. The standard normalization formula is as follows:

[0036]

[0037] in, represents the aspheric mirror sample after the i-th standard normalization process, represents the i-th aspherical mirror sample, Represents the average of all aspherical mirror samples, Represents the standard deviation of all aspherical mirror samples.

[0038] In the above formula and The calculation formula is:

[0039] ;

[0040] in, Indicates the number of aspherical mirror samples.

[0041] In the embodiment of the present invention, the above-mentioned normalization processing is mainly to adjust the numerical range of the sample data to a specific interval, usually [0, 1], through mathematical transformation. After the sample data is processed by standard normalization, its distribution pattern does not change, but the numerical range changes, and it can be restored to the original data through inverse transformation. The purpose of standard normalization processing is to eliminate the dimensional influence between different feature data, so that the contribution of each feature to the result is more balanced during subsequent model training. The data after standard normalization processing constitutes an aspheric reflector database, and the subsequent data used are all standardized data in the database.

[0042] After the sample data is normalized, the detection scheme category and detection design data in the sample data are analyzed by principal component analysis (PCA). The characteristics of the aspheric reflector itself that determine the detection scheme category can be screened out through principal component analysis. These screened features have a significant impact on the selection of the detection scheme category and are recorded as effective features. At the same time, by analyzing the detection design data, it can be obtained which parameters of the effective characteristics of the aspheric reflector significantly affect the specific selection of the detection design data, and these parameters are recorded as effective feature parameters.

[0043] The training set of the neural network is constructed using sample data that has been normalized and principal component analyzed. Each sample data in the training set consists of an input and an output. The input is: effective feature parameters; the output is: detection scheme category and detection design data.

[0044] S2: Select the basic structural properties of the required neural network according to the form of input and output in the training set. In the embodiment of the present invention, a multi-task learning neural network is selected as the neural network mapping model for autonomous decision-making and design of aspheric reflector detection schemes. In the autonomous decision-making and design of aspheric reflector detection schemes, the multi-task learning neural network can process multiple related tasks at the same time, and output the detection scheme category prediction value and the detection design data prediction value at the same time according to the input valid feature parameters.

[0045] The specific topological structure of the multi-task learning neural network includes: an input layer, a shared hidden layer, a regression task output layer, and a classification task output layer, wherein the input layer is mainly used to receive the effective characteristic parameters of the aspheric reflector samples screened out in step S1, which usually include but are not limited to aperture parameters, vertex curvature radius parameters, material characteristic parameters, and surface shape accuracy requirements. The shared hidden layer is responsible for extracting the common features of the input data, which are shared by multiple tasks and help reduce model complexity and improve generalization capabilities. The regression task output layer predicts the optimal detection design parameters of the aspheric reflector based on the input data, that is, outputs the detection design data prediction value. The classification task output layer predicts the optimal detection scheme category of the aspheric reflector based on the input data, that is, outputs the detection scheme category prediction value.

[0046] In addition, it is also necessary to customize the loss function of a multi-task learning neural network based on the characteristics of the regression task and the classification task, and optimize the performance of the regression task and the classification task at the same time. The regression task is evaluated by the mean square error (MSE) to optimize the output detection design data prediction value; the classification task is evaluated by the cross entropy loss to optimize the output detection scheme category prediction value.

[0047] The loss function of the output layer of the regression task uses the mean square error , the specific calculation formula is:

[0048]

[0049] in, represents the number of aspherical mirror samples, represents the actual value of the detection design data in the i-th aspheric mirror sample input to the neural network, that is, the input data of the neural network, It represents the predicted value of the detection design data corresponding to the i-th aspherical mirror sample input into the neural network, that is, the output data of the neural network.

[0050] The loss function of the output layer of the classification task adopts cross entropy loss , the specific calculation formula is:

[0051]

[0052] Where n represents the total number of samples, j represents the index of the detection scheme category. In multi-classification tasks, it represents the jth category. C represents the total number of detection scheme categories. The value range of j is [1, C]. The detection scheme categories include but are not limited to aberration-free detection method, computer-generated holographic detection method, compensator detection method, etc. represents the actual value of the detection scheme category of the i-th aspherical mirror sample input into the neural network, Represents the detection scheme category prediction value corresponding to the i-th aspherical mirror sample input into the neural network.

[0053] According to the loss function of the regression task output layer and the loss function of the classification task output layer, the loss function of the multi-task learning neural network can be calculated. for:

[0054] ;

[0055] in, express In the loss function The weight ratio in express In the loss function The weight ratio in and These are constants determined based on experience and can be adjusted according to actual training results.

[0056] In an embodiment of the present invention, Bayesian search is used to update the hyperparameters of the multi-task learning neural network. The hyperparameters that need to be updated include: learning rate, batch size and task weights, wherein the learning rate is used to control the step size of each iteration update of the model; the batch size is used to control the number of data samples used in each training iteration; the task weights are used to adjust the contribution of each task loss to the total loss, and the Bayesian optimization algorithm is used to gradually explore the search space, evaluate different hyperparameter combinations, collect data and update the model to guide the next search. Bayesian search is an existing adaptive hyperparameter search method, which has a fast search speed, is not easily affected by local optimal solutions, and can greatly improve the efficiency of hyperparameter search. During the training and testing process of the multi-task learning neural network, the neural network is trained using sample data in the training set, and the trained network model is temporarily saved to test whether the loss function of the temporarily saved multi-task learning neural network converges. If the loss function has not converged, the error value of the loss function is calculated, and the error value of the loss function is used as a basis for updating the hyperparameters. After the hyperparameters are updated through Bayesian search, a new round of training and testing is started to form a loop process until the loss function has converged. In this case, the multi-task learning neural network training is considered complete. At this time, the multi-task learning neural network can map the output of the best detection scheme category prediction value and the detection design data prediction value based on the input aspheric reflector effective characteristic parameters.

[0057] S3: Use the trained multi-task learning neural network to make autonomous decisions and designs on the detection scheme of the aspheric reflector to be tested. Specifically, the characteristic parameters of the aspheric reflector to be tested are normalized and the normalized characteristic parameters are input into the input layer of the multi-task learning neural network. The multi-task learning neural network automatically outputs the best detection scheme category prediction value and detection design data prediction value.

[0058] The above multi-task learning neural network for autonomous decision-making and design of high-precision detection solutions for aspheric mirrors can be directly deployed on the PC to form a software program. Please refer to Figure 3 , which is the input port interface of the multi-task learning neural network software program. According to experience and needs, the aperture, vertex curvature radius, quadratic coefficient, quartic coefficient, sextic coefficient, octadeconic coefficient, twelfth coefficient, fourteenth coefficient, sixteenth coefficient, detection accuracy, detection wavelength and reflector material are customized. Only part of the data is input during specific input. In addition, the output port is directly connected to the ZEMAX software, and the detection optical path is simulated directly by the ZEMAX software based on the output detection scheme category prediction value and detection design data prediction value.

[0059] In order to prove the effectiveness of the present invention, the following comparative experiment was conducted. A concave aspheric reflector was provided, with an aperture of 780 mm, a vertex curvature radius of 2903.22 mm, a quadratic coefficient K=-5.9, and a material of silicon carbide. The detection accuracy requirement was better than λ / 60RMS (λ=632.8 nm). For this concave aspheric reflector, the laboratory selected experienced professionals and multi-task learning neural networks for comparison. The multi-task learning neural network needs to normalize the above data standards of the aspheric reflector and input them into the trained multi-task learning neural network. The output detection scheme category is the aberration-free detection method, and the specific detection design data of the detection optical path is output. It was directly reproduced in the ZEMAX software, as shown in the following figure. Figure 4 As shown, the final detection system wavefront residual is 0.0008λRMS (λ=632.8nm), which meets the detection accuracy requirements and completes the detection scheme and detection optical path design.

[0060] By comparison, the detection scheme type and detection design data obtained by the method of the present invention are basically consistent with the detection and design schemes provided by experienced professionals, but the method of the present invention takes less time and improves the efficiency by more than 1 times, which greatly improves the detection efficiency of optical components.

[0061] Although the embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and cannot be understood as limiting the present invention. Those skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

[0062] The above specific implementations of the present invention do not constitute a limitation on the protection scope of the present invention. Any other corresponding changes and modifications made based on the technical concept of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. An autonomous decision-making and design method for a high-precision detection scheme for aspheric mirrors, characterized in that: include: S1: Collect characteristic parameters of multiple aspherical mirror samples, as well as their corresponding detection scheme categories and detection design data, perform principal component analysis on the detection scheme category and detection design data of each aspherical mirror sample, and obtain effective characteristics of the aspherical mirror that determine the detection scheme category and effective characteristic parameters that affect the detection design data; Aspheric mirrors are aspheric reflectors. The characteristic parameters of aspheric reflectors include: aperture, vertex curvature radius, quadratic coefficient, quartic coefficient, sextic coefficient, octadeconic coefficient, twelfth coefficient, fourteenth coefficient, sixteenth coefficient, detection accuracy, detection wavelength and reflector material; Categories of historical detection schemes include aberration-free detection, computerized holographic detection, and compensator detection; When the historical detection scheme category is aberration-free detection method, the detection design data include: the distance from the interferometer focus to the spherical shell, the distance from the spherical shell to the mirror to be tested, the caliber of the spherical shell, the radius of curvature, the thickness, the material, the secondary coefficient and the higher-order coefficient; When the historical detection scheme category is the computational holographic detection method, the detection design data includes: the distance from the interferometer focus to the CGH, the distance from the CGH to the mirror to be tested, the aperture, thickness, material of the CGH, and the first 37 Zernike coefficients that characterize the phase of the CGH; When the historical detection scheme category is the compensator detection method, the detection design data includes: the distance from the interferometer focus to the compensator, the distance from the compensator to the mirror to be tested, the caliber of the compensator, the radius of curvature, the thickness, the material, the secondary coefficient, and the higher-order coefficient; The effective feature parameters are used as input, and the detection scheme category and detection design data are used as output to form a training set; S2: Establish a multi-task learning neural network, including: an input layer with effective feature parameters as input, a shared hidden layer, a regression task output layer that outputs detection design data prediction values, and a classification task output layer that outputs detection scheme category prediction values; Customize the loss function of the multi-task learning neural network and select Bayesian search to update the hyperparameters of the multi-task learning neural network; Loss function for multi-task learning neural networks for: ； in, is the loss function of the output layer of the regression task, is the loss function of the output layer of the classification task, express In the loss function The weight ratio in express In the loss function The weight ratio in ; in, represents the number of aspherical mirror samples, represents the actual value of the detection design data in the i-th aspherical mirror sample, represents the predicted value of the detection design data corresponding to the i-th aspherical mirror sample; in, represents the actual value of the detection scheme category of the i-th aspherical mirror sample, represents the predicted value of the detection scheme category corresponding to the i-th aspherical mirror sample, j represents any detection scheme category, and C represents the total number of detection scheme categories; Use the training set to train the multi-task learning neural network; S3: Acquire characteristic parameters of the aspheric mirror to be tested and input them into a trained multi-task learning neural network. The multi-task learning neural network automatically outputs a prediction value of the detection scheme category and a prediction value of the detection design data.

2. The autonomous decision-making and design method for high-precision detection of aspheric mirrors according to claim 1, characterized in that: In S1, each aspherical mirror sample is first subjected to standard normalization according to the following formula, and then principal component analysis is performed: in, represents the aspheric mirror sample after the i-th standard normalization process, represents the i-th aspherical mirror sample, Represents the average of all aspherical mirror samples, Represents the standard deviation of all aspherical mirror samples.

3. The autonomous decision-making and design method of the high-precision detection scheme for aspheric mirrors as claimed in claim 2, characterized in that: In S2, the loss function of the multi-task learning neural network is Update the hyperparameters when the loss function Convergence completes the training of the multi-task learning neural network.

Images