A step-by-step alignment method for a reflective optical system based on neural network

Abstract

The application discloses a step-by-step mounting and adjusting method of a reflective optical system based on a neural network and belongs to the field of wavefront detection of reflective lens mounting and adjusting. The step-by-step mounting and adjusting method comprises the following steps: 1, an ideal mounting and adjusting imaging model of the reflective optical system is established, an initial mounting and adjusting error range is determined, and errors are generated and randomly combined; 2, a pose error is set in the model, a plurality of simulation cycles are performed, and a wavefront map-pose error data is constructed; 3, the wavefront map is taken as input, and a coarse adjustment neural network model is constructed; 4, a fine adjustment pose error is set in the model, the Zernike coefficient is taken as input, a fine adjustment neural network model is constructed and saved; and 5, in actual mounting and adjusting, the step-by-step mounting and adjusting method is applied to complete the coarse adjustment and fine adjustment processes of the lens to be mounted and adjusted. The application considers the nonlinearity between the detection parameters and the mounting and adjusting errors, directly establishes the mapping relationship between the two, and is favorable for improving the mounting and adjusting precision and efficiency of the reflective optical system.

Description

Technical Field

[0001] This invention relates to the field of assembly and adjustment of reflective optical systems utilizing wavefront detection, and more particularly to a step-by-step assembly and adjustment method for reflective optical systems based on neural networks. Background Technology

[0002] In recent years, with the development of precision machining technology, optical inspection technology, and signal processing technology, optical systems have been widely used in military and civilian fields. Reflective lenses reflect light according to the design requirements of the optical system, and reflective optical systems are characterized by large aperture, zero chromatic aberration, and long focal length. The assembly and adjustment of reflective optical systems is one of the important steps in their manufacturing process, and adopting appropriate assembly and adjustment methods is of great significance for improving the assembly and adjustment quality and efficiency of the optical system.

[0003] Lens adjustment in optical systems can be divided into coarse adjustment for larger misalignments and fine adjustment for smaller misalignments. Currently, coarse adjustment is mainly performed manually. For example, CN113917669A manually coarsely adjusts the primary and secondary mirrors in a telescope system by observing the light spot and other detection parameters in the testing equipment. CN107271150A qualitatively completes the coarse assembly of aspherical optical elements by observing interference images in the testing equipment. Establishing the relationship between misalignment and detection parameters is one of the effective ways to achieve precise and efficient adjustment of optical systems. Commonly used methods include the sensitivity matrix method and the neural network method. Among them, the sensitivity matrix method linearizes the relationship between misalignment and detection parameters. For example, CN107271150A constructs a sensitivity matrix by solving the equations of detection parameters and misalignment under multiple phases, and CN109002567A constructs a sensitivity matrix by calculating the change in detection parameters in a certain field of view caused by a unit misalignment. This method imposes certain requirements on the initial error of the optical lenses. Neural network methods establish nonlinear relationships between offset and detection parameters. For example, CN113468802B studies the relationship between offset and point spread function, ellipticity and their geometric parameters, and indirectly establishes the relationship between offset and detection parameters through neural networks. CN116699838A uses wavelet aberration data as intermediate parameters and indirectly establishes the relationship between offset and point spread function through neural networks. The accuracy of offset solution is related to the intermediate transformation relationship and the transformation accuracy.

[0004] Existing research on the assembly and adjustment methods of reflective lenses still has some shortcomings and gaps:

[0005] 1. There is limited research on quantitative lens assembly and adjustment methods in the coarse adjustment stage, and coarse adjustment still mainly relies on trial and error iteration with the assistance of manual labor and testing equipment.

[0006] 2. The nonlinear relationship between the detection parameters and the offset is not considered, which affects the accuracy of the offset calculation. At the same time, the linearization method places certain requirements on the coarse adjustment accuracy and is only suitable for adjusting the initial offset when it is small.

[0007] 3. The lack of a direct mapping relationship between the detection parameters and the assembly and adjustment parameters, and the superposition error caused by the indirect assembly and adjustment mapping relationship, affects the accuracy of the lens misalignment calculation, and ultimately affects the assembly and adjustment accuracy of the optical lens. Summary of the Invention

[0008] In view of this, the present invention provides a step-by-step assembly and adjustment method for reflective optical systems based on neural networks, corresponding to the coarse and fine adjustment processes of lenses. By directly establishing a nonlinear mapping relationship between detection parameters and lens misalignment through neural networks, the method solves the problems of qualitative coarse adjustment, initial misalignment requirements, and solution accuracy of existing assembly and adjustment methods, and provides technical support for improving the automation, efficiency, and accuracy of assembly and adjustment of reflective optical lenses.

[0009] A step-by-step assembly and adjustment method for a reflective optical system based on a neural network, characterized by comprising the following steps:

[0010] Step 1: Establish an ideal assembly and imaging model of the reflective optical system in the simulation software, determine the initial assembly and adjustment error range of the reflective lens to be assembled and adjusted in actual assembly and adjustment, and randomly generate multiple sets of pose errors within the initial error range using a programming language.

[0011] Step 2: Control the ideal imaging model of the reflective optical system using a programming language, set random pose errors, save wavefront aberration data of the imaging position, construct a set of wavefront image-pose error data, repeat the above simulation multiple times, and construct a series of wavefront image-pose error data.

[0012] Step 3: Use a neural network algorithm to construct and save the coarse-tuning neural network model in the step-by-step assembly method;

[0013] Step 4: Determine the fine-tuning error range based on the absolute error probability of the coarse-tuned neural network model's prediction of the test set. Using a programming language, randomly generate multiple sets of fine-tuning pose errors within the fine-tuning error range. Repeat Step 2 to set the fine-tuning pose error, iterate multiple times, and construct a series of Zernike polynomial coefficient-pose error data. Use a neural network algorithm to construct and save the fine-tuning neural network model in the step-by-step assembly method.

[0014] Step 5: In the actual assembly and adjustment, the step-by-step assembly and adjustment method adopts a coarse adjustment and fine adjustment neural network model. The interferometer detection value is used as the input of the coarse adjustment neural network to coarsely adjust the lens to be assembled and adjusted. Then, the interferometer detection value is used as the input of the fine adjustment neural network to finely adjust the lens to be assembled and adjusted. The fine adjustment result is detected to determine whether it meets the assembly and adjustment requirements. If it does, the assembly and adjustment of the reflective optical system is completed. If it does not, the fine adjustment process is repeated until it meets the requirements.

[0015] Furthermore, the modeling conditions of the ideal imaging assembly and adjustment model of the reflective optical system in step 1 are consistent with the actual assembly and adjustment conditions, and the initial assembly and adjustment error range is determined according to the actual assembly and adjustment conditions.

[0016] Furthermore, in step 3, the wavefront image-pose error data is divided into a training set and a test set, with the wavefront image data as the input to the coarse-tuning neural network and the pose error as the output of the coarse-tuning neural network.

[0017] Furthermore, in step 4, the Zernike polynomial coefficient-pose error data is divided into a training set and a test set, with the Zernike polynomial coefficients as the input to the fine-tuning neural network and the fine-tuning pose error as the output of the fine-tuning neural network.

[0018] Furthermore, the fine-tuned neural network model saved in step 4 is the model after hyperparameter optimization.

[0019] Beneficial effects:

[0020] 1. The step-by-step assembly and adjustment method for reflective optical systems based on neural networks proposed in this invention quantitatively describes the misalignment during the coarse adjustment stage of the lens. Combined with the mechanical structure and initial state of the lens assembly and adjustment device, it can further reduce manual steps in the assembly and adjustment process and improve the degree of automation in the assembly and adjustment.

[0021] 2. The step-by-step assembly and adjustment method for reflective optical systems based on neural networks proposed in this invention corresponds to the coarse and fine adjustment processes of lenses. It overcomes the limitation caused by the small initial assembly and adjustment error range required by existing assembly and adjustment methods, and considers the nonlinear relationship between the misalignment and the detection parameters. This improves the prediction accuracy of the misalignment while increasing the prediction range of the misalignment.

[0022] 3. The step-by-step assembly and adjustment method for reflective optical systems based on neural networks proposed in this invention directly establishes a nonlinear mapping relationship between the misalignment and the detection parameters, avoiding the superposition of errors generated during the intermediate parameter solution process, and further improving the accuracy of the misalignment solution model and the lens assembly and adjustment accuracy. Attached Figure Description

[0023] Figure 1 This is a flowchart of the step-by-step assembly and adjustment method for the reflective optical system based on neural networks according to the present invention;

[0024] Figure 2 This is the ideal assembly and imaging model of the reflective optical system of the present invention;

[0025] Among them, 1-a mirror that has been installed and adjusted, and 2-a mirror to be installed and adjusted.

[0026] Figure 3 Wavefront images under ideal conditions and with pose errors;

[0027] Figure 4 This is a flowchart illustrating the practical application of the step-by-step assembly and adjustment method for the reflective optical system based on neural networks according to the present invention. Detailed Implementation

[0028] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0029] This invention provides a step-by-step assembly and adjustment method for a reflective optical system based on a neural network, as shown in the attached figure. Figure 1 As shown, it includes the following steps:

[0030] Step 1: Establish an ideal assembly and imaging model of the reflective optical system in the simulation software, determine the initial assembly and adjustment error range of the reflective lens to be assembled and adjusted in actual assembly and adjustment, and randomly generate multiple sets of pose errors within the initial error range using a programming language.

[0031] The modeling conditions for the ideal imaging and adjustment model of the reflective optical system are consistent with the actual adjustment conditions, including the light source, various optical components of the reflective optical system, and the imaging plane. This embodiment establishes the following... Figure 2 The ideal imaging assembly model shown includes two reflective optical lenses: 1 - a mirror that has been assembled and adjusted, and 2 - a mirror to be assembled and adjusted. A light source is established based on the laser interferometer used in the actual assembly process. The mirror surface shape is set according to the lens design parameters, and the imaging plane is set at the theoretical imaging position to ensure that the model's optical path matches the actual assembly optical path.

[0032] Based on the initial position and motion accuracy of the assembly and adjustment device used in the actual assembly and adjustment process, the initial assembly and adjustment error range is determined. In this embodiment, based on the design parameters of the optical system, the range of the initial assembly and adjustment movement error is determined to be [-a, a] mm, and the range of the initial assembly and adjustment rotation error is [-b, b] °. Since the optical lens is a rotationally symmetric structure, this embodiment does not consider the pose error in the rotation direction around the optical axis of the lens.

[0033] N pose errors are randomly generated within the initial assembly and adjustment error range, and the five-degree-of-freedom pose errors are randomly combined to form N sets of pose error data.

[0034] Step 2: Control the ideal imaging model of the reflective optical system using a programming language, set random pose errors, save wavefront aberration data of the imaging position, construct a set of wavefront image-pose error data, repeat the above simulation multiple times, and construct a series of wavefront image-pose error data.

[0035] The ideal imaging assembly model is dynamically controlled using Python or other programming languages. Pose errors are read and set within the model, and wavefront aberration data at the imaging position is saved, forming a one-to-one correspondence between wavefront aberration and pose error. Wavefront aberration data can be wavefront image data or matrix text data, such as... Figure 3 The image shown is the wavefront image data saved under ideal imaging conditions and under conditions including pose error in this embodiment. The wavefront image text data is a 32×32 matrix:

[0036]

[0037] Repeat the previous simulation N times to obtain N sets of wavefront plot-pose error data.

[0038] Step 3: Use a neural network algorithm to construct and save the coarse-tuning neural network model in the step-by-step assembly method.

[0039] The wavefront image-pose error data was divided into a training set and a test set. The wavefront image data was used as the input to the coarse-tuning neural network, and the pose error was used as the output of the coarse-tuning neural network.

[0040] The N sets of wavefront image-pose error data were divided into training and testing sets in an 8:2 ratio, and the maximum and minimum normalization methods for the wavefront image data and pose error data were performed respectively:

[0041]

[0042] The assembly prediction model used in this embodiment is a convolutional neural network model, which includes convolutional layers and fully connected layers. Appropriate parameters such as the number of convolutional layers, convolutional kernel format, convolutional kernel size, number of fully connected layers, and fully connected neurons are selected. Wavefront image data is used as input and pose error is used as output to construct a coarse-tuning neural network assembly model based on wavefront images.

[0043] A coarse-tuned convolutional neural network model is constructed using Adam as the optimizer, ReLU as the activation function, and MSE as the loss function. The activation function is a function that performs a non-linear transformation on the data.

[0044]

[0045] The loss function is a function that calculates the difference between the model's actual output and its expected output.

[0046]

[0047] Save the model as the coarse - tuning neural network alignment model based on the wavefront map.

[0048] Step 4: Determine the fine - tuning error range according to the predicted absolute error probability of the test set by the coarse - tuning neural network model. Randomly generate multiple groups of fine - tuning pose errors within the fine - tuning error range through a programming language, repeat Step 2 to set the fine - tuning pose errors, perform simulation in a loop for multiple times, construct a series of Zernike polynomial coefficient - pose error data, and use the neural network algorithm to construct and save the fine - tuning neural network model in the step - by - step alignment method.

[0049] The determination of the fine - tuning error range is based on the predicted absolute error probability of the test set by the coarse - tuning neural network alignment model. Statistically analyze the absolute error probability of the pose prediction of the test set samples by the coarse - tuning neural network alignment model. In this embodiment, the probability that the absolute value of the predicted absolute error of the translation pose is <c is 100%, and the probability that the absolute value of the predicted absolute error of the rotation pose is <d is 98.8%. Therefore, the fine - tuning translation error range is determined as [-c, c] mm, and the fine - tuning rotation error range is determined as [-d, d] °.

[0050] Randomly generate M pose errors within the fine - tuning pose error range, and perform random combinations of five - degree - of - freedom pose errors to form M groups of pose error data. Similarly, dynamically control the ideal alignment model through Python or other programming languages, read and set the pose errors in the model, save the first 37 Zernike coefficients in the imaging position Zernike polynomial analysis, and form corresponding Zernike coefficient - pose error pairs. Perform simulation in a loop for M times to obtain M groups of Zernike coefficient - pose error data.

[0051] Similarly, divide the Zernike coefficient - pose error data into a training set and a test set according to a ratio of 8:2, and perform normalization pre - processing on both the Zernike coefficients and the pose errors. According to the lens alignment accuracy requirements and alignment experience, select several items from the 37 Zernike coefficients as the input of the neural network. In this embodiment, select the 1st, 4th, 5th, etc. items as the input of the fine - tuning neural network. Select appropriate neural network parameters, with the pose error as the output, to form a fine - tuning neural network alignment model based on the Zernike polynomial coefficients.

[0052] Finally, save the fine - tuning neural network model after hyperparameter optimization. In this embodiment, with the predicted absolute error probability of the pose in the test set samples as the optimization goal, use an optimization algorithm to optimize parameters such as the number of neurons, training batch size, learning rate, training iteration times, and regularization coefficient of the neural network. The average absolute value error of the pose prediction of the optimized model for the test set is <0.005, and the coefficient of determination > 0.99.

[0053] Step 5: In the actual assembly and adjustment, the step-by-step assembly and adjustment method adopts a coarse adjustment and fine adjustment neural network model. The interferometer detection value is used as the input of the coarse adjustment neural network to coarsely adjust the lens to be assembled and adjusted. Then, the interferometer detection value is used as the input of the fine adjustment neural network to finely adjust the lens to be assembled and adjusted. The fine adjustment result is detected to determine whether it meets the assembly and adjustment requirements. If it does, the assembly and adjustment of the reflective optical system is completed. If it does not, the fine adjustment process is repeated until it meets the requirements.

[0054] The practical application process of the step-by-step assembly and adjustment method for reflective optical systems based on neural networks is shown in the attached figure. Figure 4 As shown. First, the reflective optical lenses and mirrors used in the actual assembly process are fixed to the assembly device using a clamping device, constructing the actual assembly optical path consisting of an interferometer, a reflective optical system, and a mirror. After the device moves to the initial assembly position, the wavefront image information in the interferometer is normalized and input into the coarse adjustment neural network assembly model. Its output, after inverse normalization, is the coarse adjustment pose error of the optical lens to be assembled. The coarse adjustment of the optical lens to be assembled is completed by the assembly actuator. Then, the lens is fine-tuned. The Zernike polynomial coefficients in the interferometer are input into the fine adjustment neural network assembly model. Its output, after inverse normalization, is the fine adjustment pose error of the lens. The fine adjustment of the lens is completed by the precision assembly actuator. Finally, the interferometer checks whether the fine adjustment result meets the assembly imaging quality requirements, such as the imaging PV and RMS values. If it meets the requirements, the assembly process of the reflective lens is completed; if not, the fine adjustment process is repeated until the requirements are met.

[0055] Although embodiments and drawings of the present invention have been disclosed for illustrative purposes, it should be understood that these embodiments and drawings are merely examples of the principles and applications of the present invention. Various substitutions, variations and modifications are possible without departing from the spirit and scope of the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A step-by-step assembly and adjustment method for a reflective optical system based on a neural network, characterized in that, Includes the following steps: Step 1: Establish an ideal assembly and imaging model of the reflective optical system in the simulation software, determine the initial assembly and adjustment error range of the reflective lens to be assembled and adjusted in actual assembly and adjustment, and randomly generate multiple sets of pose errors within the initial error range using a programming language. Step 2: Control the ideal imaging model of the reflective optical system using a programming language, set random pose errors, save wavefront aberration data of the imaging position, construct a set of wavefront image-pose error data, repeat the above simulation multiple times, and construct a series of wavefront image-pose error data. Step 3: Use a neural network algorithm to construct and save the coarse-tuning neural network model in the step-by-step assembly method; Step 4: Based on the absolute error probability of the coarse-tuned neural network model for the test set, determine the fine-tuning error range. Use a programming language to randomly generate multiple sets of fine-tuning pose errors within the fine-tuning error range. Repeat Step 2 to set the fine-tuning pose error, and perform multiple simulations to construct a series of Zernike polynomial coefficient-pose error data. Use a neural network algorithm to construct and save the fine-tuning neural network model in the step-by-step assembly method. Step 5: In the actual assembly and adjustment, the step-by-step assembly and adjustment method adopts a coarse adjustment and fine adjustment neural network model. The interferometer detection value is used as the input of the coarse adjustment neural network to coarsely adjust the lens to be assembled and adjusted. Then, the interferometer detection value is used as the input of the fine adjustment neural network to finely adjust the lens to be assembled and adjusted. The fine adjustment result is detected to determine whether it meets the assembly and adjustment requirements. If it does, the assembly and adjustment of the reflective optical system is completed. If it does not, the fine adjustment process is repeated until it meets the requirements.

2. The step-by-step assembly and adjustment method for a neural network-based reflective optical system as described in claim 1, characterized in that, The modeling conditions of the ideal imaging assembly and adjustment model of the reflective optical system in step 1 are consistent with the actual assembly and adjustment conditions, and the initial assembly and adjustment error range is determined according to the actual assembly and adjustment conditions.

3. The step-by-step assembly and adjustment method for a neural network-based reflective optical system as described in claim 1, characterized in that, In step 3, the wavefront image-pose error data is divided into a training set and a test set. The wavefront image data is used as the input to the coarse-tuning neural network, and the pose error is used as the output of the coarse-tuning neural network.

4. The step-by-step assembly and adjustment method for a neural network-based reflective optical system as described in claim 1, characterized in that, Step 4 divides the Zernike polynomial coefficient-pose error data into a training set and a test set, using the Zernike polynomial coefficients as the input to the fine-tuning neural network and the fine-tuning pose error as the output of the fine-tuning neural network.

5. The step-by-step assembly and adjustment method for a neural network-based reflective optical system as described in claim 1, characterized in that, The fine-tuned neural network model saved in step 4 is the model after hyperparameter optimization.

Images