Off-axis four-mirror optical system intelligent installation and adjustment method based on deep learning model

Abstract

The invention discloses an intelligent assembling and adjusting method for an off-axis four-mirror optical system based on a deep learning model, and belongs to the technical field of assembling and adjusting of off-axis reflection type systems. The method comprises the following steps: constructing a data set of a misalignment amount and a full-field Zernike coefficient matrix through an optical design software interface; training by using a Transformer neural network to obtain a model capable of predicting the misalignment amount according to a Zernike coefficient matrix; in actual adjustment, point spread functions of sampling view field points are measured and fitted to obtain corresponding Zernike coefficients, a full-view-field continuous Zernike coefficient matrix is obtained through interpolation, the trained model is input, the misalignment amount of each optical element can be calculated, and then an execution mechanism is driven to complete accurate adjustment. According to the method, the installation and adjustment precision and efficiency of the off-axis four-reflector complex system are remarkably improved, the method does not need to depend on a high-precision wavefront detector, and the engineering applicability and environment robustness of the system are enhanced.

Description

Technical Field

[0001] This invention belongs to the field of assembly and adjustment technology of off-axis reflective systems, specifically relating to an intelligent assembly and adjustment method for an off-axis four-reflector optical system based on a deep learning model. Background Technology

[0002] With the ever-increasing demands for image quality, lightweight design, and large field of view in modern optical systems, off-axis reflective systems have become widely used in astronomical observation, space remote sensing, and high-end imaging due to their advantages such as chromatic aberration-free operation, lack of central obstruction, and compact structure. Among these, off-axis four-mirror systems offer more degrees of freedom and greater potential for aberration correction compared to two- or three-mirror structures; however, this also drastically increases the difficulty of assembly and adjustment. Assembly and adjustment accuracy has become a key bottleneck restricting the achievement of design specifications for such high-performance optical systems.

[0003] Traditional computer-aided assembly techniques, such as the sensitivity matrix method and the evaluation function regression method, are simple in principle, easy to implement, and widely used in engineering. The sensitivity matrix method is based on the assumption that misalignment and aberration (usually characterized by Zernike coefficients) have a linear relationship, and it inverts the misalignment by solving a system of linear equations. This method is applicable within a small misalignment range, but in actual assembly, especially when the misalignment is large and the sensitivity difference is significant, the relationship between misalignment and aberration becomes significantly nonlinear, and there is strong coupling between the degrees of freedom of the misalignment, leading to an ill-conditioned sensitivity matrix and distorted or even divergent results. The evaluation function regression method requires iterative optimization, approximating the measured aberration through simulation using optical design software. This process is time-consuming and heavily relies on the initial values ​​and convergence of the optimization algorithm, making it unsuitable for complex systems with multiple local extrema.

[0004] In existing technologies, high-precision wavefront detectors (such as Shack-Hartmann sensors) are typically used to directly acquire the wavefront information of the system to calculate the Zernike coefficients. However, these precision sensors are not only expensive, but also extremely sensitive to environmental vibrations and airflow disturbances, making it difficult to operate stably under non-ideal conditions such as telescope sites and mobile platforms, thus limiting their widespread application in engineering practice.

[0005] In recent years, deep learning technology has demonstrated tremendous potential in many engineering fields due to its powerful nonlinear fitting and feature extraction capabilities. How to introduce deep learning models into the intelligent assembly and adjustment process of complex optical systems, bypassing the dependence on precision wavefront detectors, and fundamentally solving the problems of linear assumptions and iterative inefficiency in traditional methods, to achieve high-precision, high-efficiency, and robust offset calculation, has become a pressing technical challenge in this field. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides an intelligent assembly and adjustment method for off-axis four-mirror optical systems based on a deep learning model. By utilizing the powerful nonlinear fitting capability of Transformer, a complex mapping between aberrations and misalignment is directly established, avoiding the linear approximation and iterative optimization process in traditional methods. This significantly improves the assembly and adjustment accuracy and efficiency of complex systems like off-axis four-mirror systems, and eliminates the need for a high-precision wavefront detector, thus enhancing the system's engineering applicability and environmental robustness.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for intelligent assembly and adjustment of an off-axis four-mirror optical system based on a deep learning model, comprising:

[0009] Step 1: Through the API interface of the optical design software, apply multiple combinations of misalignment to the optical system model in a programmed manner, and obtain the Zernike coefficients of discrete sampled field points under each misalignment state to construct a dataset of misalignment-theoretical Zernike coefficient matrix.

[0010] Step 2: Train a Transformer neural network model using the offset quantity-theoretical Zernike coefficient matrix dataset to obtain an offset quantity prediction model;

[0011] Step 3: For the off-axis four-reflector optical system to be assembled and adjusted, measure its actual Zernike coefficient matrix at multiple sampling field points; input the actual Zernike coefficient matrix into the misalignment prediction model to obtain the predicted misalignment parameters; adjust the pose of the optical components according to the prediction results to complete the assembly and adjustment.

[0012] Furthermore, in step 1, after obtaining the Zernike coefficients of the discrete sampled field of view points under each misalignment state, the Zernike coefficients of the discrete sampled field of view points are fitted into a theoretical Zernike coefficient matrix that covers the entire field of view continuously using a cubic interpolation algorithm.

[0013] Furthermore, in step 2, the Transformer neural network model adopts an encoder-decoder architecture:

[0014] The encoder consists of N stacked encoder layers. Each encoder layer sequentially performs multi-head self-attention, residual connection and layer normalization, feedforward network and residual connection and layer normalization again, which is used to transform the sum of the input word embedding and position encoding into a dimension-invariant context representation.

[0015] The decoder consists of N stacked decoder layers. Each decoder layer introduces an additional masked multi-head self-attention layer on top of the encoder layer structure to prevent access to future positions, and uses the context representation of the encoder output as key-value pairs through the encoder-decoder attention layer.

[0016] The final output layer of the model consists of a linear mapping layer and a Softmax function, which is used to generate a predicted sequence of offset parameters.

[0017] Furthermore, in step 3, measuring the actual Zernike coefficient matrix of multiple sampling field points specifically includes: using a camera to collect the point spread function (PSF) of each sampling field point, performing Zernike polynomial fitting on each PSF, and calculating the Zernike coefficients of each sampling point; and then generating an actual Zernike coefficient matrix that covers the entire field of view continuously through interpolation.

[0018] Furthermore, step 3 also includes a closed-loop assembly and adjustment iteration process: after adjusting the pose of the optical element according to the prediction result, the actual Zernike coefficient matrix under the current system state is remeasured and input into the misalignment prediction model again for the next round of misalignment prediction and adjustment until the system image quality index reaches the preset convergence condition.

[0019] Furthermore, the convergence condition is that the root mean square value of the system wavefront aberration is lower than a preset threshold, or the change in the predicted misalignment in two consecutive iterations is less than a set tolerance.

[0020] Furthermore, the optical design software is ZEMAX, and the API interface is called through a Python programming environment to automatically control the application of misalignment parameters and the acquisition of Zernike coefficient data.

[0021] Furthermore, the misalignment of the off-axis four-mirror optical system includes the tilt and eccentricity degrees of freedom of each mirror.

[0022] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned intelligent assembly and adjustment method for an off-axis four-reflector optical system based on a deep learning model.

[0023] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned intelligent assembly and adjustment method for an off-axis four-reflector optical system based on a deep learning model.

[0024] The beneficial effects of this invention are as follows:

[0025] High-precision misalignment calculation: Utilizing the powerful nonlinear mapping and feature extraction capabilities of the Transformer deep learning model, a complex relationship model from full-field Zernike aberration to system misalignment is directly established, completely eliminating the linear approximation assumptions relied upon by the traditional sensitivity matrix method, and significantly improving the assembly accuracy of complex multi-mirror systems such as off-axis four-mirror systems under large misalignment conditions.

[0026] The assembly and adjustment process is efficient and stable: It adopts the "one-time prediction and direct adjustment" mode, which eliminates the lengthy process of repeatedly calling optical design software and making multiple adjustments and trials in the traditional iterative optimization method. This greatly shortens the assembly and adjustment cycle and avoids the risk of the algorithm getting stuck in local optima, thus improving the determinism and efficiency of the assembly and adjustment process.

[0027] Reduced system hardware dependence and cost: Zernike coefficients are calculated directly through point spread function (PSF), eliminating the need for high-precision, high-cost wavefront detectors and their associated stable environment. This significantly reduces the system's hardware complexity and assembly / adjustment requirements, and improves the applicability and robustness of this method in engineering field applications.

[0028] Achieving global optimization of aberrations across the entire field of view: By interpolating to obtain a continuous Zernike coefficient matrix across the entire field of view as model input, the model can perform global analysis and prediction based on aberration information across the entire field of view. This overcomes the one-sidedness that may be caused by relying on a small number of sampled field points in traditional methods, thereby achieving better system-level aberration correction results.

[0029] High degree of intelligence and automation: The entire process realizes a closed loop from aberration measurement and data processing to misalignment calculation and feedback control, providing a complete and feasible technical path for the intelligent and automated assembly and adjustment of complex optical systems. Attached Figure Description

[0030] Figure 1 This is a diagram of the off-axis four-reflector structure involved in this invention;

[0031] Figure 2 This is a schematic diagram of the data interaction principle between ZEMAX and PYTHON in this invention;

[0032] Figure 3 This is a graph showing the variation of the Zernike coefficients fitted in this invention;

[0033] Figure 4 This is a basic architecture diagram of the Transformer model used in this invention;

[0034] Figure 5 This is a flowchart of the offset calculation process of the present invention;

[0035] Figure 6 This is a diagram showing the actual results of the offset calculation in this invention. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0037] The purpose of this invention is to provide an intelligent assembly and adjustment method for an off-axis four-mirror optical system based on a deep learning model. This method improves the accuracy of misalignment calculation and avoids the use of precision instruments such as wavefront sensors, thereby enhancing the system's stability in complex environments. Figure 1 The diagram shows the off-axis four-mirror structure involved in this invention. Compared to two-mirror and three-mirror structures, it has stronger aberration correction capabilities. However, with the increase in the number of mirrors, the system assembly and adjustment difficulty also increases. To achieve more intelligent and accurate assembly and adjustment of the off-axis four-mirror structure, the core of this invention lies in: establishing a dataset of misalignment and theoretical Zernike coefficient matrix through simulation, and training a Transformer neural network model; in actual assembly and adjustment, measuring the system point spread function (PSF), fitting the actual Zernike coefficients of the sampled field of view points, obtaining the actual Zernike coefficient matrix through interpolation, and inputting it into the trained model to predict the misalignment; finally, adjusting the system through the control interface and iteratively optimizing until the image quality requirements are met. Specifically, the method includes:

[0038] Step 1: Construct a dataset; use the API interface of optical design software to interact with PYTHON and other software to obtain the Zernike coefficients of the system under different misalignment states. Through cubic interpolation, obtain the theoretical Zernike coefficient matrix that is continuous across the entire field of view under different misalignment states, thus constructing a misalignment-Zernike coefficient matrix data pair.

[0039] Step 2: Input the offset quantity-theoretical Zernike coefficient matrix data into the neural network for training. After iterative training, a network model that can predict the offset quantity is obtained.

[0040] Step 3: Measure the point spread function of the actual optical system and obtain the corresponding Zernike coefficient. Input the Zernike coefficient into the trained neural network, which outputs the corresponding offset of the system. Then, use the control system to precisely adjust the spatial positions of the optical system components. Check whether the imaging quality of the adjusted optical system meets the preset requirements. If not, repeat the above steps until the imaging requirements are met.

[0041] Let the parameters of the off-axis four-reflector system designed in this invention be: entrance pupil diameter 125mm, focal length 1359mm, and full field of view. The wavelength is visible light.

[0042] Specifically, step 1 includes:

[0043] Through the interactive interface of the optical design software, data can be quickly exchanged, and the Zernike coefficients corresponding to the preset sampling field of view points can be read. Figure 2 The diagram shown is a data interaction flowchart for the simulation stage of this invention. Taking the PYTHON interface as an example, this invention first completes the design of the off-axis four-mirror optical system in ZEMAX to determine the initial design state of the system, and then uses the application programming interface (API) provided by ZEMAX to establish a connection with the external programming environment (such as PYTHON).

[0044] A control script is written in Python to programmatically generate a series of system misalignment states. The misalignment primarily refers to the tilt and eccentricity degrees of freedom of each mirror in the system. For each set of preset misalignment parameters, the script applies them to the optical model in ZEMAX via an API interface and drives the software to perform ray tracing and aberration analysis. Subsequently, the script reads multiple preset sampling field points (such as...) under this system state via the API interface. Figure 1 The coefficients of the first N terms of the Zernike polynomial corresponding to the field of view distribution shown. Each field point will obtain a Zernike coefficient vector.

[0045] Using a cubic interpolation algorithm, a continuous theoretical Zernike coefficient matrix covering the entire field of view (e.g., 6° × 6°) is generated based on the Zernike coefficient vectors of these discrete sampled field points. This matrix fully describes the aberration distribution of the system across the entire field of view under this specific misalignment state. Figure 3 An example of the fitted surface for the first 9 Zernike coefficients as a function of the field of view is shown.

[0046] Repeat the above process to generate different misalignment quantities and their corresponding theoretical Zernike coefficient matrices, which together constitute the "misalignment quantity-theoretical Zernike coefficient matrix" dataset.

[0047] In step 2, the network model for predicting the misalignment is a transformer. For example... Figure 4The diagram shows the deep learning model architecture used in this invention. The basic Transformer architecture consists of N stacked encoder and decoder layers, with multi-head self-attention as its core. In this application, the theoretical Zernike coefficient matrix or its appropriately flattened sequence is used as the model input, and the corresponding offset vector is used as the model output label. Each layer of the encoder sequentially performs multi-head self-attention, residual connections and layer normalization, feedforward network, and further residual normalization, transforming the sum of the input word embeddings and sine-cosine positional encodings into a dimension-invariant contextual representation. The decoder further introduces masked multi-head self-attention to prevent access to future positions, and uses the encoder output as the key / value pair through cross-multi-head attention. Finally, the sequence is gradually generated through linear mapping and Softmax. The Transformer has powerful parallel computing capabilities and the ability to capture complex dependencies, enabling it to process large-scale complex data.

[0048] The goal of model training is to learn a complex nonlinear mapping from the aberration space to the offset space. After sufficient iterative training, the model loss function converges, and the model parameters with the best predictive performance at this point are saved, resulting in a Transformer prediction model that can be used to solve for the offset.

[0049] In step 3, the offset calculation process is as follows: Figure 5 As shown, the specific process is as follows: For the misaligned four-reflector optical system to be installed, the point spread function (PSF) of multiple sampling field points of the actual optical system is first obtained. According to optical theory, the Fourier transform phase information of the PSF is related to the wavefront aberration. By fitting the PSF of each field point with a Zernike polynomial, the actual Zernike coefficient vector corresponding to that point can be calculated. Similar to the processing method in the simulation stage, a cubic interpolation algorithm is used to fit the actual Zernike coefficients of these discrete sampling points into an actual Zernike coefficient matrix covering the entire field of view. The actual Zernike coefficient matrix is ​​then input into the trained Transformer prediction model to calculate the misalignment. The calculated misalignment (i.e., the tilt and eccentricity adjustments required for each mirror) is then converted into control commands and returned to the optical system via a control interface. This drives adjustment mechanisms (such as piezoelectric actuators or precision displacement stages) to actively adjust the pose of the mirrors in the optical system. After adjustment, the system is checked to ensure it meets image quality requirements. If it does, the adjustment is complete; otherwise, the above steps are repeated. The actual Zernike coefficient matrix of the system in the new state is acquired and calculated again. This newly obtained actual Zernike coefficient matrix is ​​then input into the prediction model for the next round of misalignment prediction and adjustment. This constitutes a closed-loop adjustment process of "measurement-prediction-adjustment-verification" until the image quality requirements are met, at which point the adjustment process ends. Figure 6As shown, image quality requirements are typically that the system wavefront aberration RMS value reaches the design threshold (e.g., λ / 20, where λ is the wavelength), or the change in the predicted offset value between two consecutive iterations is less than the tolerance, or the maximum set number of safe iterations is reached. Figure 6 The diagram shows a comparison of the convergence effect of the system's RMS value towards the theoretical design value from the offset state to the corrected state. The left figure is the theoretical RMS of the system, the middle figure is the RMS of the system in the offset state, and the right figure is the RMS of the system after the offset is calculated and the solution value is substituted.

[0050] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned intelligent assembly and adjustment method for an off-axis four-reflector optical system based on a deep learning model.

[0051] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned intelligent assembly and adjustment method for an off-axis four-reflector optical system based on a deep learning model.

[0052] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent assembly and adjustment of an off-axis four-mirror optical system based on a deep learning model, characterized in that, include: Step 1: Through the API interface of the optical design software, apply multiple combinations of misalignment to the optical system model in a programmed manner, and obtain the Zernike coefficients of discrete sampled field points under each misalignment state to construct a dataset of misalignment-theoretical Zernike coefficient matrix. Step 2: Train a Transformer neural network model using the offset quantity-theoretical Zernike coefficient matrix dataset to obtain an offset quantity prediction model; Step 3: For the off-axis four-reflector optical system to be assembled and adjusted, measure its actual Zernike coefficient matrix at multiple sampling field points; input the actual Zernike coefficient matrix into the misalignment prediction model to obtain the predicted misalignment parameters; adjust the pose of the optical components according to the prediction results to complete the assembly and adjustment.

2. The intelligent assembly and adjustment method for an off-axis four-reflector optical system based on a deep learning model according to claim 1, characterized in that, In step 1, after obtaining the Zernike coefficients of the discrete sampled field of view points under each misalignment state, the Zernike coefficients of the discrete sampled field of view points are fitted into a theoretical Zernike coefficient matrix that covers the entire field of view continuously using a cubic interpolation algorithm.

3. The intelligent assembly and adjustment method for an off-axis four-reflector optical system based on a deep learning model according to claim 1, characterized in that, In step 2, the Transformer neural network model adopts an encoder-decoder architecture: The encoder consists of N stacked encoder layers. Each encoder layer sequentially performs multi-head self-attention, residual connection and layer normalization, feedforward network and residual connection and layer normalization again, which is used to transform the sum of the input word embedding and position encoding into a dimension-invariant context representation. The decoder consists of N stacked decoder layers. Each decoder layer introduces an additional masked multi-head self-attention layer on top of the encoder layer structure to prevent access to future positions, and uses the context representation of the encoder output as key-value pairs through the encoder-decoder attention layer. The final output layer of the model consists of a linear mapping layer and a Softmax function, which is used to generate a predicted sequence of offset parameters.

4. The intelligent assembly and adjustment method for an off-axis four-reflector optical system based on a deep learning model according to claim 1, characterized in that, In step 3, measuring the actual Zernike coefficient matrix of multiple sampling field points specifically includes: using a camera to collect the point spread function (PSF) of each sampling field point, performing Zernike polynomial fitting on each PSF, and calculating the Zernike coefficients of each sampling point; then generating an actual Zernike coefficient matrix that covers the entire field of view continuously through interpolation.

5. The intelligent assembly and adjustment method for an off-axis four-reflector optical system based on a deep learning model according to claim 1, characterized in that, Step 3 also includes a closed-loop assembly and adjustment iterative process: after adjusting the pose of the optical components according to the prediction results, the actual Zernike coefficient matrix under the current system state is remeasured and then input into the misalignment prediction model for the next round of misalignment prediction and adjustment until the system image quality index reaches the preset convergence condition.

6. The intelligent assembly and adjustment method for an off-axis four-reflector optical system based on a deep learning model according to claim 5, characterized in that, The convergence condition is that the root mean square value of the wavefront aberration of the system is lower than a preset threshold, or the change in the predicted misalignment in two consecutive iterations is less than a set tolerance.

7. The intelligent assembly and adjustment method for an off-axis four-reflector optical system based on a deep learning model according to claim 1, characterized in that, The optical design software is ZEMAX, and the API interface is called through the Python programming environment to automatically control the application of misalignment parameters and the acquisition of Zernike coefficient data.

8. The intelligent assembly and adjustment method for an off-axis four-reflector optical system based on a deep learning model according to claim 1, characterized in that, The misalignment of the off-axis four-mirror optical system includes the tilt and eccentricity degrees of freedom of each mirror.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the intelligent assembly and adjustment method for an off-axis four-reflector optical system based on a deep learning model as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the intelligent assembly and adjustment method for an off-axis four-reflector optical system based on a deep learning model as described in any one of claims 1-8.

Images