Neural network-based optical lens force-position collaborative adjustment method

Abstract

The invention discloses an optical lens force-position collaborative adjustment method based on a neural network, and relates to the technical field of reflective optical lens assembly, and the method comprises the steps: building a reflective optical system lens model, carrying out the circulation simulation, and outputting a lens simulation result; sorting lens simulation results, constructing an assembly force-assembly error data set, constructing an assembly error prediction model, and training and storing the assembly force-assembly error data set to obtain a target assembly error prediction model; detecting the Zernike coefficient of the current lens, calculating to obtain the target misalignment amount of the lens, and taking the assembly force combination applied by the current lens as the input of the target neural network model to obtain the assembly pose error of the current lens; and compensating the assembly pose error on the basis of the target misalignment amount to obtain an actual adjustment amount, and completing the force-position cooperative control of the reflective optical system. According to the invention, the automation degree of the installation and adjustment process is improved and the lens installation and adjustment precision is improved by adopting the optical lens force-position collaborative installation and adjustment method based on the neural network.

Description

Technical Field

[0001] The present invention relates to the technical field of reflective optical lens assembly, and in particular to a neural network-based optical lens force-position coordinated assembly method. Background Art

[0002] Reflective optical systems feature large apertures, high resolution, and long focal lengths. They are also simple to manufacture, can achieve diffraction limits, and are lighter than lenses. They are widely used in astronomical observation, space telemetry sensing, precision guidance, microchip lithography, and other fields. Reflective lenses are the core of optical imaging systems. Positional errors such as lens tilt and decentration caused by assembly forces can significantly affect the reflected light path, reduce energy concentration efficiency, and ultimately compromise the optical system's imaging quality. Therefore, it is necessary to minimize the impact of assembly forces on positional errors during the lens assembly phase to ensure that the lens is correctly adjusted and moved to the designed position.

[0003] At present, the research on reflective optical lens adjustment mainly focuses on optical lens adjustment devices, active optics and adaptive optics technology, computer-aided adjustment and other aspects. CN202222695025.0 proposes an optical lens adjustment device that can move in five degrees of freedom, has high adjustment accuracy and simple steps, but its adjustment process relies on manual adjustment, and there are a lot of trial and error processes in the adjustment process. CN202211472511.4 A high-precision reflector adjustment method proposes a method for installing an optical reflector on a frame by means of an adhesive, which can significantly reduce the deformation of the mirror surface and improve the surface accuracy. However, the stress and strain of the adhesive have the problem of poor consistency.

[0004] There are still some deficiencies and gaps in the existing research on reflective optical lens adjustment:

[0005] 1. The installation process of optical lenses begins with measurement and adjustment, followed by assembly. Current research can only identify or predict the misalignment of optical lenses during the measurement phase, ignoring the misalignment caused by assembly forces during the subsequent assembly phase. Displacement and deformation errors caused by assembly forces cannot be eliminated during adjustment.

[0006] 2. Unable to select parameters to qualitatively and quantitatively describe optical mirror errors. Currently, lens adjustment processes mostly rely on manual trial and error to adjust parameters, which has the disadvantages of low automation, long adjustment time, poor consistency, and difficulty in ensuring accuracy.

[0007] In order to solve these problems, a neural network-based force-position coordinated adjustment method for optical lenses is urgently needed. Summary of the Invention

[0008] To solve the above problems, this application proposes a neural network-based optical lens force-position coordinated adjustment method, including:

[0009] Step 1: Establish a reflective optical system lens model based on the actual model and assembly process of the optical lens to be assembled;

[0010] Step 2: Perform cyclic simulation on the reflective optical system lens model and output the lens simulation results;

[0011] Step 3: Arrange the lens simulation results and construct an assembly force-assembly error dataset, wherein the assembly force-assembly error dataset includes assembly force combinations and assembly posture errors;

[0012] Step 4: Construct an assembly error prediction model, train and save the assembly force-assembly error dataset to obtain a target assembly error prediction model;

[0013] Step 5: Detect the Zernike coefficient of the current lens, calculate the target misalignment of the lens, use the assembly force combination applied by the current lens as the input of the target neural network model, and obtain the assembly posture error of the current lens;

[0014] Step 6: Compensate for the assembly posture error based on the target misalignment to obtain the actual adjustment amount, completing the force-position coordinated control of the reflective optical system.

[0015] Preferably, the specific contents of establishing the reflective optical system lens model in step 1 also include:

[0016] The allowable assembly force range during the assembly process is determined, the assembly force of the bolts required for assembling the optical lens is randomly generated within the assembly force range, and the assembly force of each bolt is combined to obtain an assembly force group.

[0017] Preferably, the assembly force of each bolt is formed within the assembly force range by a normal distribution or uniform distribution rule.

[0018] Preferably, in step 2, different assembly force groups are set to perform cyclic simulation on the reflective optical system lens model;

[0019] The lens simulation results include lens deformation and rigid body displacement.

[0020] Preferably, the assembly error prediction model is constructed based on a neural network model, and the specific contents of the target assembly error prediction model obtained by training and saving the assembly force-assembly error dataset include:

[0021] During the training process, the assembly force-assembly error dataset is divided into a training set and a test set for training the assembly error prediction model.

[0022] Preferably, the assembly force combination is used as the input of the neural network model, and the assembly posture error is used as the output of the neural network model.

[0023] Preferably, the specific content of training the assembly error prediction model is:

[0024] 1) Train the weights and bias parameters of each neuron node in the neural network;

[0025] 2) Optimize the hyperparameters of the neural network to obtain the target hyperparameter group;

[0026] 3) Save the weights, bias parameters and target hyperparameter groups of each neuron node in the neural network.

[0027] Preferably, the actual adjustment amount obtained in step six is the displacement of the lens itself in five degrees of freedom. The adjustment motion actuator where the lens is located should perform corresponding actual displacement calculation according to the lens installation conditions and tooling size during actual adjustment to achieve posture adjustment of the optical lens.

[0028] In summary, the neural network-based optical lens force-position coordinated adjustment method of the present invention has the following advantages over conventional technologies:

[0029] 1. The neural network-based force-position coordinated adjustment method for optical lenses proposed in this invention takes into account the nonlinear relationship between assembly force and optical lens misalignment, can eliminate the drawbacks of existing adjustment processes, and improve the degree of automation, efficiency, and consistency of the adjustment process;

[0030] 2. The neural network-based force-position coordinated assembly method for optical lenses proposed in this invention can use a posture error prediction algorithm to predict the assembly errors of the mirror surface in five degrees of freedom at once through the bolt assembly force during a single assembly process, thereby improving the lens assembly accuracy.

[0031] 3. The neural network-based force-position coordinated assembly method of optical lenses proposed in the present invention directly establishes a nonlinear mapping relationship between assembly errors and assembly parameters through the constructed neural network model, thereby avoiding the error superposition generated in the process of solving intermediate parameters and further improving the accuracy of the error solution model.

[0032] The technical method of the present invention is further described in detail below through the accompanying drawings and examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of the neural network-based optical lens force-position coordinated adjustment method of the present invention;

[0034] Figure 2 It is an ideal model of the optical lens of the reflective optical system of the present invention;

[0035] Figure 3 is the wavefront image with pose error, without considering assembly error, and with considering assembly error. Figure 3 a in the figure is the fitting result of the mirror wavefront map before adjustment (the surface shape measured in the initial state). Figure 3 b is the shape after adjustment without considering the influence of assembly force PV=0.2773λ, Figure 3 Where c is the result of the adjustment using the force-position coordination method proposed in the present invention (i.e., considering the error caused by the assembly force) (the surface shape PV after considering the influence of the assembly force = 0.1982λ);

[0036] Figure 4 This is a flowchart of the actual use of the neural network-based optical lens force-position coordinated adjustment method of the present invention. DETAILED DESCRIPTION

[0037] The technical method of the present invention is further described below through the accompanying drawings and embodiments. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and values described in these embodiments do not limit the scope of this application.

[0038] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or uses.

[0039] Technologies, systems, and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, they should be considered part of the specification.

[0040] In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0041] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0042] The present invention provides a method for adjusting the force and position of an optical lens based on a neural network. Figure 1 As shown, the following steps are included:

[0043] Step 1: Establish a reflective optical system lens model based on the actual model and assembly process of the optical lens to be assembled;

[0044] Furthermore, the specific contents of establishing the reflective optical system lens model in step 1 also include:

[0045] Determine the allowable assembly force range during the assembly process, randomly generate the assembly force of the bolts required for assembling the optical lens within the assembly force range, and combine the assembly forces of each bolt to obtain an assembly force group. The allowable assembly force range is determined based on the actual assembly process.

[0046] Multiple bolts may be used in one installation, and the assembly force of each bolt is combined to form the assembly force combination of this installation.

[0047] Furthermore, the assembly force of each bolt is formed within the assembly force range by a normal distribution or uniform distribution rule.

[0048] Step 2: Perform cyclic simulation on the reflective optical system lens model and output the lens simulation results;

[0049] Furthermore, in step 2, different assembly force groups are set to perform cyclic simulation on the reflective optical system lens model;

[0050] The lens simulation results include lens deformation and rigid body displacement.

[0051] Step 3: Arrange the lens simulation results and construct an assembly force-assembly error dataset, wherein the assembly force-assembly error dataset includes assembly force combinations and assembly posture errors;

[0052] It can be understood that the assembly posture error is caused by the assembly force and is different from the lens error (misalignment) in the usual sense.

[0053] Step 4: Construct an assembly error prediction model, train and save the assembly force-assembly error dataset to obtain a target assembly error prediction model;

[0054] Furthermore, the assembly error prediction model is constructed based on a neural network model, and the specific contents of the target assembly error prediction model obtained by training and saving the assembly force-assembly error dataset include:

[0055] During the training process, the assembly force-assembly error dataset is divided into a training set and a test set for training the assembly error prediction model.

[0056] Furthermore, the assembly force combination is used as the input of the neural network model, and the assembly posture error is used as the output of the neural network model.

[0057] Furthermore, the specific content of the assembly error prediction model training is as follows:

[0058] Optimize the hyperparameters of the neural network to obtain the target hyperparameter group.

[0059] Step 5: Detect the Zernike coefficient of the current lens, calculate the target misalignment of the lens, use the assembly force combination applied by the current lens as the input of the target neural network model, and obtain the assembly posture error of the current lens;

[0060] It can be understood that the target misalignment is the displacement that the lens should be adjusted for in each degree of freedom.

[0061] Step 6: Compensate for the assembly posture error based on the target misalignment to obtain the actual adjustment amount, completing the force-position coordinated control of the reflective optical system.

[0062] Furthermore, the actual adjustment amount obtained in step six is the displacement of the lens itself in five degrees of freedom. The adjustment motion actuator where the lens is located should calculate the corresponding actual displacement amount according to the lens installation conditions and tooling size during actual adjustment to achieve posture adjustment of the optical lens.

[0063] Example

[0064] S1. According to the actual model and assembly process of the optical lens to be assembled, a finite element model of the optical lens assembly is constructed. In this embodiment, a secondary mirror assembly simulation model of a coaxial two-mirror optical system is used (e.g. Figure 2 ), using the assembled primary mirror as a reference. Using Python or other programming languages, generate multiple sets of bolt preload forces within a calculated, reasonable assembly force range [a, b]N according to normal or uniform distribution.

[0065] The optical lens model and finite element analysis settings are consistent with the actual model and actual assembly process. The material properties and boundary conditions of the optical lens are determined according to the actual assembly situation.

[0066] S2. Perform finite element simulation on the secondary mirror simulation model, specifically including: constructing assembly steps in the finite element simulation software, and in the pre-processing step, reasonably dividing the grid and submitting the job according to the actual relevant properties of the optical lens, such as material properties, force boundary conditions, displacement boundary conditions, etc., and using different assembly force groups for multiple simulations.

[0067] Because the optical lens in this embodiment has a rotationally symmetric structure, the mirror point cloud is extracted in the post-processing step and other posture errors except the rotation posture error in the optical axis direction are calculated, including the average displacement of all points in the X, Y, and Z directions to represent the translation of the mirror in the X, Y, and Z directions. The data of each point are linearly approximated to obtain the rotation angle in the X and Y directions.

[0068] S3. Construct a lens assembly force-assembly error data set. The assembly error is composed of the assembly force position error and the rigid body displacement. The assembly force combination set in the simulation is matched one-to-one with the assembly error in S2 to form an assembly force-assembly error data set.

[0069] In this embodiment, the bolts are tightened in a specific sequence. In particular, a diagonal tightening strategy should be adopted when multiple screws are symmetrically distributed. The five-degree-of-freedom pose errors of the secondary mirror model are calculated using the data calculation method in S1 to generate a large dataset of assembly errors caused by assembly forces on the secondary mirror. Ultimately, a large number of one-to-one corresponding assembly force-assembly error data sets are generated.

[0070] S4. Using the assembly force as input, construct and save the neural network model.

[0071] The assembly force-assembly error dataset is divided into a training set and a test set in a ratio of 8:2. Signed logarithmic transformation combined with maximum and minimum normalization is selected to process the data. In particular, the output magnitude in this embodiment is small, so the signed logarithmic transformation and maximum and minimum normalization method are used to preprocess the output assembly posture error data set. Specifically, the signed logarithmic transformation is first performed:

[0072] y=f(x,∈)=sgn(x)*log 10 (|x|+∈);

[0073] Among them, x is the original value, ∈ takes a reasonable value to prevent zero from being a true number, and then the maximum and minimum values are normalized:

[0074]

[0075] Among them, x scaled is the value after normalization, x is the value before normalization, and x max is the maximum value before normalization, x min is the minimum value before normalization.

[0076] The neural network model used in this embodiment is a fully connected neural network model, consisting of an input layer, a hidden layer, and an output layer. Appropriate parameters such as the optimizer, loss function, number of neural network layers, number of neurons, and number of iterations are selected. The assembly error prediction neural network model is constructed using the assembly force data of each bolt as input and the assembly error as output. Hyperparameter optimization of the neural network is performed to obtain a hyperparameter combination with good prediction results.

[0077] In this embodiment, an assembly error prediction neural network model is constructed using Adam as the optimizer, Relu as the activation function, and MAE as the loss function.

[0078] The loss function is a function that calculates the difference between the actual output of the model and the expected output:

[0079]

[0080] Among them, m is the number of samples, y i is the expected output, f(x i ) is the actual output of the model.

[0081] When the neural network prediction accuracy meets the requirements, save the neural network.

[0082] S5. During the actual assembly of the optical lens, the Zernike coefficients of the current lens are measured, the Zernike coefficients at this time are used to fit the wavefront diagram and calculate the peak-to-valley difference (PV value) of the current surface shape, such as Figure 3 As shown in a in the figure, the ideal misalignment of the lens is calculated based on this test value. The PV and wavefront diagram after the lens posture is adjusted according to the calculated ideal misalignment are shown in the figure below. Figure 3 As shown in b. Figure 3 A comparison shows a decrease of 28.5%.

[0083] Then, the size of the screw assembly force to be applied is determined according to the process, and the neural network model built in S4 is called to calculate the assembly error of the secondary mirror after the assembly force is applied.

[0084] S6, based on the ideal misalignment, compensate the assembly error obtained in S5 to obtain the actual adjustment amount, and send the actual adjustment amount as the input of the motion execution component to the relevant motion execution mechanism to adjust the lens posture, and detect the adjustment result, such as Figure 3 As shown in c, if it meets the requirements, the installation and adjustment of the reflective optical system is completed. If it does not meet the requirements, repeat the installation and adjustment process until it meets the requirements.

[0085] The actual use process of the neural network-based optical lens force-position collaborative adjustment method is as follows: Figure 4 shown.

[0086] First, the reflective optical lenses and mirrors used in the actual alignment process are secured to the alignment device via a clamping device, constructing an actual alignment optical path consisting of an interferometer, a reflective optical system, and a reflector. Fine-tuning begins after the optical lenses in the reflective optical system reach near their designed positions. An interferometer is used to measure the Zernike coefficients of the current optical system and calculate the ideal misalignment of the optical lenses. The assembly force of the lens's fastening bolts is then determined based on existing processes. After normalization, this force is input into an assembly error prediction neural network model. The output, after denormalization, represents the assembly error of the optical lens due to the assembly force. The assembly errors are linearly superimposed on the ideal misalignment to calculate the actual adjustment of the optical lens. The lens's position is then adjusted using a precision motion actuator. Finally, an interferometer is used to verify that the alignment results meet the required imaging quality. If so, the force-position coordinated alignment process for the reflective lens is complete. If not, the alignment process is repeated until the required quality is met.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical method of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical method to deviate from the spirit and scope of the technical method of the present invention.

Claims

1. A neural network-based method for the coordinated force and position adjustment of optical lenses, characterized in that: The following steps are involved: Step 1: Establish a reflective optical system lens model based on the actual model and assembly process of the optical lens to be assembled; Step 2: Perform cyclic simulation on the reflective optical system lens model and output the lens simulation results; Step 3: Arrange the lens simulation results and construct an assembly force-assembly error dataset, wherein the assembly force-assembly error dataset includes assembly force combinations and assembly posture errors; Step 4: Construct an assembly error prediction model, train and save the assembly force-assembly error dataset to obtain a target assembly error prediction model; Step 5: Detect the Zernike coefficient of the current lens, calculate the target misalignment of the lens, use the assembly force combination applied by the current lens as the input of the target neural network model, and obtain the assembly posture error of the current lens; Step 6: Compensate for the assembly posture error based on the target misalignment to obtain the actual adjustment amount, completing the force-position coordinated control of the reflective optical system.

2. The neural network-based force-position coordinated adjustment method for optical lenses according to claim 1, characterized in that: The specific contents of establishing the reflective optical system lens model in step 1 also include: The allowable assembly force range during the assembly process is determined, the assembly force of the bolts required for assembling the optical lens is randomly generated within the assembly force range, and the assembly force of each bolt is combined to obtain an assembly force group.

3. The neural network-based optical lens force-position coordinated adjustment method according to claim 2, characterized in that: The assembly force of each bolt is formed in the assembly force range by a normal distribution or uniform distribution law.

4. The neural network-based force-position coordinated adjustment method for optical lenses according to claim 2, characterized in that: In step 2, different assembly force groups are set to perform cyclic simulation on the reflective optical system lens model; The lens simulation results include lens deformation and rigid body displacement.

5. The neural network-based force-position coordinated adjustment method for optical lenses according to claim 4, characterized in that: The assembly error prediction model is constructed based on a neural network model, and the specific contents of the target assembly error prediction model obtained by training and saving the assembly force-assembly error dataset include: During the training process, the assembly force-assembly error dataset is divided into a training set and a test set for training the assembly error prediction model.

6. The neural network-based force-position coordinated adjustment method for optical lenses according to claim 5, characterized in that: The assembly force combination is used as the input of the neural network model, and the assembly posture error is used as the output of the neural network model.

7. The neural network-based force-position coordinated adjustment method for optical lenses according to claim 4, characterized in that: The specific content of the assembly error prediction model training is: 1) Train the weights and bias parameters of each neuron node in the neural network; 2) Optimize the hyperparameters of the neural network to obtain the target hyperparameter group; 3) Save the weights, bias parameters and target hyperparameter groups of each neuron node of the neural network.

8. The neural network-based force-position coordinated adjustment method for optical lenses according to claim 1, characterized in that: The actual adjustment amount obtained in step six is the displacement of the lens itself in five degrees of freedom. The adjustment motion actuator where the lens is located should calculate the corresponding actual displacement amount according to the lens installation conditions and tooling size during actual adjustment to achieve posture adjustment of the optical lens.

Images