Deformable mirror control method and device based on deep learning

Abstract

According to the deformable mirror control method and device based on deep learning, the deformable mirror can be efficiently and precisely controlled to precisely compensate the to-be-measured surface according to the input complex two-dimensional surface shape data, and high-precision dynamic measurement of the complex surface shape is achieved. The method comprises the following steps: (1) acquiring a data set; (2) establishing a neural network control model framework; (3) fully training the neural network control model; and (4) testing model control precision.

Description

Technical Field

[0001] The present invention relates to the technical field of photoelectric detection, and in particular to a deformable mirror control method based on deep learning, and a deformable mirror control device based on deep learning. Background Art

[0002] Deformable mirror (DM), also known as wavefront corrector, is one of the core components in adaptive optical systems and has important applications in astronomical observation, biomedicine, laser beam shaping, long-distance communication, free-form surface detection and other fields.

[0003] The deformable mirror mainly includes a surface composed of a voltage-controlled driver and a flexible metal layer. By setting the control driver voltage according to the surface shape information to be compensated, the DM surface shape can be controlled to change the optical path of wavefront transmission to dynamically correct the wavefront of the optical system. It can be seen that in the above process, the setting of the control voltage directly affects the wavefront correction accuracy of the deformable mirror. The role of the automatic control method of the deformable mirror is to automatically output the control voltage according to the characteristic vector of the target surface shape. Therefore, the accuracy and efficiency of the control method determine the accuracy and efficiency of the deformable mirror wavefront compensation. In the application of free-form surface detection, the free-form surface to be measured can be used as the target surface shape input control system to drive the deformable mirror, so that the deformable mirror can be used as a dynamic compensation mirror to achieve zero-difference interference with the surface to be measured, and the dynamic interference measurement of the free-form surface shape can be achieved without replacing the compensation mirror multiple times. At this time, the accuracy and efficiency of the deformable mirror control method also directly determine the accuracy and efficiency of free-form surface detection.

[0004] According to the above analysis, the high rate and high precision of the deformable mirror control method have an important impact on the development and application requirements of the deformable mirror. However, it is difficult for the existing deformable mirror control method to meet the requirements of control accuracy and efficiency at the same time. Commonly used deformable mirror control methods include optimization control methods and model-based control methods. When there is no prior knowledge at all, search optimization control methods can be used, such as hill climbing algorithm, random parallel gradient descent algorithm, etc. This method only optimizes the control voltage according to the compensation effect of the deformable mirror. Although it does not require modeling in advance, the convergence speed is slow, so the control efficiency is very low. In order to achieve higher control efficiency, model-based control methods are often used. However, it is very difficult to accurately obtain the control model using traditional methods. For example, the commonly used physical modeling solution method contains many necessary approximate conditions that limit the accuracy. The method of solving the control model by optimizing the influence function matrix cannot fully characterize the problems of driver coupling, nonlinear response, etc., and each control needs to be re-optimized, which is time-consuming and has low control efficiency. Therefore, a deformable mirror control method that can obtain the control model without approximation and optimization is necessary for efficient and high-precision control of the deformable mirror. Neural networks have a powerful ability to fit complex multi-dimensional nonlinear functions and can directly build high-precision control models. The model builds the data set based on the fitting learning of the data set without any approximate conditions. And the trained neural network control model can directly output the control voltage as long as the surface data is input, without the need for optimization process. Therefore, the control method based on neural networks has advantages in efficiency and accuracy.

[0005] However, the existing neural network-based models can only fit the control voltage to the input data of one-dimensional vectors or two-dimensional images with less information, such as the characterized one-dimensional surface feature vectors and circular spot images with similar distribution, but lack the nonlinear fitting ability for two-dimensional surface data with complex distribution and sufficient high-frequency information, such as the two-dimensional interference data of complex free-form surfaces. This makes it difficult to meet the actual needs of applications such as dynamic measurement of free-form surfaces that require control accuracy of deformable mirrors. Summary of the invention

[0006] In order to overcome the defects of the prior art, the technical problem to be solved by the present invention is to provide a deformable mirror control method based on deep learning, which can efficiently and accurately control the deformable mirror to accurately compensate the measured surface according to complex two-dimensional input surface data, thereby realizing high-precision dynamic measurement of complex surface shapes.

[0007] The technical solution of the present invention is: the deformable mirror control method based on deep learning comprises the following steps:

[0008] (1) Collecting data sets: Control the deformable mirror in various modes, record the control voltage as the output of the data set, use the interference optical path to measure the two-dimensional interference data of the deformable mirror under the corresponding control voltage, and then use the algorithm to remove the interference caused by the adjustment process. The processed results are recorded in a table as the input of the data set;

[0009] (2) Establish a neural network control model framework: According to the two-dimensional information feature extraction effect, select a suitable convolutional network framework to fully extract the key information in the two-dimensional surface data. For the deep network model, introduce the identity mapping and residual operation to prevent network degradation;

[0010] (3) Fully train the neural network control model: dynamically adjust the training hyperparameters in stages, observe the convergence state of the loss function curve, and when the loss function has converged in the current stage, optimize the training parameters of the next stage to jump out of the local optimum and improve the training effect and model training accuracy;

[0011] (4) Test model control accuracy: First determine the control accuracy indicators of the model in the two dimensions of voltage and surface shape, and then conduct a comprehensive analysis of the model accuracy through model testing and experimental verification. The voltage accuracy of the model test is determined by the difference between the model prediction value and the actual value through algorithm analysis. The surface shape accuracy of the experimental verification is given by the difference between the actual voltage and the predicted voltage after driving the deformable mirror. The surface shape accuracy is comprehensively evaluated in combination with the basic characteristics of the deformable mirror.

[0012] The present invention establishes a deformable mirror control model based on machine learning through the steps of collecting data sets, selecting and building models, model training, and model testing. When controlling the deformable mirror, no optimization is required, and the control efficiency is high. A convolutional network model is used to directly fit the nonlinear relationship between complex distributed two-dimensional input data and the control voltage. The input data does not need to be characterized by a characteristic polynomial, so it contains more high-frequency information, has a larger amount of information, and is easier to achieve high-precision control. The structure of the convolutional network is complex, and it is easier to fit complex nonlinear relationships, so that the model control accuracy is further improved. Therefore, the present invention has advantages in high-precision control of deformable mirrors to compensate for complex surface shapes and wavefronts, and can efficiently and high-precision control the deformable reflector to accurately compensate for the surface to be measured according to complex two-dimensional input surface shape data, thereby achieving high-precision dynamic measurement of complex surface shapes.

[0013] A deformable mirror control device based on deep learning is also provided, comprising:

[0014] An acquisition module is configured to control the deformable mirror in multiple modes, record the control voltage as the output of the data set, use an interference optical path to measure the two-dimensional interference data of the deformable mirror under the corresponding control voltage, and then remove the interference caused by the adjustment process through algorithm processing, and record the processed results in a table form as the input of the data set;

[0015] The modeling module is configured to select a suitable convolutional network framework to fully extract the key information in the two-dimensional surface data according to the effect of two-dimensional information feature extraction. For the deep network model, the operation of identity mapping and residual is introduced to prevent network degradation;

[0016] The training module is configured to dynamically adjust the training hyperparameters in stages, observe the convergence status of the loss function curve, and when the loss function has converged in the current stage, optimize the training parameters of the next stage to jump out of the local optimum, thereby improving the training effect and model training accuracy;

[0017] The accuracy control module is configured to first determine the control accuracy indicators of the model in the two dimensions of voltage and surface shape, and comprehensively analyze the model accuracy through two links: model testing and experimental verification. The voltage accuracy of the model test is determined by the difference between the model predicted value and the actual value through algorithm analysis. The surface shape accuracy of the experimental verification is given by the difference in surface shape after the deformable mirror is driven by the actual voltage and the predicted voltage. The surface shape accuracy is comprehensively evaluated in combination with the basic characteristics of the deformable mirror. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of a deformable mirror control method based on deep learning according to the present invention.

[0019] Figure 2 It is a schematic diagram of the identical residual block. Conv1 and Conv2 represent convolutional layer 1 and convolutional layer 2 respectively; BN1 and BN2 represent normalization layer 1 and normalization layer 2 respectively; Activation1 and Activation2 represent activation function layer 1 and activation function layer 2 respectively. DETAILED DESCRIPTION

[0020] like Figure 1 As shown, the deformable mirror control method based on deep learning includes the following steps:

[0021] (1) Collecting data sets: Control the deformable mirror in various modes, record the control voltage as the output of the data set, use the interference optical path to measure the two-dimensional interference data of the deformable mirror under the corresponding control voltage, and then use the algorithm to remove the interference caused by the adjustment process. The processed results are recorded in a table as the input of the data set;

[0022] (2) Establish a neural network control model framework: According to the two-dimensional information feature extraction effect, select a suitable convolutional network framework to fully extract the key information in the two-dimensional surface data. For the deep network model, introduce the identity mapping and residual operation to prevent network degradation;

[0023] (3) Fully train the neural network control model: dynamically adjust the training hyperparameters in stages, observe the convergence state of the loss function curve, and when the loss function has converged in the current stage, optimize the training parameters of the next stage to jump out of the local optimum and improve the training effect and model training accuracy;

[0024] (4) Test model control accuracy: First determine the control accuracy indicators of the model in the two dimensions of voltage and surface shape, and then conduct a comprehensive analysis of the model accuracy through model testing and experimental verification. The voltage accuracy of the model test is determined by the difference between the model prediction value and the actual value through algorithm analysis. The surface shape accuracy of the experimental verification is given by the difference between the actual voltage and the predicted voltage after driving the deformable mirror. The surface shape accuracy is comprehensively evaluated in combination with the basic characteristics of the deformable mirror.

[0025] The present invention establishes a deformable mirror control model based on machine learning through the steps of collecting data sets, selecting and building models, model training, and model testing. When controlling the deformable mirror, no optimization is required, and the control efficiency is high. A convolutional network model is used to directly fit the nonlinear relationship between complex distributed two-dimensional input data and the control voltage. The input data does not need to be characterized by a characteristic polynomial, so it contains more high-frequency information, has a larger amount of information, and is easier to achieve high-precision control. The structure of the convolutional network is complex, and it is easier to fit complex nonlinear relationships, so that the model control accuracy is further improved. Therefore, the present invention has advantages in high-precision control of deformable mirrors to compensate for complex surface shapes and wavefronts, and can efficiently and high-precision control the deformable reflector to accurately compensate for the surface to be measured according to complex two-dimensional input surface shape data, thereby achieving high-precision dynamic measurement of complex surface shapes.

[0026] Preferably, in the step (1), the control method includes a weight distribution method of multiple drivers and a limit range of the total voltage to produce deformable mirror shapes of different types and amplitudes, which are selected according to the deformable mirror driving mechanism and the characteristics of the wavefront to be compensated; the interference optical path is a Fizeau type interference optical path; the algorithm processing uses the first three terms of the Zernike polynomial to characterize the surface shape to determine the interference components related to translation and tilt, and then subtract them from the original data.

[0027] Preferably, in step (2), the specific structural parameters of the network are optimized by using an optimization algorithm based on the fitting accuracy as the goal.

[0028] Preferably, in step (2), the network framework selects a residual network or a visual geometry group network VGGNet convolutional network; and the optimization algorithm selects a genetic algorithm or an annealing algorithm.

[0029] Preferably, in step (3), the training hyperparameters include learning rate, learning rate decay parameter, optimization criterion parameter, and batch training size training parameter.

[0030] Preferably, in the step (4), the voltage accuracy index is represented by the root mean square, mean absolute deviation and relative deviation of the model output value and the true value of the data set; the surface control accuracy index is the peak-to-peak value pv and root mean square rms of the difference surface shape; and the basic characteristics include the repeatability and stability of the deformable mirror.

[0031] A person skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes the steps of the above-mentioned embodiment method, and the storage medium can be: ROM / RAM, disk, optical disk, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a deformable mirror control device based on deep learning, which is usually represented in the form of functional modules corresponding to the steps of the method. The device includes:

[0032] An acquisition module is configured to control the deformable mirror in multiple modes, record the control voltage as the output of the data set, use an interference optical path to measure the two-dimensional interference data of the deformable mirror under the corresponding control voltage, and then remove the interference caused by the adjustment process through algorithm processing, and record the processed results in a table form as the input of the data set;

[0033] The modeling module is configured to select a suitable convolutional network framework to fully extract the key information in the two-dimensional surface data according to the effect of two-dimensional information feature extraction. For the deep network model, the operation of identity mapping and residual is introduced to prevent network degradation;

[0034] The training module is configured to dynamically adjust the training hyperparameters in stages, observe the convergence status of the loss function curve, and when the loss function has converged in the current stage, optimize the training parameters of the next stage to jump out of the local optimum, thereby improving the training effect and model training accuracy;

[0035] The accuracy control module is configured to first determine the control accuracy indicators of the model in the two dimensions of voltage and surface shape, and comprehensively analyze the model accuracy through two links: model testing and experimental verification. The voltage accuracy of the model test is determined by the difference between the model predicted value and the actual value through algorithm analysis. The surface shape accuracy of the experimental verification is given by the difference in surface shape after the deformable mirror is driven by the actual voltage and the predicted voltage. The surface shape accuracy is comprehensively evaluated in combination with the basic characteristics of the deformable mirror.

[0036] Preferably, in the acquisition module, the control method includes a weight distribution method of multiple drivers and a limit range of the total voltage to produce deformable mirror shapes of different types and amplitudes, which are selected according to the deformable mirror driving mechanism and the characteristics of the wavefront to be compensated; the interference optical path is a Fizeau-type interference optical path; the algorithm processing uses the first three terms of the Zernike polynomial to characterize the surface shape to determine the interference components related to translation and tilt, and then subtract them from the original data.

[0037] Preferably, in the modeling module, the network framework selects a residual network or a visual geometry group network VGGNet; and the optimization algorithm selects a genetic algorithm or an annealing algorithm.

[0038] Preferably, in the training module, the training hyperparameters include learning rate, learning rate decay parameter, optimization criterion parameter, and batch training size training parameter.

[0039] Preferably, in the precision control module, the voltage accuracy index is represented by the root mean square, mean absolute deviation and relative deviation of the model output value and the true value of the data set; the surface shape control accuracy index is the peak-to-peak value pv and root mean square rms of the difference surface shape; and the basic characteristics include the repeatability and stability of the deformable mirror.

[0040] A specific implementation example of the present invention is described in detail below. This deformable mirror control method based on deep learning is implemented in the following manner:

[0041] The process of establishing a deformable mirror control method based on deep learning is as follows Figure 1 As shown, the specific implementation steps are:

[0042] Step 1: Collecting Datasets

[0043] In the example, the Zygo interferometer is used to measure the deformable mirror surface shape, and an attenuation plate is placed between the deformable mirror and the interferometer to improve the contrast of the interference fringes. The processed two-dimensional surface shape data saved in the example has a dimension of 223×223 and is stored in an Excel table format.

[0044] Step 2: Establish the neural network control model framework

[0045] In this example, the main framework of the network model is the residual convolutional network Resnet50, on which a fully connected layer with a dimension of 52 is added as the output. The model structure parameters are obtained by gradually optimizing the default parameters in the Keras network library as the starting values.

[0046] Step 3: Fully train the neural network control model.

[0047] In the example, the residual convolutional network is trained in two stages, with 300 training times in each stage. The training hyperparameters of the first stage are set to a learning rate of 7e-4 and a batch size of 32. In the second stage, the learning rate is reduced to 1e-6 and the batch size is increased to 128.

[0048] Step 4: Test model control accuracy

[0049] In the example, the voltage accuracy evaluation index uses the average absolute deviation and the accuracy defined by the ratio of relative deviation less than the threshold (2%, 5%). In the example, four target surface shapes with different variation amplitudes are selected from each aberration type to verify the surface accuracy, and the surface shapes with the same amplitude are selected to test the repeatability of the deformable mirror, and the effectiveness of the convolutional network model in controlling the deformable mirror is comprehensively evaluated.

[0050] The beneficial effects of the present invention are as follows:

[0051] 1. The deformable mirror control method and device based on deep learning disclosed in the present invention do not need to perform dimensionality reduction processing on the input two-dimensional surface data in advance, retain more high-frequency information, and are easier to achieve high-precision compensation of complex surface shapes.

[0052] 2. The deformable mirror control method and device based on deep learning disclosed in the present invention use a more complex deep convolutional network with residual optimization, which is easier to fit complex nonlinear relationships and further improves the model control accuracy.

[0053] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the protection scope of the technical solution of the present invention.

Claims

1. A deformable mirror control method based on deep learning, characterized in that: It includes the following steps: (1) Collecting data sets: Control the deformable mirror in various modes, record the control voltage as the output of the data set, use the interference optical path to measure the two-dimensional interference data of the deformable mirror under the corresponding control voltage, and then use the algorithm to remove the interference caused by the adjustment process. The processed results are recorded in a table as the input of the data set; (2) Establish a neural network control model framework: According to the two-dimensional information feature extraction effect, select a suitable convolutional network framework to fully extract the key information in the two-dimensional surface data. For the deep network model, introduce the identity mapping and residual operation to prevent network degradation; (3) Fully train the neural network control model: dynamically adjust the training hyperparameters in stages and observe the convergence state of the loss function curve. When the loss function has converged in the current stage, Optimize the training parameters for the next stage to escape the local optimum and improve the training effect and model training accuracy; (4) Test model control accuracy: First determine the control accuracy indicators of the model in the two dimensions of voltage and surface shape, and then conduct a comprehensive analysis of the model accuracy through model testing and experimental verification. The voltage accuracy of the model test is determined by the difference between the model prediction value and the actual value through algorithm analysis. The surface shape accuracy of the experimental verification is given by the difference between the actual voltage and the predicted voltage after driving the deformable mirror. The surface shape accuracy is comprehensively evaluated in combination with the basic characteristics of the deformable mirror.

2. The method for controlling a deformable mirror based on deep learning according to claim 1, characterized in that: In the step (1), the control method includes a weight distribution method of multiple drivers and a limit range of the total voltage to produce deformable mirror shapes of different types and amplitudes, which are selected according to the deformable mirror driving mechanism and the characteristics of the wavefront to be compensated; The interference optical path is a Fizeau type interference optical path; the algorithm processing is to use the first three terms of the Zernike polynomial to characterize the surface shape to determine the interference components related to translation and tilt, and then subtract them from the original data.

3. The method for controlling a deformable mirror based on deep learning according to claim 2, characterized in that: In the step (2), the specific structural parameters of the network are selected as optimal parameters using an optimization algorithm based on the fitting accuracy.

4. The method for controlling a deformable mirror based on deep learning according to claim 3, characterized in that: In the step (2), the network framework selects a residual network or a visual geometry group network VGGNet; the optimization algorithm selects a genetic algorithm or an annealing algorithm.

5. The method for controlling a deformable mirror based on deep learning according to claim 4, characterized in that: In step (3), the training hyperparameters include learning rate, learning rate decay parameter, optimization criterion parameter, and batch training size training parameter.

6. The method for controlling a deformable mirror based on deep learning according to claim 5, characterized in that: In the step (4), the voltage accuracy index is represented by the root mean square, mean absolute deviation and relative deviation of the model output value and the true value of the data set; the surface control accuracy index is the peak-to-peak value pv and root mean square rms of the difference surface shape; the basic characteristics include the repeatability and stability of the deformable mirror.

7. A deformable mirror control device based on deep learning, characterized in that: It includes: An acquisition module is configured to control the deformable mirror in multiple modes, record the control voltage as the output of the data set, use an interference optical path to measure the two-dimensional interference data of the deformable mirror under the corresponding control voltage, and then remove the interference caused by the adjustment process through algorithm processing, and record the processed results in a table form as the input of the data set; The modeling module is configured to select a suitable convolutional network framework to fully extract the key information in the two-dimensional surface data according to the effect of two-dimensional information feature extraction. For the deep network model, the operation of identity mapping and residual is introduced to prevent network degradation; The training module is configured to dynamically adjust the training hyperparameters in stages and observe the convergence state of the loss function curve. When the loss function has converged in the current stage, the training parameters of the next stage are optimized to jump out of the local optimum and improve the training effect and model training accuracy. The accuracy control module is configured to first determine the control accuracy indicators of the model in the two dimensions of voltage and surface shape, and conduct a comprehensive analysis of the model accuracy through two steps: model testing and experimental verification. The voltage accuracy of the model test is determined by the difference between the predicted value and the true value of the model through algorithm analysis. The surface shape accuracy of the experimental verification is given by the difference between the surface shape after the deformable mirror is driven by the true voltage and the predicted voltage. The surface shape accuracy is comprehensively evaluated in combination with the basic characteristics of the deformable mirror.

8. The deep learning-based deformable mirror control device according to claim 7, characterized in that: In the acquisition module, the control method includes a weight distribution method of multiple drivers and a limit range of the total voltage to produce deformable mirror shapes of different types and amplitudes, which are selected according to the deformable mirror driving mechanism and the characteristics of the wavefront to be compensated; The interference optical path is a Fizeau type interference optical path; the algorithm processing is to use the first three terms of the Zernike polynomial to characterize the surface shape to determine the interference components related to translation and tilt, and then subtract them from the original data.

9. The deep learning-based deformable mirror control device according to claim 8, characterized in that: In the modeling module, the network framework selects a residual network or a visual geometry group network VGGNet; the optimization algorithm selects a genetic algorithm or an annealing algorithm.

10. The deep learning-based deformable mirror control device according to claim 9, characterized in that : In the precision control module, the voltage accuracy index is represented by the root mean square, mean absolute deviation and relative deviation of the model output value and the true value of the data set; the surface shape control accuracy index is the peak-to-peak value pv and root mean square rms of the difference surface shape; the basic characteristics include the repeatability and stability of the deformable mirror.

Images