Coherent combination vortex beam wavefront aberration prediction method based on pre-correction network
Through deep learning -based pre -learning network prediction and correction CSV beam wave ahead, solve the problem of laser phase control arrays in the turbulent environment in turbulent environments, and improve the purity of the beam mode and the real -time system of the system.
Patent Information
- Application Number
- CN202411443491.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-16
- Publication Date
- 2025-01-24
AI Technical Summary
In turbulent environments, the time delay error of the laser phased array system causes the mode purity of the CSV beam to degrade, affecting its practical application.
Adopt a pre -profile network based on deep learning. By building a CSV beam atmospheric transmission model and a pre -pre -school neural network model, the correlation data of multiple historical light spots changes data predicting the evolution trend before the beam wave of the beam wave wave wave wave wave wave wave wave wave wave wave wave wave wave wave wave waves, to realize Poor real -time prediction and correction.
It effectively overcomes the delay delay of the phased array system, improves the mode purity and beam quality of the CSV beam during transmission, and improves the real-time performance and wavefront correction capabilities of the system.
Smart Images
Figure CN119358398A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of vortex beam wave technology, and in particular to a coherent synthetic vortex beam wavefront aberration prediction method based on a pre-correction network. Background Art
[0002] In recent years, with the rapid development of laser phased array adaptive optics technology, a new approach has been provided for the generation of high-power coherent synthetic vortex (CSV) beams and wavefront correction under turbulent conditions. The light field control of multiple medium-power laser beams can be achieved through laser phased arrays, which can not only correct the wavefront aberration in atmospheric transmission, but also form customized high-energy orbital angular momentum mode vortex beams on the target surface. At present, laser phased arrays have made significant progress in the research on CSV beam generation and array laser wavefront aberration correction in turbulent environments. In 2018, Kolosov et al. performed light field control on a 6-unit laser phased array system and obtained CSV beams with topological charges of +1 and -1. In 2019, Zhi Dong et al. obtained a CSV beam in the focal plane of the transform lens after near-field light field control on a 6-unit laser phased array. In 2023, Zou Fan et al. also obtained a first-order CSV beam by using a 6-unit laser phased array to perform coherent synthesis followed by spiral phase modulation. However, the wavefront distortion introduced by atmospheric turbulence will degrade the mode purity of the CSV beam, which will affect its practical application. Based on deep neural networks, it is expected to quickly calculate the required near-field phase modulation amount of the far-field spot, thereby achieving wavefront correction. In 2020, Hou Tianyue et al. used deep learning methods to realize the generation of CSV beams based on a 19-unit laser phased array. In 2021, Dan Wang et al. realized the coherent synthesis of 81 fiber lasers. In 2022, the author realized the coherent synthesis of 61 fiber lasers based on deep learning methods.
[0003] However, the impact of the real-time performance of the phased array system on the wavefront correction capability cannot be ignored. In actual systems, the response delay of the phase modulation module will result in it being able to correct only the distorted wavefront of the CSV beam in the past. Although the closed-loop bandwidth can be improved by increasing the sampling frequency of the detector, this also brings about the problem of difficulty in adapting to weak light conditions and increased system hardware costs. If the future turbulence distribution can be predicted based on the past and current turbulence distribution and converted into the correction amount of the phase modulator, the delay error of the phased array system can be effectively offset, and the correction performance of the system can be improved at the same sampling frequency.
[0004] Therefore, a new solution to the above problems needs to be proposed. Summary of the invention
[0005] The purpose of the present invention is to provide a coherent synthetic vortex beam wavefront aberration prediction method based on a pre-correction network, and then obtain a pre-correction model based on deep learning, which utilizes the correlation of multiple sets of historical spot change data to predict the evolution trend of the unit beam wavefront distortion, so as to cope with the lag problem in the CSV beam wavefront correction and further solve the technical problems raised in the background technology.
[0006] To achieve the above object, the present invention provides the following technical solution: a coherent synthetic vortex beam wavefront aberration prediction method based on a pre-correction network, comprising at least the following steps:
[0007] S1: Build a CSV beam atmospheric transmission model to achieve dynamic simulation of the atmospheric turbulence phase screen;
[0008] S2: Building a pre-correction neural network model based on deep learning, wherein the pre-correction neural network model includes a residual network module, a channel attention mechanism module and a Conv LSTM module;
[0009] S3: The corresponding coherent synthetic vortex beam wavefront aberration prediction is realized through the CSV beam atmospheric transmission model and the correction neural network model.
[0010] Furthermore, the S1 at least comprises the following steps:
[0011] Assuming that there are M sub-beams in the array, the complex amplitude CSV beam of the equivalent far-field synthesized vortex light field in cylindrical coordinates can be expressed as:
[0012]
[0013] Where U far is the far-field complex amplitude of the CSV beam, A0, ω0, k, f, J l-TM (x) and R represent the amplitude, beam waist, wave vector, focal length of the lens, modified Bessel function, and radius of the laser array of the sub-beam, respectively;
[0014] It can be seen that the light field of the CSV beam is a complex Bessel vortex light field, which is a coherent superposition of vortex light fields carrying a topological charge of l-TM, where T is an integer, indicating the order of the vortex beam, and φturb is the turbulent phase screen, which can be modeled by the atmospheric frozen turbulence hypothesis theory;
[0015] By generating atmospheric turbulence training data sets with different turbulence intensities based on the von Karman turbulence statistical theory and the HV-57 refractive index structure constant model, the pre-correction capability of the model can be improved;
[0016] SLM is further used to simulate a 1 km near-ground CSV beam atmospheric transmission model based on von Karman turbulence statistical theory.
[0017] Furthermore, the application of the pre-correction neural network model comprises at least the following steps:
[0018] A prediction neural network is constructed based on the far-field spot and defocused spot data sets through a deep learning network to achieve phase prediction and compensation of CSV beam turbulence distortion;
[0019] Use the residual network and attention mechanism module to learn the turbulence information features of the corresponding frames from the first 6 frames of spot images and their defocused images;
[0020] Then, the Conv LSTM module predicts the sub-aperture phase information of the future 8th frame, thereby realizing the corresponding coherent synthetic vortex beam wavefront aberration prediction.
[0021] Furthermore, the setting of the residual network module includes at least the following steps:
[0022] A 3×3 convolution kernel is used in the main path, and in the downsampling module, the second convolution kernel is replaced by a 2×2 convolution kernel with a stride of 2;
[0023] For shortcut connections, when the number of channels of the input and output feature maps is equal, an identity mapping is used. If they are not equal, a 1×1 convolution kernel is used for channel transformation.
[0024] In the downsampling module, a bilinear sampling method with a scaling factor equal to 0.5 is applied;
[0025] And batch normalization technology is used to standardize the output of the middle layer of the network and make the training process more stable;
[0026] And we choose Mish as the activation function instead of the traditional ReLU because Mish shows better performance in stabilizing training and improving accuracy.
[0027] The expression of the Mish activation function is as follows:
[0028] f(x)=x·tanh(softplus(x)).
[0029] Furthermore, the setting of the Conv LSTM module includes at least the following steps:
[0030] The Conv LSTM module introduces a gate mechanism to control the flow and loss of features;
[0031] The Conv LSTM module includes a forget gate, an input gate and an output gate;
[0032] The forget gate is used to forget or discard some information, accept a long-term memory (the output from the previous unit module) and decide which part to keep and which part to forget;
[0033] The input gate determines which new information is stored;
[0034] The output part of the cell state is determined by the Sigmiod function of the output gate, and then the cell state is processed by the tanh layer;
[0035] MSE is used as the loss function, and its expression is as follows:
[0036]
[0037] in, stands for the relative piston distortion index, which represents the nth relative piston distortion predicted by our network, is the standard true value of the nth relative piston distortion.
[0038] Compared with the prior art, the present invention has the following beneficial effects:
[0039] The present invention collects spot data of the laser phased array system under different turbulent conditions, uses these data to train a deep neural network model to extract features from the input historical wavefront data, predicts future wavefront distortion, and verifies the accuracy and robustness of the model through a large amount of simulation and experimental data; based on the predicted wavefront distortion, real-time phase modulation is implemented to overcome the lag problem of the existing phase control system and ensure that the generated vortex light beam maintains high-quality mode purity during transmission; finally, the experimental data required for the SLM system generation is tested to verify the accuracy of the model;
[0040] The method of the present invention can effectively improve the real-time performance and wavefront correction capability of the laser phased array system by predicting and correcting the wavefront distortion using a deep learning model, and provide strong support for long-distance and high-precision CSV beam transmission. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.
[0042] Figure 1 Schematic diagram of turbulence phase screen at different times of the present invention;
[0043] Figure 2 This is a conceptual diagram of the experimental model of the present invention;
[0044] Figure 3 It is a principle framework diagram of the pre-correction network of the present invention;
[0045] Figure 4 It is the near-field phase distribution under the turbulence intensity of the present invention and the corresponding far-field spot image in the open-loop and correction states. DETAILED DESCRIPTION
[0046] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.
[0047] A method for predicting wavefront aberration of a coherent synthetic vortex beam based on a pre-correction network comprises at least the following steps:
[0048] S1: Build a CSV beam atmospheric transmission model to achieve dynamic simulation of the atmospheric turbulence phase screen, see Figure 1 and Figure 2 ;
[0049] S2: Build a pre-correction neural network model based on deep learning. The pre-correction neural network model includes a residual network module, a channel attention mechanism module, and a Conv LSTM module. Figure 3 ;
[0050] S3: The corresponding coherent synthetic vortex beam wavefront aberration prediction is realized through the CSV beam atmospheric transmission model and the correction neural network model.
[0051] S1 includes at least the following steps:
[0052] Assuming that there are M sub-beams in the array, the complex amplitude CSV beam of the equivalent far-field synthesized vortex light field in cylindrical coordinates can be expressed as:
[0053]
[0054] Where U far is the far-field complex amplitude of the CSV beam, A0, ω0, k, f, J l-TM (x) and R represent the amplitude, beam waist, wave vector, focal length of the lens, modified Bessel function, and radius of the laser array of the sub-beam, respectively;
[0055] It can be seen that the light field of the CSV beam is a complex Bessel vortex light field, which is a coherent superposition of vortex light fields carrying a topological charge of l-TM, where T is an integer, indicating the order of the vortex beam, and φturb is the turbulent phase screen, which can be modeled by the atmospheric frozen turbulence hypothesis theory;
[0056] By generating atmospheric turbulence training data sets with different turbulence intensities based on the von Karman turbulence statistical theory and the HV-57 refractive index structure constant model, the pre-correction capability of the model can be improved;
[0057] SLM is further used to simulate a 1 km near-ground CSV beam atmospheric transmission model based on von Karman turbulence statistical theory.
[0058] The application of the pre-corrected neural network model includes at least the following steps:
[0059] A prediction neural network is constructed based on the far-field spot and defocused spot data sets through a deep learning network to achieve phase prediction and compensation of CSV beam turbulence distortion;
[0060] Use the residual network and attention mechanism module to learn the turbulence information features of the corresponding frames from the first 6 frames of spot images and their defocused images;
[0061] Then, the Conv LSTM module predicts the sub-aperture phase information of the future 8th frame, thereby realizing the corresponding coherent synthetic vortex beam wavefront aberration prediction.
[0062] The setup of the residual network module includes at least the following steps:
[0063] A 3×3 convolution kernel is used in the main path, and in the downsampling module, the second convolution kernel is replaced by a 2×2 convolution kernel with a stride of 2;
[0064] For shortcut connections, when the number of channels of the input and output feature maps is equal, an identity mapping is used. If they are not equal, a 1×1 convolution kernel is used for channel transformation.
[0065] In the downsampling module, a bilinear sampling method with a scaling factor equal to 0.5 is applied;
[0066] And batch normalization technology is used to standardize the output of the middle layer of the network and make the training process more stable;
[0067] And we choose Mish as the activation function instead of the traditional ReLU because Mish shows better performance in stabilizing training and improving accuracy.
[0068] The expression of the Mish activation function is as follows:
[0069] f(x)=x·tanh(softplus(x)).
[0070] The setup of a Conv LSTM module consists of at least the following steps:
[0071] The Conv LSTM module introduces a gate mechanism to control the flow and loss of features;
[0072] The Conv LSTM module includes a forget gate, an input gate, and an output gate;
[0073] The forget gate is used to forget or discard some information. It accepts a long-term memory (the output from the previous unit module) and decides which part to keep and which part to forget.
[0074] The input gate decides what kind of new information to store;
[0075] The output part of the cell state is determined by the Sigmiod function of the output gate, and then the cell state is processed by the tanh layer;
[0076] MSE is used as the loss function, and its expression is as follows:
[0077]
[0078] in, stands for the relative piston distortion index, which represents the nth relative piston distortion predicted by our network, is the standard true value of the nth relative piston distortion.
[0079] This example simulates five different turbulence intensity values, namely D / r 0 are 1, 3, 7, 10 and 14. The wind speed in each set of data is calculated by the Buffton wind speed model.
[0080] like Figure 2 As shown, the laser phased array is a honeycomb arrangement array with 60 array elements, an array element sub-aperture diameter of 0.5mm, a center distance between adjacent sub-apertures of 0.55mm, a wavelength of 515nm, an effective pixel number of 666×666 of the spatial light modulator, a pixel size of 12μm, a refresh rate of 500Hz, and a focusing lens focal length of 500mm. Figure 1 From left to right in the figure are the turbulence phase screens at different times, such as t, t+Δt…t+6Δt, t+7Δt, t+8Δt, where Δt = 1 / 30s. Figure 1 It can be seen that the atmospheric turbulence phase screen evolves with the change of time within a cycle, realizing the dynamic simulation of the atmospheric turbulence phase screen.
[0081] The first-level network first downsamples the original far-field spot image and the defocused spot image, and processes the input into image data with 2 channels and a size of 32*32. After four residual module feature extraction processes, the feature map structure is transformed into a size of 2*2 with 516 channels, and then the output layer maps and calculates the output result with 126 channels and a size of 1. The 126 channels correspond to 126 sub-aperture phases in turn. Finally, the result is mapped to [0,2π] through the following formula, and finally converted into data with a size of 126*1 and a channel number of 1.
[0082] f(x)=(tanh(x)+1)·π
[0083] The secondary network measures the sub-aperture wavefront phase φ for 6 consecutive frames. t □φ t+5 As input, the data size is 126*1, 6 channels. The weight matrix M is calculated through the channel attention mechanism module. c , and then multiply it channel by channel with the feature map to get the weighted feature map Finally, the Conv LSTM module outputs the sub-aperture wavefront phase of the future 8th frame, whose size is 126*1*1.
[0084] Based on the above content, the following experiments are further proposed:
[0085] In order to fully verify the prediction model, 5000 sets of data were selected for testing, and the average value was taken as the simulation result for display. In the first set of tests, the generalization ability was tested using different D / r0, that is, distorted wavefront data sets under different atmospheric turbulence intensities that had never been trained by the network, and compared with the test performance of the non-prediction method. (The system was set to calculate 2 frame delays, and the non-prediction method used the aperture phase of the 6th frame to compensate the wavefront phase of the 8th frame.)
[0086]
[0087] The experimental results are as follows Figure 4 As shown in the figure, the near-field phase distribution under turbulence intensity and the corresponding far-field spot images of the open-loop and corrected states are shown. The D / r 0 The values are 1, 3, 7, 10, and 14 (arranged from left to right). The near-field phase distribution represented by the hexagonal array shows that as D / r 0 The severity of wavefront distortion increases with the growth of D / r 0 = 14 shows the strongest turbulence condition.
[0088] As shown in the spot image, this phase distortion directly affects the far-field beam quality. Under open-loop conditions, the mode purity (MP) value drops significantly, and D / r 0= 1 and 3, the MP is 12.19% and 12.34%, respectively, indicating a serious degradation of the beam quality. However, after applying the closed-loop neural network prediction, the far-field beam spot is significantly improved, and the mode purity values for all turbulence levels are always above 90%, including D / r 0 =The most extreme case of 14.
[0089] These results demonstrate that the predictive model is able to not only correct for near-field wavefront distortions, but also maintain high mode purity and beam quality in the far field, even under adverse atmospheric conditions. This predictive approach is a substantial improvement over conventional delay-based methods, providing real-time, adaptive corrections that improve system performance over a wide range of turbulence intensities.
[0090] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.
Claims
1. A coherent synthetic vortex beam wavefront aberration prediction method based on a pre-correction network, characterized by: At least the following steps are included: S1: Build a CSV beam atmospheric transmission model to achieve dynamic simulation of the atmospheric turbulence phase screen; S2: Building a pre-correction neural network model based on deep learning, wherein the pre-correction neural network model includes a residual network module, a channel attention mechanism module and a Conv LSTM module; S3: The corresponding coherent synthetic vortex beam wavefront aberration prediction is realized through the CSV beam atmospheric transmission model and the correction neural network model.
2. The method for predicting wavefront aberration of a coherent synthetic vortex beam based on a pre-correction network according to claim 1, characterized in that: The S1 at least comprises the following steps: Assuming that there are M sub-beams in the array, the complex amplitude CSV beam of the equivalent far-field synthesized vortex light field in cylindrical coordinates can be expressed as: Where U far is the far-field complex amplitude of the CSV beam, A0, ω0, k, f, J l-TM (x) and R represent the amplitude, beam waist, wave vector, focal length of the lens, modified Bessel function, and radius of the laser array of the sub-beam, respectively; It can be seen that the light field of the CSV beam is a complex Bessel vortex light field, which is a coherent superposition of vortex light fields carrying a topological charge of l-TM, where T is an integer, indicating the order of the vortex beam, and φturb is the turbulent phase screen, which can be modeled by the atmospheric frozen turbulence hypothesis theory; By generating atmospheric turbulence training data sets with different turbulence intensities based on the von Karman turbulence statistical theory and the HV-57 refractive index structure constant model, the pre-correction capability of the model can be improved; SLM is further used to simulate a 1 km near-ground CSV beam atmospheric transmission model based on von Karman turbulence statistical theory.
3. The method for predicting wavefront aberration of a coherent synthetic vortex beam based on a pre-correction network according to claim 1, characterized in that: The application of the pre-correction neural network model comprises at least the following steps: A prediction neural network is constructed based on the far-field spot and defocused spot data sets through a deep learning network to achieve phase prediction and compensation of CSV beam turbulence distortion; Use the residual network and attention mechanism module to learn the turbulence information features of the corresponding frames from the first 6 frames of spot images and their defocused images; Then, the Conv LSTM module predicts the sub-aperture phase information of the future 8th frame, thereby realizing the corresponding coherent synthetic vortex beam wavefront aberration prediction.
4. The method for predicting wavefront aberration of a coherent synthetic vortex beam based on a pre-correction network according to claim 1, characterized in that: The setting of the residual network module includes at least the following steps: A 3×3 convolution kernel is used in the main path, and in the downsampling module, the second convolution kernel is replaced by a 2×2 convolution kernel with a stride of 2; For short-circuit connections, when the number of channels of the input and output feature maps is equal, the identity mapping is used. If they are not equal, a 1×1 convolution kernel is used for channel transformation. In the downsampling module, a bilinear sampling method with a scaling factor equal to 0.5 is applied; And batch normalization technology is used to standardize the output of the middle layer of the network and make the training process more stable; And we choose Mish as the activation function instead of the traditional ReLU because Mish shows better performance in stabilizing training and improving accuracy. The expression of the Mish activation function is as follows: f(x)=x·tanh(softplus(x)).
5. The method for predicting wavefront aberration of a coherent synthetic vortex beam based on a pre-correction network according to claim 1, characterized in that: The setting of the Conv LSTM module includes at least the following steps: The Conv LSTM module introduces a gate mechanism to control the flow and loss of features; The Conv LSTM module includes a forget gate, an input gate and an output gate; The forget gate is used to forget or discard some information, accepting a long-term memory and deciding which part to keep and which part to forget; The input gate determines which new information is stored; The output part of the cell state is determined by the Sigmiod function of the output gate, and then the cell state is processed by the tanh layer; MSE is used as the loss function, and its expression is as follows: in, stands for the relative piston distortion index, which represents the nth relative piston distortion predicted by our network, is the standard true value of the nth relative piston distortion.