Reflector antenna single-input phase recovery method based on adversarial neural network
By quickly recovering the near-field phase of the reflective plane antenna by adversarial neural network, the problems of high computational volume and sensitivity to initial guesses are solved, and fast and accurate phase recovery is achieved.
Patent Information
- Application Number
- CN202411835510.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-13
- Publication Date
- 2025-04-04
AI Technical Summary
When traditional phase iteration methods deal with near-field phase recovery of reflective plane antennas, the calculation amount is large, and the phase information cannot be quickly restored. They are sensitive to initial guesses and are easily trapped in the local optimal solution.
Adversarial neural network (GAN) is used to quickly recover the near-field phase of the reflective plane antenna, and the CGAN neural network is trained through paired data sets, and the network parameters are adjusted using the loss function to output the diameter field phase of the reflective plane antenna.
It is realized that when only one far-field amplitude is obtained, the near-field phase of the reflective plane antenna is quickly restored, the calculation amount is reduced, the detection speed is improved, and the trap of local optimal solution is avoided.
Smart Images
Figure CN119761423A_ABST
Abstract
Description
Technical Field
[0001] The method relates to the technical field of reflector antenna phase measurement, and in particular to a reflector antenna single-input phase recovery method based on an adversarial neural network. Background Art
[0002] The surface condition of the antenna not only depends on the panel processing and assembly accuracy, but is also affected by gravity, wind load and temperature deformation. Since wind load and temperature deformation have a certain degree of quasi-real-time nature, this requires that the antenna surface detection method not only has high detection accuracy but also has a fast detection speed. The traditional phase iteration method has a huge computational problem and cannot quickly recover the near-field phase information.
[0003] The performance of the traditional phase iteration method GS algorithm depends to a large extent on the selection of the initial guess. If the initial guess deviates from the vicinity of the optimal solution, the algorithm is likely to fall into the trap of the local optimal solution and cannot further explore the global optimal solution. The sensitivity to the initial guess makes the GS algorithm face certain limitations in practical applications.
[0004] Multi-image phase recovery algorithms recover the phase by utilizing information from multiple images. The Misell algorithm is a representative of this type of algorithm. It effectively reduces the algorithm's dependence on the initial guess by cleverly designing an iterative strategy. When dealing with phase recovery problems, the Misell algorithm can more flexibly adjust the search path, thereby reducing the possibility of falling into a local minimum. Although multi-image phase recovery algorithms such as the Misell algorithm have achieved remarkable results in improving reconstruction accuracy and reducing sensitivity to initial guesses, these algorithms currently do not provide rigorous mathematical proofs to ensure that they can find the global optimal solution; this means that although these algorithms perform well in practical applications, their theoretical completeness still needs to be further verified.
[0005] These traditional phase iteration methods face huge computational challenges when dealing with such problems; these methods usually require iterative calculations to gradually approach the true value of the antenna near-field phase information, but this process often consumes a lot of computing resources, resulting in slow detection speed and unable to meet the needs of fast response.
[0006] Therefore, in order to improve the efficiency and accuracy of antenna surface detection, it is urgent to develop new detection technologies or optimize existing algorithms to reduce the amount of calculation and increase the detection speed, thereby achieving rapid and accurate monitoring of the antenna surface condition.
[0007] To solve the above problems, the present invention provides a technical method for restoring antenna phase by using an adversarial neural network, aiming to quickly restore the near-field phase when obtaining the far-field amplitude of the reflector antenna. Summary of the invention
[0008] The method of the present invention uses paired data sets to train an adversarial neural network. Once the network is trained, the near-field phase can be quickly obtained when only one far-field amplitude is obtained. The method includes:
[0009] Use polynomials to simulate the reflector antenna aperture field phase, and combine with Gaussian distribution to form the reflector antenna aperture field distribution;
[0010] The far-field distribution of the reflector antenna is obtained by using the fast Fourier transform and the aperture field distribution of the reflector antenna, and the far-field amplitude of the reflector antenna is obtained. The aperture field phase and the far-field amplitude are paired to form a data set.
[0011] Design an adversarial neural network for phase recovery of reflector antennas;
[0012] The data set is sent to the designed adversarial neural network, and the network parameters are adjusted using the loss function. The network finally outputs the aperture field phase of the reflector antenna to complete the aperture field phase recovery of the reflector antenna.
[0013] Based on the above method, the use of polynomials to simulate the aperture field phase of the reflector antenna and the combination of Gaussian distribution to form the aperture field distribution of the reflector antenna include:
[0014] The polynomial simulates the phase of the aperture field of the reflector antenna. The polynomial refers to the Zernike polynomial, which uses 3rd order and 10 terms. The phase polynomial description equation is as follows:
[0015] φ(x,y)=a1Z1(x,y)+a2Z2(x,y)+a3Z3(x,y)+...+a 10 Z 10 (x,y)
[0016] The aperture field description equation is as follows:
[0017] E(x,y)=A·exp{i·φ(x,y)}
[0018] Based on the above method, the far-field distribution of the reflector antenna is obtained by using the fast Fourier and the aperture field distribution of the reflector antenna, and the far-field amplitude of the reflector antenna is obtained. The aperture field phase is paired with the far-field amplitude, and the data set includes:
[0019] The far-field distribution of the reflector antenna is calculated using the fast Fourier transform and the reflector antenna aperture field distribution as follows: F(u,v)=F{E(x,y)}, and the far-field amplitude is calculated as follows: I=|F(u,v)| 2 .
[0020] The far-field intensity image and aperture field phase image are represented as 8-bit grayscale images. To facilitate network training, the data set is normalized by minimum-maximum.
[0021] Based on the above method, the designed adversarial neural network includes:
[0022] Adversarial neural network refers to CGAN neural network, which makes full use of the "adversarial process" to train two neural networks. The two networks will compete with each other until an ideal equilibrium state is reached. One of the neural networks is called the generator network G(z|y), which contains 6 downsampling layers, 5 upsampling layers and an output layer. The downsampling layers and upsampling convolutions are activated by the LeakyReLU-Relu function and then connected in sequence. The output layer is activated by the Tanh function.
[0023] Another neural network is called the discriminator network D(x|y), which consists of two downsampling layers and an output layer. The downsampling layer convolution is activated by the LeakyReLU-Relu function and then connected in sequence. The output layer is activated by the Sigmoid function.
[0024] For each upsampling step, the corresponding upsampling layer is called to combine the output of the current layer with the feature map from the previous downsampling layer. Feature fusion helps the network integrate previously extracted details (such as features from higher resolution) into the current generated image.
[0025] Based on the above method, the loss function is specifically expressed as the sum of the following three loss functions:
[0026] For CGAN, the loss is expressed as:
[0027]
[0028] After introducing the conditional variable y, the discriminator and generator of CGAN distinguish the authenticity of real data and generated data under given conditions and generate corresponding data. The inputs of both the generator G and the discriminator D contain additional information y. In the present invention, the additional information y refers to the far-field amplitude, while the generated data x represents the near-field phase.
[0029] The mean square error (MSE) loss is expressed as:
[0030]
[0031] Represents the mean of the sum of squares of the differences between the predicted value f(x) and the target value y.
[0032] The SSIM loss is expressed as:
[0033]
[0034] where R and G represent the reconstructed phase image and the true phase image, respectively, and μ and σ are the mean and standard deviation, respectively. Extremely small constant parameters C2 and C2 are required to ensure numerical stability and prevent division by zero error.
[0035] The weighted sum of the five loss functions is expressed as:
[0036] l1=W BCE L BCE +W SSIM (1-l SSIM (R,G))+W MSE L MSE (R,G)
[0037] Among them, each loss function W MSE , W MSE , W BCE The corresponding weights are set to 2, 1, and 16 to balance the loss ratio. The Adam optimizer is used to optimize the loss function, and the learning rate is reduced by monitoring the trend of the validation loss until the losses of the training and validation datasets converge stably. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] In order to more clearly illustrate the specific implementation methods of the present invention or the technical methods in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0039] Figure 1 This is a case flow chart of the present invention.
[0040] Figure 2 Adversarial neural network architecture designed for this method.
[0041] Figure 3 Box plot for the test set
[0042] Figure 4 Box plot for the noisy test set DETAILED DESCRIPTION
[0043] In order to make the purpose, technical method and advantages of the embodiments of the present invention clearer, the technical method of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0044] To facilitate understanding of this embodiment, the embodiment of the present invention is described in detail below.
[0045] Embodiment 1:
[0046] Reference Figure 2 Once the network is trained, the near-field phase can be quickly obtained by only obtaining a far-field amplitude. The steps include the following:
[0047] Step S1, using a polynomial to simulate the aperture field phase of the reflector antenna, and combining it with a Gaussian distribution to form the aperture field distribution of the reflector antenna;
[0048] Step S2, using fast Fourier transform and the aperture field distribution of the reflector antenna to obtain the far field distribution of the reflector antenna, and obtain the far field amplitude of the reflector antenna, and pair the aperture field phase with the far field amplitude to form a data set;
[0049] Step S3: Designing an adversarial neural network
[0050] Step S4, sending the data set to the designed adversarial neural network, using the loss function to adjust the network parameters, the network finally outputs the aperture field phase of the reflector antenna, and completes the aperture field phase recovery of the reflector antenna;
[0051] Step S5: adding noise to the intensity image, and then testing the network model with intensity images of different noise levels to verify the network's resistance to noise.
[0052] Step S6: Use the GS algorithm to compare the adversarial neural network to prove the effectiveness and rapid responsiveness of the network in recovering the phase.
[0053] In step S1, the following contents are specifically included:
[0054] S11, polynomial simulation of the reflector antenna aperture field phase, the polynomial refers to the Zernike polynomial, using 3rd order 10 terms, the embodiment coefficients are shown in Table 1, and the phase polynomial description equation is as follows:
[0055] φ(x,y)=a1Z1(x,y)+a2Z2(x,y)+a3Z3(x,y)+...+a 10 Z 10 (x,y) (1)
[0056] Table 1: Experimental Zernike coefficients
[0057]
[0058] S12. The simulated antenna parameters are: main surface diameter 25 (m); secondary surface diameter: 2.5 (m); main surface focal length: 16 (m); equivalent focal length 96 (m); edge taper -16 (db); wavelength: 0.01 (m). The aperture field description equation is as follows: E(x, y) = A·exp{i·φ(x, y)}.
[0059] In step S2, the following contents are specifically included:
[0060] S21. The far-field distribution of the reflector antenna is calculated using the fast Fourier transform and the reflector antenna aperture field distribution as follows: F(u,v)=F{E(x,y)}, and the far-field amplitude is calculated as follows: I=|F(u,v)| 2 .
[0061] S22. The acquired dataset consists of 10,000 image pairs for training the proposed network, of which 9,500 pairs are used as training datasets and the rest as test datasets. The far-field intensity images and aperture field phase images are represented in 8-bit grayscale images. To facilitate network training, the dataset is normalized by min-max.
[0062] In step S3, the designed adversarial neural network is as follows:
[0063] S31. Adversarial neural network makes full use of the "adversarial process" to train two neural networks. The two networks will compete with each other until they reach an ideal equilibrium state. One of the neural networks is called the generator network G(z|y), which contains 6 downsampling layers, 5 upsampling layers and an output layer. The downsampling layers and upsampling convolutions are activated by the LeakyReLU-Relu function and then connected in sequence. The output layer is activated by the Tanh function.
[0064] S32. Another neural network is called the discriminator network D(x|y), which consists of two downsampling layers and an output layer. The downsampling layer convolution is activated by the LeakyReLU-Relu function and then connected in sequence. The output layer is activated by the Sigmoid function.
[0065] S33. For each upsampling step, call the corresponding upsampling layer to combine the output of the current layer with the feature map from the previous downsampling layer. Feature fusion can help the network integrate previously extracted details (such as features from higher resolution) into the current generated image.
[0066] In step S4, the loss function mentioned is specifically the sum of the following three loss functions:
[0067] S41. For CGAN loss, the performance is:
[0068]
[0069] After introducing the conditional variable y, the discriminator and generator of CGAN distinguish the authenticity of real data and generated data under given conditions and generate corresponding data. The inputs of both the generator G and the discriminator D contain additional information y. In the present invention, the additional information y refers to the far-field amplitude, while the generated data x represents the near-field phase.
[0070] S42, Mean Square Error (MSE) loss is expressed as:
[0071]
[0072] Represents the mean of the sum of squares of the differences between the predicted value f(x) and the target value y.
[0073] S43 and SSIM losses are expressed as:
[0074]
[0075] where R and G represent the reconstructed phase image and the true phase image, respectively, and μ and σ are the mean and standard deviation, respectively. Extremely small constant parameters C2 and C2 are required to ensure numerical stability and prevent division by zero error.
[0076] S44. The weighted sum of the five loss functions is expressed as:
[0077] l1=W BCE L BCE +W SSIM (1-l SSIM (R,G))+W MSE L MSE (R,G)
[0078] Among them, each loss function W BCE , W SSIM , W MSE , the corresponding weights are set to 2, 1, and 16 respectively to balance the loss ratio. The Adam optimizer is used to optimize the loss function. The learning rate is reduced by monitoring the trend of the validation loss until the loss of the training dataset and the validation dataset converges stably. The training process stops. The average SSIM and MSE of phase recovery on the test dataset are as follows: Figure 3 shown.
[0079] In step S5, the following contents are specifically included:
[0080] S51. The noise mentioned is Poisson noise, and its signal-to-noise ratio ranges from 28db to 29db.
[0081] S52. Poisson noise will significantly destroy the low-frequency part of the central area of the far-field intensity image.
[0082] The content of the GS algorithm mentioned in step S6 is as follows:
[0083] S61. Establish aperture field distribution, E(x,y) = A·exp{i·φ(x,y)}
[0084] S62, the aperture field Fourier transform obtains the far field distribution,
[0085] S63. Use the far-field distribution to obtain the far-field amplitude, I = |F(u,v)| 2
[0086] S64. Use normal distribution to reconstruct the aperture field distribution, E(x,y) = A·exp{i·φ′(x,y)}
[0087] S65, Fourier transform of aperture field to obtain far-field distribution, replace far-field amplitude, and reconstruct far-field distribution
[0088] S66. Inverse Fourier transform to obtain the near-field distribution, E′(x,y)=F{F′(u,v)}, and obtain the near-field phase φ(x,y)=angle(E′(x,y))
[0089] Experimental results and analysis: The simulation software used in this method is pycharm. All training and testing are performed on the RTX3080x2 (20GB) GPU, and the framework of the deep learning model is built using PyTorch.
[0090] like Figure 1 As shown in the figure, this technology tests the network model with noise and intensity images with different noise levels (SNR = 28-29db). The average SSIM and MSE of phase recovery on the noise test dataset are shown in Figure 4 As shown, it shows that CGAN still achieves good phase recovery in the case of Poisson noise.
[0091] The input and output of the simulation program are both BMP format images. To prove the superiority of the method, a control experiment was designed. The GS algorithm, which also uses a single amplitude input, was selected as the comparison method. The selected Zernike coefficients are shown in Table 1. After obtaining the aperture field phase, the aperture field is constructed by combining the Gaussian function, and the far field is obtained by Fourier transform. The far field amplitude is substituted into the trained network and the GS algorithm for 1000 iterations to obtain the near field phase, and the recovered result is compared with the original phase. The experimental results are shown in Table 2. It can be seen that compared with the GS algorithm, the algorithm designed by this method has a fast recovery speed, and the obtained near field phase MSE and SSIM indicators are high, achieving the purpose of recovering the phase. Therefore, the near field phase recovery algorithm provided by this method can achieve fast response while performing effective recovery.
[0092] Table 2: Comparison of test results
[0093]
[0094] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical methods of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical methods recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical methods deviate from the spirit and scope of the technical methods of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims
1. A single-input phase recovery method for a reflector antenna based on an adversarial neural network, characterized in that: The method comprises: Use polynomials to simulate the reflector antenna aperture field phase, and combine with Gaussian distribution to form the reflector antenna aperture field distribution; The far-field distribution of the reflector antenna is obtained by using the fast Fourier transform and the aperture field distribution of the reflector antenna, and the far-field amplitude of the reflector antenna is obtained. The aperture field phase and the far-field amplitude are paired to form a data set. Design an adversarial neural network for phase recovery of reflector antennas; The data set is sent to the designed adversarial neural network, and the network parameters are adjusted using the loss function. The network finally outputs the aperture field phase of the reflector antenna to complete the aperture field phase recovery of the reflector antenna.
2. According to claim 1, a single-input phase recovery technology for a reflector antenna using an adversarial neural network is characterized in that: The use of polynomials to simulate the aperture field phase of the reflector antenna and combining the Gaussian distribution to form the aperture field distribution of the reflector antenna includes: The polynomial simulates the phase of the aperture field of the reflector antenna. The polynomial refers to the Zernike polynomial, which uses 3rd order and 10 terms. The phase polynomial description equation is as follows: φ(x,y)=a1Z1(x,y)+a2Z2(x,y)+a3Z3(x,y)+…+a 10 Z 10 (x,y) The aperture field description equation is as follows: E(x,y)=A·exp{i·φ(x,y)}.
3. According to claim 1, a single-input phase recovery technology for a reflector antenna using an adversarial neural network is characterized in that: The fast Fourier transform and the aperture field distribution of the reflector antenna are used to obtain the far field distribution of the reflector antenna, and the far field amplitude of the reflector antenna is obtained. The aperture field phase is paired with the far field amplitude to form a data set including: The far-field distribution of the reflector antenna is calculated using the fast Fourier transform and the reflector antenna aperture field distribution. The far-field distribution is obtained by F(u,v)=F{E(x,y)}, and the far-field amplitude is obtained by I=|F(u,v)| 2 get. The far-field intensity image and aperture field phase image are represented as 8-bit grayscale images. To facilitate network training, the data set is normalized by minimum-maximum.
4. According to claim 1, a single-input phase recovery technology for a reflector antenna using an adversarial neural network is characterized in that: The designed adversarial neural network includes: Adversarial neural network refers to CGAN neural network, which makes full use of the "adversarial process" to train two neural networks. The two networks will compete with each other until an ideal equilibrium state is reached. One of the neural networks is called the generator network G(z|y), which contains 6 downsampling layers, 5 upsampling layers and an output layer. The downsampling layers and upsampling convolutions are activated by the LeakyReLU-Relu function and then connected in sequence. The output layer is activated by the Tanh function. Another neural network is called the discriminator network D(x|y), which consists of two downsampling layers and an output layer. The downsampling layer convolution is activated by the LeakyReLU-Relu function and then connected in sequence. The output layer is activated by the Sigmoid function. For each upsampling step, the corresponding upsampling layer is called to combine the output of the current layer with the feature map from the previous downsampling layer. Feature fusion helps the network integrate previously extracted details (such as features from higher resolution) into the current generated image.
5. According to claim 1, a single-input phase recovery technology for a reflector antenna using an adversarial neural network is characterized in that: The loss function is specifically expressed as the sum of the following three loss functions: For CGAN, the loss is expressed as: After introducing the conditional variable y, the discriminator and generator of CGAN distinguish the authenticity of real data and generated data under given conditions and generate corresponding data. The inputs of both the generator G and the discriminator D contain additional information y. In the present invention, the additional information y refers to the far-field amplitude, while the generated data x represents the near-field phase. The mean square error (MSE) loss is expressed as: Represents the mean of the sum of squares of the differences between the predicted value f(x) and the target value y. The SSIM loss is expressed as: where R and G represent the reconstructed phase image and the true phase image, respectively, and μ and σ are the mean and standard deviation, respectively. Extremely small constant parameters C2 and C2 are required to ensure numerical stability and prevent division by zero error. The weighted sum of the five loss functions is expressed as: l1=W BCE L BCE +W SSIM (1-l SSIM (R,G))+W MSE L MSE (R,G) Among them, each loss function W MSE , W MSE , W BCE The corresponding weights are set to 2, 1, and 16 to balance the loss ratio. The Adam optimizer is used to optimize the loss function, and the learning rate is reduced by monitoring the trend of the validation loss until the losses of the training and validation datasets converge stably.