A far-field prediction method based on neural network near-field phase reconstruction

Abstract

The application discloses a far-field prediction method based on a neural network near-field phase reconstruction, and comprises the following steps: S1, near-field simulation is performed on an antenna to be measured, near-field amplitude and phase data of the antenna to be measured are acquired, a data set is constructed, and the data set is divided into a training set and a verification set; S2, a neural network for predicting the near-field phase according to the near-field amplitude is constructed, and the neural network is trained and verified through the training set to obtain a trained neural network; S3, for the antenna to be measured not in the data set, near-field amplitude information is acquired, the trained neural network is inputted to obtain a predicted near-field phase, and an equivalent dipole source of the antenna to be measured is calculated by using a least square method; and S4, according to the equivalent dipole theory, the far field of the antenna to be measured is calculated. The application adopts the neural network to perform near-field phase recovery, then constructs an equivalent dipole source, and finally realizes far-field prediction of the antenna to be measured, and has high accuracy and timeliness.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electromagnetic waves, in particular to a far-field prediction method based on neural network near-field phase reconstruction. BACKGROUND

[0002] Near-far field conversion is an important part of antenna testing, and its core is to reconstruct the far-field information of the antenna under test (AUT) through near-field measurement data. As electronic systems become more and more complex, the full-wave method has been unable to meet the needs of electromagnetic compatibility modeling due to the huge consumption of computing resources. In the past few decades, the theory of building equivalent radiation sources has developed rapidly. Among them, the equivalent dipole theory is widely used in the construction of equivalent sources. Generally, the establishment of equivalent sources requires both the amplitude and phase information of the near-field. However, in actual testing, it is difficult or even impossible to accurately obtain the phase information of the near-field. On this basis, the method of near-field phase recovery and reconstruction of equivalent sources from non-phase near-field is proposed. There are mainly two ideas, one is to make redundant measurements, such as measuring the field amplitude information of multiple near-field scanning surfaces; the second is to use optimization algorithms to minimize the error by iteratively equivalent dipoles or near-field phases. However, both of these methods have their own shortcomings, redundant measurements require more hardware resources, and it is difficult to ensure implementation under certain measurement conditions; optimization algorithms have high requirements for computer computing power, and timeliness is difficult to guarantee. SUMMARY

[0003] The purpose of the present application is to overcome the shortcomings of the prior art and provide a far-field prediction method based on neural network near-field phase reconstruction. The neural network is used for near-field phase recovery, and then the equivalent dipole source is constructed, and finally the far-field prediction of the antenna under test is realized, which has high accuracy and timeliness.

[0004] The purpose of the present application is achieved by the following technical scheme: a far-field prediction method based on neural network near-field phase reconstruction, comprising the following steps:

[0005] S1. Simulate the near-field of the antenna under test, obtain the near-field amplitude and phase data of the antenna under test, and construct a data set, divide the data set into a training set and a validation set;

[0006] S2. Construct a neural network for predicting the near-field phase from the near-field amplitude, and train the neural network through the training set to fit the relationship between the near-field amplitude and the near-field phase, and use the validation set to verify the trained neural network, and obtain the trained neural network after verification;

[0007] S3. For the antenna to be tested not in the data set, only the near-field amplitude information is obtained by simulation, the predicted near-field phase is obtained by inputting the trained neural network, so as to obtain the complete near-field amplitude and phase information, and the equivalent dipole source of the antenna to be tested is calculated by using the least square method;

[0008] S4. According to the equivalent dipole theory, the far field of the antenna to be tested is calculated.

[0009] Preferably, the step S1 comprises:

[0010] S101. Assuming that the antenna to be tested is a circular horn antenna, the horn face radius is 1.2λ, λ is the wavelength of electromagnetic wave, the near-field sampling surface is a spherical surface with a radius R and the center of the horn face as the origin, t×t sampling points are set at equal intervals between the elevation angle θ of 1-179° and the azimuth angle of 0-342°, and the number of sampling points is denoted as m=t×t;

[0011] When the near-field sampling is performed, the near-field amplitude data and the near-field phase data are collected, the near-field amplitude data includes radial component |E r , polar angle component |E θ and azimuth angle component , and the near-field phase data φ also includes radial component φ r , polar angle component φ θ and azimuth angle component Each sampling point has a group of amplitude and phase data;

[0012] S102. Sampling at equal intervals between 1.5-2.4 GHz to obtain p frequencies, and constructing an antenna CAD model at each frequency, repeating the step S101 under the antenna CAD model at each frequency to obtain the near-field amplitude data and the phase data at each frequency, and adding them to the same data set to complete the construction of the data set;

[0013] In the data set, there are p samples in total, each sample corresponds to an antenna CAD model at a frequency, and in each sample, the feature is the near-field amplitude of the horn antenna, the dimension is m×3, and the label is the near-field phase, the dimension is also m×3;

[0014] S103. The data set is divided into a training set and a validation set.

[0015] Preferably, in the step S2, the neural network CNN for predicting the near-field phase from the near-field amplitude is a convolutional neural network, and in the training process, the neural network is trained with each sample in turn:

[0016] In training the neural network with any one sample, an adaptive momentum optimizer is adopted, the near-field amplitude of txt3 is regarded as a three-channel image of pixel t x t, input into the CNN convolutional neural network, the predicted phase of t x t x 3 is output by the CNN neural network, the loss function is calculated by using the predicted phase of t x t x 3 and the near-field phase of t x t x 3 in the label, then according to the loss function, the gradient is calculated by back propagation and the CNN neural network is updated according to the gradient descent method;

[0017] After reaching the training round, the trained neural network is verified by using the verification set, and the trained neural network is obtained after verification.

[0018] Preferably, the step S3 comprises:

[0019] S301. Set the dipoles to be placed on the spherical surface with radius R1, and set t1 x t1 points equidistantly between 10-170° of θ and 0-340° of The number of dipoles is denoted as n=t1 x t1, each dipole P has a polar angle component P θ and an azimuthal angle component Two components;

[0020] S302. For the horn antenna not in the data set, the near-field amplitude is simulated, and in the r, θ, three directions, the vector of the near-field amplitude of m sampling points is constructed according to the simulation results, denoted as |E r | m×1 , |E θ | m×1 , Therefore, the near-field amplitude is denoted as

[0021] Wherein, represents the vector of the near-field amplitude of m sampling points in the r, θ, three directions; |E ri |, |E θi |, represents the amplitude of the i-th near-field sampling point in the r, θ, three directions, and i=1, 2…m;

[0022] |E| 3m×1 is input into the trained neural network to obtain the predicted near-field phase Wherein represents the vector obtained by arranging the phase of m near-field sampling points in the r, θ, three directions; represents the phase of the i-th near-field sampling point in the r, θ, three directions, respectively;

[0023] The complete information of near-field is obtained, denoted as where E r,m×1 , E θ,m×1 , represent the complete information of near-field of m sampling points in r, θ, three directions respectively:

[0024]

[0025] E ri , E θi , represent the component of complete information of the i-th near-field sampling point in r, θ, three directions, i = 1, 2…m;

[0026] S303. Arrange n dipoles according to θ, direction to obtain column vector

[0027] where represent the polar angle component and azimuth angle component of n dipoles, P θj , represent the polar angle component and azimuth angle component of the j-th dipole, j = 1, 2…n;

[0028] S304. Equivalent dipole theory uses dipole array to replace the antenna to be tested to represent its radiation characteristics;

[0029] The electric field E 2n×1 generated by the dipole array P 3m×1 in free space is obtained by coordinate transformation, and the relationship matrix is so that T 3m×2n × P 2n×1 = E 3m×1 , where represents the relationship between the electric field component and the dipole component;

[0030] The dielectric constant in free space is ε0, the magnetic permeability is μ0, and the wave impedance is The wave number is Let the coordinates of a near-field sampling point in the spherical coordinate system be The coordinates in the rectangular coordinate system are (x Ei , y Ei , z Ei ); the coordinates of a dipole placement point in the spherical coordinate system are The coordinates in the rectangular coordinate system are (x Pj , y Pj , z PjThe distance between them is denoted as r; when the positions of the near-field sampling point and the dipole are determined, T is calculated. 3m×2n Each element in;

[0031]

[0032]

[0033]

[0034] in,

[0035]

[0036] S305. Calculate the equivalent dipole source P using near-field information. 2n×1 Using the least squares method, the objective function is:

[0037] minF(P 2n×1 )=‖E 3m×1 -T 3m×2n P 2n×1 || 2

[0038] Least square solution of dipole vector It is obtained by using the form of matrix singular value decomposition.

[0039] Preferably, step S4 includes:

[0040] The far-field sampling surface is a sphere with radius R2, centered at the center of the antenna aperture. The sampling surface is used at elevation angles θ (0–180°) and azimuth angles... Set t at equal intervals between 0 and 360° ′ 1×t ′ There are 2 sampling points, and the number of sampling points is denoted as m. ′ =t ′ 1×t ′ 2;

[0041] The relationship matrix between the dipole array and the far-field sampling points is calculated and denoted as follows: The calculation method and T 3m×2n Similar, except the number of near-field sampling points m is changed to the number of far-field sampling points m. ′ , sampling point location or (x) Ei ,y Ei ,z Ei ), i = 1, 2…m changed to or (x) ′ Ei ,y ′ Ei ,z ′Ei ), i = 1, 2…m ′ The distance r between the corresponding field and the dipole also changes accordingly;

[0042] Using matrix equations The far-field calculations are completed, and the results are used as the far-field prediction results for the antenna under test.

[0043] The beneficial effects of this invention are: this invention uses a neural network to perform near-field phase recovery, then constructs an equivalent dipole source, and finally realizes far-field prediction of the antenna under test, which has high accuracy and timeliness. Attached Figure Description

[0044] Figure 1 This is a flowchart of the method of the present invention;

[0045] Figure 2 Diagram of bottleneck structures with strides of 1 and 2;

[0046] Figure 3 This is a schematic diagram illustrating the phase recovery principle.

[0047] Figure 4 A schematic diagram of an equivalent dipole;

[0048] Figure 5 To train the loss curve;

[0049] Figure 6 This is a graph showing the near-field calculation results;

[0050] Figure 7 This is a graph showing the real and predicted far-field conditions. Detailed Implementation

[0051] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0052] like Figure 1 As shown, a far-field prediction method based on near-field phase reconstruction using neural networks includes the following steps:

[0053] S1. Perform near-field simulation on the antenna under test to obtain the near-field amplitude and phase data of the antenna under test, and construct a dataset, dividing the dataset into a training set and a validation set;

[0054] The AUT uses a conical horn antenna, and FEKO is used for near-field simulation. The geometry of the conical horn antenna is frequency-dependent, with an aperture radius of 1.2λ. A total of 800 antenna samples were simulated by sampling at equal intervals from 1.5 to 2.4 GHz, obtaining its near-field amplitude and phase data. The near-field sampling surface is a sphere with a radius of 0.6 m, with 20×20 sampling points equally spaced between 1–179° and 0–342°, i.e., m = 400. The near-field amplitude of the horn antenna in the dataset includes |E|. r |,|E θ |, Three components, phase data φ includes φ r ,φ θ , Three components.

[0055] The specific operating steps are as follows:

[0056] In CADFEKO, find the Horn-conical antenna in the components library, set the operating frequency to 1.5GHz, and obtain the model of the horn-conical antenna.

[0057] A near-field sampling surface with a radius of 0.6m and a quantity of 20×20 was set up, within the range of θ from 1 to 179°. Values ​​are taken at equal intervals between 0 and 342°;

[0058] To construct a dataset that meets the required frequency band, the operating frequency of the horn antenna needs to be scanned, with a scanning range of 1.5–2.4 GHz. The sampling method is equidistant sampling. In the software interface, under "parameter sweep: create models," the operating frequency of the conical horn antenna is selected as the scanning parameter. Before calculating the antenna's near field, FEKO first creates CAD models of 800 horn antennas at different frequencies. During the scanning process, the antenna's CAD model automatically changes accordingly with each frequency step, and then the near field simulation is completed for these 800 antenna samples sequentially. After the near field simulation calculation is completed, the near field amplitude |E| of each horn antenna is calculated. r |,|E θ |, and near-field phase φ r ,φ θ , The file will be saved to the folder where the CAD model is stored, with the file suffix.out. The near-field amplitude and phase data of 400 sampling points are extracted from the.out file in the file directory through a Python script, and the dataset is constructed. Since the near-field amplitude and phase both contain three components, the final dataset structure is the near-field amplitude of the horn antenna, with dimensions (800, 400x3), and the label is the near-field phase, also with dimensions (800, 400x3). Then, the 800 samples are divided into training and validation sets in a ratio of 8:2.

[0059] S2. Construct a neural network for predicting the near-field phase from the near-field amplitude, and train the neural network with the training set to fit the relationship between the near-field amplitude and the near-field phase. Validate the trained neural network with the validation set, and obtain the trained neural network after validation.

[0060] Neural Networks is a computational model inspired by the biological nervous system, which simulates the connection and signal transmission of brain neurons to handle complex nonlinear relationships. It has been widely applied in various fields.

[0061] (1) Structure of Neural Networks

[0062] Neural Networks is usually composed of multiple layers, each containing several neurons (nodes). These layers can be divided into three categories.

[0063] ① Input layer (Input layer): receives raw data input, and each neuron represents an input feature;

[0064] ② Hidden layer (Hidden layer): located between the input layer and the output layer, responsible for feature extraction and transformation. Complex neural networks usually contain multiple hidden layers;

[0065] ③ Output layer (Output layer): generates the final output result, and each neuron corresponds to an output category or numerical value.

[0066] Each neuron is connected to other neurons through weights and biases, and signals are processed through activation functions. Activation functions are generally nonlinear to fit the nonlinear relationship between input and output.

[0067] (2) Working principle of Neural Networks

[0068] The working principle of Neural Networks can be divided into the following steps:

[0069] Forward Propagation: Input data passes through each layer of the network, layer by layer, performing weighted and biased operations, resulting in a final output at the output layer.

[0070] Loss Calculation: By comparing the predicted results with the actual labels, the value of the loss function is calculated to measure the prediction error of the network.

[0071] Backpropagation: Based on the gradient of the loss function, the weights and biases of the network are adjusted to minimize the loss function. This process uses gradient descent or its variants of optimization algorithms.

[0072] Iterative Training: The process of forward propagation and backpropagation is repeated until the network converges or reaches the preset training rounds.

[0073] For the simulation-generated dataset, this paper uses neural networks to fit the relationship between the near-field amplitude and phase of AUT. The input feature of 400x3 near-field amplitude can be regarded as a three-channel image of 20x20 pixels, so this paper uses a mature convolutional neural network (CNN) for training and fitting.

[0074] The CNN model used in this paper is the MobileNetV2 neural network model, as shown in Figure 2 The model has been successfully applied to near-field plane measurement. MobileNetV2 neural network model follows the depth separable convolution in MobileNetV1, but unlike the latter using 1x1 convolution block for dimensionality reduction operation, the former uses 1x1 convolution block for dimensionality increase operation. MobileNetV2 neural network model introduces linear bottleneck and inverted residuals. Linear bottleneck replaces low-dimensional nonlinear activation transformation with linear transformation to avoid losing too much useful information, and inverted residuals are for residual block of ResNet network. Residual block is a funnel-shaped structure, while inverted residuals is a process of dimensionality increase and then dimensionality reduction, which is a spindle-shaped structure.

[0075] The MobileNetV2 structure used in this invention is shown in Table 1, and the network is established using Pytorch. The network input is the amplitude data |E r |,|E θ |, The network output is the predicted three-component near-field phase information of the 400 sampling points of the near field denoted as 1200 values, the phase recovery schematic diagram is shown in Figure 3 The network is trained using the adaptive momentum optimizer (Adam), with a batch size of 160, a learning rate of 0.003, and an NVIDIA RTX A5000 GPU for 100 rounds of training. The loss function used is the average mean square error (MSE) between the predicted phase of the neural network and the simulated phase φ.

[0076] Table 1 MobileNetV2 structure diagram

[0077]

[0078] S3. For the antenna to be tested that is not in the dataset, only the near-field amplitude information is obtained through simulation, and the predicted near-field phase is obtained by inputting it into the trained neural network, so as to obtain the complete near-field amplitude and phase information. The equivalent dipole source of the antenna to be tested is calculated using the least squares method;

[0079] The equivalent dipole theory uses a weighted dipole array to replace the AUT to represent its radiation characteristics, as shown in Figure 4 The larger radius sphere in the figure represents the near-field simulation or measurement sphere, and E r ,E θ , represents the three components of the near-field electric field, and the smaller radius sphere represents the sphere where the dipoles are placed, P θ , represents the two components of the dipole. Here we use the electric dipole model, to reduce the computational complexity, the dipoles are placed along the tangential direction of the sphere, and the electric field of the near-field sphere and the dipole array can be represented as follows.

[0080] T 3m×2n P 2n×1 = E 3m×1

[0081] where m represents the number of near-field sampling points, and n represents the number of dipoles. P 2n×1 is a column vector composed of the θ and components of the dipole array, E 3m×1 is a column vector composed of the r, θ and components of the near-field electric field, and T 3m×2n is the relationship matrix between them, where The six block matrices are only related to the positions of the sampling points and the dipoles.

[0082]

[0083] For AUTs not in the dataset, only near-field amplitude information is obtained by simulation (for the problem of accurately measuring near-field phase in actual testing), and according to the trained neural network obtained in the previous step, the near-field amplitude |E 3m×1 of the AUT is obtained, and the input neural network can obtain the predicted near-field phase 3m×1 , and the complete near-field is obtained. The equivalent dipole source is calculated using near-field information is an electromagnetic inverse problem, and the commonly used method is the least squares method, and the objective function is:

[0084] minF(P 2n×1 )=‖E 3m×1 -T 3m×2n P 2n×1 ‖ 2

[0085] The least squares solution of the dipole vector can be obtained by matrix singular value decomposition. In Python, the numpy library can be called to solve it.

[0086] The dipole is placed on a spherical surface with a radius of 0.1m, and 18x18 sampling points are evenly set between 10-170° of and 0-340° of , that is, n=324. According to the simulated near-field amplitude and the predicted near-field phase of the neural network, the equivalent dipole source of the AUT is easily calculated by the least squares method.

[0087] S4. Calculate the far field of the antenna under test according to the equivalent dipole theory.

[0088] According to the equivalent dipole theory, the far field of the AUT is calculated, only the relationship matrix between the dipole array and the far field sampling point is calculated, denoted as , which is similar to the relationship matrix T 3m×2n between the dipole array and the near field in the above. The calculation method of T 3m×2n is similar, only the position of the sampling point is set farther to meet the far field condition, from near field to far field. The far field distance is set to 100m (which meets the far field condition), and the sampling point is set to 37x73, that is, every 5° in 0-180° of and 0-360° of . According to the equivalent dipole source obtained in the previous step, the matrix equation can be used to calculate the far field, and the far field of the AUT is predicted.

[0089] In the embodiments of this application, the training loss curve of the neural network is as follows: Figure 5 As shown, the dark curve represents the error on the training set and the light curve represents the error on the validation set. It can be seen that the neural network fits well, with the MSE of the training set and the validation set both below 1000, and there is no serious overfitting phenomenon.

[0090] To compare existing global optimization algorithms with the proposed method, an additional 100 horn antenna samples not found in the training and validation sets were selected as the test set. An equivalent dipole source was constructed using a covariance adaptive evolution strategy (CMA-ES) and a trained neural network. Figure 6 The results demonstrate the reconstruction of the near field using a randomly selected test set of horn antenna samples. Here, E_real represents the simulated real near field, E_cma represents the near field reconstructed using the CMA-ES optimization algorithm, and E_cnn represents the near field reconstructed using the proposed method. On two perpendicular near field cross-sections, the near field reconstruction results of both methods are similar and perform well.

[0091] To compare the differences in the two methods for predicting the far field of the AUT, the error is defined as follows.

[0092]

[0093] Where E k M represents the electric field component. s M represents the true near-field amplitude data of the AUT. e This represents the near-field amplitude data obtained based on the equivalent dipole source. m is the total number of near-field sampling points, i = 1, 2…m, and…

[0094] Table 2 Comparison of Results

[0095]

[0096] As can be seen from Table 2, using neural networks for phase retrieval and constructing equivalent dipoles is significantly better than the CMA-ES algorithm, with a greater advantage in both timeliness and accuracy. Figure 7 The far-field radiation pattern of the horn antenna predicted using a neural network method is shown, where real represents the actual simulated far field and cnn represents the far field predicted using a neural network. It can be seen that it has a very high consistency with the simulation.

[0097] The foregoing is considered as illustrative of the principles of the application. Further, since numerous modifications and changes will readily occur to those skilled in the art, it is not desired to limit the application to the exact construction and practice described. Accordingly, all suitable modifications and equivalents can be resorted to falling within the scope of the invention as defined by the claims that follow and their functional equivalents.

Claims

1. A far-field prediction method based on neural network near-field phase reconstruction, characterized in that: The method comprises the following steps: S1. Near-field simulation is performed on the antenna to be tested to obtain near-field amplitude and phase data of the antenna to be tested, a data set is constructed, and the data set is divided into a training set and a verification set; S2. A neural network for predicting near-field phase from near-field amplitude is constructed, the neural network is trained through the training set to fit the relationship between the near-field amplitude and the near-field phase, the trained neural network is verified by using the verification set, and the trained neural network is obtained after the verification is passed; S3. For the antenna to be tested not in the data set, only the near-field amplitude information is obtained through simulation, the predicted near-field phase is obtained by inputting the trained neural network, so that complete near-field amplitude and phase information is obtained, and the equivalent dipole source of the antenna to be tested is calculated by using the least square method; S4. The far field of the antenna to be tested is calculated according to the equivalent dipole theory.

2. The far-field prediction method based on neural network near-field phase reconstruction according to claim 1, characterized in that: The step S1 comprises: S101. Set the antenna to be tested as a conical horn antenna, the aperture radius is 1.2λ, λ is the wavelength of electromagnetic wave, the near-field sampling surface is a spherical surface with the center of the aperture of the antenna as the origin and the radius R, and t×t sampling points are equidistantly arranged within a given pitch angle θ and azimuth angle φ range, and the number of sampling points is denoted as m=t×t. S102. Set the antenna to be tested as a conical horn antenna, the aperture radius is 1.2λ, λ is the wavelength of electromagnetic wave, the near-field sampling surface is a spherical surface with the center of the aperture of the antenna as the origin and the radius R, and t×t sampling points are equidistantly arranged within a given pitch angle θ and azimuth angle φ range, and the number of sampling points is denoted as m=t×t. In performing near-field sampling, near-field amplitude data and near-field phase data are collected, the near-field amplitude data comprising radial component |E r , polar component |E θ , and azimuthal component , the near-field phase data φ comprising radial component φ r , polar component φ θ , and azimuthal component , each sampling point having a set of amplitude and phase data; S102. Sampling at equal intervals in a given frequency range to obtain p frequencies, and constructing an antenna CAD model at each frequency, repeating step S101 under the antenna CAD model at each frequency to obtain near-field amplitude data and phase data at each frequency, and adding them to the same data set to complete the construction of the data set; In the data set, there are p samples, each sample corresponds to an antenna CAD model at a frequency, and in each sample, the feature is the near-field amplitude of the horn antenna, the dimension is m*3, and the label is the near-field phase, the dimension is also m*3; S103. The data set is divided into a training set and a verification set.

3. The far-field prediction method based on neural network near-field phase reconstruction according to claim 1, characterized in that: In the step S2, the neural network CNN for predicting near-field phase from near-field amplitude is a convolutional neural network, and in the training process, the neural network is trained by using each sample in turn: When the neural network is trained by using any sample, an adaptive momentum optimizer is used, the near-field amplitude of t*t*3 is regarded as a three-channel image of t*t pixels, the CNN convolutional neural network is input, the predicted phase of t*t*3 is output by the CNN neural network, the loss function is calculated by using the predicted phase of t*t*3 and the near-field phase of t*t*3 in the label, then the gradient is calculated by back propagation according to the loss function, and the CNN neural network is updated according to the gradient descent method; After the training round is reached, the trained neural network is verified by using the verification set, and the trained neural network is obtained after the verification is passed.

4. The far-field prediction method based on neural network near-field phase reconstruction according to claim 1, characterized in that: The step S3 comprises: S301. Let dipoles be placed on a sphere of radius R1, in a range of θ and φ for the dipoles pre-specified t1 x t1 equidistantly; the number of dipoles is noted as n = t1 x t1, each dipole P has polar angle component P θ and azimuthal angle component two components; S302. For horn antennas not in the dataset, the near-field amplitudes are simulated and plotted in r, 0, Three vectors of m sampled near-field amplitudes are constructed in the r, 0, r | m×1 , |E θ | m×1 , The near-field amplitudes are thus recorded as S303. Arrange n dipoles according to θ, directions to get column vector P θ,n×1 , represent the polar angle component and azimuth angle component of n dipoles; S304. The equivalent dipole theory uses a dipole array to replace the antenna to be tested to represent the radiation characteristics thereof; Array of dipoles P 2n×1 Electric field E generated in free space 3m×1 The electric field generated in free space by a single dipole is obtained by a coordinate transformation, whose relation matrix is T 3m×2n , such that T 3m×2n × P 2n×1 = E 3m×1 ; The permittivity in free space is ε0, the permeability is μ0, the wave impedance is Wave number The coordinates of a near-field sampling point in the spherical coordinate system are recorded as The coordinates in the rectangular coordinate system are (x Ei ,y Ei ,z Ei ); the coordinates of a dipole placement point in the spherical coordinate system are The coordinates in the rectangular coordinate system are (x Pj ,y Pj ,z Pj ), and the distance between them is recorded as r; when the positions of the near-field sampling point and the dipole are determined, each element in T 3m×2n is calculated. S305. Calculate the equivalent dipole source P using the near-field information 2n×1 The objective function is minF(P 2n×1 ) =‖E 3m×1 -T 3m×2n P 2n×1 ‖ 2 least squares solution of the dipole vector is found by means of a matrix singular value decomposition.

5. The far-field prediction method based on neural network near-field phase reconstruction according to claim 1, characterized in that: The step S4 comprises: The far field sampling surface is a spherical surface with the center of the antenna aperture as the origin and a radius of R2, and t'1xt'2 sampling points are set at equal intervals between the elevation angle θ of 0-180° and the azimuth angle φ of 0-360° The number of sampling points is recorded as m'=t'1xt'2. The relationship matrix between the dipole array and the far-field sampling points is calculated, denoted as The calculation method is similar to T 3m×2n , but the number of near-field sampling points m is changed to the number of far-field sampling points m ′ , and the sampling point positions or (x Ei ,y Ei ,z Ei ), i = 1, 2…m are changed to or (x′ Ei ,y′ Ei ,z′ Ei ), i = 1, 2…m′, and the corresponding distance r between the field and the dipole is also changed accordingly; Using matrix equations The far field is calculated and the results are used as the far field prediction of the antenna under test.

Images