A near-far field conversion method and system for unmanned aerial vehicle near-field antenna measurement

Abstract

The application discloses a near-far field conversion method and system for unmanned aerial vehicle near-field antenna measurement, which adds near-field amplitude correction or repair processing before phase recovery, which is beneficial to reducing the influence of front-end measurement distortion on subsequent complex field recovery and near-far field conversion. Moreover, joint phase recovery is carried out by using double-plane and double-polarization near-field amplitude data, and cross-plane feature exchange and sharing fusion processing is introduced in the network, which is beneficial to more fully utilizing the propagation relationship between different measurement heights, so that the recovery result is not easy to deviate from the actual physical law. Finally, by constructing complex near-field by the corrected near-field amplitude and the recovered phase, and then performing near-far field conversion, the far-field directional diagram or far-field radiation parameter can be obtained under the condition of lacking a direct phase measurement link, thereby reducing the dependence on the traditional coherent measurement mode. The conversion system can realize local reasoning and result output in the outfield measurement process, and provides a basis for subsequent online updating and engineering maintenance.

Description

Technical Field

[0001] This invention relates to the fields of antenna measurement, near-field testing, near-far-field conversion, and UAV field measurement technology, and particularly to a near-far-field conversion method and system for near-field antenna measurement of UAVs. Background Technology

[0002] While far-field measurements offer intuitive results in antenna radiation pattern measurement, they are highly demanding in terms of site distance, equipment deployment, and testing conditions, especially for large-aperture antennas, array antennas, and outdoor-mounted antennas, making practical implementation often inconvenient. In contrast, near-field measurements can acquire the field distribution over shorter distances, and the far-field radiation pattern can then be obtained through near-field-to-far-field conversion, making it a commonly used technical approach in engineering. In recent years, UAVs, due to their high mobility and rapid deployment, have also begun to be used for antenna near-field measurements, making them suitable for complex terrain, restricted sites, and temporary outdoor missions.

[0003] Most existing near-field to far-field conversion methods are based on near-field complex field, meaning that in addition to near-field amplitude, phase information is also required. These methods are relatively easy to implement under regular scanning conditions in the laboratory; however, in UAV field measurement scenarios, stable phase acquisition usually requires a coherent measurement link. Some existing solutions, such as the paper "A cost-effective tethered-UAV-based coherent near-field antenna measurement system," employ tethered UAVs or additional reference links to maintain coherence and phase availability during measurement. While these solutions can improve phase acquisition conditions to some extent, they are relatively complex in system configuration and have high equipment costs, demanding requirements for field deployment conditions, and also limit the mobility and flexibility of the UAV platform.

[0004] Given the above, processing schemes that collect near-field amplitude data without directly measuring the phase have gradually gained attention. Existing technologies, such as CN107490729A, propose using dual-plane or multi-plane amplitude data for phase recovery and then performing near-field to far-field conversion. This indicates that there are usable propagation relationships between near-field data at different measurement heights. If dual-polarization and dual-height information can be fully utilized, it will help ensure that the recovery results do not deviate from the actual propagation laws.

[0005] Furthermore, with the development of deep learning technology, existing research has begun to apply neural networks to near-field recovery, phase recovery, and phase-free near-field to far-field conversion. For example, the paper "Machine Learning Based Fourier PhaseRetrieval for Planar Near-Field Antenna Measurements" discloses directions such as learning-based phase recovery, near-field super-resolution, data assimilation, and angular spectrum propagation-assisted phase recovery. These methods provide new technical paths for near-field data-driven processing, but most are not designed for UAV field amplitude distortion scenarios, and the joint utilization of dual-plane dual-polarization propagation relationships and real-time output at field edges remain insufficient.

[0006] During actual UAV flight measurements, the acquired near-field amplitude is prone to distortion due to factors such as position jitter, attitude changes, airflow disturbances, and uneven sampling. For amplitude-only measurement schemes, these problems directly increase the difficulty of phase recovery and near-field-to-far-field conversion. Simply stitching together or processing the dual-plane information separately is insufficient to fully reflect the propagation constraints between the two measurement planes. Furthermore, existing solutions mostly focus on offline computation, neglecting on-site deployment, edge operation, and online updates, making it difficult to meet the real-time and maintainability requirements of field measurements. Therefore, it is still necessary to propose a near-field-to-far-field conversion method and system for UAV near-field antenna measurements to improve the robustness of near-field data processing, the stability of near-field-to-far-field conversion, and the efficiency of on-site output in UAV scenarios. Summary of the Invention

[0007] The purpose of this invention is to address the aforementioned shortcomings in the existing technology by providing a near-field to far-field conversion method and system for near-field antenna measurement of unmanned aerial vehicles (UAVs). This addresses issues such as the near-field amplitude being easily distorted by platform disturbances during UAV flight measurement, insufficient stability of near-field to far-field conversion under conditions lacking direct phase information, inadequate utilization of dual-plane information, and weak on-site deployment capabilities.

[0008] The technical solution of this invention to solve the aforementioned technical problem is: designing a near-field to far-field conversion method for near-field antenna measurement of unmanned aerial vehicles (UAVs), characterized in that the method includes the following steps: S1: Near-field amplitude data on a first measurement plane and a second measurement plane parallel to the aperture plane of the antenna under test are acquired by a drone. The near-field amplitude data includes at least the x-direction polarization component amplitude Ex and the y-direction polarization component amplitude Ey. The distances between the first measurement plane and the second measurement plane and the aperture plane of the antenna under test are both between 3 and 10 times the operating wavelength of the antenna under test. S2: Preprocess the near-field amplitude data of the two measurement planes in S1, and then use the pre-trained amplitude restoration network to obtain the corrected near-field amplitude data of the two measurement planes. S3: Input the corrected near-field amplitude data of the two measurement planes output by the pre-trained amplitude restoration network in S2 into the pre-trained phase recovery network to obtain the recovered phase data of the first measurement plane and the second measurement plane under the corresponding polarization components; S4: Based on the corrected near-field amplitude data output by the pre-trained amplitude repair network in S2 and the recovered phase data output by the pre-trained phase recovery network in S3, construct complex near-field data on the first and second measurement planes. Determine whether the complex near-field data of the first measurement plane is reliable based on the propagation consistency error of the second measurement plane. If reliable, proceed to S5. Otherwise, adjust the network parameters of the pre-trained amplitude repair network and / or adjust the network parameters of the pre-trained phase recovery network. Repeat steps 2-4 until reliable complex near-field data of the first measurement plane is obtained. S5: Based on the reliable first measurement plane complex near-field data obtained in S4, perform near-far field conversion to obtain the far-field radiation pattern of the antenna under test; Furthermore, this invention designs a near-far field conversion system for near-field antenna measurement of UAVs. The system is characterized by implementing the near-far field conversion method for near-field antenna measurement of UAVs as described above. Specifically, it includes a UAV, a signal source, a probe antenna, an antenna under test, a spectrum analyzer, and an edge computing unit. Its specific operation involves the UAV carrying the signal source and probe antenna flying along a preset path. The probe antenna transmits a signal, and the antenna under test receives the signal. The signal is transmitted to the spectrum analyzer via a transmission line. The spectrum analyzer transmits the amplitude data to the edge computing unit. The edge computing unit calls a trained network or directly calls an algorithm to output the far-field radiation pattern of the antenna under test.

[0009] Compared with the prior art, the present invention has at least the following beneficial effects: Firstly, this invention addresses the problem that near-field amplitude is easily affected by disturbances under UAV flight measurement conditions. By adding near-field amplitude correction or repair processing before phase recovery, it helps to reduce the impact of front-end measurement distortion on subsequent field recovery and near-far field conversion, thereby improving the stability of the results.

[0010] Secondly, this invention utilizes dual-plane, dual-polarized near-field amplitude data for joint phase recovery, and introduces cross-plane feature exchange and shared fusion processing in the network, which helps to make fuller use of the propagation relationship between different measurement heights, making the recovery results less likely to deviate from actual physical laws.

[0011] Third, by constructing a complex near field from the corrected near-field amplitude and the recovered phase, and then performing near-far field conversion, this invention can obtain far-field radiation patterns or far-field radiation parameters even in the absence of a direct phase measurement link, thereby reducing the dependence on traditional coherent measurement methods.

[0012] Fourth, this invention takes into account the needs of field deployment, and can deploy the relevant network model on edge processing devices to realize local inference and result output during the field measurement process, improve the efficiency of field application, and provide a foundation for subsequent online updates and engineering maintenance. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the steps of an embodiment of a near-field to far-field conversion method for near-field antenna measurement of unmanned aerial vehicles (UAVs) according to the present invention.

[0014] Figure 2 This is a schematic diagram illustrating the near-field and far-field conversion method for near-field antenna measurement of unmanned aerial vehicles (UAVs) according to an embodiment of the present invention, showing the UAV dual-plane near-field measurement scenario setup and near-field amplitude data acquisition.

[0015] Figure 3 This is a comparison diagram of phase recovery results and reference phase from one embodiment of a near-field to far-field conversion method for near-field antenna measurement of unmanned aerial vehicles (UAVs) according to the present invention. Figure 3 (a) in the figure represents the phase recovery result. Figure 3 (b) in the diagram is the reference phase diagram.

[0016] Figure 4 This is a comparison diagram of the far-field radiation pattern of one embodiment of a near-field to far-field conversion method for near-field antenna measurement of UAVs according to the present invention. Figure 4 (a) shows the predicted and reference radiation patterns of the phi0 plane. Figure 4 (b) in the diagram shows the predicted and reference radiation patterns of the phi90 plane. Detailed Implementation

[0017] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0018] This embodiment provides a near-field to far-field conversion method for near-field antenna measurements on UAVs. See [link to relevant documentation]. Figure 1 and Figure 2 As shown, the method includes the following steps: S1: Near-field amplitude data are acquired using a UAV on a first and a second measurement plane parallel to the aperture plane of the antenna under test. The near-field amplitude data includes at least the x-direction polarization (horizontal polarization) component amplitude Ex and the y-direction polarization (vertical polarization) component amplitude Ey. The distances between the first and second measurement planes and the aperture plane of the antenna under test are both between 3 and 10 times the operating wavelength of the antenna under test.

[0019] The operating frequency of the antenna under test is denoted as f, and the corresponding wavelength is denoted as λ. ,in c represents the speed of light. The height of the first measuring plane is denoted as . The height of the second measuring plane is denoted as The distance between the two measuring planes is denoted as ,in In one specific implementation, it is possible to... , Therefore, .

[0020] The phi0 plane radiation pattern and the phi90 plane radiation pattern represent the azimuth angles, respectively. and The far-field radiation pattern.

[0021] To reduce near-field sampling aliasing, the sampling interval along the x-direction is denoted as . The sampling interval along the y-direction is denoted as Both must satisfy the Nyquist sampling theorem:

[0022] The measurement area size of each measurement surface can be set according to the aperture size of the antenna under test and the expected near-field coverage range.

[0023] Let the equivalent aperture dimensions of the antenna under test in the x and y directions be D, respectively. x and D y The distance between the measuring plane and the aperture surface of the antenna under test is z, and the expected maximum far-field observation angle is [value missing]. Then the dimensions of the measurement area of ​​the measurement plane in the x and y directions are... L x , L y It can be set as follows:

[0024]

[0025] in, and These are edge margins in the x and y directions, respectively, used to compensate for UAV trajectory deviations, near-field energy spread, and truncation errors caused by the limited measurement area. In one specific implementation, and Desirable to Alternatively, take 10% to 30% of the equivalent aperture size of the antenna under test in the corresponding direction.

[0026] The size of the measurement area affects the near-field to far-field conversion effect. If the measurement area is too small, a strong radiation field still exists at the near-field edge, and the finite scanning surface will produce truncation errors, leading to sidelobe fluctuations in the far-field pattern, null position shifts, and increased main lobe width errors. If the measurement area is too large, it will increase the UAV scanning time, data acquisition volume, and computational load, and may introduce more environmental noise. Therefore, in practical settings, it is preferable to attenuate the near-field amplitude at the boundary of the measurement area to below a preset threshold relative to the maximum amplitude within the measurement area. This preset threshold can be between -20 dB and -30 dB. This ensures near-field to far-field conversion accuracy while balancing UAV measurement efficiency and system computational load.

[0027] First, establish the spatial relationship between the antenna under test, the first measurement plane, and the second measurement plane, and then determine the measurement areas of the two measurement planes and the UAV scanning path based on the equivalent aperture size of the antenna under test.

[0028] The scanning path can be a regular round-trip scanning path, a segmented scanning path, or other planar scanning paths suitable for UAV execution. The first and second measurement planes are preferably positioned at different distances from the antenna under test to obtain near-field amplitude information at different propagation distances.

[0029] Under preset operating conditions, the measurement link device is turned on, and then the UAV is controlled to fly along the scanning path to perform measurements on the first and second measurement planes, respectively collecting the amplitude values ​​of the x-direction polarization (horizontal polarization) component Ex and the y-direction polarization (vertical polarization) component Ey on the two measurement planes.

[0030] In this step, the measurement probe on the UAV can perform point-by-point sampling or continuous sampling according to the set sampling interval. Since the UAV is in flight, the actual sampling position may deviate from the preset trajectory. Therefore, the obtained raw near-field amplitude data may contain effects such as position jitter, attitude changes, uneven sampling, local missing measurements, and noise disturbances.

[0031] S2: Preprocess the near-field amplitude data of the two measurement planes in S1, and then use the pre-trained amplitude restoration network to obtain the corrected near-field amplitude data of the two measurement planes.

[0032] The preprocessing procedure for the near-field amplitude data of the two measurement planes in S1 is as follows: First, a regular two-dimensional grid is set according to the measurement area size of the measurement plane and the preset grid resolution. The coordinates of each sampling point in S1 and the corresponding near-field amplitude data are then mapped onto the two-dimensional grid.

[0033] When multiple sampling points exist within the same grid, the amplitudes of these multiple sampling points are aggregated. The aggregation method can be arithmetic mean, weighted average, or median aggregation. Preferably, a weighted average method is used.

[0034] in, The amplitude of the nth sampling point falling within the (i,j)th grid is given. For the corresponding weights, The amplitude is obtained after the grid is aggregated. This represents the number of sampling points within the grid.

[0035] Therefore, sampling points that are closer to the grid center, whose measurement time is closer to the target sampling time, or whose signal quality is higher are given greater weight to obtain the representative amplitude of the grid; for grid positions not covered by sampling points, amplitude filling is performed. The amplitude of the grid not covered by the sampling point can be filled with the minimum effective amplitude of the corresponding channel or a fixed amplitude lower than the minimum effective amplitude; alternatively, the missing grid location can be filled by diffusion based on the amplitude of neighboring sampled grids. Specifically, with the missing grid as the center, sampled grids are searched within a preset neighborhood, and the amplitudes of neighboring sampled grids are weighted and averaged according to spatial distance to obtain the initial filling value of the missing grid; when there are no sampled grids within the preset neighborhood, the search neighborhood is gradually expanded until at least one sampled grid is obtained or the maximum search radius is reached. For continuous missing regions, iterative filling can be performed by diffusion from the sampled region to the missing region layer by layer, so that the amplitude of the missing region is smoothly connected with the surrounding sampled region.

[0036] Thus, the four-channel amplitude matrix to be repaired on the two measurement planes is obtained. :

[0037] in, and These represent the amplitude diagrams in the x and y polarization directions on the first measurement plane, respectively. and These represent the amplitude diagrams in the x and y polarization directions on the second measurement plane, respectively.

[0038] The amplitude matrix X is subjected to amplitude clipping and normalization, and mapped to a preset numerical range to obtain a normalized amplitude matrix. Specifically, the amplitude can be cropped to a preset dB range first. Then linearly map to [-1,1], that is:

[0039] Here, clip() represents the amplitude clipping operation. and These represent the preset lower limit and upper limit of amplitude, respectively.

[0040] Simultaneously, a sampling location mask and a sampling count map are generated. The sampling location mask indicates whether the corresponding grid has a valid measurement value (i.e., whether it contains a sampling point), and the sampling count map indicates the number of times the corresponding grid is sampled. The sampling location mask B and the sampling count map C are used as auxiliary inputs. To reduce the influence of differences in sampling count scales, the sampling count map is logarithmically normalized, resulting in:

[0041] in, This represents the maximum number of samples taken in the regular two-dimensional grid of the current measurement plane.

[0042] The above normalized magnitude matrix Sampling location mask B, normalized sampling number diagram By splicing the first measurement plane and the second measurement plane respectively, we obtain the feature Z1 of the first measurement plane and the feature Z2 of the second measurement plane.

[0043] Normalized magnitude matrix Features Z1 and Z2 are used as inputs to a pre-trained amplitude restoration network to obtain corrected near-field amplitude data.

[0044]

[0045] in, and These represent the normalized amplitude diagrams in the x and y polarization directions on the first measurement plane, respectively. and These represent the normalized amplitude diagrams for the x and y polarization directions on the second measurement plane, respectively.

[0046]

[0047]

[0048] Concat() indicates concatenation along the channel dimension. and These represent the sampling position masks for the first and second measurement planes, respectively. and The diagrams show the number of samples after normalization for the first and second measurement planes, respectively.

[0049] The pre-trained amplitude restoration network includes a first measurement plane encoding branch, a second measurement plane encoding branch, a cross-plane feature exchange unit, a multi-scale fusion unit, an upsampling decoding unit, and a residual amplitude output unit.

[0050] The first measurement plane coding branch and the second measurement plane coding branch have the same structure, both containing a shallow coding layer, a middle coding layer, and a deep coding layer. The shallow coding layer is the initial convolutional unit, the middle coding layer is the first-level sampling module, and the deep coding layer is the second-level sampling module. In the first measurement plane coding branch, Z1 is input into the shallow coding layer to obtain the shallow local amplitude feature K1 of the first measurement plane. (0) In the second measurement plane encoding branch, Z2 is input into another set of initial convolutional units to obtain the shallow local amplitude features K2 of the second measurement plane. (0) The initial convolutional unit comprises a two-dimensional convolutional layer, a normalization layer, and a nonlinear activation layer connected in sequence. The first-level downsampling module reduces the feature resolution from H×W to H / 2×W / 2 and increases the number of channels; the second-level downsampling module further reduces the feature resolution to H / 4×W / 4 and further increases the number of channels. Each downsampling module comprises a two-dimensional convolutional layer with a stride of 2, a normalization layer, a nonlinear activation layer, and a residual convolutional block connected in sequence. The residual convolutional block comprises a first two-dimensional convolutional layer, a normalization layer, a nonlinear activation layer, and a second two-dimensional convolutional layer connected in sequence. The output of the second two-dimensional convolutional layer is added element-wise to the input features of the residual convolutional block to obtain the output features of the residual convolutional block.

[0051] A cross-plane feature exchange unit is set after each coding layer. Each cross-plane feature exchange unit includes a first plane update branch and a second plane update branch. For coding layers... l Local amplitude characteristics K1 of the first measurement plane output (l) Local amplitude characteristics K2 of the second measurement plane (l) First, concatenation is performed along the channel dimension to obtain cross-plane joint features:

[0052] Then, the cross-plane joint feature P (l) Input cross-plane feature exchange unit respectively lThe first and second plane update branches are provided; each update branch includes a 1×1 convolutional mapping layer, a normalization layer, a nonlinear activation layer, a residual convolutional block, and a 1×1 output convolutional layer connected in sequence, used to predict the residual update features of the current measurement plane from the joint features of the two planes:

[0053] in, and Representing cross-plane feature exchange units l The first and second plane update branches are then defined. Subsequently, the residual update features are added element-wise to the original local amplitude features of the corresponding measurement plane to obtain the updated features after cross-plane information exchange.

[0054] in, Indicates the first measurement plane in the cross-plane feature exchange unit l Updated features after feature swapping Indicates the second measurement plane in the cross-plane feature exchange unit l Updated features after feature swapping.

[0055] When the coding layer l When it is a shallow coding layer, cross-plane feature exchange unit l Output shallow layer update features; when the encoding layer l When it is a middle coding layer, cross-plane feature exchange unit l Output the middle layer to update features; when the coding layer l When it is a deep coding layer, cross-plane feature exchange unit l Output deep update features; The multi-scale fusion unit includes a shallow fusion convolution module, a deep fusion convolution module, and a shallow fusion convolution module. For the deep scale, the updated features of the first and second measurement planes are concatenated by channels and then processed by the deep fusion convolution module to obtain the deep fused features.

[0056] For the mid-level scale, the mid-level update features of the first and second measurement planes are concatenated by channels and then passed through a mid-level fusion convolution module to obtain the mid-level skip fusion features:

[0057] For shallow scales, the shallow update features of the first and second measurement planes are concatenated by channels and then processed through a shallow fusion convolution module to obtain shallow skip fusion features:

[0058] in, , , These represent the deep, middle, and shallow fusion convolutional modules, respectively. Each fusion convolutional module includes a 1×1 convolutional mapping layer, a normalization layer, a nonlinear activation layer, and a residual convolutional block connected in sequence, used to compress the number of channels after splicing and fuse complementary information from the two measurement planes.

[0059] Subsequently, the upsampling decoding unit first processes the deep fusion feature U (2) Upsampled to the mid-level feature size and fused with the mid-level skip feature U (1) By concatenating the data, we obtain the input for the middle layer decoding:

[0060] Here, Up() represents the upsampling operation. The intermediate layer decoding input D... (1) After sequentially performing convolution, normalization, nonlinear activation, and residual convolution blocks, the mid-layer decoding features R are obtained. (1) .

[0061] Then, the middle layer decoding feature R (1) Upsampled to the original grid resolution and fused with shallow skip features U (0) By concatenating the inputs, we obtain the shallow decoding input:

[0062] The shallow decoding input D (0) After being processed sequentially through convolution, normalization, nonlinear activation, and residual convolution blocks, the final decoded feature R is obtained. (0) Thus, the network progressively fuses deep global semantic information, mid-level structural information, and shallow spatial detail information, restoring the same grid resolution as the input amplitude matrix.

[0063] Finally, the residual amplitude output unit predicts the four-channel amplitude correction based on the final decoded features R(0). The residual amplitude output unit uses a 3×3 convolutional layer to map the final decoded features into four-channel residuals and limits the residual amplitude using a hyperbolic tangent function.

[0064] Where, ΔX n This represents the four-channel residual correction amount of the normalized magnitude matrix, α represents the residual scaling factor, and tanh() is used to limit the single correction magnitude.

[0065] The predicted residual correction amount ΔX n With the input normalized magnitude matrix X nElement-wise summation, followed by range clipping, yields the normalized repair magnitude matrix:

[0066] in, This is the normalized repaired amplitude matrix output by the pre-trained amplitude repair network.

[0067] The pre-trained amplitude restoration network was trained using electromagnetic simulation and flight simulation data samples. Specifically, the ideal near-field amplitude obtained from electromagnetic simulation is used as the training label, and the near-field amplitude data obtained from flight simulation under the influence of flight disturbances is used as the training sample to simulate the near-field amplitude data collected in real-world scenarios; the specific process of obtaining the pre-trained amplitude restoration network is as follows: First, a simulation model of the antenna under test and the first and second measurement planes in S1 is established using simulation. The simulation model includes the geometry, material parameters, feeding method, operating frequency, boundary conditions, measurement plane position, measurement area size, grid resolution, and near-field amplitude extraction methods for x-direction polarization and y-direction polarization of the antenna under test.

[0068] In one specific implementation, a full-wave electromagnetic simulation platform can be used to establish the model of the antenna under test, such as HFSS, CST, FEKO, or other simulation platforms capable of calculating near-field electromagnetic distribution. During simulation, the operating frequency of the antenna under test is set to f, corresponding to a wavelength of λ; the first measurement plane and the second measurement plane are respectively set at a distance from the aperture plane of the antenna under test. and The positions of both measurement planes are parallel to the aperture surface of the antenna under test; the measurement area size and grid resolution are consistent with the regular two-dimensional grid set in S2. Simulation boundaries can be radiating boundaries, absorbing boundaries, or perfectly matched layer boundaries to simulate free-space radiation conditions. After the electromagnetic simulation is completed, near-field amplitude data at the center of each grid on the first and second measurement planes are extracted to obtain amplitude maps of the x-direction polarization components and y-direction polarization components, thus forming a four-channel ideal amplitude matrix X for the two measurement planes. 0Furthermore, to simulate positional deviations and sampling gaps that occur during actual UAV flight sampling, a flight simulation platform can be used to generate disturbance sampling trajectories. In one specific implementation, a UAV flight simulation environment is established using the Gazebo platform, and the airframe parameters, flight speed, trajectory planning parameters, positioning error parameters, and weather disturbance parameters such as wind speed and direction of the DJI M300 RTK are input, causing the UAV to fly along a trajectory that is the same as or similar to the actual test mission. Based on the UAV sampling position output by Gazebo, the ideal near-field amplitude field obtained from the electromagnetic simulation is resampled to obtain near-field amplitude data affected by flight disturbances. The amplitude at the center of the regular two-dimensional grid obtained from the electromagnetic simulation is used to construct the ideal amplitude matrix X. 0 The disturbance trajectory obtained from flight simulation is used to construct the disturbed sampling points, sampling position mask B, sampling number map C, and normalized amplitude matrix X. n Then, the ideal magnitude matrix X 0 After performing the above amplitude clipping and normalization processes, the normalized amplitude matrix X is obtained. n Reference matrix X n 0 .

[0069] Then, the model parameters of the amplitude repair network are initialized. Let the amplitude repair network be F. θ (), where θ represents the trainable parameters in the first measurement plane coding branch, the second measurement plane coding branch, the cross-plane feature exchange unit, the multi-scale fusion unit, the upsampling decoding unit, and the residual amplitude output unit. For the two-dimensional convolutional layers and 1×1 convolutional mapping layers other than the residual amplitude output unit, they are randomly initialized according to their number of input channels, number of output channels, and convolutional kernel size; in one specific implementation, the Kaiming uniform initialization method is used to initialize the convolutional kernel weights to random numbers near the zero mean, and the corresponding biases are initialized to zero. For the normalization layer in the network, the scaling factor of the normalization layer is initialized to 1, and the translation bias is initialized to 0; if the normalization layer contains a running mean and a running variance, the running mean is initialized to 0, and the running variance is initialized to 1.

[0070] Furthermore, the final 3×3 convolutional layer in the residual amplitude output unit, used to generate the four-channel residual correction, is zero-initialized. This final 3×3 convolutional layer refers to the layer in the residual amplitude output unit located before the hyperbolic tangent function, used to generate the final decoded feature R. (0) This is a convolutional layer mapped to four-channel residual corrections. The weight matrix and bias vector of this final convolutional layer are denoted as W. o and b o Then initialize the command: W o =0,b o =0 At this point, the residual amplitude output unit satisfies the following in the initial training phase:

[0071] Where * denotes convolution operation, and α represents the residual scaling factor. Therefore, the initial output of the amplitude restoration network is:

[0072] That is, the network approximately performs an identity mapping at the beginning of training, and does not immediately change the normalized magnitude matrix of the input. As training iterates, the parameters of the final convolutional layer and the preceding feature extraction layer are gradually updated under the constraint of supervised loss. The network then gradually learns the amplitude residual correction caused by UAV position jitter, uneven sampling, local missing measurements, and noise perturbation. Through the above initialization method, the initial physical distribution of the input amplitude data can be preserved, and the stability of subsequent phase retrieval and near-field conversion can be avoided due to excessively large random residual output in the early stage of training.

[0073] Then, training samples are constructed. For the q-th training sample, the normalized amplitude matrix X affected by flight disturbances is... n (q) Sampling location mask B (q) Plot C of normalized sampling times n (q) As model input, the corresponding normalized ideal magnitude matrix X n0 (q) As a supervisory label, the first measurement plane input feature Z1 is constructed according to the aforementioned method. (q) Second measurement plane input feature Z2 (q) And input the amplitude repair network:

[0074] in, The normalized repaired amplitude matrix output by the amplitude repair network has the same number of channels as the input amplitude matrix, and still includes four-channel amplitude data in the x and y polarization directions of the first and second measurement planes.

[0075] In each iteration of training, a batch of training samples is selected from the training set, denoted as:

[0076] in, Let M represent the batch training sample set used in the t-th training iteration, M represent the number of training samples in the batch, and q represent the sample number in the batch training sample set. For any sample q, the normalized magnitude matrix affected by flight disturbances is... Sampling location mask Normalized sampling number graph Input the amplitude restoration network to obtain the corresponding normalized restored amplitude matrix:

[0077] in, This represents the normalized repaired amplitude matrix output by the amplitude repair network after the q-th training sample is processed. This represents the corresponding normalized ideal magnitude matrix.

[0078] Then, the amplitude reconstruction loss is calculated based on this batch of training samples. This amplitude reconstruction loss is used to constrain the repaired amplitude matrix to approximate the normalized ideal amplitude matrix. , can be represented as all The average value, where, This represents the average absolute error calculated for the four amplitude channels and each grid position.

[0079] To further maintain the consistency of the near-field amplitude spatial variation trend, a gradient consistency loss is constructed. Let... and Let represent the spatial difference gradient operators along the x and y directions, respectively. Then, the gradient consistency loss... It can be represented as all The average value.

[0080] Therefore, the total training loss of the amplitude repair network in the t-th training iteration is:

[0081] in, Let λt represent the total batch training loss for the t-th training iteration, and λ1 and λ2 represent the weight coefficients of the magnitude reconstruction loss and the gradient consistency loss, respectively.

[0082] Subsequently, the backpropagation algorithm was used to calculate the total batch training loss. Gradient of network parameter θ:

[0083] The network parameters are then updated using a gradient optimization algorithm. Preferably, the AdamW optimizer is used for parameter updates.

[0084] in, This represents the network parameters before the t-th iteration. Let represent the network parameters after the t-th iteration, η represent the learning rate, and ω represent the weight decay coefficient. During training, a learning rate decay strategy can be used to gradually decrease the learning rate with each training epoch, thereby improving the model's convergence stability.

[0085] During training, the training samples are divided into a training set and a validation set. After each training epoch, the validation loss is calculated on the validation set, and the model parameters with the minimum validation loss are saved. Training termination conditions include at least one of the following: the number of training epochs reaches a preset maximum number of epochs; the validation set loss no longer decreases for several consecutive epochs; or both the training loss and the validation loss are less than a preset threshold. After training, the network parameters with the minimum validation set loss are selected. The trained amplitude repair network is obtained. .

[0086] The trained amplitude restoration network is a pre-trained amplitude restoration network used to receive the amplitude matrix to be restored, the sampling position mask and the sampling number map during the actual measurement stage, and output the corresponding corrected near-field amplitude data.

[0087] S3: Input the corrected near-field amplitude data of the two measurement planes output by the pre-trained amplitude restoration network in S2 into the pre-trained phase recovery network to obtain the phase data of the first measurement plane and the second measurement plane under the corresponding polarization components.

[0088] The corrected near-field amplitude data includes the amplitude data of the first measurement plane and the second measurement plane under the x and y polarization components, denoted as:

[0089] in, This indicates the correction of near-field amplitude data. and These represent the normalized corrected amplitude data of the x and y polarization components on the first measurement plane, respectively. and These represent the normalized corrected amplitude data of the x and y polarization components on the second measurement plane, respectively. H and W represent the number of sampling points in the longitudinal and transverse directions of the regular two-dimensional grid, respectively.

[0090] Input data for constructing a pre-trained phase retrieval network: First, obtain the coordinate feature map: Let the horizontal and vertical coordinates corresponding to the i-th row and j-th column in the regular two-dimensional grid of the measurement plane be x and x, respectively. j and y i Construct horizontal coordinate graphs, vertical coordinate graphs, and radial position graphs based on the coordinate information:

[0091] Among them, P x (i,j) represents the normalized lateral coordinate feature of the (i,j)th grid point, P y (i,j) represents the normalized longitudinal coordinate feature of the (i,j)-th grid point, Pr (i,j) represents the normalized radial position feature of the (i,j)th grid point; s xy and s r These are the coordinate normalization scales, calculated as follows:

[0092] Here, ε is a very small positive number set to prevent the denominator from being zero. This yields the coordinate characteristic diagram:

[0093] in, This represents the measurement plane coordinate feature map, and Concat() represents the concatenation of channel dimensions.

[0094] Secondly, regarding X cn Low-frequency and high-frequency information are decomposed. The low-frequency amplitude component and the high-frequency amplitude component are represented as follows:

[0095] Among them, X L X represents the low-frequency amplitude component. H Indicates the high-frequency amplitude component. Let σ represent the two-dimensional Gaussian smoothing operator, and let σ represent the Gaussian smoothing parameter.

[0096] The low-frequency amplitude component is decomposed according to the measurement height as follows:

[0097] The high-frequency amplitude component is separated according to the measurement height as follows:

[0098] Among them, X L,1 and X L,2 X represents the low-frequency dual-polarization amplitude characteristics of the first and second measurement planes, respectively. H,1 and X H,2 These represent the high-frequency dual-polarization amplitude characteristics of the first and second measurement planes, respectively.

[0099] Furthermore, mean and difference characteristics between the two measurement planes are constructed. For the low-frequency amplitude component, we have:

[0100] For high-frequency amplitude components, we have:

[0101] in, and These represent the mean amplitude characteristics of the two measurement planes under low-frequency and high-frequency conditions, respectively, and are used to characterize the common amplitude structure of the two measurement planes; ΔX L and ΔX H These represent the difference amplitude characteristics of two measurement planes under low-frequency and high-frequency conditions, respectively, and are used to characterize the amplitude variation information between the two measurement heights.

[0102] Therefore, the inputs to the first measurement plane low-frequency coding branch, the second measurement plane low-frequency coding branch, the first measurement plane high-frequency coding branch, and the second measurement plane high-frequency coding branch in the phase retrieval network are respectively:

[0103] Among them, Z L,1 Z represents the input of the low-frequency coding branch of the first measurement plane. L,2 Z represents the input of the low-frequency coding branch of the second measurement plane. H,1 Z represents the input of the high-frequency coding branch of the first measurement plane. H,2 This represents the input to the high-frequency coding branch of the second measurement plane.

[0104] The phase recovery network includes a first measurement plane low-frequency coding branch, a second measurement plane low-frequency coding branch, a first measurement plane high-frequency coding branch and a second measurement plane high-frequency coding branch, a frequency feature fusion unit, a cross-plane feature exchange unit, a shared fusion unit, a shared upsampling unit, and a polarization phase decoding unit.

[0105] Each low-frequency coding branch and high-frequency coding branch contains multiple frequency coding layers, Z L,1 Z L,2 Z H,1 Z H,2 These correspond to the inputs of the first-level low-frequency coding layer and the first-level high-frequency coding layer of the first and second measurement planes, respectively; let m = 1, 2, representing the measurement plane number. l =1,⋯,N represents the frequency coding level, then the m-th measurement plane is in the _____. l The low-frequency coding features and high-frequency coding features output by the frequency coding layer are denoted as follows:

[0106] in, Indicates the first l The coding mapping of the low-frequency coding layer, Indicates the first lThe encoding mapping of the first-level frequency coding layer is as follows: The first-level frequency coding layer consists of a convolutional layer, a normalization layer, a nonlinear activation layer, and a residual convolutional block connected in sequence. Subsequent frequency coding layers, based on the structure of the first-level frequency coding layer, connect a downsampling module after the residual convolutional block. The output of the previous-level frequency coding layer serves as the input of the next-level frequency coding layer. The first-level frequency coding layer is defined as a shallow coding layer, outputting shallow-scale frequency coding features; the first 2 - |N / 2| frequency coding layers are defined as intermediate coding layers, with their last layer outputting intermediate-scale frequency coding features (|| represents the rounding operation); the remaining frequency coding layers are defined as deep coding layers, with their last layer outputting deep-scale frequency coding features.

[0107] For shallow-scale, mid-scale, and deep-scale frequency-coded features, they are processed by a frequency feature fusion unit, a cross-plane feature exchange unit, and a shared fusion unit, respectively. The frequency feature fusion unit fuses low-frequency and high-frequency coded features at each scale to obtain the frequency fused features of the m-th measurement plane.

[0108] in, The frequency feature fusion unit represents the scale r. r Indicates shallow, intermediate, or deep scale. The unit first concatenates low-frequency and high-frequency encoded features at scale r of the m-th measurement plane. This concatenates low-frequency features, high-frequency features, their difference features, and their product features at the same scale to obtain joint features for generating gate weights. Then, a gated mapping layer consisting of a 1×1 convolutional layer and a sigmoid function is used to obtain gate weights between 0 and 1, which are then multiplied element-wise with the high-frequency features. Finally, the low-frequency features and the gated high-frequency features are concatenated and fused to obtain the frequency fusion features. This enhances effective local details while preserving the global structure. The gated mapping layer includes a 1×1 convolutional layer and a sigmoid function connected in sequence. The 1×1 convolutional layer is used to map the input joint features to gated features with the same size as the features to be adjusted. The sigmoid function is used to map the gated features to a range of 0 to 1 to obtain the gated weights.

[0109] After obtaining the frequency fusion features of the two measurement planes, the cross-plane feature exchange unit performs bidirectional information exchange at each scale. For each scale... r Based on the frequency fusion characteristics, joint features from the first measurement plane to the second measurement plane and from the second measurement plane to the first measurement plane are constructed:

[0110] in, This represents the cross-plane joint feature used to update the features of the first measurement plane. This represents the cross-plane joint feature used to update the features of the second measurement plane, where * indicates element-wise multiplication.

[0111] The cross-plane feature exchange units generate update values ​​based on the aforementioned joint features, and inject these update values ​​into the original frequency fusion features of the corresponding measurement plane using a gating method. Specifically, for the first... There are measurement planes, among which Joint features across planes Input convolution update mapping This yields the cross-plane update amount; simultaneously, the cross-plane joint features are... The input consists of a gated mapping layer composed of a 1×1 convolutional layer and a sigmoid function, resulting in gate weights with the same size as the update amount and values ​​ranging from 0 to 1. After multiplying the gating weights element-wise by the update amount, the resulting feature is fused with the original frequency. Element-wise summation, followed by residual convolution for refinement, yields the updated features after cross-plane information exchange:

[0112] in, and These represent the updated features of the first and second measurement planes after cross-plane information exchange, respectively. and These represent the convolutional update mappings of the first and second measurement planes, respectively. The convolutional update mappings map the cross-plane joint features to an update amount that is the same as the number of channels of the corresponding original frequency fused features through a two-dimensional convolutional layer, a normalization layer, and a non-linear activation layer. and These represent the gating weights of the first and second measurement planes, respectively. This represents a residual convolution thinning operation at scale r, which processes the gated updated features through at least one residual convolution block. Here, * denotes element-wise multiplication.

[0113] Then, the shared fusion unit fuses the updated features of the two measurement planes at the same scale to obtain shared features:

[0114] Among them, S (r) s represents the shared features at scale r. (r) () represents a shared fusion unit, which concatenates the features of two measurement planes, their absolute difference features, and their product features into channels, and then completes channel compression and information fusion through 1×1 convolution, normalization, nonlinear activation, and residual convolution blocks.

[0115] Let r = 0, 1, 2, corresponding to the shallow, medium, and deep scales, respectively. In the shared upsampling unit, the deep scale shares feature S. (2) First, the bottleneck residual module is input. This module further extracts deep-scale shared features while maintaining feature resolution. It then adds the processed features element-wise to the input features via residual connections to obtain the bottleneck residual features. Subsequently, the bottleneck residual features are input to the channel attention module. This module first performs global average pooling on each channel of the bottleneck residual features to obtain global statistical features for each channel. Then, it maps channel weights to these global statistical features and uses the Sigmoid function to obtain the weights for each channel. Finally, it multiplies the channel weights with the bottleneck residual features channel-wise to obtain the deep global shared features G. (2) Then, the deep globally shared features G are... (2) Upsampled to the mid-scale and sharing features S with the mid-scale (1) By performing skip connection fusion, the middle-layer shared decoding feature G is obtained. (1) :

[0116] Where Up() represents the upsampling operation, D s (1) () indicates the intermediate shared decoding convolution module, which will upsample the deep global shared features G. (2) Sharing features S with the mid-level scale (1) Channel concatenation is performed, and the concatenated features are sequentially processed through a 2D convolutional layer, a normalization layer, a nonlinear activation layer, and a residual convolutional block to obtain the middle-layer shared decoding features G. (1) ; Subsequently, the middle layer shared decoding feature G (1) These features are combined with the mid-scale update features of the first and second measurement planes, respectively, to form the first measurement plane decoding input feature D1 and the second measurement plane decoding input feature D2:

[0117] in, and These are the update features of the first and second measurement planes at the mesoscale, respectively, D1. (1) () and D2 (1)() represent the mid-level decoding convolutional modules of the first and second measurement planes, respectively. The two mid-level decoding convolutional modules have the same structure and independent parameters. They both include a two-dimensional convolutional layer, a normalization layer, a nonlinear activation layer and a residual convolutional block connected in sequence, which are used to fuse the mid-level shared decoding features and the mid-level scale update features of the corresponding measurement plane.

[0118] For each measurement plane, the polarization phase decoding unit is set to E respectively. x Polarization phase decoder and E y Polarization phase decoder. Using the m-th measurement plane, polarization component p∈{E} x E y For example, its polarization phase decoder uses the decoding input feature D of the m-th measurement plane. m and shallow-scale update features of the same measurement plane This is the input. The shallow-scale update features... This corresponds to the features of the measurement plane after cross-plane information exchange at a shallow scale, used to supplement the phase decoder with local amplitude edges, detailed textures, and spatial location information at the original mesh resolution. The polarization phase decoder will then... m After upsampling to a shallow scale, the features are updated together with those at the shallow scale. Channel concatenation is performed, followed by residual convolution for thinning, to obtain polarization phase decoding features:

[0119] Among them, Q m,p D represents the phase decoding characteristics of the m-th measurement plane under polarization component p. m,p (0) () indicates the corresponding shallow polarization phase decoding module. Subsequently, the polarization phase decoder determines the polarization phase based on Q... m,p Output low-frequency phase representation and residual phase representation.

[0120] Subsequently, the low-frequency phase output branch is based on the phase decoding feature Q. m,p The low-frequency phase components of two channels are output through a two-dimensional convolutional layer, and the two low-frequency phase components are vector normalized to obtain a two-dimensional unit vector:

[0121] in, and Let the low-frequency phase cosine component and low-frequency phase sine component of the polarization component p on the m-th measurement plane be represented respectively. The two normalized components satisfy:

[0122] Calculate the low-frequency phase based on the low-frequency phase cosine component and the low-frequency phase sine component:

[0123] Meanwhile, the residual phase prediction branch outputs a constrained residual phase:

[0124] in, This represents the low-frequency phase of the polarization component p on the m-th measurement plane. R represents the corresponding residual phase. m,p This represents the residual phase of the output. ρ This represents the residual phase scaling factor. The final recovered phase is:

[0125] Here, `wrap()` represents the phase wrapping operation, used to constrain the phase within the range (-π, π]. Thus, the phase retrieval network ultimately outputs the dual-polarized phase data for the first and second measurement planes:

[0126] in, This represents the recovered phase data of the first measurement plane. This represents the recovered phase data of the second measurement plane.

[0127] The training data for the pre-trained phase retrieval network comes from the training data of the pre-trained amplitude restoration network; the training data originates from measurement samples of electromagnetic simulations, flight simulations, or known complex near-field data. For each training sample, the first and second measurement planes are extracted at E... x E y Near-field amplitude data under polarization components are used, and amplitude data to be repaired due to flight disturbances is generated according to the aforementioned amplitude perturbation construction method; after processing by the amplitude repair module, the normalized correction amplitude matrix X is obtained. cn (q) Simultaneously, the corresponding true phase label is calculated based on the complex near-field data of this sample:

[0128] Where q represents the training sample number, and Let E and represent the true phase labels of the first and second measurement planes in the q-th training sample, respectively. Given the complex electric field E... m,p,0 (q) The true phase is then calculated using the following formula:

[0129] Where m∈{1,2} represents the measurement plane number, p∈{Ex E y} represents the polarization component, and arg() represents the complex phase operation. Then, according to X... cn (q) Measure the plane coordinate information and condition information, and construct the phase retrieval network input as described above:

[0130] in, Z represents the dataset input to the phase retrieval network for the q-th training sample. L,1 (q) Z L,2 (q) Z H,1 (q) Z H,2 (q) These represent the low-frequency and high-frequency coding layer inputs, respectively, constructed from the normalized correction amplitude matrix, mean feature, difference feature, and coordinate condition feature map.

[0131] Next, the parameters of the phase retrieval network model are initialized. Let the phase retrieval network be:

[0132] Wherein, ψ represents the trainable parameters of the low-frequency coding layer, high-frequency coding layer, frequency feature fusion unit, cross-plane feature exchange unit, shared fusion unit, shared upsampling unit, and polarized phase decoding unit in the phase retrieval network. The trainable parameters include the kernel weights and biases of the two-dimensional convolutional layer and the 1×1 convolutional mapping layer, the scaling coefficients and translation biases of the normalization layer, the parameters of the gated mapping layer, and the parameters of the phase decoding output layer.

[0133] When initializing the trainable parameters of the phase retrieval network, the two-dimensional convolutional layers and 1×1 convolutional mapping layers in the low-frequency coding layer, high-frequency coding layer, frequency feature fusion unit, cross-plane feature exchange unit, shared fusion unit, shared upsampling unit, and polarized phase decoding unit are randomly initialized according to their number of input channels, number of output channels, and convolutional kernel size. Specifically, the convolutional kernel weights are initialized using the Kaiming uniform initialization method, and the corresponding biases are initialized to zero. For the normalization layer in the network, the scaling factor of the normalization layer is initialized to 1, and the translation bias is initialized to 0. If the normalization layer contains a running mean and a running variance, the running mean is initialized to 0, and the running variance is initialized to 1. For the convolutional weights in the aforementioned gated mapping layer, the same random initialization method as the convolutional layers is used, and the bias of the gated mapping layer is initialized to zero, so that the gated weights are in the non-saturated range in the initial stage of training, which facilitates subsequent gradient updates.

[0134] Furthermore, for the low-frequency phase output branch and residual phase output branch in the polarization phase decoding unit, the weights of their terminal convolutional layers are initialized using a small-amplitude random initialization method, and the bias is initialized to zero, so that the phase output will not produce excessive random shifts in the initial training stage. If a pre-trained phase retrieval model is used as the pre-trained model, the parameters of the pre-trained model are used as the initial parameters, and some parameters are updated or frozen as needed during subsequent training.

[0135] Input the q-th training sample into the phase recovery network to obtain the predicted phase:

[0136] in, and These represent the predicted phases of the first and second measurement planes output by the phase retrieval network, respectively.

[0137] To mitigate the computational discontinuity issues caused by phase periodicity, the predicted and true phases are converted into phasor form. For any phase φ, its phasor representation is:

[0138] Therefore, the phase-monitoring loss is constructed:

[0139] Simultaneously, a phase gradient loss is constructed:

[0140] And low-frequency phase loss and residual phase loss are constructed based on low-frequency phase and residual phase respectively:

[0141]

[0142] in, This represents the predicted low-frequency phase of the m-th measurement plane. This indicates the low-frequency phase corresponding to the actual phase label. Indicates the predicted residual phase. This represents the true residual phase.

[0143] Furthermore, to constrain the physical consistency between the two measurement planes, a complex near field of the first measurement plane is constructed based on the predicted phase and corrected amplitude of the first measurement plane, and then propagated to the second measurement plane through the angular spectrum propagation operator to obtain the propagated predicted phase. Similarly, the propagation prediction phase can be obtained by backpropagating from the second measurement plane to the first measurement plane. The loss of propagation consistency can be expressed as:

[0144] Therefore, the total training loss of the phase retrieval model is:

[0145] Where μ1, μ2, μ3, μ4, and μ5 are the weights of phase supervision loss, phase gradient loss, low-frequency phase loss, residual phase loss, and propagation consistency loss, respectively.

[0146] In each training iteration, a batch of training samples is input into the phase retrieval network, and the total training loss is calculated. The gradient of the loss function with respect to the network parameter ψ is calculated using the backpropagation algorithm.

[0147] Subsequently, the network parameters are updated using a gradient optimization algorithm. Preferably, the AdamW optimizer is used for parameter updates:

[0148] Where, ψ t Let ψ represent the phase retrieval network parameters before the t-th iteration. t+1 Let represent the phase retrieval network parameters after the t-th iteration update, η represent the learning rate, and ω represent the weight decay coefficient. A learning rate decay strategy can be used during training to gradually decrease the learning rate with each training epoch.

[0149] During training, the samples are divided into training and validation sets. After each training epoch, the validation loss is calculated on the validation set, and the model parameters with the minimum validation loss are saved. Training termination conditions include at least one of the following: the number of training epochs reaches a preset maximum number of epochs; the validation set loss no longer decreases for several consecutive epochs; or both the training loss and validation loss are less than a preset threshold. After training, the network parameter ψ with the minimum validation set loss is selected. ∗ The trained phase retrieval network is obtained:

[0150] The trained phase retrieval network is the pre-trained phase retrieval network, which is used in the actual application stage to receive the normalized correction amplitude matrix and its coordinate condition features output by the amplitude repair module, and output the recovered phase data of the first measurement plane and the second measurement plane under the x and y polarization components.

[0151] It should be noted that the training process for the aforementioned amplitude reconstruction model and phase retrieval model is an offline training process, and the model parameters are not updated during actual measurement applications. The training data and the actual data to be predicted have the same data organization format, both including four-channel near-field amplitude data of the first and second measurement planes under the Ex and Ey polarization components, and using the same grid resolution, channel arrangement, amplitude normalization method, and measurement plane coordinate feature construction method.

[0152] Specifically, the training samples constructed during the training phase can be represented as:

[0153] in, Let X represent the training sample set, q represent the training sample ID, and N represent the number of training samples; n (q) B represents the normalized magnitude matrix of the disturbance-affected values ​​in the q-th training sample; (q) Indicates the sampling location mask; C n (q) Represents the normalized sampling count; X n0 (q) Φ represents the corresponding ideal clean normalized magnitude matrix; 1,0 (q) and Φ 2,0 (q) These represent the true phase labels of the first and second measurement planes, respectively; P u (q) This indicates the coordinate characteristics of the measurement plane and the optional condition characteristics.

[0154] The perturbation magnitude matrix X in the training data n (q) This model is used to simulate the effects that may occur during actual UAV measurements, such as trajectory deviation, local missed sampling, sparse sampling, amplitude noise, channel gain drift, and altitude disturbance. The near-field amplitude data acquired in the practical application phase S1 has the same data format and physical meaning as the disturbed amplitude matrix in the training phase. Therefore, the trained amplitude reconstruction model and phase recovery model can be used to process the near-field amplitude data to be predicted actually acquired in S1.

[0155] After offline training is completed, the trained amplitude reconstruction model and phase retrieval model are obtained, which are represented as follows:

[0156] Among them, F θ ∗ () represents the trained amplitude reconstruction model, θ ∗ G represents the model parameters after training; ψ∗ () represents the trained phase retrieval model, ψ ∗ This represents the model parameters after training.

[0157] In the actual measurement application stage, the raw near-field amplitude data obtained in S1 is mapped to a regular grid and normalized to obtain the amplitude matrix X to be repaired. n Simultaneously, a sampling location mask B and a normalized sampling count map C are generated. n Then, X n B and C n Input the trained amplitude reconstruction model:

[0158] in, This represents the normalized repaired amplitude matrix output by the amplitude reconstruction model. If subsequent steps require corrected near-field amplitude data in dB format, then... Perform inverse normalization to obtain the corrected near-field amplitude data X. c If the phase retrieval network directly receives the normalized amplitude, then let:

[0159] Among them, X cn This represents the normalized correction magnitude matrix of the input phase retrieval network.

[0160] Subsequently, according to X cn And measure the planar coordinate features, and construct the phase retrieval network input as described above:

[0161] in, Z represents the data set input to the phase retrieval network in the practical application stage. L,1 Z L,2 Z H,1 Z H,2 These represent the low-frequency encoded input of the first measurement plane, the low-frequency encoded input of the second measurement plane, the high-frequency encoded input of the first measurement plane, and the high-frequency encoded input of the second measurement plane, respectively; P u Represents the coordinate condition feature map.

[0162] The above inputs are fed into the trained phase retrieval model to obtain the recovered phases of the first and second measurement planes under the x and y polarization components:

[0163] in, This represents the recovered phase data of the first measurement plane. This represents the recovered phase data of the second measurement plane. Therefore, the corrected near-field amplitude data X obtained in practical applications...c and recover phase data , Together they serve as inputs for the subsequent construction of complex near-field data in S4.

[0164] S4: Based on the corrected near-field amplitude data output by the pre-trained amplitude repair network in S2 and the recovered phase data output by the pre-trained phase recovery network in S3, construct complex near-field data on the first and second measurement planes. Determine whether the complex near-field data of the first measurement plane is reliable based on the propagation consistency error of the second measurement plane. If reliable, proceed to S5. Otherwise, adjust the network parameters of the pre-trained amplitude repair network and / or adjust the network parameters of the pre-trained phase recovery network, and repeat steps 2-4 until reliable complex near-field data of the first measurement plane is obtained.

[0165] First, the dB-form corrected near-field amplitude data is converted into linear amplitude data. For any polarization component p, where Its linear amplitude can be expressed as:

[0166] in, This represents the corrected dB amplitude of the polarization component p on the measurement plane z. This represents the corresponding linear amplitude.

[0167] Then, the linear amplitude and the recovered phase are combined to obtain the complex electric field data under the corresponding polarization component:

[0168] in, This represents the complex near-field electric field of the polarization component p on the measurement plane z, where j is the imaginary unit. This corresponds to the phase output of the phase recovery network.

[0169] Therefore, the complex near-field data of the first measurement plane can be expressed as:

[0170] The complex near-field data of the second measurement plane can be expressed as:

[0171] Then, based on the complex near-field data E of the polarization component p on the first measurement plane... p,z1 The propagation is carried out through the angular spectrum propagation operator to the second measurement plane, resulting in the propagation prediction complex field of the second measurement plane:

[0172] Where, p∈{E x E y}, Represents a two-dimensional Fourier transform. This represents the two-dimensional inverse Fourier transform. H(k) represents the distance between the first and second measuring planes. x ,k y ,Δz) represents the angular spectrum propagation transfer function.

[0173] After obtaining the propagation prediction complex field, its phase is extracted as the propagation consistency reference phase of the second measurement plane:

[0174] Furthermore, the propagation consistency reference phase is recombine with the linear amplitude of the second measurement plane to obtain an amplitude-constrained second measurement plane reference complex field:

[0175] in, This represents the linear amplitude of the polarization component p on the second measurement plane. This represents the reference complex field composed of the propagation phase of the first measurement plane and the linear amplitude of the second measurement plane.

[0176] To determine whether the phase retrieval result satisfies the dual-plane propagation consistency requirement, the propagation consistency error of the second measurement plane is calculated. It consists of two planes It consists of half of the sum of and . Among them, E p,z2 This represents the complex near-field data directly constructed from the output phase of the linear amplitude and phase retrieval network of the second measurement plane. A very small positive number set to prevent the denominator from being zero. If ε prop If the value is less than a preset threshold, it indicates that the complex near-field data from the first measurement plane, after propagation to the second measurement plane, has good consistency with the measured amplitude and recovered phase of the second measurement plane. Therefore, the complex near-field data from the first measurement plane is considered reliable and can be used as input for subsequent near-field to far-field conversion. If ε prop If the value exceeds the preset threshold, it indicates that the current phase retrieval result may have significant uncertainty. Therefore, adjust the network parameters of the pre-trained amplitude repair network and / or the network parameters of the pre-trained phase retrieval network, repeating steps 2-4 until ε... prop If the value is less than a preset threshold, reliable complex near-field data of the first measurement plane can be obtained.

[0177] S5: Based on the reliable first measurement plane complex near-field data obtained in S4, perform near-far field conversion to obtain the far-field radiation pattern of the antenna under test.

[0178] Specifically, a planar near-far field conversion is performed using the reliable first measurement plane complex near-field data obtained in step S4. The first measurement plane complex near-field data includes E x and E y The two tangential complex electric field components are denoted as:

[0179] Among them, E x,z1 (i,j) represents the Ex-polarized complex electric field at the (i,j)-th grid point on the first measurement plane, E y,z1 (i,j) represents the position E at the (i,j)th grid point on the first measurement plane. y The polarized complex electric field, i=1,2,⋯,H, j=1,2,⋯,W, represents the coordinate variables of the grid points.

[0180] First, the grid spacing is calculated based on the regular two-dimensional grid coordinates of the first measurement plane. Let the horizontal coordinate sequence be x1, x2, ..., x... W The vertical coordinate sequence is y1, y2, ..., y H The horizontal and vertical grid spacings are respectively:

[0181] When the grid is sampled at equal intervals, it can also be written as:

[0182] Where Δx represents the horizontal sampling interval and Δy represents the vertical sampling interval, both in meters.

[0183] Then, the operating wavelength and free-space wavenumber are calculated based on the operating frequency f of the antenna under test:

[0184] Where λ represents the operating wavelength, c represents the speed of light in a vacuum, f represents the operating frequency of the antenna under test, and k0 represents the free-space wavenumber. If the operating frequency is given in GHz, it is first converted to Hz:

[0185] And substitute:

[0186] Next, the complex near-field data of the first measurement plane is zero-filled. In this embodiment, a fixed zero-fill of 4 times is used. Let the original number of near-field sampling points be H×W, and the number of sampling points after zero-filling be:

[0187] Among them, H p and W pThese represent the number of longitudinal and lateral sampling points after zero-filling, respectively. Zero-filling involves padding the edges of the original complex near-field matrix with zeros, without altering the complex field values ​​at the original measurement points. The resulting complex near-field matrix after zero-filling is:

[0188] Among them, E x,z1 p and E y,z1 p These represent E after zero padding. x E y Polarized complex near-field matrix.

[0189] In this implementation, no additional window function is applied to the near-field data; that is, a rectangular window is used.

[0190] Therefore, the near-field data input for the Fourier transform are:

[0191] in, and This represents the complex near-field data used for plane spectrum calculations.

[0192] Then, to , Two tangential complex electric field components , Perform two-dimensional Fourier transforms to obtain the plane spectral components:

[0193] Among them, F x (k x ,k y ) and F y (k x ,k y ) represent E respectively x E y The plane spectrum corresponding to the component, F 2D () denotes a two-dimensional Fourier transform, k x and k y These represent the horizontal and vertical spatial wavenumbers, respectively.

[0194] After zero-filling, the space wavenumber sampling points are determined according to the following formula:

[0195] in:

[0196] Furthermore, the longitudinal wavenumber component is calculated based on the free-space propagation conditions:

[0197] Where, k z (k x ,k y ) represents the longitudinal spatial wavenumber component, k x= k y= For those satisfying k x 2 +k y 2 ≤k0 2 The spectral points and their corresponding propagating wave components participate in the far-field calculation; for evanescent wave components that do not meet this condition, they are not considered as effective far-field radiation components in the far-field pattern calculation.

[0198] Based on the transverse and longitudinal spectral components, the longitudinal electric field spectral component can be obtained:

[0199] Among them, F z (k x ,k y ) represents the longitudinal spectral component derived from the transverse spectral component.

[0200] Then, spectral values ​​for the corresponding directions are extracted from the plane wave spectrum based on the far-field observation angle θ and azimuth angle φ. This implementation outputs two principal tangential radiation patterns, namely the φ=0° tangential and φ=90° tangential patterns. For a given observation angle θ, the space wavenumber correspondence is as follows:

[0201] For a tangent with φ=0°, we have:

[0202] The corresponding far-field angular and azimuth components are:

[0203] Therefore, the far-field amplitude of the φ=0° section is:

[0204] Similarly, the far-field amplitude of the φ=90° section is:

[0205] Finally, the far-field amplitude is converted to dB form:

[0206] in, This represents the far-field radiation pattern under the azimuth section φ, where ε is a very small positive number to prevent taking the logarithm of zero, and its value is no greater than 10. -6 .

[0207] To facilitate radiation pattern comparison, the far-field radiation pattern under each cross-section is normalized relative to its own maximum value:

[0208] Thus, the normalized far-field patterns of the two principal tangent planes, φ=0° and φ=90°, are obtained:

[0209] Based on the far-field radiation pattern, the far-field radiation parameters of the antenna under test can be further extracted, including the main lobe direction, half-power beamwidth, sidelobe level, and null position.

[0210] Furthermore, this invention provides a near-field to far-field conversion system for near-field antenna measurement of unmanned aerial vehicles (UAVs). This system specifically includes a UAV, a signal source, a probe antenna, an antenna under test (UTP), a spectrum analyzer, and an edge computing unit. Specifically, the UAV carries the signal source and probe antenna and flies along a preset path. The probe antenna transmits a signal, and the UTP receives the signal. The signal is transmitted to the spectrum analyzer via a transmission line. The spectrum analyzer then transmits the amplitude data to the edge computing unit. The edge computing unit calls a trained network or directly calls an algorithm to output the far-field radiation pattern of the UTP.

[0211] It should be noted that the specific heights, measurement ranges, sampling intervals, polarization forms, model structure details, and edge processor models of the first and second measurement planes in the above embodiments can all be adjusted according to actual application scenarios, and this invention does not limit them. Any technical solution based on dual-plane dual-polarization near-field amplitude data, which undergoes correction or repair, phase recovery, complex near-field construction, and near-far-field conversion to obtain far-field results, and can perform on-site inference on an edge device, is applicable to the technical concept of this invention.

[0212] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A near-field to far-field conversion method for near-field antenna measurement of unmanned aerial vehicles, characterized in that, The method includes the following steps: S1: Acquire near-field amplitude data on a first measurement plane and a second measurement plane parallel to the aperture plane of the antenna under test using a drone. The near-field amplitude data includes at least the x-direction polarization component amplitude E. x and the amplitude of the polarization component in the y direction E y The distances between the first and second measurement planes and the aperture plane of the antenna under test are both between 3 and 10 times the operating wavelength of the antenna under test; let the equivalent aperture dimensions of the antenna under test in the x and y directions be D respectively. x and D y The distance between the measuring plane and the aperture surface of the antenna under test is z, and the expected maximum far-field observation angle is [value missing]. Then the dimensions of the measurement area of ​​the measurement plane in the x and y directions are... L x , L y Set it according to the following formula: ， ， in, and These are the margins in the x and y directions, respectively; S2: Preprocess the near-field amplitude data of the two measurement planes in S1, and then use the pre-trained amplitude restoration network to obtain the corrected near-field amplitude data of the two measurement planes. The preprocessing procedure for the near-field amplitude data of the two measurement planes in S1 is as follows: First, a regular two-dimensional grid is set according to the measurement area size of the measurement plane and the preset grid resolution. The coordinates of each sampling point in S1 and the corresponding near-field amplitude data are mapped to the above two-dimensional grid. For cases where there are multiple sampling points in the same grid, the amplitudes of multiple sampling points are aggregated. The aggregation method can be arithmetic mean, weighted average, or median aggregation. For grid positions not covered by sampling points, amplitude filling is performed. Thus, the four-channel amplitude matrices to be repaired for the two measurement planes are obtained. : ， in, and These represent the amplitude diagrams in the x and y polarization directions on the first measurement plane, respectively. and These represent the amplitude diagrams in the x and y polarization directions on the second measurement plane, respectively. The amplitude matrix X is subjected to amplitude clipping and normalization, and mapped to a preset numerical range to obtain a normalized amplitude matrix. Simultaneously, a sampling location mask and a sampling count map are generated. The sampling location mask indicates whether the corresponding grid has a valid measurement value, and the sampling count map indicates the number of times the corresponding grid is sampled. The sampling location mask B and the sampling count map C are used as auxiliary inputs. To reduce the influence of differences in sampling count scales, the sampling count map is logarithmically normalized, resulting in: ， in, This represents the maximum number of samples taken within the regular two-dimensional grid of the current measurement plane. The above normalized magnitude matrix Sampling location mask B, normalized sampling number diagram By splicing the first measurement plane and the second measurement plane respectively, we obtain the feature Z1 of the first measurement plane and the feature Z2 of the second measurement plane; Normalized magnitude matrix Features Z1 and Z2 are used as inputs to a pre-trained amplitude restoration network to obtain corrected near-field amplitude data. ， in, and These represent the normalized amplitude diagrams in the x and y polarization directions on the first measurement plane, respectively. and These represent the normalized amplitude diagrams for the x and y polarization directions on the second measurement plane, respectively. ， ， Concat() indicates concatenation along the channel dimension. and These represent the sampling position masks for the first and second measurement planes, respectively. and These represent the normalized sampling counts for the first and second measurement planes, respectively. The pre-trained amplitude restoration network includes a first measurement plane coding branch, a second measurement plane coding branch, a cross-plane feature exchange unit, a multi-scale fusion unit, an upsampling decoding unit, and a residual amplitude output unit; The first measurement plane coding branch and the second measurement plane coding branch have the same structure, both containing a shallow coding layer, a middle coding layer, and a deep coding layer. The shallow coding layer is the initial convolutional unit, the middle coding layer is the first-level sampling module, and the deep coding layer is the second-level sampling module. In the first measurement plane coding branch, Z1 is input into the shallow coding layer to obtain the shallow local amplitude feature K1 of the first measurement plane. (0) In the second measurement plane encoding branch, Z2 is input into another set of initial convolutional units to obtain the shallow local amplitude features K2 of the second measurement plane. (0) The initial convolutional unit includes a two-dimensional convolutional layer, a normalization layer, and a non-linear activation layer connected in sequence; the first-level downsampling module reduces the feature resolution from H×W to H / 2×W / 2 and increases the number of channels; the second-level downsampling module further reduces the feature resolution to H / 4×W / 4 and further increases the number of channels; each downsampling module includes a two-dimensional convolutional layer with a stride of 2, a normalization layer, a non-linear activation layer, and a residual convolutional block connected in sequence, the residual convolutional block includes a first two-dimensional convolutional layer, a normalization layer, a non-linear activation layer, and a second two-dimensional convolutional layer connected in sequence, the output of the second two-dimensional convolutional layer is added element-wise to the input features of the residual convolutional block to obtain the output features of the residual convolutional block; A cross-plane feature exchange unit is set after each coding layer, and each cross-plane feature exchange unit includes a first plane update branch and a second plane update branch; for coding layers l Local amplitude characteristics K1 of the first measurement plane output (l) Local amplitude characteristics K2 of the second measurement plane (l) First, concatenation is performed along the channel dimension to obtain cross-plane joint features: ， Then, the cross-plane joint feature P (l) Input cross-plane feature exchange unit respectively l The first and second plane update branches are provided; each update branch includes a 1×1 convolutional mapping layer, a normalization layer, a nonlinear activation layer, a residual convolutional block, and a 1×1 output convolutional layer connected in sequence, used to predict the residual update features of the current measurement plane from the joint features of the two planes: ， in, and Representing cross-plane feature exchange units l The first and second plane update branches are then used; subsequently, the residual update features are added element-wise to the original local amplitude features of the corresponding measurement plane to obtain the update features after cross-plane information exchange: ， in, Indicates the first measurement plane in the cross-plane feature exchange unit l Updated features after feature swapping Indicates the second measurement plane in the cross-plane feature exchange unit l Updated features after feature swapping; When the coding layer l When it is a shallow coding layer, cross-plane feature exchange unit l Output shallow layer update features; when the encoding layer l When it is a mid-level coding layer, the cross-plane feature exchange unit l Output the middle layer to update features; when the coding layer l When it is a deep coding layer, cross-plane feature exchange unit l Output deep update features; The multi-scale fusion unit includes a shallow fusion convolution module, a deep fusion convolution module, and a shallow fusion convolution module. For the deep scale, the updated features of the first and second measurement planes are concatenated by channels and then processed by the deep fusion convolution module to obtain the deep fused features. ， For the mid-level scale, the mid-level update features of the first and second measurement planes are concatenated by channels and then passed through a mid-level fusion convolution module to obtain the mid-level skip fusion features: ， For shallow scales, the shallow update features of the first and second measurement planes are concatenated by channels and then processed through a shallow fusion convolution module to obtain shallow skip fusion features: ， in, , , These represent deep, middle, and shallow fusion convolutional modules, respectively. Each fusion convolutional module includes a 1×1 convolutional mapping layer, a normalization layer, a nonlinear activation layer, and a residual convolutional block connected in sequence, used to compress the number of channels after splicing and fuse complementary information from the two measurement planes. Subsequently, the upsampling decoding unit first processes the deep fusion feature U (2) Upsampled to the mid-level feature size and fused with the mid-level skip feature U (1) By concatenating the data, we obtain the input for the middle layer decoding: ， Where Up() represents the upsampling operation; the middle layer decoding input D (1) After sequentially performing convolution, normalization, nonlinear activation, and residual convolution blocks, the mid-layer decoding features R are obtained. (1) ; Then, the middle layer decoding feature R (1) Upsampled to the original grid resolution and fused with shallow skip features U (0) By concatenating the inputs, we obtain the shallow decoding input: ， The shallow decoding input D (0) After being processed sequentially through convolution, normalization, nonlinear activation, and residual convolution blocks, the final decoded feature R is obtained. (0) ; Finally, the residual amplitude output unit predicts the four-channel amplitude correction amount based on the final decoded feature R(0); the residual amplitude output unit uses a 3×3 convolutional layer to map the final decoded feature into a four-channel residual, and limits the residual amplitude using a hyperbolic tangent function: ， Where, ΔX n This represents the four-channel residual correction amount of the normalized magnitude matrix, α represents the residual scaling factor, and tanh() is used to limit the single correction magnitude. The predicted residual correction amount ΔX n With the input normalized magnitude matrix X n Element-wise summation, followed by range clipping, yields the normalized repair magnitude matrix: ， in, The normalized repaired magnitude matrix output by the pre-trained magnitude repair network; S3: Input the corrected near-field amplitude data of the two measurement planes output by the pre-trained amplitude restoration network in S2 into the pre-trained phase recovery network to obtain the recovered phase data of the first measurement plane and the second measurement plane under the corresponding polarization components; The corrected near-field amplitude data includes the amplitude data of the first measurement plane and the second measurement plane under the x and y polarization components, denoted as: ， in, This indicates the correction of near-field amplitude data. and These represent the normalized corrected amplitude data of the x and y polarization components on the first measurement plane, respectively. and These represent the normalized corrected amplitude data of the x and y polarization components on the second measurement plane, respectively; H and W represent the number of sampling points of the regular grid in the longitudinal and transverse directions, respectively. Input data for constructing a pre-trained phase retrieval network: First, obtain the coordinate feature map: Let the horizontal and vertical coordinates corresponding to the i-th row and j-th column in the regular two-dimensional grid of the measurement plane be x and x, respectively. j and y i Construct horizontal coordinate graphs, vertical coordinate graphs, and radial position graphs based on the coordinate information: ， Among them, P x (i,j) represents the normalized lateral coordinate feature of the (i,j)th grid point, P y (i,j) represents the normalized longitudinal coordinate feature of the (i,j)-th grid point, P r (i,j) represents the normalized radial position feature of the (i,j)th grid point; s xy and s r These are the coordinate normalization scales, calculated as follows: ， Here, ε is a very small positive number set to prevent the denominator from being zero; thus, the coordinate characteristic diagram is obtained: ， in, This represents the measurement planar coordinate feature map; Concat() represents concatenation of channel dimensions. Secondly, regarding X cn Low-frequency and high-frequency information are decomposed; the low-frequency amplitude component and the high-frequency amplitude component are represented as follows: ， Among them, X L X represents the low-frequency amplitude component. H Indicates the high-frequency amplitude component. denoted as a two-dimensional Gaussian smoothing operator, where σ represents the Gaussian smoothing parameter; The low-frequency amplitude component is decomposed according to the measurement height as follows: ， The high-frequency amplitude component is separated according to the measurement height as follows: ， Among them, X L,1 and X L,2 X represents the low-frequency dual-polarization amplitude characteristics of the first and second measurement planes, respectively. H,1 and X H,2 These represent the high-frequency dual-polarization amplitude characteristics of the first and second measurement planes, respectively. Furthermore, the mean and difference characteristics between the two measurement planes are constructed; for the low-frequency amplitude component, we have: ， For high-frequency amplitude components, we have: ， in, and These represent the mean amplitude characteristics of the two measurement planes under low-frequency and high-frequency conditions, respectively, and are used to characterize the common amplitude structure of the two measurement planes; ΔX L and ΔX H These represent the difference amplitude characteristics of two measurement planes under low-frequency and high-frequency conditions, respectively, and are used to characterize the amplitude variation information between the two measurement heights; Therefore, the inputs to the first measurement plane low-frequency coding branch, the second measurement plane low-frequency coding branch, the first measurement plane high-frequency coding branch, and the second measurement plane high-frequency coding branch in the phase retrieval network are respectively: ， Among them, Z L,1 Z represents the input of the low-frequency coding branch of the first measurement plane. L,2 Z represents the input of the low-frequency coding branch of the second measurement plane. H,1 Z represents the input of the high-frequency coding branch of the first measurement plane. H,2 This represents the input to the high-frequency coding branch of the second measurement plane; The phase recovery network includes a first measurement plane low-frequency coding branch, a second measurement plane low-frequency coding branch, a first measurement plane high-frequency coding branch, a second measurement plane high-frequency coding branch, a frequency feature fusion unit, a cross-plane feature exchange unit, a shared fusion unit, a shared upsampling unit, and a polarization phase decoding unit. Each low-frequency coding branch and high-frequency coding branch contains multiple frequency coding layers, Z L,1 Z L,2 Z H,1 Z H,2 These correspond to the inputs of the first-level low-frequency coding layer and the first-level high-frequency coding layer of the first and second measurement planes, respectively; let m = 1, 2, representing the measurement plane number. l =1,⋯,N represents the frequency coding level, then the m-th measurement plane is in the _____. l The low-frequency coding features and high-frequency coding features output by the frequency coding layer are denoted as follows: ， in, Indicates the first l The coding mapping of the low-frequency coding layer, Indicates the first l The encoding mapping of the first-level frequency coding layer consists of a convolutional layer, a normalization layer, a nonlinear activation layer, and a residual convolutional block connected in sequence. Subsequent frequency coding layers, based on the structure of the first-level frequency coding layer, connect a downsampling module after the residual convolutional block. The output of the previous-level frequency coding layer serves as the input of the next-level frequency coding layer. The first-level frequency coding layer is defined as a shallow coding layer, outputting shallow-scale frequency coding features. The first 2 - |N / 2| frequency coding layers are defined as intermediate coding layers, with their last layer outputting intermediate-scale frequency coding features. || represents the rounding operation. The remaining frequency coding layers are defined as deep coding layers, with their last layer outputting deep-scale frequency coding features. For shallow-scale, mid-scale, and deep-scale frequency-coded features, they are processed by a frequency feature fusion unit, a cross-plane feature exchange unit, and a shared fusion unit, respectively. The frequency feature fusion unit fuses low-frequency and high-frequency coded features at each scale to obtain the frequency fused features of the m-th measurement plane. ， in, This represents a frequency feature fusion unit at scale r; r Indicates shallow scale, mesoscale, or shallow scale. The unit first concatenates the low-frequency and high-frequency encoded features at scale r of the m-th measurement plane to obtain joint features for generating gate weights. Then, it uses a gated mapping layer consisting of a 1×1 convolutional layer and a sigmoid function to obtain gate weights between 0 and 1, which are then multiplied element-wise with the high-frequency features. Finally, the low-frequency features are concatenated and fused with the gated high-frequency features to obtain the frequency fusion features. This enhances effective local details while preserving the global structure. The gated mapping layer includes a 1×1 convolutional layer and a sigmoid function connected in sequence. The 1×1 convolutional layer is used to map the input joint features to gated features with the same size as the features to be adjusted. The sigmoid function is used to map the gated features to a range of 0 to 1 to obtain the gated weights. After obtaining the frequency fusion features of the two measurement planes respectively, the cross-plane feature exchange unit performs bidirectional information exchange at each scale; for each scale... r Based on the frequency fusion characteristics, joint features from the first measurement plane to the second measurement plane and from the second measurement plane to the first measurement plane are constructed: ， Among them, J1 (r) J2 represents the cross-plane joint feature used to update the features of the first measurement plane. (r) This represents the cross-plane joint feature used to update the features of the second measurement plane, where * indicates element-wise multiplication; The cross-plane feature exchange unit generates update values ​​based on the aforementioned joint features, and injects them into the original features of the corresponding measurement plane through a gating method to obtain the updated features: ， in, and These represent the updated features of the first and second measurement planes after cross-plane information exchange, respectively. and These represent the convolution update mappings of the first and second measurement planes, respectively. and These represent the gating weights of the first and second measurement planes, respectively. This indicates a residual convolution thinning operation, which processes the gated updated features through at least one residual convolution block to obtain the updated features of the corresponding measurement plane; here, * indicates element-wise multiplication. Then, the shared fusion unit fuses the updated features of the two measurement planes at the same scale to obtain shared features: ， Among them, S (r) s represents the shared features at scale r. (r) () represents a shared fusion unit, which concatenates the features of the two measurement planes, their absolute difference features, and their product features into channels, and then completes channel compression and information fusion through 1×1 convolution, normalization, nonlinear activation, and residual convolution blocks. Let r = 0, 1, 2, corresponding to the shallow, medium, and deep scales, respectively; in the shared upsampling unit, the deep scale shares feature S. (2) First, the bottleneck residual module is input. This module further extracts deep-scale shared features while maintaining feature resolution. The processed features are then added element-wise to the input features via residual connections to obtain the bottleneck residual features. Subsequently, the bottleneck residual features are input to the channel attention module. This module first performs global average pooling on each channel of the bottleneck residual features to obtain global statistical features for each channel. Then, channel weights are mapped onto these global statistical features, and the weights for each channel are obtained using the Sigmoid function. Finally, the channel weights are multiplied channel-wise by the bottleneck residual features to obtain the deep global shared features G. (2) Then, the deep globally shared features G (2) Upsampled to the mid-scale and sharing features S with the mid-scale (1) By performing skip connection fusion, the middle-layer shared decoding feature G is obtained. (1) : ， Where Up() represents the upsampling operation, D s (1) () indicates the intermediate shared decoding convolution module, which will upsample the deep global shared features G. (2) Sharing features S with the mid-level scale (1) Channel concatenation is performed, and the concatenated features are sequentially processed through a 2D convolutional layer, a normalization layer, a nonlinear activation layer, and a residual convolutional block to obtain the middle-layer shared decoding features G. (1) ; Subsequently, the middle layer shared decoding feature G (1) These features are combined with the mid-scale update features of the first and second measurement planes, respectively, to form the first measurement plane decoding input feature D1 and the second measurement plane decoding input feature D2: ， in, and These are the update features of the first and second measurement planes at the mesoscale, respectively, D1. (1) () and D2 (1) () represent the middle-layer decoding convolutional modules of the first and second measurement planes, respectively. The two middle-layer decoding convolutional modules have the same structure and independent parameters. They both include a two-dimensional convolutional layer, a normalization layer, a nonlinear activation layer and a residual convolutional block connected in sequence, which are used to fuse the middle-layer shared decoding features and the middle-layer scale update features of the corresponding measurement plane. For each measurement plane, the polarization phase decoding unit is set to E respectively. x Polarization phase decoder and E y Polarization phase decoder; for the m-th measurement plane, the polarization component p∈{E x E y }, its polarization phase decoder uses the decoding input feature D of the m-th measurement plane. m and shallow-scale update features of the same measurement plane For input; the polarization phase decoder will use D m After upsampling to a shallow scale, the features are updated together with those at the shallow scale. Channel concatenation is performed, followed by residual convolution for thinning, to obtain polarization phase decoding features: ， Among them, Q m,p D represents the phase decoding characteristics of the m-th measurement plane under polarization component p. m,p (0) () indicates the corresponding shallow polarization phase decoding module; subsequently, the polarization phase decoder determines the polarization phase decoding module based on Q. m,p Output low-frequency phase representation and residual phase representation; Subsequently, the low-frequency phase output branch is based on the phase decoding feature Q. m,p The low-frequency phase components of two channels are output through a two-dimensional convolutional layer, and the two low-frequency phase components are vector normalized to obtain a two-dimensional unit vector: ， in, and Let the low-frequency phase cosine component and low-frequency phase sine component of the polarization component p on the m-th measurement plane be represented respectively. The two normalized components satisfy: ， Calculate the low-frequency phase based on the low-frequency phase cosine component and the low-frequency phase sine component: ， Meanwhile, the residual phase prediction branch outputs a constrained residual phase: ， in, This represents the low-frequency phase of the polarization component p on the m-th measurement plane. R represents the corresponding residual phase. m,p This represents the residual phase representation of the output. ρ This represents the residual phase scaling factor; the final recovered phase is: ， Here, `wrap()` represents the phase wrapping operation, used to constrain the phase within the range of (-π, π]. Thus, the phase retrieval network ultimately outputs the dual-polarized phase data for the first and second measurement planes. ， in, This represents the recovered phase data of the first measurement plane. This represents the recovered phase data of the second measurement plane; S4: Based on the corrected near-field amplitude data output by the pre-trained amplitude repair network in S2 and the recovered phase data output by the pre-trained phase recovery network in S3, construct complex near-field data on the first and second measurement planes. Determine whether the complex near-field data of the first measurement plane is reliable based on the propagation consistency error of the second measurement plane. If reliable, proceed to S5. Otherwise, adjust the network parameters of the pre-trained amplitude repair network and / or adjust the network parameters of the pre-trained phase recovery network. Repeat steps 2-4 until reliable complex near-field data of the first measurement plane is obtained. First, the dB-form corrected near-field amplitude data is converted into linear amplitude data; for any polarization component p, where p∈{E x E y Its linear magnitude can be expressed as: ， in, This represents the corrected dB amplitude of the polarization component p on the measurement plane z. This represents the corresponding linear amplitude; Then, the linear amplitude and the recovered phase are combined to obtain the complex electric field data under the corresponding polarization component: ， in, This represents the complex near-field electric field of the polarization component p on the measurement plane z, where j is the imaginary unit. The corresponding phase output by the phase recovery network; Therefore, the complex near-field data of the first measurement plane can be expressed as: ， The complex near-field data of the second measurement plane can be expressed as: ， Then, based on the complex near-field data E of the polarization component p on the first measurement plane... p,z1 The propagation is carried out through the angular spectrum propagation operator to the second measurement plane, resulting in the propagation prediction complex field of the second measurement plane: ， Where, p∈{E x E y }, Represents a two-dimensional Fourier transform. This represents the two-dimensional inverse Fourier transform. H(k) represents the distance between the first and second measuring planes. x ,k y ,Δz) represents the angular spectrum propagation transfer function; After obtaining the propagation prediction complex field, its phase is extracted as the propagation consistency reference phase of the second measurement plane: ， Furthermore, the propagation consistency reference phase is recombine with the linear amplitude of the second measurement plane to obtain an amplitude-constrained second measurement plane reference complex field: ， in, This represents the linear amplitude of the polarization component p on the second measurement plane. This represents the reference complex field composed of the propagation phase from the first measurement plane and the linear amplitude of the second measurement plane. To determine whether the phase retrieval result satisfies the dual-plane propagation consistency requirement, the propagation consistency error of the second measurement plane is calculated. It consists of two planes It consists of half of E and E; where E p,z2 This represents the complex near-field data directly constructed from the output phase of the linear amplitude and phase retrieval network of the second measurement plane. A very small positive number set to prevent the denominator from being zero; if ε prop If the value is less than a preset threshold, it indicates that the complex near-field data from the first measurement plane, after propagation to the second measurement plane, has good consistency with the measured amplitude and recovered phase of the second measurement plane. Therefore, the complex near-field data from the first measurement plane is considered reliable and can be used as input for subsequent near-field to far-field conversion. If ε prop If the value exceeds the preset threshold, it indicates that the current phase retrieval result may have significant uncertainty. Therefore, adjust the network parameters of the pre-trained amplitude repair network and / or the network parameters of the pre-trained phase retrieval network, repeating steps 2-4 until ε... prop If the value is less than a preset threshold, reliable complex near-field data of the first measurement plane can be obtained. S5: Based on the reliable first measurement plane complex near-field data obtained in S4, perform near-far field conversion to obtain the far-field radiation pattern of the antenna under test; Specifically, a planar near-far field conversion is performed using the reliable first measurement plane complex near-field data obtained in step S4. The first measurement plane complex near-field data includes E x and E y The two tangential complex electric field components are denoted as: ， Among them, E x,z1 (i,j) represents E at the (i,j)th grid point on the first measurement plane. x Polarized complex electric field, E y,z1 (i,j) represents the position E at the (i,j)th grid point on the first measurement plane. y The polarized complex electric field, i=1,2,⋯,H, j=1,2,⋯,W, represents the coordinate variables of the grid points; First, the grid spacing is calculated based on the regular two-dimensional grid coordinates of the first measurement plane; let the horizontal coordinate sequence be x1, x2, ..., x W The vertical coordinate sequence is y1, y2, ..., y H The horizontal and vertical grid spacings are respectively: ， Where Δx represents the horizontal sampling interval and Δy represents the vertical sampling interval, both in meters; Then, the operating wavelength and free-space wavenumber are calculated based on the operating frequency f of the antenna under test: ， Where λ represents the operating wavelength, c represents the speed of light in a vacuum, f represents the operating frequency of the antenna under test, and k0 represents the free space wavenumber; if the operating frequency is given in GHz, it is first converted to Hz: ， And substitute: ， Next, the complex near-field data of the first measurement plane is zero-filled, thus obtaining the zero-filled complex near-field matrix: ， Among them, E x,z1 p and E y,z1 p These represent E after zero padding. x E y Polarized complex near-field matrix, H p and W p These represent the number of longitudinal and lateral sampling points after zero-filling, respectively; the zero-filling operation fills the edges of the original complex near-field matrix with zeros without changing the complex field values ​​at the original measurement points; No additional window function is applied to the near-field data; that is, a rectangular window is used. ， Therefore, the near-field data input for the Fourier transform are: ， in, , This represents the complex near-field data used for plane spectrum calculations; Then, to , Two tangential complex electric field components , Perform two-dimensional Fourier transforms to obtain the plane spectral components: ， Among them, F x (k x ,k y ) and F y (k x ,k y ) represent E respectively x E y The plane spectrum corresponding to the component, F 2D {} denotes a two-dimensional Fourier transform, k x and k y These represent the horizontal and vertical spatial wavenumbers, respectively. After zero-filling, the space wavenumber sampling points are determined according to the following formula: ， in: ， Furthermore, the longitudinal wavenumber component is calculated based on the free-space propagation conditions: ， Where, k z (k x ,k y ) represents the longitudinal spatial wavenumber component, k x= k y= For those satisfying k x 2 +k y 2 ≤k0 2 The spectral points and their corresponding propagating wave components participate in the far-field calculation; for evanescent wave components that do not meet this condition, they are not considered as effective far-field radiation components in the far-field pattern calculation. Based on the transverse and longitudinal spectral components, the longitudinal electric field spectral component can be obtained: ， Among them, F z (k x ,k y () represents the longitudinal spectral component derived from the transverse spectral component; Then, spectral values ​​for the corresponding directions are extracted from the plane wave spectrum based on the far-field observation angle θ and azimuth angle φ; for a given observation angle θ, the corresponding space wavenumbers are: ， For a tangent with φ=0°, we have: ， The corresponding far-field angular and azimuth components are: ， Therefore, the far-field amplitude of the φ=0° section is: ， Similarly, the far-field amplitude of the φ=90° section is: ， Finally, the far-field amplitude is converted to dB form: ， in, This represents the far-field radiation pattern under the azimuth section φ, where ε is a very small positive number to prevent taking the logarithm of zero.

2. The near-field to far-field conversion method for near-field antenna measurement of UAVs according to claim 1, characterized in that, In S1, when the UAV acquires near-field amplitude data on the first and second measurement planes, the sampling interval along the x-direction is denoted as... The sampling interval along the y-direction is denoted as Both must satisfy the Nyquist sampling theorem: .

3. The near-field to far-field conversion method for near-field antenna measurement of UAVs according to claim 1, characterized in that, In S1, the edge margin of the measurement area of ​​the measurement plane. to Alternatively, take 10% to 30% of the equivalent aperture size of the antenna under test in the corresponding direction.

4. The near-field to far-field conversion method for near-field antenna measurement of UAVs according to claim 1, characterized in that, In S1, the near-field amplitude at the boundary of the measurement area of ​​the measurement plane decays to below a preset threshold relative to the maximum amplitude within the measurement area.

5. The near-field to far-field conversion method for near-field antenna measurement of UAVs according to claim 1, characterized in that, In S2, the amplitude of the grid not covered by the sampling point can be filled with the minimum effective amplitude of the corresponding channel or a fixed amplitude lower than the minimum effective amplitude; or, the missing grid position can be filled by diffusion based on the amplitude of the neighboring sampled grids; specifically, with the missing grid as the center, sampled grids are searched within a preset neighborhood, and the amplitudes of the neighboring sampled grids are weighted and averaged according to the spatial distance to obtain the initial filling value of the missing grid; when there are no sampled grids in the preset neighborhood, the search neighborhood is gradually expanded until at least one sampled grid is obtained or the maximum search radius is reached; for continuous missing regions, iterative filling can be carried out in a way that diffuses layer by layer from the sampled region to the missing region, so that the amplitude of the missing region is smoothly connected with the surrounding sampled region.

6. The near-field to far-field conversion method for near-field antenna measurement of UAVs according to claim 1, characterized in that, In S2, the pre-trained amplitude restoration network is trained using electromagnetic simulation and flight simulation data samples. Specifically, the ideal near-field amplitude obtained from electromagnetic simulation is used as the training label, and the near-field amplitude data obtained from flight simulation under flight disturbances is used as the training samples to simulate near-field amplitude data collected in real-world scenarios. The specific process for obtaining the pre-trained amplitude restoration network is as follows: First, a simulation model of the antenna under test and the first and second measurement planes in S1 is established using simulation. The simulation model includes the geometry, material parameters, feeding method, operating frequency, boundary conditions, measurement plane position, measurement area size, grid resolution, and near-field amplitude extraction methods for x-direction polarization and y-direction polarization of the antenna under test. A full-wave electromagnetic simulation platform was used to establish a model of the antenna under test. During the simulation, the operating frequency of the antenna under test was set to f, and the corresponding wavelength was set to λ. The first measurement plane and the second measurement plane were respectively set at a distance from the aperture plane of the antenna under test. and The positions of the two measurement planes are parallel to the aperture surface of the antenna under test; the measurement area size and grid resolution are consistent with the regular two-dimensional grid set in S2; the simulation boundary adopts a radiating boundary, an absorbing boundary, or a perfectly matched layer boundary to simulate free space radiation conditions; after the electromagnetic simulation is completed, the near-field amplitude data at the center position of each grid on the first and second measurement planes are extracted to obtain the amplitude maps of the polarization components in the x and y directions, respectively, thus forming the four-channel ideal amplitude matrix X of the two measurement planes. 0 Furthermore, to simulate the positional deviations and sampling gaps that occur during actual UAV flight sampling, a flight simulation platform is used to generate a disturbance sampling trajectory. A UAV flight simulation environment is established using the Gazebo platform, and the UAV's body parameters, flight speed, trajectory planning parameters, positioning error parameters, and wind speed, wind direction, and weather disturbance parameters are input, causing the UAV to fly along the same trajectory as the actual test mission. Based on the UAV sampling position output by Gazebo, the ideal near-field amplitude field obtained from the electromagnetic simulation is resampled to obtain near-field amplitude data affected by flight disturbances. The amplitude data at the center of the regular two-dimensional grid obtained from the electromagnetic simulation is used to construct the ideal amplitude matrix X. 0 The disturbance trajectory obtained from flight simulation is used to construct the disturbed sampling points, sampling position mask B, sampling number map C, and normalized amplitude matrix X. n Then, the ideal magnitude matrix X 0 Amplitude clipping and normalization are performed to obtain the normalized amplitude matrix X. n Reference matrix X n 0 ; Then, the model parameters of the amplitude repair network are initialized; let the amplitude repair network be F. θ (•), where θ represents the trainable parameters in the first measurement plane coding branch, the second measurement plane coding branch, the cross-plane feature exchange unit, the multi-scale fusion unit, the upsampling decoding unit, and the residual amplitude output unit; for the two-dimensional convolutional layers and 1×1 convolutional mapping layers other than the residual amplitude output unit, they are randomly initialized according to their number of input channels, number of output channels, and convolutional kernel size. Specifically, the Kaiming uniform initialization method is used to initialize the convolutional kernel weights to random numbers near the zero mean and initialize the corresponding biases to zero; for the normalization layers in the network, the scaling factor of the normalization layer is initialized to 1 and the translation bias is initialized to 0; if the normalization layer contains running mean and running variance, the running mean is initialized to 0 and the running variance is initialized to 1; Furthermore, the final 3×3 convolutional layer in the residual amplitude output unit, used to generate the four-channel residual correction, is zero-initialized; the final 3×3 convolutional layer refers to the layer in the residual amplitude output unit located before the hyperbolic tangent function, used to generate the final decoded feature R. (0) The convolutional layer is mapped to a four-channel residual correction layer; the weight matrix and bias vector of this terminal convolutional layer are denoted as W. o and b o Then initialize the command: W o =0,b o =0, At this point, the residual amplitude output unit satisfies the following in the initial training phase: ， Where * represents the convolution operation and α represents the residual scaling factor; therefore, the initial output of the magnitude repair network is: ， That is, the network approximately performs an identity mapping at the beginning of training, and does not immediately change the normalized magnitude matrix of the input. As training iterates, the parameters of the final convolutional layer and the previous feature extraction layer are gradually updated under the supervision loss constraint. The network then gradually learns the amplitude residual correction amount caused by UAV position jitter, uneven sampling, local missing measurements and noise disturbance. Through the above initialization method, the initial physical distribution of the input amplitude data can be preserved, and the stability of subsequent phase recovery and near-far field conversion can be avoided due to excessive random residual output in the early stage of training. Then, training samples are constructed; for the q-th training sample, the normalized amplitude matrix X affected by flight disturbances is... n (q) Sampling location mask B (q) Plot C of normalized sampling times n (q) As model input, the corresponding normalized ideal magnitude matrix X n0 (q) As a supervisory label; construct the first measurement plane input feature Z1 according to the aforementioned method. (q) Second measurement plane input feature Z2 (q) And input the amplitude repair network: ， in, The normalized repaired amplitude matrix output by the amplitude repair network has the same number of channels as the input amplitude matrix, and still includes four-channel amplitude data in the x and y polarization directions of the first and second measurement planes. In each iteration of training, a batch of training samples is selected from the training set, denoted as: ， in, Let M represent the batch training sample set used in the t-th training iteration, M represent the number of training samples in the batch, and q represent the sample number in the batch training sample set; for any sample q, the normalized amplitude matrix affected by flight disturbances is... Sampling location mask Normalized sampling number graph Input the amplitude restoration network to obtain the corresponding normalized restored amplitude matrix: ， in, This represents the normalized repaired amplitude matrix output by the amplitude repair network after the q-th training sample is processed. This represents the corresponding normalized ideal magnitude matrix; Then, the amplitude reconstruction loss is calculated based on this batch of training samples; the amplitude reconstruction loss is used to constrain the repaired amplitude matrix to approximate the normalized ideal amplitude matrix. It can be represented as all The average value; where, This represents the average absolute error calculated over the four amplitude channels and each grid position. To further maintain the consistency of the near-field amplitude spatial variation trend, a gradient consistency loss is constructed; let... and Let represent the spatial difference gradient operators along the x and y directions, respectively, then the gradient consistency loss... Can be represented as all The average value; Therefore, the total training loss of the amplitude repair network in the t-th training iteration is: ， in, Let λt represent the total batch training loss of the t-th training iteration, and λ1 and λ2 represent the weight coefficients of the magnitude reconstruction loss and the gradient consistency loss, respectively. Subsequently, the backpropagation algorithm was used to calculate the total batch training loss. Gradient of network parameter θ: ， The network parameters are updated using a gradient optimization algorithm; a learning rate decay strategy is adopted during training to gradually reduce the learning rate with each training round. During training, the training samples are divided into a training set and a validation set. After each training round, the validation loss is calculated on the validation set, and the model parameters with the minimum validation loss are saved. Training termination conditions include at least one of the following: the number of training rounds reaches a preset maximum number of rounds; the validation set loss no longer decreases for several consecutive rounds; or both the training loss and the validation loss are less than a preset threshold. After training, the network parameters with the minimum validation set loss are selected. The trained amplitude repair network is obtained. .

7. The near-field to far-field conversion method for near-field antenna measurement of UAVs according to claim 1, characterized in that, In S3, the training data source for the pre-trained phase retrieval network is the training data of the pre-trained amplitude restoration network; for each training sample, the first measurement plane and the second measurement plane are extracted at E. x E y Near-field amplitude data under polarization components, and generate amplitude data to be repaired affected by flight disturbances according to the aforementioned amplitude disturbance construction method; After processing by the amplitude repair module, the normalized correction amplitude matrix X is obtained. cn (q) Simultaneously, the corresponding true phase label is calculated based on the complex near-field data of this sample: ， Where q represents the training sample number, and Let E represent the true phase labels of the first and second measurement planes in the q-th training sample, respectively; if the complex electric field is known to be E... m,p,0 (q) The true phase is then calculated using the following formula: ， Where m∈{1,2} represents the measurement plane number, p∈{E x E y } represents the polarization component, and arg() represents the complex phase take operation; then, according to X cn (q) Measure the plane coordinate information and condition information, and construct the phase retrieval network input as described above: ， in, Z represents the dataset input to the phase retrieval network for the q-th training sample. L,1 (q) Z L,2 (q) Z H,1 (q) Z H,2 (q) These represent the low-frequency and high-frequency coding layer inputs, respectively, constructed from the normalized correction amplitude matrix, mean feature, difference feature, and coordinate condition feature map; Next, the parameters of the phase retrieval network model are initialized; let the phase retrieval network be: ， Where ψ represents the trainable parameters of the low-frequency coding layer, high-frequency coding layer, frequency feature fusion unit, cross-plane feature exchange unit, shared fusion unit, shared upsampling unit, and polarization phase decoding unit in the phase retrieval network; When initializing the trainable parameters of the phase retrieval network, the two-dimensional convolutional layers and 1×1 convolutional mapping layers in the low-frequency coding layer, high-frequency coding layer, frequency feature fusion unit, cross-plane feature exchange unit, shared fusion unit, shared upsampling unit, and polarized phase decoding unit are randomly initialized according to their number of input channels, number of output channels, and convolutional kernel size. In one specific implementation, the convolutional kernel weights are initialized using the Kaiming uniform initialization method, and the corresponding biases are initialized to zero. For the normalization layer in the network, the scaling factor of the normalization layer is initialized to 1, and the translation bias is initialized to 0. If the normalization layer contains a running mean and a running variance, the running mean is initialized to 0, and the running variance is initialized to 1. For the convolutional weights in the gated mapping layer, the same random initialization method as the convolutional layer is used, and the gated mapping layer bias is initialized to zero. Furthermore, for the low-frequency phase output branch and residual phase output branch in the polarization phase decoding unit, the weights of their terminal convolutional layers are initialized using a small-amplitude random initialization method, and the bias is initialized to zero, so that the phase output will not produce too large random shifts in the initial stage of training; if the pre-trained phase retrieval model is used as the pre-trained model, the parameters of the pre-trained model are used as the initial parameters, and some parameters are updated or frozen as needed during subsequent training. Input the q-th training sample into the phase recovery network to obtain the predicted phase: ， in, and These represent the predicted phases of the first and second measurement planes output by the phase retrieval network, respectively. To mitigate the computational discontinuity issues caused by phase periodicity, the predicted and true phases are converted into phasor form; for any phase φ, its phasor representation is: ， Therefore, the phase-monitoring loss is constructed: ， Simultaneously, a phase gradient loss is constructed: ， And low-frequency phase loss and residual phase loss are constructed based on low-frequency phase and residual phase respectively: ， ， in, This represents the predicted low-frequency phase of the m-th measurement plane. This indicates the low-frequency phase corresponding to the actual phase label. Indicates the predicted residual phase. Indicates the true residual phase; Furthermore, to constrain the physical consistency between the two measurement planes, a complex near field of the first measurement plane is constructed based on the predicted phase and corrected amplitude of the first measurement plane, and then propagated to the second measurement plane through the angular spectrum propagation operator to obtain the propagated predicted phase. Similarly, the propagation prediction phase can be obtained by backpropagating from the second measurement plane to the first measurement plane. The loss of propagation consistency can be expressed as: ， Therefore, the total training loss of the phase retrieval model is: ， Where μ1, μ2, μ3, μ4, and μ5 are the weights of phase supervision loss, phase gradient loss, low-frequency phase loss, residual phase loss, and propagation consistency loss, respectively. In each training iteration, a batch of training samples is input into the phase retrieval network, and the total training loss is calculated. The gradient of the loss function with respect to the network parameter ψ is calculated using the backpropagation algorithm. ， Subsequently, the network parameters are updated using a gradient optimization algorithm; during training, a learning rate decay strategy is adopted to gradually reduce the learning rate with each training round. During training, the samples are divided into training and validation sets. After each training round, the validation loss is calculated on the validation set, and the model parameter with the minimum validation loss is saved. Training termination conditions include at least one of the following: the number of training rounds reaches a preset maximum number of rounds; the validation set loss no longer decreases for several consecutive rounds; or both the training loss and the validation loss are less than a preset threshold. After training, the network parameter ψ with the minimum validation set loss is selected. ∗ The trained phase retrieval network is obtained. .

8. The near-field to far-field conversion method for near-field antenna measurement of UAVs according to claim 1, characterized in that, In S5, when zero-filling the complex near-field data of the first measurement plane, a fixed zero-fill of 4 times is used; assuming the original number of near-field sampling points is H×W, the number of sampling points after zero-filling is: .

9. A near-field to far-field conversion system for near-field antenna measurement of unmanned aerial vehicles, characterized in that, This system is used to implement the near-field to far-field conversion method for near-field antenna measurement of UAVs as described in any one of claims 1-8. Specifically, it includes a UAV, a signal source, a probe antenna, an antenna under test, a spectrum analyzer, and an edge computing unit. The specific working method is as follows: the UAV carries the signal source and the probe antenna and flies along a preset path. The probe antenna transmits the signal, and the antenna under test receives the signal. The signal is transmitted to the spectrum analyzer through a transmission line. The spectrum analyzer transmits the amplitude data to the edge computing unit. The edge computing unit calls the trained network or directly calls the algorithm to output the far-field radiation pattern of the antenna under test.

Images