Near-field scanning rapid imaging device for radio frequency performance of multi-channel electrically tunable base station antenna

Abstract

The invention provides a near-field scanning rapid imaging device for the radio frequency performance of a multi-channel electrically tunable base station antenna, and relates to the technical field of radio frequency antenna testing, the device comprises a scanning assembly and a data processing module, the scanning assembly is composed of a radio frequency probe installed on a two-dimensional precision displacement platform and a vector network analyzer connected with the radio frequency probe, and the data processing module is connected with the scanning assembly. The scanning module is used for performing point-by-point scanning in an antenna near-field area along a preset plane grid path and collecting scattering parameters representing electromagnetic field characteristics at sampling points to form a near-field scanning data set; and the data processing module receives the data set, executes fast Fourier transform operation, and converts near-field data into far-field radiation characteristics so as to generate a far-field radiation pattern representing the radio frequency performance of the antenna. And the data processing module applies a preset static compensation matrix to the near-field data before conversion, so that the mutual coupling effect between the probe and the antenna is inhibited, and the accuracy of a measurement result is improved.

Description

Technical Field

[0001] This application relates to the field of radio frequency antenna testing technology, and more specifically, to a near-field scanning fast imaging device for the radio frequency performance of a multi-channel electrically tunable base station antenna. Background Technology

[0002] As a core component of mobile communication systems, base station antennas directly affect signal coverage, system capacity, and interference suppression capabilities through their radio frequency (RF) performance. With the rapid development of 5G and future 6G communication technologies, base station antennas are gradually evolving towards multi-channel, large-scale electrically tunable arrays. These antennas typically possess wide bandwidth, multi-beam characteristics, and dynamic adjustment features, placing higher demands on the precise testing of their RF performance indicators such as radiation pattern, gain, and sidelobe level.

[0003] Existing RF performance testing methods mainly include far-field testing and traditional anechoic chamber testing. Far-field testing typically requires setting up a large test area in a region where the antenna beam is fully deployed, resulting in low testing efficiency due to limitations in site conditions and cost. While anechoic chamber far-field testing can simulate a free-space environment, it has strict requirements for equipment layout and testing procedures, and a long testing cycle. For multi-channel electrically tunable base station antennas, due to the large number of ports and complex beam combinations, traditional far-field testing struggles to complete full-coverage measurements within a limited time.

[0004] Therefore, how to provide a near-field scanning device that can rapidly measure and image the radio frequency performance of multi-channel electrically tunable base station antennas within a limited space has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application provides a near-field scanning rapid imaging device for multi-channel electrically adjustable base station antenna radio frequency performance, comprising:

[0006] The scanning component includes:

[0007] Radio frequency probe mounted on a two-dimensional precision displacement platform;

[0008] and a vector network analyzer, connected to the radio frequency probe;

[0009] The two-dimensional precision displacement platform is used to drive the radio frequency probe to scan along a preset planar grid path in the near field region of the multi-channel electrically tunable base station antenna under test. At the same time, the vector network analyzer measures and collects scattering parameters characterizing the electromagnetic field properties at each sampling point of the grid path to form a near field scanning dataset.

[0010] A data processing module is connected to the scanning component, and the data processing module is configured as follows:

[0011] The near-field scan dataset is received, and a fast Fourier transform operation is performed on the near-field scan dataset to complete the near-field to far-field transformation and generate a far-field radiation pattern characterizing the antenna's radio frequency performance.

[0012] Optionally, the data processing module is further configured to apply a preset, static compensation matrix to correct the near-field scanning dataset before performing the near-field to far-field transformation operation, so as to suppress the mutual coupling effect between the probe and the antenna under test.

[0013] Optional, including:

[0014] The scanning module includes a scanning component that scans the near-field region of a multi-channel electrically adjustable base station antenna, generates near-field scanning data, and transmits the near-field scanning data to a compensation module.

[0015] An antenna parameter module is used to store and provide the target parameters of the multi-channel electrically adjustable base station antenna to the compensation module.

[0016] The compensation module includes an electromagnetic suppression structure and a large language model. The electromagnetic suppression structure includes an electromagnetic bandgap structure or a split-ring resonator, which is used to suppress the mutual coupling components and their cascaded interaction with the edge scattering components in the near-field scanning data, and input the suppression result as suppression data into the large language model.

[0017] The large language model generates a block strategy and adaptive parameters based on the target parameters, and calculates the sub-mutual scattering matrix by dividing the near-field scanning data and the suppressed data into blocks.

[0018] The result merging module is used to splice the sub-mutually coupled scattering matrices to generate a mutually coupled scattering matrix;

[0019] The processing module is used to perform matrix solving on the near-field scanning data based on the mutual scattering matrix, and generate a radio frequency performance image through far-field transformation.

[0020] Optionally, calculating the sub-mutual scattering matrix includes:

[0021] The target parameters are converted into serialized prompts and input into the large language model;

[0022] The large language model generates a segmentation strategy based on the serialization prompts, wherein the segmentation strategy includes instructions for dividing the near-field scan data into sub-regions.

[0023] The large language model generates adaptive parameters based on the serialization prompts, wherein the adaptive parameters include prompts for adjusting the calculations;

[0024] Based on the block segmentation strategy, the near-field scanning data and the suppression data are divided into blocks to generate block data;

[0025] The segmented data is processed based on the adaptive parameters to calculate the sub-mutual scattering matrix.

[0026] Optionally, converting the target parameters into serialization prompts and inputting them into the large language model includes:

[0027] The target parameters are obtained from the antenna parameter module, wherein the target parameters include the scan data size, compensation sub-block capacity, electromagnetic frequency range, and dielectric constant gradient of the multi-channel electrically tunable base station antenna;

[0028] The target parameters are converted into serialization hints, wherein the serialization hints include a numerical sequence of the scan data size, a limit sequence of the compensating sub-block capacity, and a physical model description of the electromagnetic frequency range and dielectric constant gradient;

[0029] The serialization prompt is input into the large language model, wherein the large language model generates adaptive parameters containing electromagnetic constraints based on the physical model description.

[0030] Optionally, the near-field scanning data and the suppression data are divided into blocks based on the block segmentation strategy to generate block data, including:

[0031] Based on the block-segmentation strategy, edge scattering sensitive regions in the near-field scanning data are identified, wherein the block-segmentation strategy includes sub-region partitioning instructions, and the sub-region partitioning instructions integrate boundary mutual coupling constraints.

[0032] According to the sub-region division instruction, the near-field scanning data is dynamically divided into blocks to generate sub-scan data, wherein the dynamic block division integrates the wavelength shortening scattering distribution;

[0033] Based on the block-segmentation strategy, the suppression data is divided into corresponding blocks to generate sub-suppression data, wherein the sub-suppression data corresponds to the edge values ​​of the mutually coupled components in the sub-scan data;

[0034] The sub-scan data and the sub-suppression data are combined to generate block data.

[0035] Optionally, concatenating the sub-mutually coupled scattering matrices to generate a mutually coupled scattering matrix includes:

[0036] The compensation module receives the sub-mutual scattering matrix and identifies a first boundary phase region based on the sub-mutual scattering matrix, wherein the first boundary phase region includes a sequence of phase difference values ​​between the sub-mutual scattering matrices;

[0037] A first weight sequence is calculated based on the first boundary phase region, wherein the first weight sequence is generated based on the edge reflection distribution of the sub-mutual scattering matrix;

[0038] The first weight sequence is applied to fuse and stitch the sub-mutually coupled scattering matrices to generate a first preliminary mutually coupled scattering matrix, wherein the fusion and stitching includes a weighted superposition operation of the sub-mutually coupled scattering matrices;

[0039] The first preliminary mutual scattering matrix is ​​subjected to array response verification, wherein the array response verification calculates the response consistency index of the first preliminary mutual scattering matrix, and the first preliminary mutual scattering matrix is ​​adjusted based on the response consistency index to generate the mutual scattering matrix.

[0040] Optionally, the near-field scanning data is solved by matrix transformation based on the mutual scattering matrix to generate a radio frequency performance image through far-field transformation, including:

[0041] The mutual-coupled scattering matrix is ​​received from the result merging module, and the near-field scanning data is received from the scanning module;

[0042] Based on the mutual scattering matrix, a preliminary matrix solution is performed on the near-field scanning data to generate a first solution matrix, wherein the preliminary matrix solution includes the inverse operation of the mutual scattering matrix;

[0043] A first distortion index is calculated based on the first solution matrix, wherein the first distortion index is generated based on the response consistency index and includes the multiple scattering deviation value of the first solution matrix;

[0044] The first solution matrix is ​​adjusted by applying the first distortion index to generate a second solution matrix, wherein the adjustment includes an adaptive regularization operation of the first solution matrix;

[0045] A far-field transformation is performed on the second solution matrix to generate the radio frequency performance image, wherein the far-field transformation includes the Fourier integral operation of the second solution matrix.

[0046] Optionally, performing a far-field transformation on the second solution matrix to generate the RF performance image includes:

[0047] The processing module receives the second solution matrix and calculates a first far-field integral sequence based on the second solution matrix, wherein the first far-field integral sequence includes the Fourier integral result of the second solution matrix;

[0048] A first radiation pattern region is identified based on the first far-field integral sequence, wherein the first radiation pattern region includes a polarization deviation value sequence in the first far-field integral sequence;

[0049] A first compensation sequence is generated based on the first radiation pattern region, wherein the first compensation sequence includes the multiple scattering correction value of the polarization deviation value sequence;

[0050] The first far-field integral sequence is adjusted by applying the first compensation sequence to generate a second far-field integral sequence, wherein the adjustment includes a weighted correction operation on the first far-field integral sequence.

[0051] The second far-field integral sequence is image-mapped to generate the radio frequency performance image, wherein the image-mapping includes a two-dimensional projection operation of the second far-field integral sequence.

[0052] Optionally, converting the target parameter into a serialization hint includes:

[0053] The target parameters are received from the antenna parameter module, and the numerical data of the scan data size, the limitation data of the compensation sub-block capacity, the range data of the electromagnetic frequency range, and the gradient data of the dielectric constant gradient are extracted.

[0054] The numerical data of the scanned data size is serialized to generate a first numerical sequence, wherein the first numerical sequence includes the dimension array of the scanned data size;

[0055] The limitation data of the compensation sub-block capacity is serialized to generate a first limitation sequence, wherein the first limitation sequence includes an array of boundary thresholds for the compensation sub-block capacity;

[0056] The range data of the electromagnetic frequency range and the gradient data of the dielectric constant gradient are subjected to physical model serialization conversion to generate a first physical model sequence, wherein the first physical model sequence includes the spectrum vector of the electromagnetic frequency range and the gradient vector of the dielectric constant gradient.

[0057] The first numerical sequence, the first constraint sequence, and the first physical model sequence are combined to generate the serialization hint, wherein the serialization hint includes an integrated parameter sequence array.

[0058] Compared with existing technologies, this application achieves compensation for the combined effects of mutual coupling and edge scattering by integrating electromagnetic suppression structures and a large language model. Specifically, the device utilizes antenna parameter modules to provide relevant parameters, generates block-based strategies and adaptive parameters from the large language model, performs block processing on near-field scanning data, calculates sub-mutual coupling scattering matrices, and generates a complete mutual coupling scattering matrix through a result merging module for matrix solving and far-field transformation in the processing module. This scheme improves the efficiency and accuracy of imaging processing and is suitable for large-scale RF data scenarios. Simultaneously, the adaptive block mechanism reduces computational overhead and adapts to the wavelength shortening and scattering inhomogeneity issues in the millimeter-wave band. Furthermore, the introduction of electromagnetic suppression structures reduces parasitic field interference and improves the accuracy of radiation patterns. Overall, the device enhances the real-time performance of batch testing and is suitable for the optimization and diagnostics of 5G / 6G antenna arrays. Attached Figure Description

[0059] Figure 1 A schematic diagram of a near-field scanning fast imaging device for the radio frequency performance of a multi-channel electrically adjustable base station antenna provided in an embodiment of this application;

[0060] Figure 2 A flowchart illustrating a method for calculating a sub-mutual scattering matrix provided in this application embodiment;

[0061] Figure 3 A flowchart illustrating a method for inputting target parameters into a large language model, as provided in this application embodiment;

[0062] Figure 4 This is a flowchart illustrating a method for generating a mutually coupled scattering matrix, as provided in an embodiment of this application.

[0063] Explanation of reference numerals in the attached diagram: 10, scanning module; 20, antenna parameter module; 30, compensation module; 40, result merging module; 50, processing module. Detailed Implementation

[0064] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0065] This application provides a near-field scanning fast imaging device for the radio frequency performance of a multi-channel electrically adjustable base station antenna, comprising:

[0066] The scanning component includes:

[0067] Radio frequency probe mounted on a two-dimensional precision displacement platform;

[0068] and a vector network analyzer, connected to the radio frequency probe;

[0069] The two-dimensional precision displacement platform is used to drive the radio frequency probe to scan along a preset planar grid path in the near field region of the multi-channel electrically tunable base station antenna under test. At the same time, the vector network analyzer measures and collects scattering parameters characterizing the electromagnetic field properties at each sampling point of the grid path to form a near field scanning dataset.

[0070] A data processing module is connected to the scanning component, and the data processing module is configured as follows:

[0071] Receive the near-field scan dataset and perform a fast Fourier transform operation on the near-field scan dataset to complete the near-field to far-field transformation;

[0072] This generates a far-field radiation pattern characterizing the radio frequency performance of the antenna.

[0073] Optionally, the data processing module is further configured to apply a preset, static compensation matrix to correct the near-field scanning dataset before performing the near-field to far-field transformation operation, so as to suppress the mutual coupling effect between the probe and the antenna under test.

[0074] In practice, the scanning components include an RF probe mounted on a two-dimensional precision displacement platform and a vector network analyzer electrically connected to it. The two-dimensional precision displacement platform, driven by a control system, precisely moves the RF probe along a preset planar grid path, enabling it to scan point-by-point in the near-field region of the multi-channel electrically tunable base station antenna under test. The vector network analyzer measures and acquires scattering parameters (including amplitude and phase information) characterizing the electromagnetic field at each scanning point, thus forming a complete near-field scanning dataset.

[0075] The data processing module receives the near-field scanning dataset transmitted by the scanning component and performs a Fast Fourier Transform (FFT) operation on the dataset to transform it from the near field to the far field, thereby generating a far-field radiation pattern characterizing the RF performance of the antenna under test. With this device, antenna RF performance imaging can be rapidly completed under limited space conditions, and key parameters such as main lobe, sidelobes, beamwidth, and sidelobe levels can be output.

[0076] Before performing the near-field to far-field transformation, the data processing module is configured to preprocess the original near-field scan dataset. Specifically, a preset static compensation matrix is ​​applied to correct the near-field scan data to suppress the mutual coupling effect between the probe and the antenna under test. This compensation matrix can be obtained through prior modeling or calibration experiments and is used to correct the additional coupling components introduced by the probe. After this correction, the physical accuracy of the near-field data is improved, and the far-field radiation pattern generated by the subsequent transformation can more realistically reflect the intrinsic RF performance of the antenna.

[0077] This approach can significantly improve testing efficiency while maintaining testing accuracy, making it particularly suitable for rapid RF performance measurement and engineering applications of multi-channel electrically tunable base station antennas.

[0078] See Figure 1 The diagram shown is a schematic of a near-field scanning rapid imaging device for the radio frequency performance of a multi-channel electrically tunable base station antenna provided in an embodiment of this application. The device includes:

[0079] The scanning module 10 includes a scanning component, which scans the near-field region of the multi-channel electrically adjustable base station antenna, generates near-field scanning data, and transmits the near-field scanning data to the compensation module 30.

[0080] Antenna parameter module 20 is used to store and provide the target parameters of the multi-channel electrically adjustable base station antenna to the compensation module 30;

[0081] The compensation module 30 includes an electromagnetic suppression structure and a large language model. The electromagnetic suppression structure includes an electromagnetic bandgap structure or a split ring resonator, which is used to suppress the mutual coupling components and their cascaded interaction with the edge scattering components in the near-field scanning data, and input the suppression result as suppression data into the large language model.

[0082] The large language model generates a block strategy and adaptive parameters based on the target parameters, and calculates the sub-mutual scattering matrix by dividing the near-field scanning data and the suppressed data into blocks.

[0083] The result merging module 40 is used to splice the sub-mutually coupled scattering matrices to generate a mutually coupled scattering matrix;

[0084] The processing module 50 is used to perform matrix solving on the near-field scanning data based on the mutual scattering matrix, and generate a radio frequency performance image through far-field transformation.

[0085] In specific implementation, the scanning module 10 internally includes a scanning component, such as an open-circuit waveguide probe mounted on a two-dimensional precision displacement platform driven by a stepper motor. During operation, this scanning component performs point-by-point scanning along a preset planar grid path, typically at a distance of several wavelengths from the antenna surface, within the near-field region of the multi-channel electrically tunable base station antenna under test. At each sampling point, a vector network analyzer (VNA) connected to the probe measures and records the complex transmission coefficient between the probe and the antenna under test, such as the S21 parameter. These complex data, including amplitude and phase information, together constitute the near-field scanning data. After the scan is completed, the scanning module 10 transmits the dataset to the compensation module 30.

[0086] The antenna parameter module 20 in the device can be a solid-state drive or other non-volatile digital storage unit. Its function is to provide the prior information required for calculation by the compensation module 30. This module stores the key target parameters of the multi-channel electrically adjustable base station antenna under test. These parameters are the physical basis for the subsequent adaptive adjustment of the compensation algorithm.

[0087] The compensation module 30 consists of an electromagnetic suppression structure and a large language model. In this embodiment, the electromagnetic suppression structure is a physical device, specifically an electromagnetic bandgap (EBG) structure with periodic metal patches etched on the substrate of the antenna under test, or a split-ring resonator (SRR) array arranged between antenna elements. This structure, through its frequency selectivity, suppresses surface wave coupling between antenna elements at the physical level, thereby weakening the mutual coupling components in the near-field scanning data, especially suppressing the complex cascading interaction between the components and the scattering components generated by the antenna edge structure. The measurement signal processed by this structure is acquired, forming suppressed data, and input into the large language model along with the original near-field scanning data.

[0088] The large language model in compensation module 30 can be a pre-trained model based on a Transformer architecture, configured to perform a planning task rather than a direct solution. It receives target parameters from antenna parameter module 20 and generates a set of partitioning strategies and adaptive parameters to guide subsequent numerical computations. The partitioning strategy is specifically a set of coordinate boundaries defining how the near-field scan data matrix is ​​divided, while the adaptive parameters may be a series of numerical computation suggestions for each sub-region, such as iterative convergence thresholds or regularization coefficients. Subsequently, the numerical computation unit within compensation module 30, for example, a solver based on the method of moments (MoM), partitions the input near-field scan data and suppression data into blocks according to the strategy generated by the large language model, and computes each block independently, thereby obtaining multiple sub-mutually coupled scattering matrices.

[0089] The result merging module 40 receives the sub-mutual scattering matrices calculated in the preceding steps. This module performs a matrix stitching operation, specifically, based on the spatial location information provided by the block strategy, assembling the sub-mutual scattering matrices as block elements into a sparse, large matrix whose size corresponds to the complete near-field scan data, thereby generating a complete mutual scattering matrix.

[0090] Processing module 50 performs final compensation and imaging. This module receives the complete cross-coupling scattering matrix from result merging module 40 and the raw near-field scan data from scanning module 10. Based on this cross-coupling scattering matrix, processing module 50 performs matrix operations on the near-field scan data. Specifically, this operation solves a large system of linear equations to mathematically eliminate the influence of cross-coupling effects on the raw data. After obtaining compensated near-field distribution data that is closer to the ideal state, processing module 50 applies a far-field transform to it, for example, by performing a two-dimensional fast Fourier transform (2D-FFT), ultimately calculating a radio frequency performance image characterizing the antenna performance, such as a two-dimensional or three-dimensional far-field radiation pattern.

[0091] In this way, by integrating the electromagnetic suppression structure and the large language model, compensation for the combined effects of mutual coupling and edge scattering is achieved. Specifically, the device utilizes the antenna parameter module 20 to provide relevant parameters, generates a block strategy and adaptive parameters from the large language model, performs block processing on the near-field scanning data, calculates the sub-mutual coupling scattering matrix, and generates the complete mutual coupling scattering matrix through the result merging module 40, which is then used for matrix solving and far-field transformation in the processing module 50. This scheme improves the efficiency and accuracy of imaging processing and is suitable for large-scale RF data scenarios. At the same time, the adaptive block mechanism reduces computational overhead and adapts to the wavelength shortening and scattering inhomogeneity problems in the millimeter-wave band. In addition, the introduction of the electromagnetic suppression structure reduces parasitic field interference and improves the accuracy of the radiation pattern. Overall, the device enhances the real-time performance of batch testing and is suitable for the optimization and diagnosis of 5G / 6G antenna arrays.

[0092] As an optional implementation method, see [link to implementation details]. Figure 2 The flowchart below illustrates a method for calculating a sub-mutually coupled scattering matrix, as provided in this application embodiment, including steps S101 to S105, wherein:

[0093] S101: Convert the target parameters into serialized prompts and input them into the large language model;

[0094] S102: The large language model generates a segmentation strategy based on the serialization prompts, wherein the segmentation strategy includes instructions for dividing the near-field scan data into sub-regions;

[0095] S103: The large language model generates adaptive parameters based on the serialization prompts, wherein the adaptive parameters include prompts for adjusting the calculations;

[0096] S104: Based on the block segmentation strategy, the near-field scanning data and the suppression data are divided into blocks to generate block data;

[0097] S105: Process the block data based on the adaptive parameters to calculate the sub-mutual scattering matrix.

[0098] To address the technical challenge of effectively converting high-level physical parameters into inputs that are understandable to large language models and ensuring that the output computational strategies can be accurately executed by downstream numerical computation units, this application further optimizes the process of calculating the sub-mutual scattering matrix in the aforementioned compensation module 30.

[0099] In practice, this process begins with a structured data conversion step, which converts the target parameters provided by the antenna parameter module 20 into serialized prompts and provides them as input to the large language model. This step ensures that the numerical values ​​and concepts of the physical world, such as the antenna's geometry and material properties, can be accurately translated into a text format that the large language model can process.

[0100] For example, a set of target parameters, such as {"antenna size": [0.5, 0.5], "frequency": 28e9}, can be formatted as a JSON (JavaScript Object Notation) text string. This string is then embedded in a broader text template containing task instructions to form the final, complete serialized prompt, which is then fed into a large language model.

[0101] Upon receiving the serialization prompt, the large language model generates two types of structured output based on its internal representation and reasoning capabilities. The first type of output is a chunking strategy, which contains specific instructions for dividing sub-regions. These instructions are not vague suggestions, but explicit, machine-readable commands.

[0102] The second type of output is adaptive parameters, which include hints for adjusting the calculations. These hints are specific numerical or configuration suggestions for the calculation process of each sub-region.

[0103] For example, for one of the sub-regions mentioned above, the corresponding adaptive parameters might be {"regularization coefficient": 0.01,"convergence threshold": 1e-6}, while for another sub-region that may contain more complex edge effects, the parameters might be {"regularization coefficient": 0.05, "convergence threshold": 1e-7}.

[0104] Next, the numerical computation unit within the compensation module 30, such as the aforementioned MoM solver, begins to execute the strategy for generating the large language model. First, this unit parses the sub-region partitioning instructions in the partitioning strategy and, based on these precise coordinate indices, performs matrix slicing operations on the complete near-field scan data matrix and suppression data matrix, thereby generating block data that corresponds one-to-one with each sub-region.

[0105] Finally, for each data block, the numerical computation unit processes it based on the corresponding adaptive parameters generated by the large language model. For example, when calculating the sub-mutual scattering matrix of the first sub-region, the computation unit reads its corresponding adaptive parameters and sets the regularization coefficient inside the solver to 0.01 and the convergence threshold to 1e-6. By performing this customized processing flow based on adaptive parameters on all data blocks, all sub-mutual scattering matrices are finally calculated.

[0106] In this way, by serializing physical parameters into structured prompts, a clear and unambiguous communication protocol is established, ensuring that the large language model can accurately understand the physical context of the computational task. Secondly, the strategies generated by the large language model, including specific instructions and parameters, enable its high-level planning to be executed accurately and repeatedly by downstream numerical computation units, avoiding the uncertainty caused by ambiguous instructions. This complete process from parameters to prompts to specific strategies allows the entire device to dynamically and intelligently adjust its core computational process according to the specific conditions of different antennas, improving computational efficiency and automation while ensuring accuracy.

[0107] For example, a large language model can employ a Transformer-based encoder-decoder architecture. Specifically, the model may contain 12 encoder layers and 12 decoder layers, each with 8 self-attention heads, for a total of approximately 110 million parameters. The model is pre-trained on a dataset containing general scientific literature, programming code, and specialized electromagnetics papers to enable it to perform basic logical reasoning and code generation.

[0108] Furthermore, to ensure the large language model is capable of handling this task, it can be fine-tuned for a specific domain. In the training dataset used for fine-tuning, each sample consists of a set of antenna target parameters and corresponding optimal block-splitting strategies and adaptive parameters pre-designed by electromagnetic simulation experts. The simulation data can be generated using commercial electromagnetic simulation software (such as CST Studio Suite or ANSYS HFSS), covering near-field distribution data of multi-channel electrically tunable base station antenna arrays of different sizes, frequencies, and substrate materials. The training process employs supervised learning methods, aiming to minimize the difference between the strategy generated by the large language model and the expert strategy; for example, a cross-entropy loss function can be used. The Adam optimizer is used for training, with an initial learning rate set to 0.001, and 100 iterations are performed.

[0109] For example, at runtime, the serialization prompts input to the large language model are specially engineered. For instance, a complete prompt might include the following text structure:

[0110] Task: Generate a computational strategy for antenna near-field compensation.

[0111] Input parameters: {"Antenna size_m": [0.1, 0.1], "Frequency_GHz": 28, "Dielectric constant": 3.5,"Number of scan points":}. Output the chunking strategy and adaptive parameters in JSON format. The large language model understands the task description and parameters in the prompt through its attention mechanism and generates a structured JSON output, such as: {"Chunking strategy": {"Type": "Coordinate index", "Boundary": [, ,...]}, "Adaptive parameters": [{"Region index": 0, "Regularization coefficient": 0.01}, {"Region index": 1, "Regularization coefficient": 0.05},...]}. This mechanism ensures the accuracy and machine readability of both input and output.

[0112] Understandably, the principle behind this invention's use of a large language model to generate computational strategies lies in the fact that mutual coupling and edge scattering effects in the near field of an antenna are highly complex and nonlinear physical processes. Traditional fixed algorithms struggle to adapt to subtle variations under different antenna configurations. However, a large language model fine-tuned with domain data can learn deep, nonlinear mappings between antenna physical parameters such as frequency and dielectric constant and the optimal computational strategy from a large number of samples. The optimal computational strategy could be, for example, determining where to perform more refined partitioning to capture edge effects. This is similar to the model learning some implicit laws of electromagnetic wave propagation and scattering, thereby generating a customized numerical solution that is superior to a general fixed algorithm for a specific problem. Essentially, this is a data-driven, heuristic process of algorithm selection and parameter optimization.

[0113] As an alternative implementation, the large language model can be a base model with powerful general reasoning and code generation capabilities. These models are typically based on the Transformer architecture and have acquired the ability to understand complex instructions and generate structured outputs through pre-training on massive amounts of text and code data.

[0114] For example, in different embodiments of the present invention, the following models may be used as the base model: such as GPT-4 or its subsequent versions, Gemini 1.5 Pro, Llama series models.

[0115] As an optional implementation method, see [link to implementation details]. Figure 3 The flowchart provided in this application embodiment illustrates a method for inputting target parameters into a large language model, including steps S201 to S203, wherein:

[0116] S201: Obtain the target parameters from the antenna parameter module 20, wherein the target parameters include the scan data size, compensation sub-block capacity, electromagnetic frequency range, and dielectric constant gradient of the multi-channel electrically tunable base station antenna;

[0117] S202: Convert the target parameters into serialization hints, wherein the serialization hints include a numerical sequence of the scan data size, a limit sequence of the compensating sub-block capacity, and a physical model description of the electromagnetic frequency range and dielectric constant gradient;

[0118] S203: Input the serialization prompt into the large language model, wherein the large language model generates adaptive parameters containing electromagnetic constraints based on the physical model description.

[0119] The technical problem this embodiment aims to solve is how to transform discrete, purely numerical engineering parameters into an input format that allows a Large Language Model (LLM) to understand their underlying physical meaning and inherent constraints. Simply inputting a string of numbers into an LLM may not allow the model to understand whether these numbers represent antenna size or frequency, or to grasp the physical relationship between them, resulting in a lack of physical realism in the generated strategy.

[0120] In a specific implementation, a set of specific target parameters is obtained from the antenna parameter module 20. In this embodiment, this set of target parameters is limited to at least the following: the scanning data size of the multi-channel electrically tunable base station antenna, such as a 128×128 matrix, representing the number and layout of sampling points for near-field scanning; the capacity of the compensation sub-block, such as 1024, representing the maximum number of matrix units that the downstream numerical calculation unit can process in a single operation, which is a hardware or algorithmic performance limitation; the electromagnetic frequency range, such as 28 GHz to 30 GHz; and the dielectric constant gradient, which describes the spatial variation of the dielectric constant of the antenna substrate material.

[0121] Next, the step of converting the target parameters into serialization hints is performed. This step is not a simple listing of numbers, but rather maps parameters of different properties to different types of serialization representations:

[0122] For the scanned data size, convert it into a numerical sequence.

[0123] For the compensation sub-block capacity, convert it into a limit sequence, explicitly informing the LLM that its generated block strategy must comply with this computational resource limit.

[0124] For the electromagnetic frequency range and dielectric constant gradient, these are transformed into a context-rich physical model description. This is a textual description that not only provides numerical values ​​but also explains their physical meaning. For example, {"frequency_GHz": 28, "dielectric_description": "Substrate permittivity varies linearly from 2.2 at the center to 2.5 at the edges."}. This description places the pure numbers (28, 2.2, 2.5) within a physical framework.

[0125] Finally, this constructed serialized cue, containing numerical sequences, constraint sequences, and a physical model description, is input into the large language model. Thanks to learning from similar physical descriptions during fine-tuning, the large language model is able to parse the cue. In particular, it understands the physical model description of "the dielectric constant changing linearly from the center to the edge" and infers that this means the electromagnetic wavelength in the antenna edge region will be shorter, and the field variation will be more drastic. Therefore, when generating adaptive parameters, the model naturally incorporates instructions containing electromagnetic constraints, such as assigning smaller sizes or higher computational precision to sub-blocks in the edge region, to accurately capture this non-uniform scattering effect.

[0126] In this way, by introducing physical model descriptions, abstract physical concepts are conveyed to the large language model, improving the physical accuracy of the model generation strategy. By transforming hardware or algorithmic constraints into explicit sequences of constraints, the strategy generated by LLM is ensured to be executable in reality, avoiding resource overflow. This enables LLM to generate adaptive parameters containing electromagnetic constraints based on an understanding of physical laws, making the entire compensation process more intelligent and refined.

[0127] As an optional implementation, the near-field scanning data and the suppression data are divided into blocks based on the block segmentation strategy to generate block data, including:

[0128] Based on the block-segmentation strategy, edge scattering sensitive regions in the near-field scanning data are identified, wherein the block-segmentation strategy includes sub-region partitioning instructions, and the sub-region partitioning instructions integrate boundary mutual coupling constraints.

[0129] According to the sub-region division instruction, the near-field scanning data is dynamically divided into blocks to generate sub-scan data, wherein the dynamic block division integrates the wavelength shortening scattering distribution;

[0130] Based on the block-segmentation strategy, the suppression data is divided into corresponding blocks to generate sub-suppression data, wherein the sub-suppression data corresponds to the edge values ​​of the mutually coupled components in the sub-scan data;

[0131] The sub-scan data and the sub-suppression data are combined to generate block data.

[0132] This embodiment aims to address the issues of wasted computational resources and distortion of physical effects caused by traditional uniform partitioning when processing large near-field scanning data in blocks. Traditional uniform partitioning methods either fail to accurately capture the complex electromagnetic effects in areas such as antenna edges due to excessively coarse meshes, or result in a large amount of unnecessary redundant calculations in areas with gentle field changes due to excessively fine meshes.

[0133] In practical implementation, based on a block-based strategy, edge-scattering sensitive regions in the near-field scanning data are identified. The core sub-region partitioning instructions in the LLM-generated block-based strategy are not merely a set of segmentation coordinates, but rather an intelligent instruction set integrating boundary coupling constraints. For example, based on the antenna geometry provided by the antenna parameter module 20, this instruction set explicitly marks the surrounding area of ​​the entire scanning data matrix, such as the outermost 10% of rows and columns, as edge-scattering sensitive regions. Simultaneously, the boundary coupling constraints in the instructions provide specific numerical suggestions for the boundary processing of these regions, ensuring the accuracy of subsequent sub-matrix calculations.

[0134] Secondly, the numerical calculation unit dynamically divides the near-field scan data into blocks according to the sub-region division instructions to generate sub-scan data. This dynamic block division is a non-uniform, physically-aware segmentation process that incorporates the physical phenomenon of wavelength-shortened scattering distribution. Specifically, for the previously identified edge-scattering sensitive regions, the discontinuity of the antenna substrate at the edges leads to changes in the effective dielectric constant, resulting in wavelength shortening of surface waves and a more pronounced field gradient. The sub-region division instructions will then command a finer division in these areas, for example, dividing them into multiple 8×8 pixel sub-blocks. For the non-sensitive regions at the antenna center, the instructions will use a coarser division, such as 32×32 pixel sub-blocks. In this way, the entire near-field scan data matrix is ​​non-uniformly segmented, resulting in a set of sub-scan data of varying sizes.

[0135] Furthermore, based on the same segmentation strategy, the suppressed data is segmented accordingly to generate sub-suppressed data. The numerical calculation unit uses the exact same coordinate index as when segmenting the near-field scan data to cut the suppressed data matrix obtained after processing by electromagnetic suppression structures such as EBG or SRR. Each generated sub-suppressed data corresponds physically to the edge values ​​of the mutual coupling components in its corresponding sub-scan data, providing physically preprocessed and purer boundary condition information for subsequent calculation of the mutual coupling scattering matrix of that sub-region.

[0136] The sub-scan data and the sub-suppression data are combined to generate the final block data. For each divided sub-region, its corresponding sub-scan data and sub-suppression data are packaged into a unified data structure.

[0137] For example, for the i-th sub-region, the generated block data i is a data pair consisting of two matrices, namely the sub-scan data i and the sub-suppression data i.

[0138] In this way, by identifying edge-scattering sensitive areas and dynamically dividing them into blocks, intelligent on-demand allocation of computing resources is achieved, improving computational efficiency while ensuring accurate capture of physical effects in key areas such as antenna edges. The method integrates physical constraints such as wavelength shortening, making the block division strategy no longer a blind geometric segmentation but an adaptive process based on electromagnetic principles, thereby improving the accuracy of the final imaging results. By co-blocking and combining the raw scan data and suppressed data, more complete and reliable input information is provided for subsequent submatrix calculations.

[0139] As an optional implementation method, see [link to implementation details]. Figure 4 The flowchart below illustrates a method for generating a mutually coupled scattering matrix according to an embodiment of this application, including steps S301 to S304, wherein:

[0140] S301: Receive the sub-mutual scattering matrix from the compensation module 30, and identify a first boundary phase region based on the sub-mutual scattering matrix, wherein the first boundary phase region includes a sequence of phase difference values ​​between the sub-mutual scattering matrices;

[0141] S302: Calculate a first weight sequence based on the first boundary phase region, wherein the first weight sequence is generated based on the edge reflection distribution of the sub-mutually coupled scattering matrix;

[0142] S303: The first weight sequence is applied to fuse and stitch the sub-mutually coupled scattering matrices to generate a first preliminary mutually coupled scattering matrix, wherein the fusion and stitching includes a weighted superposition operation of the sub-mutually coupled scattering matrices;

[0143] S304: Perform array response verification on the first preliminary mutual scattering matrix, wherein the array response verification calculates the response consistency index of the first preliminary mutual scattering matrix, and adjusts the first preliminary mutual scattering matrix based on the response consistency index to generate the mutual scattering matrix.

[0144] This embodiment aims to address the technical problem of accurately fusing multiple sub-mutually coupled scattering matrices into a physically continuous and self-consistent complete matrix. Simply directly stitching these independently calculated sub-matrices together in blocks easily introduces non-physical phase jumps and amplitude discontinuities at the boundaries of sub-regions. These abrupt changes at the stitching seams act as error sources, amplified in subsequent matrix solving, ultimately leading to severe artifacts in the generated RF performance image and distorting the true radiation characteristics. This embodiment solves this problem through a refined stitching method that incorporates weighted fusion and response verification.

[0145] Regarding the above S301:

[0146] In practice, the result merging module 40 first receives all calculated sub-mutually coupled scattering matrices from the compensation module 30. Then, the module analyzes any two spatially adjacent sub-matrices, such as sub-matrices A and B, along their common boundary. Specifically, the module extracts the complex values ​​of a column (or row) of elements at the boundary of sub-matrices A, and the complex values ​​of the corresponding column (or row) of elements at the adjacent boundary of sub-matrices B, and calculates the element-wise phase difference between them. All these phase difference values ​​calculated at the boundary together constitute the first boundary phase region, which is represented by a sequence of phase difference values. This sequence visually quantifies the degree of phase discontinuity at the splicing boundary.

[0147] Regarding the above S302:

[0148] In practice, a first weight sequence for subsequent fusion operations is calculated based on the identified boundary phase regions. This weight sequence is calculated based on the edge reflection distribution of the sub-mutually coupled scattering matrix.

[0149] Specifically, the magnitudes of the off-diagonal elements in the edge regions of each submatrix are analyzed. If a submatrix has strong edge reflection distribution, meaning the magnitudes of the off-diagonal elements near the edge are large, it indicates that the coupling effect between this sub-region and other regions is truncated more severely at the boundary. Its boundary data should be assigned lower weights during fusion to reduce its disruption to the overall continuity. Conversely, submatrixes with weak edge reflection are assigned higher weights to their boundary data. This method generates a first weight sequence that corresponds one-to-one with all boundary points.

[0150] Regarding the above S303:

[0151] In practice, the calculated first weight sequence is used to fuse and stitch together all sub-mutually coupled scattering matrices. This process is not a simple block stitching, but a weighted superposition operation.

[0152] For example, in the boundary region between two submatrices, the module creates a smooth transition region where the new matrix element values ​​are determined by the weighted average of the corresponding boundary elements of the two submatrices, and the weights are the first weight sequence calculated in the previous step.

[0153] By performing this operation on all boundaries, the phase jumps identified by the first boundary phase region can be effectively smoothed out, thereby generating a first preliminary mutual-coupled scattering matrix that transitions smoothly at the boundaries.

[0154] Regarding S104 above:

[0155] To ensure the global physical consistency of the entire matrix, an array response verification is required on the generated initial mutual-coupled scattering matrix. This verification process is accomplished by calculating a response consistency index.

[0156] For example, a known, ideal excitation vector, such as simulating a plane wave incident from the front of the antenna, can be multiplied by the initial cross-scattering matrix to obtain a theoretical array response. Then, the response vector is analyzed to determine if there are any high-frequency spatial oscillations or outliers that do not conform to physical laws. The response consistency index quantifies the degree of these anomalies. If this index exceeds a preset threshold, the module will adjust the first initial cross-scattering matrix, for example, by applying a low-pass spatial filter or an iterative smoothing algorithm, until the response consistency index meets the requirements. The matrix obtained after this verification and adjustment is the final, high-quality cross-scattering matrix.

[0157] In this way, by identifying and weighting the fusion boundary regions, non-physical stitching gaps introduced by independently calculating sub-matrices are effectively eliminated, ensuring the continuity and accuracy of the final matrix. Determining weights based on edge reflection distribution enables the fusion process to intelligently handle the boundary characteristics of different sub-regions, improving the physical rationality of the stitching. The final array response verification and adjustment steps provide a global quality assurance for the entire stitching process, enhancing the reliability and robustness of the final generated mutually coupled scattering matrix, laying the foundation for subsequent high-precision radio frequency performance imaging.

[0158] As an optional implementation, the near-field scanning data is solved by matrix transformation based on the mutual scattering matrix, and a radio frequency performance image is generated by far-field transformation, including:

[0159] The mutual scattering matrix is ​​received from the result merging module 40, and the near-field scanning data is received from the scanning module 10;

[0160] Based on the mutual scattering matrix, a preliminary matrix solution is performed on the near-field scanning data to generate a first solution matrix, wherein the preliminary matrix solution includes the inverse operation of the mutual scattering matrix;

[0161] A first distortion index is calculated based on the first solution matrix, wherein the first distortion index is generated based on the response consistency index and includes the multiple scattering deviation value of the first solution matrix;

[0162] The first solution matrix is ​​adjusted using the first distortion index to generate a second solution matrix, wherein the adjustment includes an adaptive regularization operation on the first solution matrix;

[0163] The second solution matrix is ​​subjected to a far-field transformation to generate the radio frequency performance image, wherein the far-field transformation includes the Fourier integral operation of the second solution matrix.

[0164] This embodiment aims to address the technical problem of overcoming the inherent instability of numerical computation and suppressing residual errors when using the finally generated mutual coupling scattering matrix for compensation solutions. Theoretically, the mutual coupling effect can be eliminated through a simple matrix inversion operation. However, in practical applications, due to the presence of measurement noise and the small residual errors that may be introduced by the aforementioned stitching and fusion process, direct matrix inversion is often an ill-conditioned problem. This ill-conditioned problem is extremely sensitive to noise and errors, and its solution often contains severely amplified, non-physical artifacts, thus failing to obtain a clear and accurate RF performance image.

[0165] In practice, the processing module 50 first retrieves the fused and verified complete mutual scattering matrix from the result merging module 40, and then retrieves the original, uncompensated near-field scanning data from the scanning module 10.

[0166] Processing module 50 first performs a preliminary matrix solution. This step is a standard inverse operation application, such as solving a system of linear equations using the conjugate gradient method or direct inversion. In the x, C is the mutual scattering matrix, and v is the near-field scan data vector. "" represents the multiplication operation between a matrix and a vector, that is, multiplying matrix C by vector x to obtain vector v. The solved x is the preliminary estimate of the ideal uncoupled near-field distribution, denoted as the first solution matrix.

[0167] After obtaining the initial solution, the module does not directly use it for imaging, but instead performs a quality assessment. The module calculates a first distortion index based on the first solved matrix. This index calculation is complex: on one hand, it is based on the aforementioned response consistency index, i.e., using the first solved matrix as a new excitation source to inversely calculate the response that should be produced on the antenna array and assess the physical consistency of that response; on the other hand, it also specifically calculates the multiple scattering bias value of the first solved matrix. This bias value is obtained through two-dimensional spectral analysis of the first solved matrix, aiming to identify and quantify high-frequency spatial ripples or non-physical oscillations characteristic of high-order multiple scattering effects that are not fully compensated between the probe and the antenna under test.

[0168] Furthermore, processing module 50 applies the calculated first distortion index to fine-tune the first solution matrix. This adjustment is an adaptive regularization operation, such as Tikhonov regularization. Unlike traditional methods that use fixed regularization parameters, the adaptive aspect lies in the fact that the strength of the regularization is spatially modulated by the first distortion index.

[0169] Specifically, in regions where the first distortion index indicates significant distortion, such as areas with high multiple scattering bias values, the module applies stronger regularization to suppress artifacts; while in regions with less distortion, weaker regularization is applied to preserve realistic details to the greatest extent possible. After this adaptive adjustment, a second solution matrix is ​​generated, in which noise and artifacts are effectively suppressed and physical details are more realistic.

[0170] Finally, the processing module 50 takes this optimized second solution matrix, which represents the final, high-fidelity compensated near-field distribution, as input and performs a far-field transformation on it. Mathematically, this transformation is a Fourier integral operation, and in practical calculations, it is typically implemented using an efficient two-dimensional Fast Fourier Transform (2D-FFT) algorithm. The result of the transformation is the final, clear, and accurate RF performance image, such as the far-field radiation pattern.

[0171] As an alternative implementation, it is also necessary to address how to eliminate far-field image distortion caused by near-field residual errors during the final far-field transformation and imaging stages. Although the aforementioned steps have provided fine compensation and adjustment for the near-field data, some weak, spatially highly correlated residual errors, such as those caused by probe polarization leakage or incompletely modeled higher-order scattering effects, may still exist. During far-field transformation using Fourier integration, these small near-field errors can propagate and may be amplified in specific far-field regions, manifesting as non-physical artifacts in the radiation pattern, such as unrealistic cross-polarization lobes or sidelobe lift, thus affecting the accurate assessment of the antenna's true performance.

[0172] In this embodiment, the specific steps include: receiving the second solution matrix from the processing module 50, and calculating the first far-field integral sequence based on the second solution matrix.

[0173] In practice, processing module 50 first receives the high-fidelity second solution matrix, adjusted by adaptive regularization, from the previous step. Then, the module performs a standard Fast Fourier Transform (FFT) operation on this matrix, for example, using a two-dimensional FFT. The result is a preliminary calculation of the antenna's far-field radiation characteristics, forming the first far-field integral sequence. This sequence is a complex matrix representing the far-field amplitude and phase at different azimuth and elevation angles.

[0174] Furthermore, the first radiation pattern region is identified based on the first far-field integral sequence.

[0175] In practice, the initial far-field results are analyzed to identify areas where distortion may exist. Specifically, the module analyzes the polarization purity of each point in the first far-field integral sequence based on the physical characteristics of the antenna design itself, such as its expected dominant polarization direction. For an antenna designed for vertical polarization, its horizontal polarization components in each far-field direction are calculated. Regions with abnormally high horizontal polarization components are identified as the first radiation pattern region. The abnormal cross-polarization values ​​of all points within this region collectively constitute the polarization deviation value sequence.

[0176] Furthermore, a first compensation sequence is generated based on the first radiation pattern region.

[0177] After identifying regions with polarization bias, a first compensation sequence for correction is generated. This sequence is generated based on the physical assumption that significant polarization bias in the far field often originates from incompletely compensated multiple scattering components in the near field. Using a pre-defined physical model or lookup table, the polarization bias values ​​in the far field are inversely mapped to a set of multiple scattering correction values. These correction values ​​aim to counteract the effects of residual multiple scattering propagating to the far field.

[0178] Furthermore, the first compensation sequence is applied to adjust the first far-field integral sequence to generate a second far-field integral sequence.

[0179] In practice, the generated first compensation sequence is used to adjust the initial far-field results. This adjustment is a weighted correction operation.

[0180] Specifically, the correction value is modulated by a spatial weighting function that reaches its maximum value at the center of the first radiation pattern region and then smoothly decays outwards. This ensures that the correction operation is concentrated in the region with the most severe problems, while avoiding the introduction of new abrupt changes at the region boundaries. After this weighted correction, the polarization bias is suppressed, resulting in a purer and more physically accurate second far-field integral sequence.

[0181] Furthermore, the second far-field integral sequence is image-mapped to generate the radio frequency performance image.

[0182] In practice, this finalized second far-field integral sequence is image-mapped. This step includes: first, calculating the magnitude of each complex value in the sequence and converting it to decibels (dB); second, normalizing the result; and finally, performing a two-dimensional projection operation to project the three-dimensional spherical radiation data onto a two-dimensional plane, generating a radio frequency performance image that can be intuitively interpreted by the user, such as a two-dimensional cross-sectional view of the E-plane and H-plane, or a pseudo-color image of the entire visible hemisphere.

[0183] As an optional implementation, converting the target parameter into a serialization prompt includes:

[0184] The target parameters are received from the antenna parameter module 20, and the numerical data of the scan data size, the limitation data of the compensation sub-block capacity, the range data of the electromagnetic frequency range, and the gradient data of the dielectric constant gradient are extracted.

[0185] The numerical data of the scanned data size is serialized to generate a first numerical sequence, wherein the first numerical sequence includes the dimension array of the scanned data size;

[0186] The limitation data of the compensation sub-block capacity is serialized to generate a first limitation sequence, wherein the first limitation sequence includes an array of boundary thresholds for the compensation sub-block capacity;

[0187] The range data of the electromagnetic frequency range and the gradient data of the dielectric constant gradient are subjected to physical model serialization conversion to generate a first physical model sequence, wherein the first physical model sequence includes the spectrum vector of the electromagnetic frequency range and the gradient vector of the dielectric constant gradient.

[0188] The first numerical sequence, the first constraint sequence, and the first physical model sequence are combined to generate the serialization hint, wherein the serialization hint includes an integrated parameter sequence array.

[0189] This embodiment aims to address the technical challenge of ensuring standardization, repeatability, and machine readability during the conversion of target parameters into serialized prompts. While the aforementioned embodiments proposed the concept of converting parameters into numerical sequences, constraint sequences, and physical model descriptions, the lack of a systematic conversion and organization process can lead to inconsistent prompt formats and unstable LLM responses.

[0190] In practical implementation, the device receives target parameters from the antenna parameter module 20 and extracts and classifies these parameters. Specifically, the module extracts the following from the target parameters: numerical data representing the size of the scan data array; limiting data representing the capacity of the compensation sub-block representing the processing capability of the computing unit; range data representing the electromagnetic frequency range of the operating frequency; and gradient data representing the dielectric constant gradient representing the properties of the substrate material.

[0191] Next, the numerical data of the scanned data dimensions is serialized to generate the first numerical sequence. This transformation converts a human-readable dimension descriptor into a machine-readable array of dimensions.

[0192] Next, the constraint data for the capacity of the compensation sub-block is serialized to generate the first constraint sequence. This conversion transforms a single constraint value (1024) into a boundary threshold array containing that value.

[0193] Subsequently, the range data of the electromagnetic frequency range and the gradient data of the dielectric constant are subjected to physical model serialization transformation to generate the first physical model sequence. This step converts the physical parameters into a vectorized expression:

[0194] Convert the electromagnetic frequency range (28 GHz) into a spectrum vector, for example, [28e9].

[0195] Convert the dielectric constant gradient (a linear change from 2.2 to 2.5) into a gradient vector, for example, ["linear", 2.2, 2.5], where the first element defines the gradient type and subsequent elements are key parameters.

[0196] These two vectors together constitute the first physical model sequence.

[0197] Finally, the generated first numerical sequence, first constraint sequence, and first physical model sequence are combined to generate the final serialization hint. This combination process integrates all serialized arrays and vectors into a unified, structured data object, forming an integrated parameter sequence array.

[0198] For example, the final serialization hint can be a JSON object with the following format:

[0199] {"task": "generate_strategy", "params": {"scan_dims": , "block_capacity": , "physics": {"freq_vector": [28e9], "diel_vector": ["linear",2.2, 2.5]}}}.

[0200] This lowers the barrier for LLM to understand complex physical and engineering constraints, thereby ensuring that it can generate high-quality computational strategies more stably and accurately.

[0201] As an optional implementation, the step of dynamically dividing the near-field scanning data into blocks according to the sub-region division instruction to generate sub-scan data includes:

[0202] The sub-region partitioning instruction is received from the block partitioning strategy, and a first scattering vector is calculated based on the sub-region partitioning instruction, wherein the first scattering vector includes the wavelength distribution sequence of the near-field scanning data;

[0203] A first block boundary sequence is generated based on the first scattering vector, wherein the first block boundary sequence includes an array of mutual coupling constraint values ​​of the sub-region partitioning instruction;

[0204] The first block boundary sequence is applied to the near-field scan data to perform block operations, generating a first sub-scan block sequence, wherein the block operations include the region segmentation operation of the near-field scan data;

[0205] The first sub-scan block sequence is integrated to generate the sub-scan data, wherein the integration includes the sequential arrangement operation of the first sub-scan block sequence.

[0206] This embodiment aims to address the algorithm implementation problem: how the numerical computation unit can accurately translate the abstract partitioning instructions, which include physical constraints, provided by the Large Language Model (LLM), into specific segmentation operations on the near-field scan data matrix. Without a clear computational path and only a human-defined concept, the accuracy and repeatability of dynamic segmentation are difficult to guarantee. This embodiment transforms the high-level strategies of LLM into precise and executable segmentation boundaries through a step-by-step, data-driven computational process.

[0207] In practice, the numerical computation unit receives sub-region partitioning instructions from the block strategy generated by the LLM. It then calculates the first scattering vector based on these instructions. This step transforms the physical insights from the LLM into quantitative metrics. For example, the sub-region partitioning instructions might include a note such as "Note the edge wavelength shortening effect caused by the dielectric constant gradient." Accordingly, the numerical computation unit performs a two-dimensional phase gradient analysis on the entire near-field scan data, calculates the local spatial frequency at each sampling point, and thereby derives the effective wavelength distribution sequence at that point. This sequence, the first scattering vector, quantifies the actual propagation characteristics of electromagnetic waves at various points on the antenna surface.

[0208] Next, based on the first scattering vector, a first block boundary sequence is generated. This step utilizes the quantized physical indices from the previous step to determine the optimal dividing line. The numerical computation unit extracts an array of mutual coupling constraint values ​​from the sub-region partitioning instructions; for example, this array might specify that "the mutual coupling truncation error at any sub-block boundary should be less than -40dB." Then, the unit uses the first scattering vector (wavelength distribution sequence) as input and employs optimization algorithms such as dynamic programming to find a set of partitioning boundaries that satisfy the mutual coupling constraints while preferentially placing them in regions with gentle wavelength changes to avoid cutting into regions with sharp field gradients. This optimized, non-uniformly distributed set of boundary coordinates constitutes the first block boundary sequence.

[0209] Next, the first block boundary sequence is applied to perform block segmentation on the near-field scan data. This operation is a precise region segmentation operation. The numerical calculation unit uses the first block boundary sequence as a cutting instruction to slice the complete near-field scan data matrix, generating a set of matrix blocks of varying sizes and with clear physical meanings, namely the first sub-scan block sequence.

[0210] Finally, the first sub-scan block sequence is integrated to generate the final sub-scan data. This integration is a sequential arrangement operation; for example, all the cut sub-scan blocks are stored in a list or structured array according to their spatial positions in the original matrix. This integrated data structure is the final sub-scan data that can be processed in parallel downstream.

[0211] This enhances the feasibility and reproducibility of the entire technical solution, allowing block decisions to be directly driven by the physical characteristics of the measurement data itself. This is more accurate and adaptive than simply relying on geometric location for partitioning, ensuring that the sub-problems after partitioning are physically benign, and laying the foundation for subsequent high-precision calculation of the sub-coupling matrix.

[0212] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A near-field scanning rapid imaging device for multi-channel electrically adjustable base station antenna radio frequency performance, characterized in that, include: The scanning component includes: Radio frequency probe mounted on a two-dimensional precision displacement platform; and a vector network analyzer, connected to the radio frequency probe; The two-dimensional precision displacement platform is used to drive the radio frequency probe to scan along a preset planar grid path in the near field region of the multi-channel electrically tunable base station antenna under test. At the same time, the vector network analyzer measures and collects scattering parameters characterizing the electromagnetic field properties at each sampling point of the grid path to form a near field scanning dataset. A data processing module is connected to the scanning component, and the data processing module is configured as follows: The near-field scan dataset is received, and a fast Fourier transform operation is performed on the near-field scan dataset to complete the near-field to far-field transformation and generate a far-field radiation pattern characterizing the antenna's radio frequency performance.

2. The near-field scanning rapid imaging device for the radio frequency performance of a multi-channel electrically adjustable base station antenna according to claim 1, characterized in that, The data processing module is further configured to apply a preset, static compensation matrix to correct the near-field scanning dataset before performing the near-field to far-field transformation operation, so as to suppress the mutual coupling effect between the probe and the antenna under test.

3. The near-field scanning rapid imaging device for the radio frequency performance of a multi-channel electrically adjustable base station antenna according to claim 1, characterized in that, include: The scanning module includes a scanning component that scans the near-field region of a multi-channel electrically adjustable base station antenna, generates near-field scanning data, and transmits the near-field scanning data to a compensation module. An antenna parameter module is used to store and provide the target parameters of the multi-channel electrically adjustable base station antenna to the compensation module. The compensation module includes an electromagnetic suppression structure and a large language model. The electromagnetic suppression structure includes an electromagnetic bandgap structure or a split-ring resonator, which is used to suppress the mutual coupling components and their cascaded interaction with the edge scattering components in the near-field scanning data, and input the suppression result as suppression data into the large language model. The large language model generates a block strategy and adaptive parameters based on the target parameters, and calculates the sub-mutual scattering matrix by dividing the near-field scanning data and the suppressed data into blocks. The result merging module is used to splice the sub-mutually coupled scattering matrices to generate a mutually coupled scattering matrix; The processing module is used to perform matrix solving on the near-field scanning data based on the mutual scattering matrix, and generate a radio frequency performance image through far-field transformation.

4. The near-field scanning rapid imaging device for the radio frequency performance of a multi-channel electrically adjustable base station antenna according to claim 3, characterized in that, The calculation of the sub-mutual scattering matrix includes: The target parameters are converted into serialized prompts and input into the large language model; The large language model generates a segmentation strategy based on the serialization prompts, wherein the segmentation strategy includes instructions for dividing the near-field scan data into sub-regions. The large language model generates adaptive parameters based on the serialization prompts, wherein the adaptive parameters include prompts for adjusting the calculations; Based on the block segmentation strategy, the near-field scanning data and the suppression data are divided into blocks to generate block data; The segmented data is processed based on the adaptive parameters to calculate the sub-mutual scattering matrix.

5. The near-field scanning rapid imaging device for the radio frequency performance of a multi-channel electrically adjustable base station antenna according to claim 4, characterized in that, Converting the target parameters into serialized prompts and inputting them into the large language model includes: The target parameters are obtained from the antenna parameter module, wherein the target parameters include the scan data size, compensation sub-block capacity, electromagnetic frequency range, and dielectric constant gradient of the multi-channel electrically tunable base station antenna; The target parameters are converted into serialization hints, wherein the serialization hints include a numerical sequence of the scan data size, a limit sequence of the compensating sub-block capacity, and a physical model description of the electromagnetic frequency range and dielectric constant gradient; The serialization prompt is input into the large language model, wherein the large language model generates adaptive parameters containing electromagnetic constraints based on the physical model description.

6. The near-field scanning rapid imaging device for the radio frequency performance of a multi-channel electrically adjustable base station antenna according to claim 5, characterized in that, Based on the segmentation strategy, the near-field scanning data and the suppression data are segmented into blocks to generate segmented data, including: Based on the block-segmentation strategy, edge scattering sensitive regions in the near-field scanning data are identified, wherein the block-segmentation strategy includes sub-region partitioning instructions, and the sub-region partitioning instructions integrate boundary mutual coupling constraints. According to the sub-region division instruction, the near-field scanning data is dynamically divided into blocks to generate sub-scan data, wherein the dynamic block division integrates the wavelength shortening scattering distribution; Based on the block-segmentation strategy, the suppression data is divided into corresponding blocks to generate sub-suppression data, wherein the sub-suppression data corresponds to the edge values ​​of the mutually coupled components in the sub-scan data; The sub-scan data and the sub-suppression data are combined to generate block data.

7. The near-field scanning rapid imaging device for the radio frequency performance of a multi-channel electrically adjustable base station antenna according to claim 3, characterized in that, The process of concatenating the sub-mutually coupled scattering matrices to generate a mutually coupled scattering matrix includes: The compensation module receives the sub-mutual scattering matrix and identifies a first boundary phase region based on the sub-mutual scattering matrix, wherein the first boundary phase region includes a sequence of phase difference values ​​between the sub-mutual scattering matrices; A first weight sequence is calculated based on the first boundary phase region, wherein the first weight sequence is generated based on the edge reflection distribution of the sub-mutual scattering matrix; The first weight sequence is applied to fuse and stitch the sub-mutually coupled scattering matrices to generate a first preliminary mutually coupled scattering matrix, wherein the fusion and stitching includes a weighted superposition operation of the sub-mutually coupled scattering matrices; The first preliminary mutual scattering matrix is ​​subjected to array response verification, wherein the array response verification calculates the response consistency index of the first preliminary mutual scattering matrix, and the first preliminary mutual scattering matrix is ​​adjusted based on the response consistency index to generate the mutual scattering matrix.

8. The near-field scanning rapid imaging device for the radio frequency performance of a multi-channel electrically adjustable base station antenna according to claim 7, characterized in that, Based on the mutual scattering matrix, matrix solving is performed on the near-field scanning data, and a radio frequency performance image is generated through far-field transformation, including: The mutual-coupled scattering matrix is ​​received from the result merging module, and the near-field scanning data is received from the scanning module; Based on the mutual scattering matrix, a preliminary matrix solution is performed on the near-field scanning data to generate a first solution matrix, wherein the preliminary matrix solution includes the inverse operation of the mutual scattering matrix; A first distortion index is calculated based on the first solution matrix, wherein the first distortion index is generated based on the response consistency index and includes the multiple scattering deviation value of the first solution matrix; The first solution matrix is ​​adjusted using the first distortion index to generate a second solution matrix, wherein the adjustment includes an adaptive regularization operation on the first solution matrix; The second solution matrix is ​​subjected to a far-field transformation to generate the radio frequency performance image, wherein the far-field transformation includes the Fourier integral operation of the second solution matrix.

9. The near-field scanning rapid imaging device for the radio frequency performance of a multi-channel electrically adjustable base station antenna according to claim 8, characterized in that, Performing a far-field transformation on the second solution matrix to generate the RF performance image includes: The processing module receives the second solution matrix and calculates a first far-field integral sequence based on the second solution matrix, wherein the first far-field integral sequence includes the Fourier integral result of the second solution matrix; A first radiation pattern region is identified based on the first far-field integral sequence, wherein the first radiation pattern region includes a polarization deviation value sequence in the first far-field integral sequence; A first compensation sequence is generated based on the first radiation pattern region, wherein the first compensation sequence includes the multiple scattering correction value of the polarization deviation value sequence; The first far-field integral sequence is adjusted by applying the first compensation sequence to generate a second far-field integral sequence, wherein the adjustment includes a weighted correction operation on the first far-field integral sequence. The second far-field integral sequence is image-mapped to generate the radio frequency performance image, wherein the image-mapping includes a two-dimensional projection operation of the second far-field integral sequence.

10. The near-field scanning rapid imaging device for the radio frequency performance of a multi-channel electrically adjustable base station antenna according to claim 5, characterized in that, Converting the target parameter into a serialization prompt includes: The target parameters are received from the antenna parameter module, and the numerical data of the scan data size, the limitation data of the compensation sub-block capacity, the range data of the electromagnetic frequency range, and the gradient data of the dielectric constant gradient are extracted. The numerical data of the scanned data size is serialized to generate a first numerical sequence, wherein the first numerical sequence includes the dimension array of the scanned data size; The limitation data of the compensation sub-block capacity is serialized to generate a first limitation sequence, wherein the first limitation sequence includes an array of boundary thresholds for the compensation sub-block capacity; The range data of the electromagnetic frequency range and the gradient data of the dielectric constant gradient are subjected to physical model serialization conversion to generate a first physical model sequence, wherein the first physical model sequence includes the spectrum vector of the electromagnetic frequency range and the gradient vector of the dielectric constant gradient. The first numerical sequence, the first constraint sequence, and the first physical model sequence are combined to generate the serialization hint, wherein the serialization hint includes an integrated parameter sequence array.

Images