A multi-beam synthesis method and device based on a cooperative null generation strategy

Abstract

This application provides a multi-beam synthesis method and apparatus based on a cooperative null generation strategy, belonging to the field of antenna design technology. The method provided in this application involves: randomly generating design specifications within the beam direction range; constructing a mask matrix based on the design specifications; constructing a multi-beam synthesis model; inputting the mask matrix into the multi-beam synthesis model; training the multi-beam synthesis model to output the amplitude and phase excitations for each beam; constructing a target mask matrix based on beam performance specifications in the actual application scenario; inputting the target mask matrix into the trained multi-beam synthesis model to output the amplitude and phase excitations for each array element for different beams; loading the amplitude and phase excitations onto the multi-beam phased array to generate a multi-beam far-field radiation pattern. The multi-beam synthesis method and apparatus based on a cooperative null generation strategy provided in this application are used to reduce the loss of equivalent isotropic radiated power while effectively suppressing multi-beam interference.

Description

Technical Field

[0001] This application relates to the field of antenna design technology, and in particular to a multi-beam synthesis method and apparatus based on a cooperative null generation strategy. Background Technology

[0002] Multi-beam phased arrays, with their ability to simultaneously generate multiple independent high-gain directional beams, offer advantages such as high flexibility and a wide scanning angle, and have been widely applied in many key fields such as point-to-multipoint broadcasting, satellite communication, and 5G wireless communication. In scenarios such as low Earth orbit satellite communication user terminals, multi-beam phased arrays can meet the needs of broadband transmission, access point communication, and multi-point emergency communication, providing an efficient solution for dynamic communication environments and becoming one of the core technologies supporting simultaneous data transmission by multiple users.

[0003] For typical single-beam phased array systems, existing algorithms generate nulls at interference points to improve the beam's anti-interference capability. However, for multi-beam phased array systems, each beam needs to independently generate nulls along the main lobe direction of other beams to achieve a high signal-to-interference ratio (SNR) and reduce interference in the main lobe region. Applying multiple null constraints simultaneously to each beam inevitably disrupts its radiation characteristics, leading to problems such as decreased main lobe gain, increased beamwidth, and reduced main lobe gain, especially when there are many beams. Furthermore, when multiple beams generate nulls in the same direction, these nulls are not simply probabilistically superimposed in space; their phase inconsistencies can cause undesirable coherent superposition of electromagnetic fields in that direction, amplifying interference energy in local areas and deteriorating anti-interference performance. This makes it difficult for existing algorithms to achieve high main lobe gain and anti-interference performance when dealing with multi-beam shaping.

[0004] On the other hand, among the existing phased array synthesis methods based on deep learning, some models need to be retrained when the target changes, which limits their applicability. Others are limited by the performance of traditional algorithms on which the training dataset depends, making it difficult to simultaneously meet the performance requirements of multi-beamforming and the real-time operation requirements, and failing to effectively solve the problem of balancing inter-beam interference suppression and radiation efficiency. Summary of the Invention

[0005] In view of this, this application provides a multi-beam synthesis method and apparatus based on a cooperative null generation strategy, which can reduce the loss of equivalent isotropic radiated power while effectively suppressing multi-beam interference.

[0006] Specifically, this application is implemented through the following technical solution:

[0007] The first aspect of this application provides a multi-beam synthesis method based on a cooperative null generation strategy, the method comprising:

[0008] The beam direction range of the multi-beam phased array is determined, and design parameters are randomly generated within the beam direction range. Based on the design parameters, the main lobe center region, main lobe region, side lobe level, and null region of each beam are defined. The side lobe level is used as a boundary constraint, and a mask matrix is ​​constructed based on the main lobe center region, main lobe region, side lobe level, and null region of each beam. The null region of each beam is the set of the main lobe regions of other beams in the multi-beam array.

[0009] A multi-beam synthesis model is constructed based on a residual network and a Transformer module. The mask matrix is ​​used as a training set and input into the multi-beam synthesis model to train it to output the amplitude excitation and phase excitation of each beam. The residual network learns the local spatial features of the mask matrix and outputs a local feature map. The Transformer module learns the global spatial features of the mask matrix based on the local feature map and outputs a global feature map.

[0010] Construct a target mask matrix based on beam performance indicators in actual application scenarios;

[0011] The target mask matrix is ​​input into a trained multi-beam synthesis model. The trained multi-beam synthesis model outputs the amplitude excitation and phase excitation of each array element in the target mask matrix for different beams. The amplitude excitation and the phase excitation are loaded into a multi-beam phased array to generate a multi-beam far-field radiation pattern. Beam synthesis is achieved based on the multi-beam far-field radiation pattern.

[0012] A second aspect of this application provides a multibeam synthesis device based on a collaborative null-hole generation strategy, the device comprising a construction module, a training module, and a processing module;

[0013] The construction module is used to determine the beam direction range of the multi-beam phased array, randomly generate design parameters within the beam direction range, define the main lobe center region, main lobe region, side lobe level, and null region of each beam based on the design parameters, and construct a mask matrix based on the main lobe center region, main lobe region, side lobe level, and null region of each beam, using the side lobe level as a boundary constraint; wherein, the null region of each beam is the set of the main lobe regions of other beams in the multi-beam array;

[0014] The training module is used to construct a multi-beam synthesis model based on a residual network and a Transformer module. The mask matrix is ​​input into the multi-beam synthesis model as a training set, and the multi-beam synthesis model is trained to output the amplitude excitation and phase excitation of each beam. The residual network learns the local spatial features of the mask matrix and outputs a local feature map. The Transformer module learns the global spatial features of the mask matrix based on the local feature map and outputs a global feature map.

[0015] The construction module is also used to construct a target mask matrix based on beam performance indicators in actual application scenarios;

[0016] The processing module is used to input the target mask matrix into a trained multi-beam synthesis model. The trained multi-beam synthesis model outputs the amplitude excitation and phase excitation of each array element in the target mask matrix for different beams. The amplitude excitation and the phase excitation are loaded into the multi-beam phased array to generate a multi-beam far-field radiation pattern. Beam synthesis is then achieved based on the multi-beam far-field radiation pattern.

[0017] The multi-beam synthesis method and apparatus based on the cooperative null generation strategy provided in this application jointly optimizes all beams, enabling each beam to focus on forming a high-gain main lobe, while all other beams collaboratively generate nulls within the main lobe region of the target beam, thereby globally minimizing interference signal power. By combining the cooperative null generation strategy with a residual network and a Transformer module fusion model, this method solves the difficulties in balancing inter-beam interference suppression and equivalent omnidirectional radiation power assurance in multi-beam phased array synthesis, as well as the inability to simultaneously achieve real-time performance and synthesis performance. This results in a multi-beam synthesis effect with high radiation efficiency, low interference, and real-time response. Specifically, in the mask matrix construction stage, by defining the null region of each beam as a set of main lobe regions of other beams, prior features of cooperative nulls are injected into the model. This allows the model to clearly define the optimization direction of multi-beam cooperative interference suppression from the early training stage, avoiding the radiated power loss caused by traditional independent null generation, and providing a feature foundation for cooperative null generation strategies. In terms of model structure, the residual network is integrated with the Transformer module. To address the deficiency of traditional models in simultaneously capturing local spatial features and global dependencies, multiple residual blocks of the residual network extract the local spatial correlation of the main lobes and null regions of each beam in the mask matrix, ensuring the local rationality of array unit excitation. Then, the multi-head self-attention mechanism of the Transformer module is used to capture long-range distances between different beams. Dependence on spatiality enables global collaborative optimization of multiple beams, solving the gradient explosion problem in deep network training and improving the model's optimization performance in collaborative null generation. This allows each beam to maintain high radiated power while effectively suppressing interference through collaborative null generation from other beams. The target mask matrix construction and excitation generation steps accurately convert performance indicators based on actual scenario requirements, enabling the trained model to directly output amplitude and phase excitations adapted to actual needs without iterative optimization. While maintaining the equivalent omnidirectional radiated power of a single beam, it significantly improves the signal-to-interference ratio of multi-beam phased arrays. Furthermore, it completes the main computational load during model training, requiring only a single forward propagation to generate results during inference, meeting real-time synthesis requirements and enhancing the practical value of multi-beam phased arrays. Attached Figure Description

[0018] Figure 1 A flowchart of an embodiment of the multi-beam synthesis method based on a collaborative zero-traps generation strategy provided in this application;

[0019] Figure 2 This is a schematic diagram of the structure of a second embodiment of the multibeam synthesis device based on the collaborative zero-trap generation strategy provided in this application. Detailed Implementation

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0021] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0022] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0023] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0024] Figure 1 This is a flowchart of an embodiment of the multi-beam synthesis method based on a cooperative null generation strategy provided in this application. Please refer to... Figure 1 The method provided in this embodiment may include:

[0025] S101. Determine the beam direction range of the multi-beam phased array, randomly generate design parameters within the beam direction range, define the main lobe center region, main lobe region, side lobe level, and null region of each beam based on the design parameters, construct a mask matrix based on the main lobe center region, main lobe region, side lobe level, and null region of each beam using the side lobe level as a boundary constraint; wherein, the null region of each beam is the set of the main lobe regions of other beams in the multi-beam array.

[0026] Specifically, considering the application scenarios and hardware performance limitations of multi-beam phased arrays, the azimuth and elevation angle ranges that the beams can cover are clearly defined to ensure that the beam direction is within the scanning range achievable by the equipment, providing a reasonable boundary for subsequent random generation of design specifications. Within the determined beam direction range, the main lobe center direction of each beam is randomly selected, i.e., the target pointing direction of the beam. Simultaneously, the main lobe width, side lobe level constraints, and null depth constraints corresponding to each beam are randomly generated. The random generation of design specifications must meet actual communication performance requirements. For example, the main lobe width must ensure the directional transmission accuracy of the beam, the side lobe level constraints must control signal leakage, and the null depth constraints must ensure the suppression effect on interference. All randomly generated specifications must be within a reasonable range achievable in engineering, avoiding exceeding the performance limits of the equipment.

[0027] Furthermore, using the randomly generated main lobe center as the core point, a very small region is designated as the main lobe center. This main lobe center region is the core part where beam energy is most concentrated, used to clearly define the beam's precise pointing target. Based on the main lobe center region, and according to the randomly generated main lobe width, a rectangular region is designated as the main lobe region. This region is the main energy radiation area of ​​the beam, and it is necessary to ensure that the beam energy within this region meets transmission requirements. The main lobe regions of all other beams in the multi-beam array are combined to form the null region of the current beam. That is, the current beam needs to form cooperative nulls in the main lobe regions of other beams to suppress mutual interference between different beams and ensure signal purity when each beam operates independently. It should be noted that the other beams in the multi-beam array refer to all beams except the current beam. The null is not generated independently by a single beam, but rather formed by the cooperative action of the remaining beams corresponding to this main lobe region.

[0028] Furthermore, a two-dimensional matrix is ​​assigned as a sub-mask for each beam, the size of which matches the sampling resolution of the beam pattern. In the sub-mask, the center region of the main lobe is assigned the highest gray value, the main lobe region is assigned the second highest gray value, the null region is assigned the lowest gray value, and the remaining regions are assigned intermediate gray values ​​based on sidelobe level constraints. The sub-masks of all beams are combined along the third dimension to form a mask matrix. The matrix values ​​are controlled within a specific range and integerized to reduce data storage requirements while providing clear feature identifiers for the neural network.

[0029] Furthermore, the multi-beam phased array supports space-division duplexing while utilizing the maximum available antenna aperture, by N x *N y Composed of individual array elements, forming N B There are j-th beams, and each antenna element is connected to a phase shifter and an attenuator. The far-field pattern of the j-th beam can be expressed by the following formula:

[0030] ;

[0031] in, For the j-th beam in Far-field radiation pattern at the location;

[0032] The element pattern of the m-th row and n-th column is shown.

[0033] The horizontal distance between the m-th row and n-th column element and the first element;

[0034] The perpendicular distance between the m-th row and n-th column element and the first element;

[0035] The operating wavelength for a multi-beam phased array;

[0036] , The pitch angle, It is the azimuth angle;

[0037] , For amplitude incentive, This is for phase excitation.

[0038] Furthermore, the implementation steps for defining the main lobe center region, main lobe region, and null region for each beam based on the aforementioned design specifications include:

[0039] (1) The design parameters include at least the beam pointing and the main lobe width, and the main lobe center region of each beam is determined based on the beam pointing;

[0040] (2) Determine the main lobe region of each beam based on the main lobe width;

[0041] (3) Determine the sidelobe level based on the main lobe region;

[0042] (4) Determine the null region of each beam based on the set of main lobe regions in other beams in the multi-beam system.

[0043] Specifically, using the beam pointing of each beam as the sole reference, a very small area of ​​the main lobe center region is delineated. This region is the core point where the beam energy is most concentrated. Taking the main lobe center region as the midpoint, the main lobe region of the beam is delineated by expanding outwards according to the main lobe width. The main lobe regions of all beams other than the current beam in the multi-beam array are collected as the null region of the current beam, which is used to clarify the specific range of interference suppression.

[0044] Furthermore, in the sampling coordinate system of the beam pattern, a very small rectangular region is defined as the main lobe center region, with the coordinate point corresponding to the beam direction as the core. This region is the core part where the beam energy is most concentrated, used to clearly define the precise target of the beam and provide a clear core feature identifier for the neural network. The delineation of the main lobe region is based on the main lobe width and the main lobe center region. With the main lobe center region as the midpoint, the region is symmetrically extended in the azimuth and elevation directions according to the randomly generated main lobe width to form a complete main lobe region. For example, when the main lobe width is set to 13 degrees, it is extended 6.5 degrees to each side of the main lobe center. This region corresponds to a coverage area where the beam energy drops to within half of the peak value, covering a range of -3 dB, ensuring that the main energy of the beam is concentrated in this region.

[0045] Furthermore, in the construction of the mask matrix of a multi-beam phased array, the determination of the sidelobe level is based on the main lobe region as the reference boundary. First, all spatial locations outside the main lobe region of each beam are identified as potential sidelobe distribution areas. Then, according to the preset sidelobe suppression requirements and combined with the multi-beam cooperative interference control target, differentiated gray values ​​are assigned to this region. If the preset sidelobe level is low and energy leakage needs to be strictly suppressed, then the spatial locations outside the main lobe region and the zero-point region are assigned low gray values ​​close to 0 to transmit strong suppression signals. If the sidelobe level constraint is relatively loose, then an intermediate gray value between the main lobe region and the zero-point region is assigned to balance energy utilization and interference control. Through this gray value quantization method, the performance requirements of the sidelobe level are transformed into the feature identifiers of the mask matrix, enabling the neural network to accurately identify the sidelobe regions that need to control energy leakage and the suppression intensity.

[0046] Furthermore, the null region is defined using the mutually exclusive main lobe principle. For any beam in a multi-beam array, its null region is the union of the main lobe regions of all other beams. For example, in a four-beam phased array, the null region of beam 1 is the set of the main lobe regions of beams 2, 3, and 4. By forcing all other beams in the current beam to form cooperative nulls in the main energy radiation region of the current beam, mutual interference between beams is suppressed at the source, ensuring the signal purity when each beam operates independently.

[0047] By using beam pointing and main lobe width as core design metrics, the main lobe center region and main lobe region of each beam are precisely defined, and the main lobe regions of other beams are collected as the current beam null region. Based on a clear and precise beam pointing and effective coverage, mutual interference between multiple beams can be specifically suppressed. At the same time, in conjunction with a multi-beam synthesis model based on residual networks and Transformer modules, it not only ensures the high equivalent omnidirectional radiation power of each beam, but also significantly improves the signal-to-interference ratio, achieving an effective balance between interference suppression and radiation efficiency. Moreover, it supports flexible adjustment of beam direction without repeated training, meeting the engineering application requirements of real-time multi-beam synthesis.

[0048] Furthermore, the implementation steps for constructing the mask matrix based on the main lobe center region, main lobe region, side lobe level, and null region of each beam include:

[0049] (1) Determine the number of row vectors of the mask matrix based on the number of array elements of the multi-beam phased array in the horizontal direction, and determine the number of column vectors of the mask matrix based on the number of array elements of the multi-beam phased array in the vertical direction.

[0050] Specifically, the mask matrix corresponds to the desired radiation pattern and serves as the input to the neural network. Its value is determined by the number of elements in the multi-beam phased array and the number of downsampling operations in the neural network. Assuming the number of downsampling operations in the neural network is N, the number of array elements in the horizontal direction of the multi-beam phased array is counted, and then multiplied... The number of row vectors in the mask matrix; count the number of array cells in the vertical direction and multiply it. The number of column vectors in the mask matrix ensures that the row and column dimensions of the mask matrix match the distribution of the array cells.

[0051] Furthermore, the number of rows and columns in the mask matrix is ​​directly determined by the physical structure of the multi-beam phased array and the number of downsampling operations in the neural network. For example, in an 8×8 four-beam phased array with 8 array elements in both the horizontal and vertical directions and a neural network with 3 downsampling layers, the number of rows and columns in the mask matrix is ​​set to 64. For a 16×16 four-beam phased array, the number of rows and columns is set to 128 each. This ensures that the global feature map output by the mask matrix after passing through the neural network accurately corresponds to the position of each element in the array, establishing a spatial mapping relationship for subsequent allocation, allowing the neural network to directly associate the mask features with the excitation parameters of the array elements.

[0052] (2) According to the design specifications, the gray values ​​of the main lobe center region, main lobe region, side lobe level and null region of each beam are allocated to form a two-dimensional mask for each beam;

[0053] Specifically, the design specifications for each beam are clearly defined, including beam pointing, main lobe width, side lobe level constraints, and null depth requirements, which serve as the core basis for grayscale value allocation. For each beam, the main lobe center region, main lobe region, side lobe region, and null region are delineated according to the design specifications. The main lobe center region is assigned the highest grayscale value, the main lobe region is assigned the secondary grayscale value, the side lobe region is assigned the corresponding grayscale value according to the side lobe level constraints, and the null region is assigned the lowest grayscale value. The grayscale value distribution of each region is integrated to form a two-dimensional mask for a single beam, ensuring a one-to-one correspondence between regional features and grayscale values, thus meeting the requirements of neural network recognition.

[0054] Furthermore, the identification intensity of each region is clearly defined based on design specifications. The main lobe center region, as the core where beam energy is most concentrated, is assigned the highest gray value of 255 to highlight its key directional characteristics. The main lobe region, which is 3dB less than the maximum beam value, is assigned a slightly higher gray value, slightly below 255. Based on sidelobe level constraints, intermediate gray values ​​between 0 and 255 are assigned, distinguishing it from the main lobe center region while also indicating its energy radiation range. Null regions are assigned the lowest gray value of 0, clearly defining them as interference suppression areas. For example, if the main lobe center of a beam points at (30°, 30°) and the main lobe width is 13°, then in the two-dimensional mask, the coordinate point corresponding to this direction is assigned a value of 255. The main lobe region extending 6.5° outwards from this point is assigned the second highest gray value, the regions corresponding to the main lobes of other beams are assigned 0, and the remaining regions are assigned intermediate gray values, ensuring that the characteristics of each region are clearly distinguishable in the mask.

[0055] The main lobe center region is assigned a maximum gray value of 255 to highlight its key position as the core of beam energy and provide the model with clear directional characteristics. The main lobe region is assigned an intermediate gray value between 100 and 200 to identify it as the main energy radiation area. The value can be fine-tuned according to the equivalent isotropic radiation power requirement. The sidelobe region is assigned gray values ​​according to the sidelobe level constraint. The lower the sidelobe level, the stricter the constraint, and the closer the gray value is to the null region, which transmits a strong suppression signal. If the constraint is loose, the gray value can be appropriately increased to balance energy utilization and interference control. The null region is uniformly assigned a minimum gray value of 0 to clearly identify it as a strong interference suppression area, ensuring that the model learns the characteristic that the energy in this area needs to be completely suppressed. After the gray values ​​of each region are assigned, the gray values ​​are filled into the coordinate positions of the corresponding two-dimensional matrix according to the sampling coordinate system of the beam pattern to form a two-dimensional mask for a single beam. The mask matrix not only clearly presents the spatial structure of the beam (the location and range of the main lobe, side lobes, and nulls), but also quantifies the performance requirements of each region through grayscale values, providing accurate feature input for the subsequent multi-beam synthesis model to learn the mapping relationship between region features and excitation parameters.

[0056] Furthermore, based on beam pattern indices randomly generated within a specified directional range, and combined with predefined sidelobe levels and null depth constraints, a mask corresponding to each beam is constructed. Using this dataset for model training, the network learns a multi-beam array synthesis model capable of generating beams under any angular combination within the range, requiring only one forward inference to generate the excitation coefficients. The specific generation process is as follows:

[0057] Consider the j-th beam pointing (u j ,v j The main lobe width is d. j Then the main lobe region M of the j-th beam j for:

[0058] ;

[0059] in, For directional range;

[0060] (u j ,v j ) represents the pointing range of the j-th beam;

[0061] The width of the main lobe.

[0062] N B The directions of the beams are sorted from largest to smallest to ensure a consistent spatial sequence of mask generation, facilitating stable convergence of model training. Then, the mask U of the j-th beam... j It can be obtained through the following formula:

[0063] ;

[0064] Wherein, NDL is the null depth, representing the ratio of the power of other beams in the main lobe region of the j-th beam to the power of the main lobe of the j-th beam; Let be the sidelobe level of the j-th beam;

[0065] N B The masks of each beam are combined along the third dimension to construct a three-dimensional mask matrix M, M∈C. 8Nx×8Ny×NB Its value ranges from 0 to 255. After calculation, it is rounded and stored as an 8-bit integer array, which significantly reduces the storage requirements of the dataset and facilitates training on large-scale datasets.

[0066] It is important to note that the mask is not used to reproduce the actual beam pattern, but rather to provide explicit learning features for the neural network, such as the main lobe region, side lobe levels, and null locations. For large-scale arrays, the actual main lobe becomes extremely narrow, and using its true width would compress these features into only a few pixels, making them difficult for the network to capture. Therefore, the main lobe width of the mask remains consistent across different array sizes, with only the sampling resolutions fx and fy adjusted. In this embodiment, the main lobe width of the mask is fixed at 13°.

[0067] (3) Stack the two-dimensional masks of each beam in three dimensions according to the number of row vectors and the number of column vectors to obtain the mask matrix.

[0068] Specifically, with the number of row vectors and column vectors as fixed dimensions, the two-dimensional masks of all beams are stacked sequentially along the third dimension to form a three-dimensional mask matrix, ensuring that the matrix structure is compatible with the input requirements of the subsequent neural network.

[0069] Furthermore, the 3D stacking process uses the number of row and column vectors as a fixed dimension, sequentially stacking the 2D masks of each beam to form a third dimension. For example, in a four-beam phased array, the 2D mask size of each beam is 64×64 (corresponding to an 8×8 array). After stacking four 64×64 2D masks along the third dimension, a 64×64×4 3D mask matrix is ​​finally formed. This stacking method preserves the spatial feature information of each beam, while distinguishing different beams through the third dimension. This allows the neural network to learn the regional constraint relationships of all beams at once, providing complete feature input for subsequent collaborative null generation and multi-beam synthesis. The values ​​of the stacked mask matrix are integerized into 8-bit integer arrays, significantly reducing data storage overhead while ensuring no loss of feature information, adapting to the training needs of large-scale datasets.

[0070] By determining the number of row and column vectors in the mask matrix based on the number of array elements in the horizontal and vertical directions of the multi-beam phased array, the mask matrix is ​​ensured to accurately correspond to the array's physical structure. Differential grayscale values ​​are then assigned to the main lobe center region, main lobe region, side lobe region, and null region of each beam according to design specifications to form a two-dimensional mask. Finally, the two-dimensional masks are stacked in three dimensions according to their corresponding row and column numbers to obtain the mask matrix. This allows the mask matrix to clearly and accurately carry the spatial characteristics and constraint information of each beam, laying a solid foundation for the subsequent residual network and Transformer module to efficiently learn local and global spatial characteristics and accurately output amplitude and phase excitations. Simultaneously, it ensures a high degree of compatibility between the mask matrix and the input requirements of the multi-beam synthesis model, facilitating rapid model convergence and achieving high signal-to-interference ratio and high equivalent isotropic radiation power in multi-beam synthesis. Furthermore, it can be directly used for model training and inference without additional format conversion, improving the efficiency and accuracy of multi-beam synthesis.

[0071] S102. Construct a multi-beam synthesis model based on a residual network and a Transformer module. Input the mask matrix as a training set into the multi-beam synthesis model and train the multi-beam synthesis model to output the amplitude excitation and phase excitation of each beam. The residual network learns the local spatial features of the mask matrix and outputs a local feature map. The Transformer module learns the global spatial features of the mask matrix based on the local feature map and outputs a global feature map.

[0072] Specifically, a multi-beam synthesis model integrating a residual network and a Transformer module is constructed. The residual network is responsible for extracting local spatial features, while the Transformer module is responsible for capturing global spatial features. The constructed mask matrix is ​​divided into training and validation sets proportionally. The training set is used for model parameter learning, and the validation set is used to evaluate the model's generalization ability. The training set is input into the model, and the residual network extracts the local spatial features of the mask matrix and outputs a local feature map. Then, the Transformer module learns the global spatial features and outputs a global feature map. Based on the global feature map, the model outputs the predicted amplitude and phase excitation values ​​for each beam. The prediction error is calculated using a composite loss function, and the model parameters are optimized through backpropagation until the model converges.

[0073] Furthermore, in the model structure design, the residual network consists of multiple residual blocks, including downsampling modules and convolutional layers. The residual blocks solve the gradient vanishing problem in deep network training through skip connections. The downsampling module compresses the feature dimension while retaining key information. The convolutional layers extract local spatial features such as the main lobe shape and local null distribution of individual beams in the mask matrix. The Transformer module is embedded in the output stage of the residual network and adopts a multi-head self-attention mechanism to capture the long-distance dependencies between the main lobes and nulls of different beams, achieving global feature fusion. During training, the mask matrix is ​​input into the model in batches. The residual network performs convolution operations and feature extraction on each batch of mask matrices and outputs a local feature map reflecting the local feature correlation. After reshaping, the local feature maps are input into the Transformer module. A multi-head self-attention mechanism calculates the global correlation between features, which is then processed by a position-feedforward network to output a global feature map that integrates global information, achieving an effective combination of local and global features. The amplitude and phase excitations output by the model are compared with the ideal target values. The composite loss function includes sidelobe level loss, equivalent isotropic radiation power loss, and signal-to-interference ratio loss, quantifying the sidelobe suppression effect, beam radiation capability, and anti-interference performance, respectively. The loss error is propagated layer by layer to each layer of the model using a backpropagation algorithm, adjusting parameters such as convolutional kernel weights and attention coefficients to continuously reduce the loss value. During training, the model performance is periodically evaluated using a validation set. When the validation set loss value stabilizes and meets preset requirements, training is stopped, and the trained model is saved.

[0074] Furthermore, the steps for training the multi-beam synthesis model to output the amplitude and phase excitation of each beam include:

[0075] (1) The mask matrix is ​​input into the multi-beam synthesis model, the residual network extracts the local spatial features of the mask matrix, the Transformer module extracts the global spatial features of the mask matrix, and outputs a global feature map;

[0076] Specifically, the constructed mask matrix is ​​input into the multi-beam synthesis model. The residual network in the model extracts local spatial features such as the main lobe shape and local region correlation of a single beam in the mask matrix. The Transformer module captures global spatial features such as long-distance dependence between different beams based on local features, and together outputs a global feature map that integrates local and global information.

[0077] Furthermore, after the mask matrix is ​​input into the model, local spatial features are first extracted by a residual network. The residual network consists of multiple residual blocks, which, through sliding window operations in convolutional layers, capture the local shape and positional relationships of the main lobe center region, main lobe region, and null regions of a single beam in the mask matrix. Simultaneously, skip connections are used to address the gradient vanishing problem in deep network training, ensuring the effectiveness of local feature extraction. The local feature map output by the residual network is input into the Transformer module. The Transformer module, through a multi-head self-attention mechanism, calculates the association weights of each feature point with all other feature points, thereby capturing the long-distance dependencies between the main lobe regions of different beams and the null regions of other beams, achieving global feature fusion. For example, the Transformer module can identify the association between the main lobe region of beam 1 and the null regions of beams 2, 3, and 4, providing feature support for collaborative null generation. The Transformer module outputs a global feature map, which retains the local spatial features of a single beam while fusing global association information between multiple beams, laying the foundation for the subsequent generation of excitation parameters.

[0078] (2) The global feature map is split according to the channel dimension of the global feature map, and the split global feature map is mapped to obtain the initial values ​​of amplitude excitation and phase excitation for each beam.

[0079] Specifically, the global feature map is split according to the channel dimension to obtain feature sub-maps corresponding to the number of beams and excitation type. Mapping operations are then performed on the split feature sub-maps to generate initial amplitude excitation values ​​and initial phase excitation values ​​for each beam.

[0080] Furthermore, the channel dimension of the global feature map contains the amplitude and phase-related feature information of all beams. When splitting according to the channel dimension, every two channels correspond to the amplitude and phase features of one beam. For example, channel 1 and channel 5 correspond to the amplitude and phase features of beam 1, channel 2 and channel 6 correspond to the amplitude and phase features of beam 2, and so on. After splitting, an amplitude feature submap and a phase feature submap equal to the number of beams are obtained. A mapping operation is performed on the amplitude feature submap, and the feature values ​​are mapped to the [0,1] interval through an activation function to obtain the initial amplitude excitation value for each beam. The initial amplitude excitation value corresponds to the signal amplitude adjustment reference of the array unit. A mapping operation is performed on the phase feature submap, and the feature values ​​are mapped to the [-π, π] interval through an activation function to obtain the initial phase excitation value for each beam. The initial phase excitation value corresponds to the signal phase adjustment reference of the array unit. During the mapping operation, the weight parameters of the neural network are adjusted to ensure that the generated initial excitation value can initially meet the design indicators such as the main lobe pointing and main lobe width of the beam, providing a reasonable starting point for subsequent model training and optimization.

[0081] (3) Construct a total loss function based on sidelobe level, equivalent isotropic radiation power and signal-to-interference ratio, and adjust the model parameters of the multi-beam integrated model through backpropagation of the total loss function until the termination condition is met to complete the training of the multi-beam integrated model.

[0082] Specifically, a total loss function is constructed that includes sidelobe level loss, equivalent isotropic radiated power loss, and signal-to-interference ratio loss. The error between the model's predicted value and the ideal target value is calculated using the total loss function. The error is then propagated layer by layer using the backpropagation algorithm. Model parameters such as the convolution kernel weights of the residual network and the attention coefficients of the Transformer module are adjusted. The training process is repeated until the preset termination condition is reached, thus completing the training of the multi-beam integrated model.

[0083] Furthermore, a loss function based on a collaborative null generation strategy is constructed to guide the optimization of neural network parameters during model training. This involves introducing a signal-to-interference ratio (SIR) loss focused on collaborative interference suppression among multiple beams. Through a dynamic weight allocation mechanism, the strategy requirement of accurately covering the main lobe regions of other beams with the null region of each beam is transformed into a quantifiable optimization objective. During model training, the loss function first calculates the superposition value of the radiated energy of other beams within the main lobe region of each beam, quantifying the deviation between the actual interference suppression effect and the preset null depth. Then, combined with the main lobe energy concentration index, the multi-head self-attention coefficient of the Transformer module is preferentially adjusted through backpropagation to strengthen the model's learning of the global correlation between the main lobe region and the null region. Simultaneously, the convolutional kernel weights of the residual network are optimized to guide the amplitude and phase excitation parameters of the array units to converge towards the direction of in-phase superposition of energy in the main lobe and out-of-phase cancellation of interference in the null region. This achieves the collaborative optimization objective of significantly improving the SIR while maintaining high equivalent omnidirectional radiated power among multiple beams.

[0084] Furthermore, the three components of the total loss function target different performance indicators: sidelobe level loss ensures that the sidelobe levels of each beam are not boosted, reducing the impact of noise from non-interference directions on beam performance; equivalent isotropic radiation power loss ensures that each beam focuses on forming a high-gain main lobe; and the signal-to-interference ratio (SIR) loss function, when calculating the SIR of each beam, first performs complex domain superposition of the complex far-field radiation patterns of all other beams within the main lobe region of that beam, and then calculates the total power of the superposition result as the interference signal power. This accurately reflects the coherent superposition effect caused by phase inconsistency among multiple beams in the main lobe region, guiding all other beams to... The beams collaboratively generate nulls within the main lobe region of the target beam to globally minimize interference signal power. The total loss function combines the three loss components through weighted summation, and the weight coefficients can be flexibly adjusted according to the performance requirements of the actual application scenario. For example, in satellite communication scenarios, the weight of the equivalent isotropic radiation power loss can be appropriately increased to ensure energy supply for long-distance communication. During model training, the value of the total loss function is first calculated, and then the loss error is propagated from the output layer to the input layer layer by layer through the backpropagation algorithm. Based on the error gradient, the convolutional kernel weights, bias terms, attention coefficients of the Transformer module, and fully connected layer weights of the residual network are adjusted. After each parameter adjustment, the mask matrix is ​​re-inputted for forward inference, and a new loss function value is calculated. This process is repeated. The training termination condition can be set to a preset maximum number of iterations or the loss function value on the validation set tending to stabilize. When the termination condition is reached, the model parameters are no longer adjusted, and the model can stably output amplitude and phase excitations that meet the performance indicators. The multi-beam integrated model training is then complete.

[0085] The mask matrix is ​​input into the multi-beam synthesis model that integrates the residual network and the Transformer module. The residual network accurately extracts local spatial features, while the Transformer module effectively captures global spatial features and outputs a global feature map. The global feature map is then split along the channel dimension and mapped to obtain the initial values ​​of amplitude and phase excitation for each beam. Finally, a total loss function is constructed based on sidelobe level, equivalent isotropic radiated power, and signal-to-interference ratio (SIR). The model parameters are iteratively adjusted through backpropagation until the termination condition is met. This allows the model to fully learn the mapping relationship between beam spatial features and excitation parameters, ensuring high equivalent isotropic radiated power and low sidelobe level for each beam. It also significantly improves the SIR through collaborative optimization, effectively balancing interference suppression and radiation efficiency. At the same time, it avoids the iterative computation redundancy of traditional methods, allowing the model to support flexible beam direction adjustment without retraining after training, thus achieving real-time, efficient, and high-performance multi-beam synthesis.

[0086] Furthermore, the implementation steps for constructing the total loss function based on sidelobe level, equivalent isotropic radiated power, and signal-to-interference ratio include:

[0087] 3.1 Calculate the sidelobe level loss, equivalent isotropic radiated power loss, and signal-to-interference ratio loss of the multi-beam phased array respectively;

[0088] Specifically, the core purpose of sidelobe level loss is to ensure that the sidelobe energy of the beam is effectively suppressed, avoiding interference with other signals. The calculation involves obtaining the sidelobe level predicted by the model for each beam and comparing it with the preset target sidelobe level. The sidelobe level loss is obtained by calculating the absolute value of the predicted sidelobe level and the target sidelobe level for each beam, and then averaging these absolute values ​​across all beams. This calculation method directly reflects the deviation between the overall sidelobe suppression effect and the target, ensuring that the sidelobe level of each beam meets the constraints and preventing excessively high sidelobe levels in individual beams from affecting overall performance. The sidelobe level loss can be calculated using the following formula:

[0089] ;

[0090] in, The sidelobe level of the predicted j-th beam;

[0091] This represents the actual sidelobe level of the j-th beam;

[0092] Number of beams;

[0093] Furthermore, the equivalent isotropic radiated power directly determines the long-distance communication capability of the beam, and its loss calculation must focus on ensuring the radiation intensity of the main lobe in the target direction. In the calculation, instead of using the traditional maximum beam energy value, the field strength amplitude at the center of the target main lobe is used as the core parameter. Combined with the input power of the array elements and the integral of the beam's radiated energy, logarithmic operations are used to quantify and predict the difference between the equivalent isotropic radiated power and the ideal value. Then, this difference value is averaged across all beams to obtain the equivalent isotropic radiated power loss. This design forces the model to ensure that the main lobe accurately points in the target direction, while maximizing the radiated energy in the main lobe region to meet the energy requirements of long-distance communication. The equivalent isotropic radiated power loss can be calculated using the following formula:

[0094] ;

[0095] in, Number of beams;

[0096] For directional range;

[0097] (u j ,v j ) represents the pointing range of the j-th beam;

[0098] Number of beamlines;

[0099] For beam column vectors;

[0100] For the j-th beam in (u j ,v j Far-field radiation pattern at ( );

[0101] For the j-th beam in Far-field radiation pattern at [location].

[0102] Furthermore, the signal-to-interference ratio (SIR) loss is used to evaluate the interference suppression effect between beams, and its core is to reflect the effectiveness of the cooperative null generation mechanism. During calculation, firstly, the radiated energy of all other beams within the main lobe region of each beam is summed in the complex domain. Then, the ratio of the radiated energy of the main lobe region itself to the interference energy is calculated, i.e., the SIR. The SIR loss is obtained by taking the logarithm of the SIR of all beams and averaging the results. This accurately captures the interference level of other beams in the current main lobe region, ensuring that the model, through cooperative null generation, reduces the loss of equivalent isotropic radiated power while effectively suppressing multi-beam interference. The SIR loss can be calculated using the following formula:

[0103] ;

[0104] in, Number of beams;

[0105] Let be the signal-to-interference ratio of the j-th beam.

[0106] 3.2 Determine the weight of each loss based on the performance priority requirements of the multi-beam phased array;

[0107] Specifically, the allocation of weighting coefficients must be closely aligned with the performance priorities of the actual application scenario. For instance, in long-distance transmission scenarios such as low-Earth orbit satellite communication, equivalent isotropic radiated power is crucial for ensuring communication link stability and should be given the highest weight. Signal-to-interference ratio (SIR) directly affects the signal quality of concurrent transmissions by multiple users, and its weight can be flexibly adjusted according to the interference environment. Sidelobe level loss has a relatively low weight to ensure that sidelobe energy is suppressed without significantly impacting core performance. The setting of weighting coefficients requires multiple adjustments to ensure they match actual engineering needs and achieve a balanced optimization of different performance indicators.

[0108] 3.3 The total loss function is obtained by weighting and summing the sidelobe level loss, equivalent isotropic radiated power loss, and signal-to-interference ratio loss based on the weights.

[0109] Specifically, the total loss function can be expressed by the following formula:

[0110] ;

[0111] in, , , These are the corresponding weights;

[0112] This is due to sidelobe level loss;

[0113] This is the equivalent isotropic radiation power loss;

[0114] This is the signal-to-interference ratio loss.

[0115] The sidelobe level loss, equivalent isotropic radiated power loss, and signal-to-interference ratio (SIR) loss of a multi-beam phased array are calculated separately to accurately quantify the deviations of each performance index from the target value. Then, appropriate weights are assigned to the three losses according to the performance priority requirements of the actual application scenario. Finally, a total loss function is constructed by weighted summation. This allows the multi-beam synthesis model to accurately focus on core performance requirements during training, effectively suppressing sidelobe energy and ensuring the equivalent isotropic radiated power for long-distance beam transmission. It also significantly improves the SIR through a collaborative null mechanism, achieving a dynamic balance between interference suppression and radiation efficiency. At the same time, it avoids the model training from getting trapped in local optima, ensuring that the trained model can stably output amplitude and phase excitations that meet the performance requirements of multiple scenarios, supporting real-time and efficient multi-beam synthesis.

[0116] Furthermore, the residual network includes multiple residual blocks, which include N downsampled residual blocks and M basic residual blocks. The mask matrix is ​​downsampled based on the N downsampled residual blocks.

[0117] Based on the M basic residual blocks, a two-dimensional convolution operation is performed on the downsampled mask matrix to extract local spatial features and output the local feature map.

[0118] Specifically, a residual network is constructed, consisting of N downsampled residual blocks and M basic residual blocks. The downsampled residual blocks are configured with convolutional layers with strides, while the basic residual blocks are configured with two-dimensional convolutional layers with fixed strides. The downsampled and basic residual blocks are alternately set. The mask matrix is ​​input into the residual network and downsampled through the N downsampled residual blocks to compress the spatial dimension of the mask matrix and retain core local features. Two-dimensional convolutional operations are performed through the M basic residual blocks to extract more refined local spatial features, including the main lobe shape of a single beam, region boundaries, and local positional correlations. After processing by all residual blocks, a local feature map containing complete local spatial feature information is output, providing a foundation for the subsequent Transformer module to extract global features. For example, in one embodiment, after a downsampled residual block is downsampled, the downsampled mask matrix is ​​convolved by two basic residual blocks, and then processed based on another downsampled residual block. This process is repeated to achieve alternating settings of downsampled residual blocks and basic residual blocks to process the mask matrix. It should be noted that the downsampled residual blocks and basic residual blocks are set according to actual needs during the alternation process. In this embodiment, there is no limitation on them. The total number of downsampled residual blocks is N, and the total number of basic residual blocks is M.

[0119] Furthermore, the structure design of the residual network needs to be adapted to the feature extraction requirements of the mask matrix. The core of the downsampling residual block is to configure a convolutional layer with a stride of 2, which is used to achieve downsampling. The value of N is usually determined based on the initial size of the mask matrix and the size of the target feature map. The basic residual block uses a two-dimensional convolutional layer with a stride of 1, focusing on feature extraction rather than dimensionality compression. The value of M needs to ensure the sufficiency of local feature extraction. Both types of residual blocks contain skip connections to alleviate the gradient vanishing problem in deep network training and ensure the stability and effectiveness of feature extraction. After the mask matrix is ​​input, it alternately passes through N downsampling residual blocks and M basic residual blocks. Each downsampling residual block passes through a convolutional layer with a stride of 2, compressing the row and column dimensions of the mask matrix to half of their original values. After N downsampling operations, the spatial resolution of the mask matrix is ​​reduced to 1 / (2^N) of the initial value, while preserving the core positional features of the main lobe center region, the main lobe region, and the null region.

[0120] Furthermore, each basic residual block extracts features from local regions through two-dimensional convolution operations. The convolutional layer slides through each local window of the mask matrix, capturing fine information such as the main lobe shape and contour of a single beam, the boundary features between the main lobe and null regions, and the local correlations between adjacent beam regions. For example, convolution operations identify the high gray-level clustering features in the central region of the main lobe, the gray-level gradient features in the main lobe region, and the low gray-level continuous features in the null regions. At the same time, it mines the spatial proximity relationships and boundary transition features of different regions, achieving deep extraction of local spatial features. After alternating processing by N downsampled residual blocks and M basic residual blocks, all local spatial features are fully extracted and integrated, finally outputting a local feature map. The spatial dimension of the local feature map is consistent with the downsampled mask matrix, and the channel dimension, after being expanded by the convolutional layer, can carry rich local feature information, including both the regional features of a single beam and the local correlation features between beams. This lays a solid foundation for the subsequent Transformer module to capture global spatial dependencies, ensuring that the model can simultaneously take into account the accuracy of local features and the synergy of global features.

[0121] Furthermore, given a size of 8Nx×8Ny×N B The input mask matrix M, after being processed by these residual blocks, outputs a size of Nx×Ny×2N. BThe results show that the residual network employs a two-dimensional convolutional output layer, preserving the spatial correlation between array unit excitations and avoiding the spatial structure loss caused by one-dimensional flattening, thus achieving more efficient learning and better training performance. Specifically, N downsampling residual blocks with convolutional layers of stride 2 are used to downsample the input data. The downsampling process ensures that the final output contains the same amount of information as the number of antenna units. Correspondingly, the sampling sizes are set to fx = 8Nx and fy = 8Ny, matching the initial spatial resolution of the input data. The output layer generates a size of Nx × Ny × 2N. B The feature map is mapped to the range [-1, 1] after passing through the tanh activation function.

[0122] Furthermore, the Transformer module includes a multi-head self-attention layer and a position-wise feedforward network. The Transformer module receives the local feature map, learns the long-range spatial dependencies in the mask matrix based on the multi-head self-attention layer, and outputs global features.

[0123] The location-wise feedforward network performs dimensional expansion and compression transformation on the global features, and outputs a global feature map.

[0124] Specifically, the Transformer module receives the local feature map output by the residual network, which contains local spatial feature information of the mask matrix. The local feature map is processed by a multi-head self-attention layer to learn the long-range spatial dependencies between the main lobe regions and null regions of different beams in the mask matrix, and outputs preliminary global features. The preliminary global features are input into a position-wise feedforward network, first expanded in dimension by a fully connected layer, then processed by an activation function, and finally compressed and transformed in dimension by another fully connected layer, outputting a global feature map that fuses the global dependencies.

[0125] Furthermore, the input to the Transformer module is the local feature map extracted by the residual network. This local feature map already contains fine local information such as the main lobe shape of a single beam and local region associations. After downsampling, its dimension matches the distribution of array units, laying the foundation for subsequent global feature learning. The core role of the multi-head self-attention layer is to capture long-distance spatial dependencies. First, the two-dimensional local feature map is reshaped into a one-dimensional sequence, with each sequence element corresponding to a 128-dimensional feature vector containing local feature information of the corresponding spatial location. Subsequently, the multi-head self-attention layer calculates the association weights of each sequence element with all other elements through multiple parallel attention heads. For example, the association strength between the main lobe region features of beam 1 and the null region features of beams 2, 3, and 4. In this way, the model can learn the collaborative relationships between key regions of different beams across spatial distances, overcoming the limitation of the residual network that can only capture local features, and outputting preliminary global features that integrate global dependency information. The position-wise feedforward network is used to perform dimensional transformation and nonlinear optimization on the preliminary global features. First, a fully connected layer expands the feature dimension from 128 to 512, broadening the feature representation space and capturing more complex feature interaction patterns. Then, a ReLU activation function is used to introduce non-linearity, enhancing the model's fitting ability. Finally, a second fully connected layer compresses the dimension back to 128, preserving key feature information while reducing computational complexity. After processing, the one-dimensional sequence is reshaped into a two-dimensional global feature map with the same spatial dimension as the input local feature map. This feature map contains both local spatial details and incorporates global collaborative relationships between multiple beams, providing comprehensive feature support for accurate prediction of subsequent amplitude and phase excitations. The entire process introduces only moderate computational overhead, yet significantly improves the model's ability to model global spatial relationships, facilitating collaborative optimization of multiple beams.

[0126] To further improve model performance, a Transformer module employing a multi-head self-attention mechanism is integrated. This mechanism allows each feature element to adaptively attend to all other feature elements, enabling the model to capture local structures and long-range dependencies. In array pattern synthesis, the main lobe characteristics are intrinsically coupled with distant regions such as side lobes and nulls. By introducing the Transformer, these global coupling effects are effectively modeled, achieving more accurate multi-beam coordination and improving the control precision of main lobe gain and cooperative null layout. Multi-head self-attention requires flattening the two-dimensional feature map into a one-dimensional sequence and calculating pairwise interactions between all elements, resulting in high computational complexity. Therefore, the proposed network integrates only one Transformer block in the output stage. After compressing the intermediate feature map and extracting key spatial information from the residual block, the downsampled feature representation is input into the Transformer. This design allows the model to capture global contextual dependencies while ensuring that the additional computational overhead introduced by the Transformer is far lower than that of the residual block.

[0127] Furthermore, the input feature map of size Nx×Ny×128 is reshaped into a sequence of length Nx*Ny, each labeled as a 128-dimensional feature vector. This sequence is processed by a multi-head self-attention layer, enabling each array unit to attend to all other units and capture global spatial relationships. The output is then passed through a position-wise feedforward network, which includes a fully connected layer that expands the channel dimension from 128 to 512, a ReLU activation function, and another fully connected layer that maps the channel dimension back to 128. Residual connections and normalization operations are used throughout the module. Finally, the transformed sequence is reshaped back to Nx×Ny×128, providing a feature map with global contextual information for subsequent convolutional layers. Since the Transformer operates on compressed features, its computational cost is still significantly lower than the preceding residual blocks, while effectively modeling long-range dependencies.

[0128] S103. Construct a target mask matrix based on the beam performance indicators in the actual application scenario.

[0129] Specifically, based on the communication requirements of the actual application scenario, the performance indicators of the target beam are defined, including the number of beams, main lobe center direction, main lobe width, upper limit of side lobe level, and null depth requirement. Based on the target performance indicators, the main lobe center region, main lobe region, and null region of each target beam are defined respectively, and the null region is the set of main lobe regions of other target beams. Combining the above description, a sub-mask is constructed for each target beam, and then combined to form a target mask matrix, ensuring that the format of the target mask matrix is ​​consistent with the training set mask matrix.

[0130] Furthermore, the determination of target performance indicators must be closely integrated with the actual application scenario. For example, in a satellite communication scenario, the number of target beams and the direction of the main lobe center need to be determined based on the satellite orbit and user distribution; the main lobe width corresponding to the equivalent omnidirectional radiation power needs to be set according to the communication distance requirements; the upper limit of the sidelobe level and the null depth requirements need to be set according to the space interference environment to ensure that the beam can effectively resist interference while meeting the energy requirements of long-distance communication; the regional definition rules of the target beam are consistent with those in step S101, the main lobe center region is delineated with the direction of the target main lobe center as the core; the main lobe region is delineated according to the target main lobe width to define the decibel energy coverage range; the null region summarizes the main lobe regions of all other target beams to clarify the specific direction of interference that the target beam needs to suppress; when constructing the submask, the gray value assignment rules of the training set are used to ensure the consistency of feature identification. The main lobe center region, the main lobe region, and the null region are respectively assigned corresponding gray values, and then the submasks of all target beams are combined along the third dimension to form a three-dimensional target mask matrix. The matrix's size, numerical range, and other format are completely consistent with the training set mask matrix, ensuring that the model can process it directly without additional format conversion.

[0131] S104. Input the target mask matrix into the trained multi-beam synthesis model. The trained multi-beam synthesis model outputs the amplitude excitation and phase excitation of each array element in the target mask matrix for different beams. Load the amplitude excitation and the phase excitation into the multi-beam phased array to generate a multi-beam far-field radiation pattern. Beam synthesis is realized based on the multi-beam far-field radiation pattern.

[0132] Specifically, the constructed target mask matrix is ​​input into the trained multi-beam synthesis model. After one forward inference, the model outputs the amplitude excitation and phase excitation of each array element corresponding to different target beams. The output amplitude excitation is applied to the attenuator of the multi-beam phased array, and the phase excitation is applied to the phase shifter to adjust the signal amplitude and phase of each antenna element. The antenna elements generate radiated signals according to the applied excitation parameters. The radiated signals of multiple elements are superimposed to form a multi-beam far-field pattern, thus realizing beam synthesis.

[0133] Furthermore, after the target mask matrix is ​​input into the model, the model does not need to be retrained and can directly perform inference through the learned mapping relationship. The residual network quickly extracts the local spatial features of the target mask matrix, and the Transformer module captures the global correlation between target beams. Combined with the optimization experience learned during training, it quickly outputs accurate amplitude and phase excitations. The entire inference process does not require iteration, meeting real-time requirements. After the amplitude excitation is loaded onto the attenuator, the attenuator adjusts the output signal amplitude of each antenna element according to the excitation value, ensuring that the energy of the beam in the main lobe region meets the communication requirements, while suppressing sidelobe energy. After the phase excitation is loaded onto the phase shifter, the phase shifter adjusts the signal phase of each antenna element, so that the radiated signals of all elements are superimposed in phase in the main lobe direction, enhancing the main lobe energy, and canceling out phase in the null region, suppressing interference. All antenna elements work synchronously according to the loaded excitation parameters, and the radiated signals are superimposed in space to form a multi-beam far-field pattern. The energy in the main lobe region of each target beam is concentrated, meeting the communication transmission requirements; the energy in the null region is effectively suppressed, reducing inter-beam interference. The resulting multi-beam far-field pattern meets the performance indicators of actual application scenarios, achieving efficient and interference-resistant beam synthesis.

[0134] Furthermore, the steps for loading the amplitude excitation and the phase excitation onto the multi-beam phased array to generate a multi-beam far-field radiation pattern include:

[0135] (1) The amplitude excitation and the phase excitation are applied to each array element in the multi-beam phased array to form an excitation matrix;

[0136] Specifically, the amplitude and phase excitation of each beam output from the multi-beam synthesis model are loaded onto each array element of the multi-beam phased array to form an excitation matrix containing the amplitude and phase information of all elements.

[0137] Furthermore, amplitude excitation and phase excitation are the core control parameters of a multi-beam phased array array element, corresponding to the attenuator and phase shifter adjustment requirements of the element, respectively. The amplitude excitation matrix and phase excitation matrix of each beam output from the model are loaded into the corresponding array element according to their row and column positions. Each array element needs to receive amplitude and phase excitation commands from all beams. For example, each element of an 8×8 four-beam phased array will receive four sets of amplitude excitation values ​​and four sets of phase excitation values. These excitation information are integrated according to beam dimension and element position to form a complex excitation matrix with dimensions of the number of horizontal array elements × the number of vertical array elements × the number of beams. Each element in the matrix contains the combined amplitude and phase information of the corresponding element and beam, providing a foundation for subsequent far-field radiation calculations.

[0138] (2) Determine the calculation method according to the application scenario of the multi-beam phased array, and calculate the excitation matrix based on the calculation method to obtain the far-field radiation information of each beam;

[0139] Specifically, ideal scenarios are suitable for theoretical verification or situations without the influence of non-ideal factors such as electromagnetic coupling. In this case, the inverse Fourier transform is chosen as the calculation method because it is computationally efficient and can quickly obtain the far-field radiation pattern from the excitation matrix without considering the mutual influence between elements. Practical engineering scenarios are suitable for actual arrays with non-ideal factors such as electromagnetic coupling and inconsistent radiation characteristics of elements. In this case, a calculation method based on full-wave simulation data is chosen. It is necessary to extract the far-field electric field component data of each element from the high-frequency structure simulation software in advance to ensure that the calculation process can reflect the actual effects of electromagnetic coupling and other factors.

[0140] Furthermore, the specific implementation steps include:

[0141] 2.1 If the application scenario is an ideal multi-beam phased array scenario, the excitation matrix is ​​calculated based on the two-dimensional inverse Fourier transform to obtain the far-field radiation information of each beam;

[0142] Specifically, based on the inverse Fourier transform, the excitation matrix of each beam is used as input, and the far-field radiation intensity distribution of the beam is directly obtained through two-dimensional inverse Fourier transform. Since the ideal scenario assumes that all elements have the same radiation characteristics, the calculation process can ignore the individual differences of elements and quickly output far-field radiation information. For example, for each beam of an 8×8 four-beam phased array, the energy distribution data of its main lobe and side lobes can be obtained in milliseconds using this method.

[0143] 2.2 If the application scenario is a real engineering scenario, the far-field electric field component of each array unit is extracted from the full-wave simulation, and the far-field radiation information of each beam is obtained by electric field superposition and radiation intensity conversion.

[0144] Specifically, the pre-stored far-field electric field component data for each array element is retrieved. Then, the amplitude and phase excitations in the excitation matrix are combined with the electric field component data of the corresponding element through complex number operations, and the results are superimposed to obtain the total far-field electric field component for each beam. The total electric field component is then converted into radiation intensity using the conversion formula between electric field strength and radiation intensity, ultimately yielding the far-field radiation information for each beam. This process fully considers the differences in element radiation characteristics caused by electromagnetic coupling, resulting in calculation results that more closely approximate the actual operating state of the array.

[0145] For ideal multi-beam phased array scenarios, a two-dimensional inverse Fourier transform is used to operate on the excitation matrix. For practical engineering scenarios, the far-field electric field components of the array elements are extracted from full-wave simulation and calculated through electric field superposition and radiation intensity conversion. This enables accurate scenario-based solutions for far-field radiation information. In ideal scenarios, it balances computational efficiency and theoretical accuracy, quickly outputting beam radiation characteristics. In practical scenarios, it effectively incorporates non-ideal factors such as electromagnetic coupling, making the calculation results consistent with engineering realities. Ultimately, it provides reliable data support for the accurate generation of multi-beam far-field radiation patterns, helping multi-beam integration achieve an optimized balance between high equivalent omnidirectional radiation power and high signal-to-interference ratio in different scenarios.

[0146] (3) Integrate the far-field radiation information of each beam to generate the multi-beam far-field pattern.

[0147] Specifically, the integration process needs to preserve the independent radiation characteristics of each beam while also showcasing the interaction effects between beams. The far-field radiation information of each beam is mapped using a unified spatial coordinate system and superimposed to form a complete multi-beam far-field pattern. The pattern must clearly show the main lobe center position, main lobe width, side lobe level, and null suppression effect in the main lobe regions of other beams for each beam. For example, the far-field pattern of a four-beam phased array will visually demonstrate that the main lobes of the four beams point to preset directions, and that the radiation energy of other beams within each main lobe region is effectively suppressed. This ultimately forms a pattern with independently distributed main lobes and precise null suppression of interference, providing a direct basis for subsequent verification of beam integration effects.

[0148] Amplitude and phase excitations are applied to each array element of a multi-beam phased array to form an excitation matrix. The inverse Fourier transform or calculation method based on full-wave simulation data is determined according to the ideal or practical engineering scenario. The excitation matrix is ​​then used to perform targeted calculations to obtain the far-field radiation information of each beam. Finally, the far-field radiation information of all beams is integrated to generate a multi-beam far-field radiation pattern. This ensures accurate matching between the far-field radiation calculation and the application scenario, guaranteeing efficient calculation in ideal scenarios while also taking into account non-ideal factors such as electromagnetic coupling in practical scenarios. The resulting radiation pattern clearly presents the main lobe pointing, energy distribution, and null suppression effect of each beam, achieving synergistic optimization of high equivalent omnidirectional radiation power and high signal-to-interference ratio. This provides an intuitive and reliable basis for the performance verification and practical application of multi-beam integration.

[0149] This embodiment presents a multi-beam synthesis method based on a cooperative null generation strategy, achieving a precise balance between interference suppression and radiation efficiency in multi-beam phased arrays. Through a unique cooperative null generation mechanism, the cooperative null region of each beam is defined as a set of main lobe regions of other beams. Combined with the construction of a differentiated grayscale mask matrix, explicit interference suppression prior features are injected into the model, fundamentally avoiding the equivalent omnidirectional radiation power loss caused by traditional independent null generation. This ensures that each beam maintains high radiation energy while achieving targeted suppression of interference from other beams. The model architecture, which integrates residual networks and Transformer modules, not only accurately extracts local spatial features of the mask matrix through multi-level residual blocks of the residual network, ensuring the local rationality of array element excitation, but also captures multiple beams using the multi-head self-attention mechanism of the Transformer module. It overcomes long-distance spatial dependence and achieves global collaborative optimization, effectively solving the problems of gradient training in deep networks and global correlation modeling in multi-beam systems, significantly improving the signal-to-interference ratio. Furthermore, it boasts outstanding advantages in scenario adaptation and real-time response. For ideal scenarios, it uses two-dimensional inverse Fourier transform to efficiently calculate far-field radiation information, while for actual engineering scenarios, it incorporates full-wave simulation data to address non-ideal factors such as electromagnetic coupling. Moreover, the model completes the main computational load during the training phase, and only requires a single forward propagation to output accurate excitation parameters during the inference phase. It can flexibly adjust the beam direction without repeated training, meeting the real-time and high-performance requirements of multi-user concurrent transmission in scenarios such as satellite communication, and greatly enhancing the practical value of multi-beam phased arrays.

[0150] Corresponding to the aforementioned embodiment of a multi-beam synthesis method based on a cooperative null generation strategy, this application also provides an embodiment of a multi-beam synthesis device based on a cooperative null generation strategy.

[0151] Figure 2 This is a schematic diagram of the structure of Embodiment 2 of the multi-beam synthesis device based on the cooperative null generation strategy provided in this application. Please refer to... Figure 2 The apparatus provided in this embodiment includes a construction module 210, a training module 220, and a processing module 230;

[0152] The construction module 210 is used to determine the beam direction range of the multi-beam phased array, randomly generate design parameters within the beam direction range, define the main lobe center region, main lobe region, side lobe level, and null region of each beam based on the design parameters, construct a mask matrix based on the main lobe center region, main lobe region, side lobe level, and null region of each beam using the side lobe level as a boundary constraint; wherein, the null region of each beam is the set of the main lobe regions of other beams in the multi-beam array;

[0153] The training module 220 is used to construct a multi-beam synthesis model based on a residual network and a Transformer module. The mask matrix is ​​input into the multi-beam synthesis model as a training set, and the multi-beam synthesis model is trained to output the amplitude excitation and phase excitation of each beam. The residual network learns the local spatial features of the mask matrix and outputs a local feature map. The Transformer module learns the global spatial features of the mask matrix based on the local feature map and outputs a global feature map.

[0154] The construction module 210 is also used to construct a target mask matrix based on the beam performance indicators in the actual application scenario;

[0155] The processing module 230 is used to input the target mask matrix into a trained multi-beam synthesis model. The trained multi-beam synthesis model outputs the amplitude excitation and phase excitation of each array element in the target mask matrix for different beams. The amplitude excitation and the phase excitation are loaded into the multi-beam phased array to generate a multi-beam far-field radiation pattern. Beam synthesis is then achieved based on the multi-beam far-field radiation pattern.

[0156] The apparatus of this embodiment can be used to perform... Figure 1 The steps of the method embodiment shown are similar in principle and process, and will not be repeated here.

[0157] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0158] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0159] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A multi-beam synthesis method based on a co-zerotone generating strategy, characterized in that, The method includes: The beam direction range of the multi-beam phased array is determined, and design parameters are randomly generated within the beam direction range. Based on the design parameters, the main lobe center region, main lobe region, side lobe level, and null region of each beam are defined. The side lobe level is used as a boundary constraint, and a mask matrix is ​​constructed based on the main lobe center region, main lobe region, side lobe level, and null region of each beam. The null region of each beam is the set of the main lobe regions of other beams in the multi-beam array. A multi-beam synthesis model is constructed based on a residual network and a Transformer module. The mask matrix is ​​used as a training set and input into the multi-beam synthesis model to train it to output the amplitude excitation and phase excitation of each beam. The residual network learns the local spatial features of the mask matrix and outputs a local feature map. The Transformer module learns the global spatial features of the mask matrix based on the local feature map and outputs a global feature map. Construct a target mask matrix based on beam performance indicators in actual application scenarios; The target mask matrix is ​​input into a trained multi-beam synthesis model. The trained multi-beam synthesis model outputs the amplitude excitation and phase excitation of each array element in the target mask matrix for different beams. The amplitude excitation and the phase excitation are loaded into a multi-beam phased array to generate a multi-beam far-field radiation pattern. Beam synthesis is achieved based on the multi-beam far-field radiation pattern.

2. The method of claim 1, wherein, The definition of the main lobe center region, main lobe region, side lobe level, and null region for each beam based on the design specifications includes: The design specifications include at least beam pointing and main lobe width, and the main lobe center region of each beam is determined based on the beam pointing. The main lobe region of each beam is determined based on the main lobe width. The sidelobe level is determined based on the main lobe region; The null region of each beam is determined based on the set of main lobe regions in other beams in a multi-beam system.

3. The method of claim 1, wherein, The construction of the mask matrix based on the main lobe center region, main lobe region, side lobe level, and null region of each beam includes: The number of row vectors of the mask matrix is ​​determined based on the number of array elements of the multi-beam phased array in the horizontal direction, and the number of column vectors of the mask matrix is ​​determined based on the number of array elements of the multi-beam phased array in the vertical direction. According to the design specifications, grayscale values ​​are assigned to the main lobe center region, main lobe region, side lobe level, and null region of each beam to form a two-dimensional mask for each beam. The two-dimensional masks of each beam are stacked in three dimensions according to the number of row vectors and the number of column vectors to obtain the mask matrix.

4. The method of claim 1, wherein, The residual network learns the local spatial features of the mask matrix and outputs a local feature map, including: The residual network comprises multiple residual blocks, including N downsampled residual blocks and M basic residual blocks, and the mask matrix is ​​downsampled based on the N downsampled residual blocks; Based on the M basic residual blocks, a two-dimensional convolution operation is performed on the downsampled mask matrix to extract local spatial features and output the local feature map.

5. The method of claim 1, wherein, The Transformer module learns the global spatial features of the mask matrix based on the local feature map and outputs a global feature map, including: The Transformer module includes a multi-head self-attention layer and a position-wise feedforward network. The Transformer module receives the local feature map, learns the long-range spatial dependencies in the mask matrix based on the multi-head self-attention layer, and outputs global features. The location-wise feedforward network performs dimensional expansion and compression transformation on the global features, and outputs a global feature map.

6. The method of claim 1, wherein, The training of the multi-beam synthesis model outputs the amplitude and phase excitation of each beam, including: The mask matrix is ​​input into the multi-beam synthesis model, the residual network extracts the local spatial features of the mask matrix, the Transformer module extracts the global spatial features of the mask matrix, and outputs a global feature map. The global feature map is split according to the channel dimension, and a mapping operation is performed on the split global feature map to obtain the initial values ​​of amplitude excitation and phase excitation for each beam. A total loss function is constructed based on sidelobe level, equivalent isotropic radiated power, and signal-to-interference ratio. The model parameters of the multi-beam synthesis model are adjusted by backpropagation through the total loss function until the termination condition is met, thus completing the training of the multi-beam synthesis model.

7. The method of claim 1, wherein, The amplitude excitation and the phase excitation are applied to a multi-beam phased array to generate a multi-beam far-field radiation pattern, including: The amplitude excitation and the phase excitation are applied to each array element in the multi-beam phased array to form an excitation matrix; The calculation method is determined according to the application scenario of the multi-beam phased array, and the excitation matrix is ​​calculated based on the calculation method to obtain the far-field radiation information of each beam. The far-field radiation information of each beam is integrated to generate the multi-beam far-field radiation pattern.

8. The method of claim 6, wherein, The total loss function constructed based on sidelobe level, equivalent isotropic radiated power, and signal-to-interference ratio includes: Calculate the sidelobe level loss, equivalent isotropic radiated power loss, and signal-to-interference ratio loss of the multi-beam phased array, respectively. The weight of each loss is determined based on the performance priority requirements of the multi-beam phased array; The total loss function is obtained by weighting and summing the sidelobe level loss, equivalent isotropic radiated power loss, and signal-to-interference ratio loss based on the weights.

9. The method of claim 7, wherein, The step of determining the calculation method based on the application scenario of the multi-beam phased array, and calculating the excitation matrix based on the calculation method to obtain the far-field radiation information of each beam includes: If the application scenario is an ideal multi-beam phased array scenario, the excitation matrix is ​​calculated based on the two-dimensional inverse Fourier transform to obtain the far-field radiation information of each beam. If the application scenario is a real engineering scenario, the far-field electric field component of each array unit is extracted from the full-wave simulation, and the far-field radiation information of each beam is obtained by electric field superposition and radiation intensity conversion.

10. A multi-beam synthesis apparatus based on a co-located null generating strategy, characterized by, The device includes a construction module, a training module, and a processing module; The construction module is used to determine the beam direction range of the multi-beam phased array, randomly generate design parameters within the beam direction range, define the main lobe center region, main lobe region, side lobe level, and null region of each beam based on the design parameters, and construct a mask matrix based on the main lobe center region, main lobe region, side lobe level, and null region of each beam, using the side lobe level as a boundary constraint; wherein, the null region of each beam is the set of the main lobe regions of other beams in the multi-beam array; The training module is used to construct a multi-beam synthesis model based on a residual network and a Transformer module. The mask matrix is ​​input into the multi-beam synthesis model as a training set, and the multi-beam synthesis model is trained to output the amplitude excitation and phase excitation of each beam. The residual network learns the local spatial features of the mask matrix and outputs a local feature map. The Transformer module learns the global spatial features of the mask matrix based on the local feature map and outputs a global feature map. The construction module is also used to construct a target mask matrix based on beam performance indicators in actual application scenarios; The processing module is used to input the target mask matrix into a trained multi-beam synthesis model. The trained multi-beam synthesis model outputs the amplitude excitation and phase excitation of each array element in the target mask matrix for different beams. The amplitude excitation and the phase excitation are loaded into the multi-beam phased array to generate a multi-beam far-field radiation pattern. Beam synthesis is then achieved based on the multi-beam far-field radiation pattern.

Images