Nonlinear beamforming device, method, medium and system based on gated cyclic unit

By combining a circular array and a GRU neural network into a nonlinear beamforming device, the poor performance of traditional beamforming in complex environments is solved, achieving high-resolution and robust signal processing effects.

CN123122609APending Publication Date: 2026-11-1710TH RES INST OF CETC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611613253.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Traditional beamforming techniques perform poorly under complex conditions such as non-stationary noise, small number of array elements, and model mismatch, while existing deep learning methods suffer from information loss and high computational complexity.

Method used

An end-to-end nonlinear beamforming system is constructed by using a gated cyclic unit-based nonlinear beamforming device, combined with a circular array structure, a differential beamforming module, a GRU neural network module, and an attention-weighted fusion module. By learning nonlinear mapping and dynamically adjusting weights through the GRU network, rapid suppression of nonstationary noise and wideband processing are achieved.

Benefits of technology

It achieves high-resolution beamforming with a small number of array elements, is robust to array element errors and complex environments, reduces online computational complexity, and improves signal quality and anti-interference capability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123122609A_ABST
    Figure CN123122609A_ABST
Patent Text Reader

Abstract

This invention discloses a nonlinear beamforming device, method, medium, and system based on a gated cyclic unit (GRU), belonging to the field of sensor array signal processing. The method includes the following steps: acquiring multi-channel time-domain signals received by each element of a uniform circular array; performing differential beamforming processing on the multi-channel time-domain signals to generate multi-channel differential feature signals containing spatial characteristic information; inputting the multi-channel differential feature signals into a pre-trained neural network (GRU) to output the features of the differential channels; outputting a set of attention weights from the differential channel features output by the GRU network, and fusing these weights with the multi-channel spatial features extracted by the differential beamforming module element-wise; extracting temporal information from the fused features and mapping it to complex beamforming weights; and using the complex beamforming weights to perform a weighted summation of the multi-channel time-domain signals to output the enhanced target signal. This invention effectively improves signal quality and anti-interference capability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of sensor array signal processing technology, and more specifically, to a nonlinear beamforming device, method, medium, and system based on a gated cyclic unit. Background Technology

[0002] Traditional beamforming techniques, such as Delay-and-Sum (DAS), Minimum Variance Distortionless Response (MVDR), and Generalized Sidelobe Canceller (GSC), can effectively enhance the target direction signal and suppress interference under ideal conditions. However, as application scenarios become increasingly complex, these traditional methods based on linear models and statistical assumptions have revealed many limitations.

[0003] (1) Insufficient noise suppression capability: Traditional beamformers mostly assume that noise is stationary or changes slowly, and its statistical characteristics (such as the covariance matrix) can be accurately estimated through long-term observation. However, in real-world environments, noise is often highly non-stationary. For example, sudden traffic noise, intermittent human voice interference, or rapidly changing industrial noise. The statistical characteristics of these non-stationary noises change drastically over time, making it difficult for traditional methods to update their model parameters in real time and accurately. Adaptive algorithms such as MVDR rely on accurate estimation of the noise covariance matrix. In non-stationary noise environments, the lag and inaccuracy of the estimation will severely weaken their interference suppression capability, and may even lead to beam pointing errors, aligning the null point with the target signal, causing signal distortion.

[0004] (2) Inconsistent performance between low and high frequencies: The performance of a beamformer is closely related to the signal frequency and the array aperture. For a fixed-size array, its beamwidth is inversely proportional to the signal wavelength. This means that in the low-frequency band, due to the longer wavelength, the main lobe of the beam will widen, resulting in a decrease in spatial resolution and a weakening ability to suppress interference. Conversely, in the high-frequency band, the main lobe of the beam narrows, the resolution increases, but it is more sensitive to array calibration errors and model mismatch, and is prone to performance instability. This inconsistency between low-frequency and high-frequency performance makes it difficult for traditional beamformers to maintain stable performance in broadband signal processing. For example, in speech signal processing, the beamforming effects of low-frequency components (such as vowels) and high-frequency components (such as consonants) differ significantly, which will affect the final speech quality and intelligibility.

[0005] (3) Strong dependence on the number of array elements: The performance of traditional high-resolution beamforming algorithms (such as MVDR and MUSIC) depends heavily on the number of array elements. Increasing the number of array elements can increase the aperture of the array, thereby obtaining a narrower main lobe and lower sidelobe level, improving spatial resolution and interference suppression capability. However, in practical applications, the number of array elements is often limited due to factors such as cost, size and power consumption. Under the condition of a small number of array elements, the performance of traditional algorithms will drop sharply, making it difficult to meet the needs of high-performance applications. For example, on portable devices or small drones, only a small number of microphones or antennas can be deployed, which makes it difficult for traditional methods to achieve effective beamforming in complex environments.

[0006] (4) Limitations of the Linear Model Assumption: Most traditional beamforming algorithms are based on the linear system assumption, i.e., the array output is a linear weighted sum of the signals received by each array element. This assumption simplifies the problem, allowing the algorithm to be implemented by solving linear equations or optimization problems. However, actual signal propagation and reception processes often involve nonlinear factors. For example, nonlinear distortion of power amplifiers, coherent signal superposition caused by multipath effects, and nonlinear response of the sensor itself. In these nonlinear scenarios, the linear model assumption no longer holds, and the performance of traditional linear beamformers is severely affected, failing to achieve the theoretically optimal effect. Deep learning technology, especially deep neural networks, provides a new approach to solving these problems due to their powerful nonlinear mapping capabilities.

[0007] In recent years, deep learning, especially deep neural networks (DNNs), has been widely used in beamforming and has shown the potential to surpass traditional methods. However, existing deep learning-based beamforming methods still have some shortcomings that limit further performance improvements.

[0008] (1) Limitations of the post-filter architecture: Many existing deep learning beamforming methods adopt a two-stage architecture of "beamforming + post-filtering". In this architecture, the first stage is usually a traditional beamformer (such as MVDR) for preliminary spatial filtering; the second stage is a deep learning model (such as DNN or CNN) as a post-filter to further suppress noise and enhance the signal of the beamformer output. Although this step-by-step processing method is effective, it still has problems such as information loss, non-optimal independent cascaded structure, and high computational complexity.

[0009] (2) Lack of a unified nonlinear beamforming framework: Although existing research has applied network structures such as DNN, CNN, and RNN to beamforming, most methods still treat deep learning models as a "black box" to replace a specific module in traditional algorithms (such as weight calculation or post-filtering). While these methods utilize the nonlinear fitting capabilities of neural networks, they fail to construct a unified, end-to-end nonlinear beamforming framework. For example, some methods use DNNs to fit the mapping relationship between the received signal covariance matrix and the optimal weight vector, but this is still within the framework of the traditional linear MVDR algorithm and has not completely escaped the constraints of linear models. Other methods, although proposing end-to-end structures, have relatively simple network designs and fail to fully utilize the complex temporal and spatial correlations of signals, especially when dealing with non-stationary noise, resulting in insufficient utilization of temporal information. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide a nonlinear beamforming device, method, medium and system based on a gated cyclic unit. It is applicable to engineering applications such as airborne / vehicle-mounted cockpit voice communication / interaction equipment, radio reconnaissance equipment, and underwater acoustic detection equipment that require high-precision beamforming in wide-bandwidth, strong interference and resource-constrained scenarios, and can effectively improve signal quality and anti-interference capability.

[0011] The objective of this invention is achieved through the following approach: A nonlinear beamforming device based on a gated cyclic unit includes: a circular array structure, a differential beamforming module, a GRU neural network module, an attention-weighted fusion module, and an output module; The circular array structure is composed of multiple omnidirectional array elements evenly distributed on a circle, and is used to acquire multi-channel time-domain signals. The differential beamforming module is used to receive multi-channel time-domain signals from the circular array and generate differential feature signals containing spatial feature information through differential operations. The GRU neural network module is used to learn a nonlinear mapping from differential feature signals to optimal beamforming weights through the GRU neural network. The attention-weighted fusion module is used to analyze the historical signal state and dynamically generate attention weights, which reflect the importance of each differential channel in suppressing noise and enhancing the target signal at the current moment. Then, the weights are fused with the spatial features extracted by the differential beamforming module element by element. The output module is used to extract temporal information from the fused features and map it into complex beamforming weights; the complex beamforming weights are used to perform weighted summation of the multi-channel time-domain signals to output the enhanced target signal.

[0012] Furthermore, the element spacing of the circular array structure is set to half a wavelength.

[0013] Furthermore, the differential operation specifically calculates the difference between the received signals of adjacent or symmetrical array elements, suppresses low-frequency noise and far-field coherent interference, and enhances the spatial characteristic information of the signal.

[0014] Furthermore, in the GRU neural network module, the nonlinear mapping from the differential signal to the optimal beamforming weights learned by the GRU neural network specifically includes: capturing long-term dependencies in the time series by introducing a gating mechanism; the GRU network takes the time-series differential signal output by the differential beamforming module as input, utilizes its time-series modeling capability to dynamically track the changing trend of non-stationary noise, and predicts the optimal complex weight vector in real time; these weights are then applied to the original array signal, and the final beamforming output is obtained through subsequent weighted summation.

[0015] Furthermore, the GRU neural network is trained in an end-to-end manner, and the loss function is directly related to the quality of the final output signal, thereby achieving globally optimal nonlinear beamforming.

[0016] A nonlinear beamforming method based on a gated cyclic unit, using a nonlinear beamforming device based on a gated cyclic unit as described in any of the preceding claims, includes the following steps: Acquire the multi-channel time-domain signals received by each element on a uniform circular array; Differential beamforming is performed on the multi-channel time-domain signal to generate a multi-channel differential feature signal containing spatial feature information; The multi-channel differential feature signal is input into a pre-trained neural network GRU, and the features of the differential channels are output. The differential channel features output by the GRU network output a set of attention weights, which reflect the importance of each differential channel in suppressing noise and enhancing the target signal at the current time. These weights are then fused element-wise with the multi-channel spatial features extracted by the differential beamforming module. Temporal information is extracted from the fused features and mapped to complex beamforming weights; the complex beamforming weights are then used to perform weighted summation of the multi-channel time-domain signals to output the enhanced target signal.

[0017] Furthermore, the differential beamforming process specifically involves zero-point constrained differential beamforming.

[0018] Furthermore, spatial feature information specifically includes spatial gradient information.

[0019] A computer-readable storage medium storing a computer program that, when loaded by a processor, executes the nonlinear beamforming method based on a gated cyclic unit as described above.

[0020] A nonlinear beamforming system based on a gated cyclic unit includes a nonlinear beamforming device based on a gated cyclic unit as described in any of the preceding claims.

[0021] The beneficial effects of this invention include: This invention integrates circular array differential beamforming technology with GRU deep neural networks to construct an end-to-end nonlinear beamformer, achieving significant beneficial effects: In terms of performance, it achieves rapid suppression of non-stationary noise and consistent wideband spatial processing, obtaining high resolution even with a small number of array elements; in terms of robustness, the system exhibits strong robustness and adaptability to array element errors and complex environmental changes; and in terms of system implementation, it bypasses the complex matrix inversion operations in traditional algorithms, reducing online computational complexity.

[0022] This invention provides a technical solution for achieving high-performance, high-reliability, and complex-environment-adaptive beamforming in complex real-world scenarios for airborne / vehicle-mounted cockpit voice communication / interaction devices, radio reconnaissance devices, underwater acoustic detection devices, and other similar applications. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 for Beamform at time; Figure 2 for Beamform at time; Figure 3 for Beamform at time; Figure 4 for Beamform at time; Figure 5 This is a flowchart illustrating the overall steps of the method in an embodiment of the present invention. Detailed Implementation

[0025] All features disclosed in all embodiments of this specification, or steps in all methods or processes implied in the disclosure, may be combined and / or extended or replaced in any way, except for mutually exclusive features and / or steps.

[0026] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0027] To address the poor performance of traditional beamforming methods under complex conditions such as non-stationary noise, limited array element count, and model mismatch, this invention proposes a nonlinear beamforming scheme based on a gated cyclic unit (GRU) circular array. At the front end, a circular array combined with differential beamforming achieves high directivity and frequency consistency, while simultaneously outputting a differential signal containing spatial information as a feature. At the back end, a GRU-gated cyclic unit neural network receives the features and performs timing modeling and nonlinear mapping processing, ultimately outputting the optimal beamforming weights and the enhanced signal. This invention achieves front-end and back-end fusion and global optimization, directly mapping the original multi-channel signal to the output signal.

[0028] Specifically, as a first aspect of the present invention, a nonlinear beamforming device is provided, which integrates differential beamforming technology of a ring array with a gated recurrent unit (GRU) deep neural network to construct a unified, end-to-end nonlinear beamformer. The device comprises four core modules: a ring array structure, a differential beamforming module, a GRU neural network module, an attention-weighted fusion module, and an output module.

[0029] (1) Circular Array Structure: A circular array consisting of M omnidirectional array elements uniformly distributed on a circle is adopted. Compared with linear arrays, circular arrays have 360-degree omnidirectional coverage in the horizontal plane, and their beam pattern has rotational symmetry in the principal axis direction, which gives them a natural advantage in target tracking and omnidirectional scanning applications. The element spacing is usually designed to be half a wavelength to avoid the appearance of grating lobes. The geometry of the circular array facilitates subsequent differential operations, because array elements at any two symmetrical positions can easily form differential pairs to extract the spatial gradient information of the signal.

[0030] (2) Differential Beamforming Module: This module is the spatial information preprocessing unit of the system. It receives M raw signals from the circular array and generates a series of higher-order differential signals through differential operations. The essence of differential operations is to calculate the difference between the received signals of adjacent or symmetrical array elements. This is equivalent to a spatial high-pass filter, which can effectively suppress low-frequency noise and far-field coherent interference, while enhancing the spatial gradient information of the signal. For example, the first-order difference can be represented as the difference between the signals of adjacent array elements, while higher-order differences are the differential operations performed again on the lower-order differential signals. These differential signals not only contain the amplitude and phase information of the original signals, but more importantly, they encode the spatial location information of the signal source, providing more discriminative feature inputs for the subsequent neural network.

[0031] (3) GRU Neural Network Module: This module is the core decision-making unit of the system, responsible for learning the nonlinear mapping from the differential signal to the optimal beamforming weights. Unlike traditional DNNs or CNNs, this invention uses a GRU network. GRU is a special type of recurrent neural network that effectively captures long-term dependencies in time series by introducing gating mechanisms (update gate and reset gate). In this system, the GRU network takes the time-series differential signal output by the differential beamforming module as input, utilizes its powerful time-series modeling capabilities to dynamically track the changing trends of non-stationary noise, and predicts the optimal complex weight vector in real time. These weights are then applied to the original array signal, and the final beamforming output is obtained through subsequent weighted summation. The entire network is trained end-to-end, and the loss function is directly related to the final output signal quality (such as signal-to-noise ratio and speech quality score), thereby achieving globally optimal nonlinear beamforming.

[0032] (4) Attention-weighted fusion module: This module analyzes the historical signal state through a GRU network and dynamically generates a set of attention weights. These weights reflect the importance of each differential channel in suppressing noise and enhancing the target signal at the current moment. Subsequently, these weights are fused element-wise with the multi-channel spatial features extracted by the differential beamforming module. Its core function is to achieve deep coupling between differential processing and neural networks, enabling the system to adaptively adjust the spatial filtering strategy according to the real-time acoustic environment, thereby synergistically optimizing noise suppression and signal fidelity.

[0033] (5) Output module: extracts time-series information from the fused features and maps it to complex beamforming weights; uses the complex beamforming weights to perform weighted summation on the multi-channel time-domain signals, thereby outputting the enhanced target signal.

[0034] In a more specific implementation, based on the device described in any of the above embodiments, in the annular array structure, it is further assumed that the uniform circular array has Each array element, at time The received signal vector is: ; Typically, the number of array elements is taken. , , [.] T C represents the transpose of a vector. M Let M represent the M-dimensional complex space.

[0035] In the differential beamforming module, a differentiable differential preprocessing layer is further designed, that is, a layer containing... Learnable difference layers for different spatial difference modes, output Individual differential feature channels: ; in, for Features of each differential channel The number of differential feature channels in the design; To form a weight matrix for trainable complex differential beams, The initial values ​​can be designed using traditional zero-point constraints, but they will be jointly optimized with subsequent networks during training. It is the bias vector. Each row can be viewed as a differential beamformer with a specific zero-point mode, but unlike traditional fixed designs, its parameters can be optimized through data learning. Typically, the number of differential modes is taken. , , , .

[0036] In the GRU neural network module, the following GRU encoding network is further designed: [The network incorporates] differential features. Input the GRU coding network and update the hidden state: ; in, Indicates GRU network processing, express The real part, express The imaginary part, This indicates the transpose operation of a vector. [...] T This represents the transpose of a vector. For the GRU encoder at time The hidden state, For the GRU encoder at time The hidden state, The hidden layer dimension encodes the historical information of the acoustic environment up to the current moment; These are the parameters for the GRU encoding network. Typically, they are... The real and imaginary parts are concatenated to form a 6-dimensional real vector, which is then input into the GRU encoding network, and the hidden layer dimension is taken. Then the GRU encoder outputs the hidden state. .

[0037] In the attention-weighted fusion module, the attention weight generation process is further as follows: based on the current GRU hidden state... Generate through an attention generation network dimensional attention weight vector : ; in: , , All of these are trainable parameters. Let R be the attention parameter dimension. N for N 3D real space, It can be guaranteed Weight Indicates at time No. The importance of each differential feature channel to the current environment is considered. When a sudden disturbance occurs in a certain direction, the attention mechanism automatically enhances the suppression of the differential feature channel in that direction. Typically, [the following is taken as an example]. , , , .

[0038] Furthermore, the detailed process of attention-weighted fusion is as follows: Attention weights are applied to the difference features to obtain the fused feature vector: ; in, This represents element-wise multiplication, a process that can adjust the contribution of each difference channel in real time according to the environment. Typically, when and hour, .

[0039] Furthermore, the detailed process of generating beamforming weights for the GRU decoding network is as follows: The fused features... The input is fed into the GRU decoding network to further extract temporal information, which is then finally mapped to real-valued beamforming weights. ; ; in, , All of these are training parameters. To output complex weights, These are the parameters for the GRU decoding network. for t Real-valued beamforming weights at time -1. Typically, take... ,when hour, , , , .

[0040] During the beamforming output process, time-varying weights generated by the GRU are used to beamform the original array signal, resulting in an enhanced signal: ; in, Weights are assigned to the recombined complex beams. Represents the imaginary unit. This is for the final beam output. Typically, when hour, .

[0041] Beamforming is performed according to the above process, assuming... , The radius of the circular array is wave speed is Then the arc length between adjacent array elements Assuming the ratio of arc length to wavelength is Then the incident wavelength Incident frequency . The range of values ​​is generally as follows: Typically, for radio reconnaissance scenarios, a typical set of parameters is taken as follows: , If it is 0.5, then , , For airborne / vehicle-mounted voice communication scenarios, a typical set of parameters is as follows: , If it is 0.18, then , , The following airborne / vehicle-mounted voice communication scenarios are used as examples to simulate different scenarios. The beam patterns obtained in the five cases are as follows: Figure 1 , Figure 2 , Figure 3 and Figure 4 As shown, given the radius of the circular array, incident waves of different wavelengths exhibit similar beam responses. Tests were conducted using audio mixed with ordinary male speech and throat singing as the input raw speech.

[0042] To further clarify, end-to-end training and loss function: the trainable parameter set of the entire system is... End-to-end training is employed, minimizing the following loss function: ; in, It is the loss function of the main task. It is an optional regularization term. It is the first t A clean target signal at all times. The number of training samples, It is the regularization coefficient.

[0043] As a second aspect of the present invention, in the interdisciplinary field of signal processing and artificial intelligence, a nonlinear adaptive beamforming method is provided, particularly suitable for nonlinear adaptive beamforming methods for acoustic and radio sensor arrays. This technical solution can solve common bottleneck problems in sensor array signal processing, such as low-frequency-high-frequency performance imbalance, poor non-stationary noise suppression, and poor robustness in complex scenarios. The inventive concept deeply integrates spatial differential beamforming of a Uniform Circular Array (CUA) with a Gated Recurrent Unit (GRU) deep neural network. By constructing an end-to-end deep learning framework, the multi-channel signals received by the array are directly mapped to optimal beamforming weights, thereby achieving effective enhancement of the target signal and precise suppression of interference / noise. Specifically, based on the device described in any of the above embodiments, such as... Figure 5 As shown, the steps include: Step (1): Obtain the multi-channel time-domain signal received by each element on the uniform circular array.

[0044] Step (2) Perform zero-point constrained differential beamforming processing on the multi-channel time-domain signal to generate a multi-channel differential feature signal containing spatial gradient information.

[0045] Step (3) inputs the multi-channel differential feature signal into the pre-trained gated recurrent unit network (GRU) and outputs the features of the differential channels.

[0046] Step (4) outputs a set of attention weights based on the differential channel features output by the GRU network. These weights reflect the importance of each differential channel in suppressing noise and enhancing the target signal at the current moment. These weights are then fused element-wise with the multi-channel spatial features extracted by the differential beamforming module. The fused features are then input into the GRU decoding network to further extract temporal information and finally mapped to complex beamforming weights.

[0047] Step (5): Use the complex beamforming weights to perform weighted summation on the multi-channel time-domain signals and output the enhanced target signal.

[0048] It should be noted that, within the scope of protection defined in the claims of this invention, the following embodiments can be combined and / or extended or replaced in any logical manner from the above specific embodiments, such as the disclosed technical principles, disclosed technical features or implicitly disclosed technical features.

[0049] Example 1 A nonlinear beamforming device based on a gated cyclic unit includes: a circular array structure, a differential beamforming module, a GRU neural network module, an attention-weighted fusion module, and an output module; The circular array structure is composed of multiple omnidirectional array elements evenly distributed on a circle, and is used to acquire multi-channel time-domain signals. The differential beamforming module is used to receive multi-channel time-domain signals from the circular array and generate differential feature signals containing spatial feature information through differential operations. The GRU neural network module is used to learn a nonlinear mapping from differential feature signals to optimal beamforming weights through the GRU neural network. The attention-weighted fusion module is used to analyze the historical signal state and dynamically generate attention weights, which reflect the importance of each differential channel in suppressing noise and enhancing the target signal at the current moment. Then, the weights are fused with the spatial features extracted by the differential beamforming module element by element. The output module is used to extract temporal information from the fused features and map it into complex beamforming weights; the complex beamforming weights are used to perform weighted summation of the multi-channel time-domain signals to output the enhanced target signal.

[0050] Example 2 Based on Example 1, the element spacing of the circular array structure is set to half a wavelength.

[0051] Example 3 Based on Example 1, the differential operation specifically calculates the difference between the received signals of adjacent or symmetrical array elements, suppresses low-frequency noise and far-field coherent interference, and enhances the spatial characteristic information of the signal.

[0052] Example 4 Based on Example 1, in the GRU neural network module, the nonlinear mapping from differential feature signals to optimal beamforming weights learned by the GRU neural network specifically includes: capturing long-term dependencies in the time series by introducing a gating mechanism; the GRU network takes the time-series differential signal output by the differential beamforming module as input, utilizes its time-series modeling capability to dynamically track the changing trend of non-stationary noise, and predicts the optimal complex weight vector in real time; these weights are then applied to the original array signal, and the final beamforming output is obtained through subsequent weighted summation.

[0053] Example 5 Based on Example 4, the GRU neural network is trained in an end-to-end manner, and the loss function is directly related to the quality of the final output signal, thereby achieving globally optimal nonlinear beamforming.

[0054] Example 6 A nonlinear beamforming method based on a gated cyclic unit, using any one of Embodiments 1 to 5, includes the following steps: Acquire the multi-channel time-domain signals received by each element on a uniform circular array; Differential beamforming is performed on the multi-channel time-domain signal to generate a multi-channel differential feature signal containing spatial feature information; The multi-channel differential feature signal is input into a pre-trained neural network GRU, and the features of the differential channels are output. The differential channel features output by the GRU network output a set of attention weights, which reflect the importance of each differential channel in suppressing noise and enhancing the target signal at the current time. These weights are then fused element-wise with the multi-channel spatial features extracted by the differential beamforming module. Temporal information is extracted from the fused features and mapped to complex beamforming weights; the complex beamforming weights are then used to perform weighted summation of the multi-channel time-domain signals to output the enhanced target signal.

[0055] Example 7 Based on Example 6, the differential beamforming process is specifically a zero-point constrained differential beamforming process.

[0056] Example 8 Based on Example 6, the spatial feature information specifically includes spatial gradient information.

[0057] Example 9 A computer-readable storage medium storing a computer program that, when loaded by a processor, executes the nonlinear beamforming method based on a gated cyclic unit as described in Embodiment 6.

[0058] Example 10 A nonlinear beamforming system based on a gated cyclic unit includes the nonlinear beamforming device based on a gated cyclic unit as described in any one of Embodiments 1 to 5.

[0059] The units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0060] According to one aspect of the present invention, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above.

[0061] In another aspect, embodiments of the present invention also provide a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods described in the above embodiments.

Claims

1. A nonlinear beamforming device based on a gated cyclic unit, characterized in that, include: The system comprises a circular array structure, a differential beamforming module, a GRU neural network module, an attention-weighted fusion module, and an output module. The circular array structure is composed of multiple omnidirectional array elements evenly distributed on a circle, and is used to acquire multi-channel time-domain signals. The differential beamforming module is used to receive multi-channel time-domain signals from the circular array and generate differential feature signals containing spatial feature information through differential operations. The GRU neural network module is used to learn a nonlinear mapping from differential feature signals to optimal beamforming weights through the GRU neural network. The attention-weighted fusion module is used to analyze the historical signal state and dynamically generate attention weights, which reflect the importance of each differential channel in suppressing noise and enhancing the target signal at the current moment. The weights and spatial features extracted by the differential beamforming module are then fused element-wise. The output module is used to extract temporal information from the fused features and map it into complex beamforming weights; the complex beamforming weights are used to perform weighted summation of the multi-channel time-domain signals to output the enhanced target signal.

2. The nonlinear beamforming device based on a gated cyclic unit according to claim 1, characterized in that, The element spacing of the circular array structure is set to half a wavelength.

3. The nonlinear beamforming device based on a gated cyclic unit according to claim 1, characterized in that, The differential operation specifically calculates the difference between the received signals of adjacent or symmetrical array elements, suppresses low-frequency noise and far-field coherent interference, and enhances the spatial characteristic information of the signal.

4. The nonlinear beamforming device based on a gated cyclic unit according to claim 1, characterized in that, In the GRU neural network module, the nonlinear mapping from differential feature signals to optimal beamforming weights learned by the GRU neural network specifically includes: capturing long-term dependencies in the time series by introducing a gating mechanism; the GRU network takes the time-series differential signal output by the differential beamforming module as input, utilizes its time-series modeling capability to dynamically track the changing trend of non-stationary noise, and predicts the optimal complex weight vector in real time; these weights are then applied to the original array signal, and the final beamforming output is obtained through subsequent weighted summation.

5. The nonlinear beamforming device based on a gated cyclic unit according to claim 4, characterized in that, The GRU neural network is trained in an end-to-end manner, and the loss function is directly related to the quality of the final output signal, thereby achieving globally optimal nonlinear beamforming.

6. A nonlinear beamforming method based on gated cyclic units, characterized in that, The nonlinear beamforming device based on a gated cyclic unit according to any one of claims 1 to 5 includes the following steps: Acquire the multi-channel time-domain signals received by each element on a uniform circular array; Differential beamforming is performed on the multi-channel time-domain signal to generate a multi-channel differential feature signal containing spatial feature information; The multi-channel differential feature signal is input into a pre-trained neural network GRU, and the features of the differential channels are output. The differential channel features output by the GRU network output a set of attention weights, which reflect the importance of each differential channel in suppressing noise and enhancing the target signal at the current time. These weights are then fused element-wise with the multi-channel spatial features extracted by the differential beamforming module. Temporal information is extracted from the fused features and mapped to complex beamforming weights; the complex beamforming weights are then used to perform weighted summation of the multi-channel time-domain signals to output the enhanced target signal.

7. The nonlinear beamforming method based on a gated cyclic unit according to claim 6, characterized in that, The differential beamforming process specifically involves zero-point constrained differential beamforming.

8. The nonlinear beamforming method based on a gated cyclic unit according to claim 6, characterized in that, Spatial feature information specifically includes spatial gradient information.

9. A computer-readable storage medium, characterized in that, A computer program is stored in a readable storage medium, which, when loaded by a processor, executes the nonlinear beamforming method based on a gated cyclic unit as described in claim 6.

10. A nonlinear beamforming system based on a gated cyclic unit, characterized in that, The nonlinear beamforming device based on a gated cyclic unit, as described in any one of claims 1 to 5, is included.