A physical information embedded complex domain attention network DOA estimation method and system
By embedding physical information into a complex domain attention network, and utilizing augmented complex tensors and self-attention mechanisms, the accuracy degradation of traditional DOA estimation algorithms under multipath conditions and the localization error of deep learning models are solved, achieving high-precision and low-complexity DOA estimation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TAIZHOU UNIV
- Filing Date
- 2026-04-10
- Publication Date
- 2026-08-04
AI Technical Summary
In complex wireless propagation scenarios, traditional DOA estimation algorithms suffer from decreased accuracy under strong multipath conditions, and deep learning models lack physical constraints, leading to positioning errors and high computational complexity, which cannot meet real-time requirements.
A complex domain attention network with embedded physical information is adopted. By augmenting the complex tensor input, complex convolution operation and self-attention mechanism, combined with the physical constraint of the steering vector, multipath suppression and direct wave feature extraction are achieved. Sparse soft thresholding is used for DOA estimation.
Achieving high-precision and robust DOA estimation in complex environments reduces computational complexity, meets real-time processing requirements, and improves angular resolution and positioning accuracy.
Smart Images
Figure CN122508024A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary fields of wireless communication, radar detection, and array signal processing, and specifically to a method and system for DOA estimation in complex domain attention networks with embedded physical information. Background Technology
[0002] Direction of Arrival (DOA) estimation is a core technology in array signal processing, wireless communication, and radar detection. Its goal is to quickly and accurately calculate the angle of incidence of a signal received by an array antenna. In complex wireless propagation scenarios such as urban canyons, low-altitude flight, and vehicle-mounted detection, signals are reflected multiple times by buildings, the ground, and obstacles, resulting in significant multipath effects. The receiver simultaneously acquires direct waves and a large number of coherent reflected waves. Signal coherence and environmental noise pose severe challenges to DOA estimation.
[0003] Traditional DOA estimation algorithms, represented by subspace methods such as MUSIC and ESPRIT, heavily rely on accurate estimation of signal statistical properties and ideal calibration of array manifolds. Under strong multipath conditions, the direct and reflected waves are highly coherent, directly causing a rank deficit in the signal covariance matrix, leading to a sharp decline in the estimation accuracy of subspace algorithms or even complete failure. Although spatial smoothing techniques can alleviate the coherent signal problem to some extent, they sacrifice the effective aperture of the array, significantly reducing angular resolution. Furthermore, traditional algorithms are sensitive to noise distribution; the covariance matrix estimation error increases dramatically at low signal-to-noise ratios, and they require eigenvalue decomposition and two-dimensional spectral peak search, resulting in high computational complexity and long processing times, failing to meet real-time engineering requirements.
[0004] In recent years, data-driven deep learning has provided new approaches to DOA estimation, but existing convolutional neural network and deep neural network models have significant shortcomings. Most models directly decompose the complex covariance matrix into real and imaginary inputs, disrupting the inherent amplitude-phase coupling physical properties of the complex domain and making it difficult to learn the fine spatial structure formed by phase interference. Purely data-driven "black box" networks lack prior physical constraints on array signals, making them prone to overfitting to specific multipath distributions in the training set. When the multipath statistical characteristics of the testing environment change, their generalization ability rapidly declines. More critically, in strongly reflective scenarios, reflected wave energy is often higher than that of the direct wave. Existing algorithms lack a direct wave feature identification mechanism, generally identifying the strongest spectral peak as the signal source direction, leading to serious localization errors. Summary of the Invention
[0005] The purpose of this invention is to propose a complex domain attention network DOA estimation method and system with embedded physical information, so as to achieve high-precision and robust DOA estimation of direct waves in complex environments.
[0006] According to a first aspect of the present disclosure, a method for estimating the Data Address Allocation (DOA) of a complex domain attention network with embedded physical information is provided, comprising the following steps: The far-field narrowband signal received by the array is preprocessed to obtain the sample covariance matrix and bispectral slice. The real and imaginary parts of the covariance matrix and the real and imaginary parts of the bispectral slice are stacked along the channel dimension to form an augmented complex tensor. The augmented complex tensor is input into the complex domain feature extraction module, and feature extraction is completed through complex convolution operation and modulus-preserving complex activation function, preserving the amplitude-phase coupling characteristics of the signal. The extracted complex features are input into the multipath perception spatial attention module, and the original attention weights are obtained based on the self-attention mechanism. The multipath suppression matrix is generated by combining the physical constraint mask of the guide vector, and adaptive suppression is performed on non-physically related regions. The attention-weighted features are mapped to an angular grid to obtain the initial spatial spectrum, which is then sparsified by sparse soft thresholding. The confidence scores of direct waves at each angle are obtained based on energy and spatial continuity. The signals are sorted by confidence score and the final DOA estimation results are output according to the preset number of sources.
[0007] In one embodiment, the sample covariance matrix is represented as: in, This indicates the conjugate transpose. Represents the array received signal vector. Indicates the number of snapshots; The bispectral slice is obtained by performing a two-dimensional discrete Fourier transform on the third-order cumulant of the array signal, and is used to extract the high-order phase coupling features of the signal; the third-order cumulant of the array signal is defined as: in, and The amount of time delay, For mathematical expectation; Performing a two-dimensional discrete Fourier transform on the third-order cumulants yields a bispectrum: Using finite-delay snapshots for solving, the bispectral density is simplified to a discrete form: in, The signal time-domain vector The Fourier transform result, To indicate complex conjugate, and It represents two-dimensional frequency components.
[0008] Therefore, the augmented complex tensor is: in, This indicates stacking along the channel dimension. To take the real part of a complex number, This indicates taking the imaginary part of a complex number.
[0009] In one embodiment, the complex convolution operation is implemented as follows: for the input complex feature map... and complex convolution kernel Its output is defined as: in, , These represent the real and imaginary parts of the input feature map, respectively. , These are the real and imaginary parts of the convolution kernel, respectively. The imaginary unit, To augment complex tensors, This is a bias term.
[0010] The modulus-preserving complex activation function is: in, The output feature map of the complex convolution operation. It is a linear rectified activation function. This is a bias term.
[0011] In one embodiment, a multipath suppression matrix is generated by combining the guide vector physical constraint mask, specifically as follows: Based on the guiding vector, any two array elements can be obtained. The theoretical coherence coefficient between them is: in, Let m be the component of the array guiding vector corresponding to the array element m. The component corresponding to array element n in the array steering vector; Iterate through all possible incident angles and count all array element pairs. Corresponding coherence coefficient The template for the global theoretical coherence coefficient matrix is obtained as follows: in, The total number of angle samples. It is a set containing all the sampled angles that have been traversed; Let the coherence threshold be... Generate physical constraint mask for: in, The region marked as non-physically relevant corresponds to spurious coherence such as multipath reflection; The region marked as physically feasible relevance corresponds to the actual relevance of the direct wave.
[0012] Further generate the multipath suppression matrix : in, These are the angle-dependent suppression weights.
[0013] In one embodiment, the attention weight matrix of the multipath sensing spatial attention module for: in, It represents the Hadamah accumulation. The learnable inhibition strength coefficient; For querying the matrix, The key matrix; Key matrix Dimensions This is a multipath suppression matrix; based on this, adaptive suppression is performed on non-physically related regions to achieve directional attenuation.
[0014] In one embodiment, the attention-weighted features are: Will Mapping to an angular grid yields the initial spatial spectrum. : in, It is a non-linear activation function. Indicates will Flattened into a one-dimensional vector, The weight matrix is a learnable angle mapping. For bias terms; For the initial spatial spectrum Soft thresholding is performed to obtain the energy values of the sparsified spatial spectrum: in, The learnable sparse threshold.
[0015] In one embodiment, the direct wave confidence score is obtained as follows: in, For angle The corresponding neighborhood region in the angle grid This represents the number of angular grid cells within this neighborhood. for The The diagonal elements represent the concentration of spatial correlation between array elements.
[0016] According to a second aspect of the present disclosure, a complex domain attention network DOA estimation system with embedded physical information is provided, comprising: The data preprocessing module preprocesses the far-field narrowband signal received by the array to obtain the sample covariance matrix and bispectral slice. The real and imaginary parts of the covariance matrix and the real and imaginary parts of the bispectral slice are stacked along the channel dimension to form an augmented complex tensor. The complex domain feature extraction module takes the augmented complex tensor as input, and completes feature extraction through complex convolution operation and modulus-preserving complex activation function, thus preserving the amplitude-phase coupling characteristics of the signal. The multipath sensing spatial attention module inputs the extracted complex features into the multipath sensing spatial attention module, obtains the original attention weights based on the self-attention mechanism, and generates a multipath suppression matrix by combining the physical constraint mask of the guide vector, and performs adaptive suppression on non-physically related regions. The sparse spectrum reconstruction module maps the attention-weighted features to an angular grid to obtain the initial spatial spectrum, and then achieves spectral sparsity through sparse soft thresholding. The DOA estimation and output module obtains the confidence scores of direct waves at various angles based on energy and spatial continuity, sorts them by confidence score, and outputs the final DOA estimation results according to the preset number of sources.
[0017] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the aforementioned physical information embedding complex domain attention network DOA estimation method.
[0018] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the aforementioned method for estimating the Data of Aspects (DOA) of a complex domain attention network with embedded physical information.
[0019] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: 1. By fusing the sample covariance matrix with the augmented complex tensor input of bispectral slices, and combining second-order statistics with high-order phase coupling features, Gaussian noise is effectively suppressed and coherent multipath signals are distinguished. This solves the performance failure problem caused by the rank deficiency of the covariance matrix in traditional subspace algorithms, avoids the array aperture loss of spatial smoothing technology, and maintains high angular resolution in strong multipath scenarios.
[0020] 2. By employing convolution operations that follow the rules of complex algebra and a modulus-preserving activation function, only the modulus is adjusted and the phase information is strictly preserved, avoiding phase distortion caused by the separation of real and imaginary parts. The fine structure of phase interference is learned, which greatly improves the feature extraction accuracy and estimation robustness under low signal-to-noise ratio.
[0021] 3. By introducing a physical constraint mask for the guide vector in the multipath perception spatial attention module, the propagation law of the array signal is incorporated into the model, breaking the limitations of the pure data-driven "black box", suppressing overfitting, maintaining stable performance even when the multipath characteristics of the test environment change, and the generalization ability is significantly better than existing deep learning methods.
[0022] 4. A direct wave confidence scoring mechanism that integrates spectral energy and spatial autocorrelation is proposed. High confidence is assigned only to angles with high energy and high spatial continuity, effectively distinguishing direct waves from multipath reflected waves, completely solving the problem of misjudgment of reflected waves in strong reflection scenarios, and significantly improving the positioning accuracy in complex environments.
[0023] 5. It eliminates the need for time-consuming feature decomposition and spectral peak search in traditional algorithms. The end-to-end network architecture has low computational complexity and fast inference speed, meeting the real-time processing requirements of scenarios such as wireless communication base stations and vehicle radar, and is suitable for complex environments in multiple fields. Attached Figure Description
[0024] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.
[0025] Figure 1 A framework diagram of a complex domain attention network for DOA estimation method embedded with physical information; Figure 2 This is a schematic diagram showing the DOA estimation results under different signal-to-noise ratios. Detailed Implementation
[0026] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.
[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0028] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0029] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.
[0030] Example 1: like Figure 1 As shown, this embodiment provides a complex domain attention network DOA estimation method with embedded physical information. The core idea is to construct an augmented complex signal model, design a true complex domain convolutional neural network, introduce a multipath-aware spatial attention mechanism, and train it using a hybrid physical constraint loss function. This achieves adaptive filtering of multipath interference and high-precision localization of direct waves. Specifically, it includes the following steps: S1. Preprocess the far-field narrowband signal received by the array to obtain the sample covariance matrix and bispectral slice. Stack the real and imaginary parts of the covariance matrix and the real and imaginary parts of the bispectral slice along the channel dimension to form an augmented complex tensor. Specifically, assuming a uniform linear array contains Each array element receives A far-field narrowband signal, in The array received signal vector at time t is represented as: (1) in, Complex gain including path loss and reflection coefficient, For corresponding angles The guiding vector, As the source signal, It is additive white Gaussian noise.
[0031] The sample covariance matrix can be expressed as: (2) in, This indicates the conjugate transpose.
[0032] Bispectral slicing involves truncating the bispectrum at a specific frequency dimension, removing redundant frequency information, and retaining core phase coupling features. This is used to capture the nonlinear phase correlation between multipath signals, providing the network with high-order features to distinguish between direct waves and multipath interference, thus improving model performance in complex interference scenarios. The foundation of bispectrum slicing is the third-order cumulant, which is defined as follows: (3) in, and The amount of time delay, Equation (3) represents the third-order correlation characteristics of the signal. The third-order cumulative quantity of Gaussian noise is always zero, which is the core reason for bispectral anti-Gaussian interference.
[0033] By performing a two-dimensional discrete Fourier transform on the third-order cumulant, the bispectrum can be obtained: (4) In practical engineering, finite-delay snapshots are used to solve the problem, which simplifies it to a discrete form: (5) in, The signal time-domain vector The Fourier transform result, To indicate complex conjugate, and It represents two-dimensional frequency components.
[0034] Unlike traditional methods that only use the sample covariance matrix, this invention constructs an augmented complex tensor as network input, which not only includes second-order statistics but also preserves higher-order phase information. Specifically, the input tensor is composed of the real and imaginary parts of the sample covariance matrix and the real and imaginary parts of a selected bispectral slice: (6) in, This indicates stacking along the channel dimension. To take the real part of a complex number, This represents taking the imaginary part of a complex number. This preprocessing method utilizes the ability of bispectral capture of non-Gaussian signal characteristics and phase coupling information, significantly enhancing the resolution of coherent multipath signals.
[0035] S2. Input the augmented complex tensor into the complex domain feature extraction module, and complete the feature extraction through complex convolution operation and the modulus-preserving complex activation function, preserving the amplitude-phase coupling characteristics of the signal; The complex domain feature extraction module employs complex convolution operations for the input complex feature map. and complex convolution kernel Its output is defined as: (7) in, and They represent the real part and the imaginary part, respectively. The term represents the bias term. It can be seen that the operation of equation (7) follows the rules of complex algebra, avoiding the information fragmentation caused by the separation of real and imaginary parts.
[0036] The activation function is a modulus-preserving complex activation function: (8) This operation only adjusts the modulus of the complex number while strictly preserving its phase information, effectively avoiding the phase distortion caused by traditional ReLU.
[0037] S3. Input the extracted complex features into the multipath perception spatial attention module, obtain the original attention weights based on the self-attention mechanism, and generate a multipath suppression matrix by combining the physical constraint mask of the guide vector to perform adaptive suppression on non-physically related regions; The Multipath Aware Spatial Attention Module (MASA) is based on a self-attention mechanism and integrates prior physical knowledge of array signal processing to achieve adaptive learning of spatial correlation between array elements and precise suppression of multipath interference.
[0038] The core design logic of the multipath sensing spatial attention module is as follows: based on the self-attention mechanism to calculate the spatial correlation between array elements, a physical constraint, a steering vector prior mask, is introduced to adaptively suppress non-physically strongly correlated regions caused by multipath interference, retaining only the true spatial correlation features corresponding to the direct wave. The specific implementation steps are as follows: (1) The input feature map is obtained through tensor dimension adaptation. Furthermore, the query matrix required for the self-attention mechanism is obtained through feature mapping. Key matrix and self-attention value matrix : (9) in, , and The complex weight matrix is a learnable matrix. , and This is a bias term.
[0039] The original attention weight matrix is obtained according to the basic rules of self-attention: (10) in, Key matrix The dimension is used to scale the matrix inner product result. The Softmax function is used to normalize the weight values to the [0,1] interval, realizing the probabilistic representation of the correlation.
[0040] (2) Introduce physical constraint mask Constructing a multipath suppression matrix Calculate any two array elements based on the guiding vector. The theoretical coherence coefficient between them is: (11) This coefficient represents the angle. Below, array element and The theoretical correlation strength when receiving direct waves, with a value range of [0,1].
[0041] Iterate through all possible incident angles and count all array element pairs. Corresponding coherence coefficient The template for the global theoretical coherence coefficient matrix is obtained as follows: (12) in, The total number of angle samples. It is a set containing all the sampled angles that were traversed.
[0042] Let the coherence threshold be... It can generate physical constraint masks. for: (13) in, The region marked as non-physically relevant corresponds to spurious coherence such as multipath reflection; The region marked as physically feasible relevance corresponds to the actual relevance of the direct wave.
[0043] Further generate the multipath suppression matrix : (14) in, The angle-dependent suppression weights can be adaptively learned from the training data.
[0044] By performing a Hadamard product operation between the original attention weights and the multipath suppression matrix, targeted attenuation of non-physically relevant regions can be achieved. (15) in, It represents the Hadamah accumulation. The suppression strength coefficient is a learnable parameter. This mechanism can adaptively learn the spatial correlation between array elements and assign low weights to non-physical correlations generated by multipath interference, thereby effectively filtering out multipath interference.
[0045] S4. Map the attention-weighted features to an angular grid to obtain the initial spatial spectrum, and then perform spectral sparsification through sparse soft thresholding. To address the technical problem of misidentifying reflected waves as signal source directions due to their higher energy levels compared to direct waves in multipath strong reflection scenarios, this invention proposes a direct wave confidence discrimination strategy based on a joint energy-spatial continuity approach. The final attention-weighted feature from the MASA module can be expressed as: (16) Will Mapping onto a preset angle grid yields the initial spatial spectrum. : (17) in, Indicates will Flattened into a one-dimensional vector, As a learnable angle mapping weight matrix, it learns a nonlinear mapping from spatial features to angles. This is a bias term.
[0046] To achieve spectral sparsity, a learnable sparsity threshold is introduced for the initial spectrum. After applying a soft threshold, we can obtain: (18) in, As a learnable sparse threshold, it is optimized through backpropagation to automatically adapt to different signal-to-noise ratio scenarios. It can be seen that Equation (18) only retains intensities exceeding the threshold. By setting the spectral peaks to zero and eliminating noise and multipath pseudo-peaks, a sparse spatial spectrum can be achieved.
[0047] use Combination For each angle to be detected in the angle grid The confidence score for direct waves can be defined as: (19) in, For angle The corresponding neighborhood region in the angle grid This represents the number of angular grid cells within this neighborhood. for The The diagonal elements represent the concentration of spatial correlation between array elements. The higher the value of the diagonal element, the more stable the spatial correlation structure of the signal corresponding to that angle.
[0048] S5. Based on energy and spatial continuity, obtain the confidence scores of direct waves at each angle, sort them by confidence, and output the final DOA estimation results according to the preset number of sources.
[0049] The direct wave confidence score of each candidate angle is obtained by equation (19). After sorting by confidence in descending order and filtering by the preset number of information sources, the accurate and stable DOA estimation result is finally obtained.
[0050] To verify the effectiveness of this invention, simulation experiments were conducted under various complex scenarios. The experimental setup employed a 10-element uniform linear array with an element spacing of half a wavelength and a working carrier frequency of 700MHz. The dataset was constructed using the Saleh-Valenzuela channel model to simulate complex multipath environments, where each direct wave is accompanied by 1 to 5 reflection paths. The reflection path angles follow a Gaussian distribution centered on the direct wave angle, and the reflection coefficients follow a Rayleigh distribution. The training data covered a signal-to-noise ratio range of -5dB to 25dB, and the comparison algorithms included the MUSIC algorithm, the ESPRIT algorithm, and a deep learning algorithm based on real-valued CNNs.
[0051] like Figure 2 As shown, within a signal-to-noise ratio (SNR) range of 0 dB to 20 dB, this invention demonstrates significantly superior performance compared to the contrasting algorithms in the DOA estimation task. Specifically, the root mean square error (RMSE) of this invention is consistently significantly lower than that of real-valued CNN, MUSIC, and ESPRIT algorithms across the entire SNR range, with a particularly pronounced advantage under low SNR conditions, indicating stronger noise resistance. As the SNR gradually increases, the RMSE of this invention steadily decreases, further validating its robustness under different noise environments. These results demonstrate that this invention, through its designed deep learning model, can more effectively extract signal features, overcoming the inherent weakness of traditional algorithms in being sensitive to coherent signals in complex multipath environments, while also exhibiting efficient utilization of phase information.
[0052] Example 2: This embodiment provides a complex domain attention network DOA estimation system with embedded physical information, including: The data preprocessing module preprocesses the far-field narrowband signal received by the array to obtain the sample covariance matrix and bispectral slice. The real and imaginary parts of the covariance matrix and the real and imaginary parts of the bispectral slice are stacked along the channel dimension to form an augmented complex tensor. The complex domain feature extraction module takes the augmented complex tensor as input, and completes feature extraction through complex convolution operation and modulus-preserving complex activation function, thus preserving the amplitude-phase coupling characteristics of the signal. The multipath sensing spatial attention module inputs the extracted complex features into the multipath sensing spatial attention module, obtains the original attention weights based on the self-attention mechanism, and generates a multipath suppression matrix by combining the physical constraint mask of the guide vector, and performs adaptive suppression on non-physically related regions. The sparse spectrum reconstruction module maps the attention-weighted features to an angular grid to obtain the initial spatial spectrum, and then achieves spectral sparsity through sparse soft thresholding. The DOA estimation and output module obtains the confidence scores of direct waves at various angles based on energy and spatial continuity, sorts them by confidence score, and outputs the final DOA estimation results according to the preset number of sources.
[0053] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.
[0054] Example 3: An electronic device is provided for running the aforementioned "a method for estimating the Data of Aspect (DOA) of a complex domain attention network with embedded physical information". The electronic device includes a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S5 of the method described in Embodiment 1, specifically including but not limited to: S1. Preprocess the far-field narrowband signal received by the array to obtain the sample covariance matrix and bispectral slice. Stack the real and imaginary parts of the covariance matrix and the real and imaginary parts of the bispectral slice along the channel dimension to form an augmented complex tensor. S2. Input the augmented complex tensor into the complex domain feature extraction module, and complete the feature extraction through complex convolution operation and the modulus-preserving complex activation function, preserving the amplitude-phase coupling characteristics of the signal; S3. Input the extracted complex features into the multipath perception spatial attention module, obtain the original attention weights based on the self-attention mechanism, and generate a multipath suppression matrix by combining the physical constraint mask of the guide vector to perform adaptive suppression on non-physically related regions; S4. Map the attention-weighted features to an angular grid to obtain the initial spatial spectrum, and then perform spectral sparsification through sparse soft thresholding. S5. Based on energy and spatial continuity, obtain the confidence scores of direct waves at each angle, sort them by confidence, and output the final DOA estimation results according to the preset number of sources.
[0055] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage (acquiring observation data and outputting inversion results) through a communication interface. The above hardware configuration does not constitute a limitation of the present invention.
[0056] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to execute the method steps S1 to S5 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.
[0057] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.
[0058] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
[0059] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.
Claims
1. A method for estimating DOA (Domain of Analysis) in a complex domain attention network with embedded physical information, characterized in that, Includes the following steps: The far-field narrowband signal received by the array is preprocessed to obtain the sample covariance matrix and bispectral slice. The real and imaginary parts of the covariance matrix and the real and imaginary parts of the bispectral slice are stacked along the channel dimension to form an augmented complex tensor. The augmented complex tensor is input into the complex domain feature extraction module, and feature extraction is completed through complex convolution operation and a modulus-preserving complex activation function, preserving the amplitude-phase coupling characteristics of the signal. The extracted complex features are input into the multipath perception spatial attention module, and the original attention weights are obtained based on the self-attention mechanism. The multipath suppression matrix is generated by combining the physical constraint mask of the guide vector, and adaptive suppression is performed on non-physically related regions. The attention-weighted features are mapped to an angular grid to obtain the initial spatial spectrum, which is then sparsified by sparse soft thresholding. The confidence scores of direct waves at each angle are obtained based on energy and spatial continuity. The signals are sorted by confidence score and the final DOA estimation results are output according to the preset number of sources.
2. The method for DOA estimation in complex domain attention networks with embedded physical information according to claim 1, characterized in that, The sample covariance matrix is expressed as follows: ； in, This indicates the conjugate transpose. This represents the array received signal vector. Indicates the number of snapshots; The bispectral slice is obtained by performing a two-dimensional discrete Fourier transform on the third-order cumulant of the array signal, and is used to extract the high-order phase coupling features of the signal; the third-order cumulant of the array signal is defined as: ； in, and The amount of time delay, For mathematical expectation; Performing a two-dimensional discrete Fourier transform on the third-order cumulants yields a bispectrum: ； Using finite-delay snapshots for solving, the bispectral density is simplified to a discrete form: ； in, The signal time-domain vector The Fourier transform result, To indicate complex conjugate, and It consists of two-dimensional frequency components; Therefore, the augmented complex tensor is: ； in, This indicates stacking along the channel dimension. To take the real part of a complex number, This indicates taking the imaginary part of a complex number.
3. The method for estimating DOA in a complex domain attention network with embedded physical information according to claim 1, characterized in that, The complex convolution operation is implemented as follows: for the input complex feature map... and complex convolution kernel Its output is defined as: ； in, , These represent the real and imaginary parts of the input feature map, respectively. , These are the real and imaginary parts of the convolution kernel, respectively. The imaginary unit, To augment complex tensors, For bias terms; The modulus-preserving complex activation function is: ； in, The output feature map of the complex convolution operation. It is a linear rectified activation function. This is a bias term.
4. The method for DOA estimation in a complex domain attention network with embedded physical information according to claim 1, characterized in that, The multipath suppression matrix is generated by combining the physical constraint mask of the guide vector, specifically as follows: Based on the guiding vector, any two array elements can be obtained. The theoretical coherence coefficient between them is: ； in, Let m be the component of the array guiding vector corresponding to the array element m. The component corresponding to array element n in the array steering vector; Iterate through all possible incident angles and count all array element pairs. Corresponding coherence coefficient The template for the global theoretical coherence coefficient matrix is obtained as follows: ； in, The total number of angle samples. It is a set containing all the sampled angles that have been traversed; Let the coherence threshold be... Generate physical constraint mask for: ； in, The region marked as non-physically relevant corresponds to spurious multipath reflection coherence. The region marked as physically feasible relevance corresponds to the actual relevance of the direct wave. Further generate the multipath suppression matrix : ； in, These are the angle-dependent suppression weights.
5. The method for DOA estimation in a complex domain attention network with embedded physical information according to claim 1, characterized in that, The attention weight matrix of the multipath sensing spatial attention module for: ； in, It represents the Hadamah accumulation. The learnable inhibition strength coefficient; For querying the matrix, The key matrix; Key matrix Dimensions This is a multipath suppression matrix; based on this, adaptive suppression is performed on non-physically related regions to achieve directional attenuation.
6. The method for DOA estimation in a complex domain attention network with embedded physical information according to claim 1, characterized in that, The attention-weighted features are: ； Will Mapping to an angular grid yields the initial spatial spectrum. : ； in, It is a non-linear activation function. Indicates will Flattened into a one-dimensional vector, The weight matrix is a learnable angle mapping. For bias terms; For the initial spatial spectrum Soft thresholding is performed to obtain the energy values of the sparsified spatial spectrum: ； in, The learnable sparse threshold.
7. The method for DOA estimation in a complex domain attention network with embedded physical information according to claim 6, characterized in that, The method for obtaining the direct wave confidence score is as follows: ； in, For angle The corresponding neighborhood region in the angle grid This represents the number of angular grid cells within this neighborhood. for The The diagonal elements represent the concentration of spatial correlation between array elements.
8. A complex domain attention network DOA estimation system with embedded physical information, characterized in that, include: The data preprocessing module preprocesses the far-field narrowband signal received by the array to obtain the sample covariance matrix and bispectral slice. The real and imaginary parts of the covariance matrix and the real and imaginary parts of the bispectral slice are stacked along the channel dimension to form an augmented complex tensor. The complex domain feature extraction module takes the augmented complex tensor as input, and completes feature extraction through complex convolution operation and modulus-preserving complex activation function, thus preserving the amplitude-phase coupling characteristics of the signal. The multipath sensing spatial attention module inputs the extracted complex features into the multipath sensing spatial attention module, obtains the original attention weights based on the self-attention mechanism, and generates a multipath suppression matrix by combining the physical constraint mask of the guide vector, and performs adaptive suppression on non-physically related regions. The sparse spectrum reconstruction module maps the attention-weighted features to an angular grid to obtain the initial spatial spectrum, and then achieves spectral sparsity through sparse soft thresholding. The DOA estimation and output module obtains the confidence scores of direct waves at various angles based on energy and spatial continuity, sorts them by confidence score, and outputs the final DOA estimation results according to the preset number of sources.
9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the complex domain attention network DOA estimation method for physical information embedding as described in any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements a complex domain attention network DOA estimation method for embedding physical information as described in any one of claims 1-7.