A DOA estimation method for non-gaussian colored noise interference scene

Abstract

This invention relates to the field of array signal processing, and particularly to a DOA estimation method for non-Gaussian colored noise interference scenarios. The method includes: acquiring signals under non-Gaussian spatial colored noise interference via an antenna receiving array; calculating the covariance matrix based on the received array data, and constructing a training set, a validation set, and labels based on the covariance matrix; constructing a DOA estimation model, which includes a self-attention-enhanced covariance matrix reconstruction module and a channel attention-enhanced angle estimation module connected in sequence; constructing a training loss function for the DOA estimation model, the loss function of the covariance matrix reconstruction module including signal subspace loss components, diagonal element loss components, off-diagonal element line loss components, and phase loss components; and training the DOA estimation model using a supervised offline training method based on the training set, validation set, and training loss function. This invention exhibits stronger robustness and higher DOA estimation accuracy in mixed environments with low signal-to-noise ratio non-Gaussian colored noise.

Description

Technical Field

[0001] This invention relates to the field of array signal processing, and in particular to a DOA estimation method for non-Gaussian colored noise interference scenarios. Background Technology

[0002] Direction-of-arrival (DOA) estimation is a core research area in array signal processing, with wide applications in sonar detection, wireless communication, and radar positioning. Its core objective is to process signals received by a sensor array to calculate spatial parameters such as the azimuth angle of one or more incident sources, thereby achieving spatial spectrum sensing and localization of targets. Model-driven high-resolution DOA estimation algorithms, such as the Multiple Signal Classification (MUSIC) algorithm and the Rotation Invariant Subspace (ESPRIT) algorithm, leverage the orthogonality between the signal and noise subspaces. Under ideal signal models and environmental assumptions, they can achieve estimation accuracy close to the Cramer-Rao lower bound.

[0003] However, in the complex and open electromagnetic environments of modern wireless communication and radar detection, signals received by array antennas are susceptible to interference from non-Gaussian colored noise. Due to the scarcity of spectrum resources, spectrum sharing has become the norm, and the overlap of secondary and primary user frequency bands leads to non-fixed frequency band interference mixed into the received signal. This type of interference exhibits both spatial correlation and non-Gaussian characteristics, breaking the assumption of independent and identically distributed Gaussian white noise. Furthermore, malicious intrusions by non-cooperative systems further exacerbate the severity of the electromagnetic environment. Non-cooperative systems randomly transmit interference within the bandwidth of the intruding target, and their heavy-tailed non-Gaussian characteristics, combined with multipath propagation and array element coupling resulting in spatial colored characteristics, further disrupt the ideal structure of the received signal covariance matrix. These interferences cause problems such as spectral peak splitting, increased spurious peaks, and a significant decrease in resolution in DOA estimation algorithms based on the Gaussian white noise assumption, making it difficult to meet the requirements of high-precision angle estimation in practical applications. A severe challenge is faced in balancing accuracy improvement and robustness enhancement.

[0004] Therefore, under the current technological background, this invention aims to solve the problems of accuracy and noise robustness in DOA estimation under complex real-world environments. It proposes a DOA estimation method for non-Gaussian colored noise interference scenarios that combines high accuracy, strong robustness, and high efficiency. This provides theoretical support and technical reference for improving the practical application performance of array signal processing systems and meets the urgent needs of related fields for high-precision DOA estimation under complex real-world environments. Summary of the Invention

[0005] Traditional DOA estimation algorithms suffer from spectral peak splitting, increased spurious peaks, and significant resolution degradation under non-Gaussian colored noise interference, making it difficult to balance accuracy improvement with robustness enhancement. Therefore, this invention provides a DOA estimation method specifically for non-Gaussian colored noise interference scenarios. This invention primarily utilizes a self-attention-enhanced covariance matrix reconstruction module to learn the mapping relationship between the covariance matrix and the ideal covariance matrix. It extracts angle-related features using a channel attention-enhanced ConvNeXt-based angle estimation module. Furthermore, the loss function considers signal subspace loss components, diagonal element loss components, off-diagonal element line loss components, and phase loss components, thereby effectively suppressing non-Gaussian colored noise interference, restoring the ideal structure of the covariance matrix, significantly improving the accuracy and robustness of DOA estimation, and achieving stable high-resolution angle estimation in complex electromagnetic environments.

[0006] The technical means employed in this invention are as follows:

[0007] A method for DOA estimation in non-Gaussian colored noise interference scenarios includes the following steps: Signals affected by non-Gaussian spatial colored noise are collected through an antenna receiving array. Using the first antenna in the antenna receiving array as a reference element, a narrowband signal model is established based on the signal to obtain the array received data; Based on the data received by the array, the covariance matrix is ​​calculated, and a training set, a validation set, and labels are constructed based on the covariance matrix. A DOA estimation model is constructed, which includes a self-attention-enhanced covariance matrix reconstruction module and a channel attention-enhanced angle estimation module connected in sequence. The training loss function for constructing the DOA estimation model includes the loss function of the self-attention enhanced covariance matrix reconstruction module and the loss function of the channel attention enhanced angle estimation module. The loss function of the self-attention enhanced covariance matrix reconstruction module includes signal subspace loss components, diagonal element loss components, off-diagonal element line loss components, and phase loss components. A supervised offline training method is used to train the DOA estimation model based on the training set, validation set, labels, and the training loss function. Based on the trained DOA estimation model, the direction of arrival (DOA) is estimated.

[0008] Furthermore, the formula for calculating the received signal of the array is as follows: ,

[0009] in, For array to receive signals, For guiding vectors, For the first K A far-field narrowband signal, K For the number of information sources, k As the first serial number, To interfere with the amplitude, c ( t () represents the interference component, T represents the number of snapshots acquired, and the formula for calculating the guide vector is:

[0010] in, j The imaginary unit, λ The wavelength at the carrier frequency, , c At the speed of light, f For carrier frequency, d For sensor spacing, The angle of signal incidence. N This represents the number of sensors.

[0011] Furthermore, the training set for training the self-attention-enhanced covariance matrix reconstruction module includes a covariance matrix, and the label includes an ideal covariance matrix; The training set used to train the channel attention-enhanced angle estimation module includes a reconstructed covariance matrix and labels including the true angle.

[0012] Furthermore, the formula for calculating the covariance matrix is ​​as follows:

[0013] in, Let covariance matrix be the variance matrix. As the guide vector, Let be the signal covariance matrix. To interfere with the amplitude, This is the spatial correlation coefficient. This represents the index difference between array elements.

[0014] Furthermore, the formula for calculating the training loss function is as follows:

[0015] in, To train the loss function, The loss weights for the loss function of the self-attention enhanced covariance matrix reconstruction module. The loss function for the self-attention-enhanced covariance matrix reconstruction module is... The loss weights for the loss function of the channel attention enhancement angle estimation module. The loss function for the channel attention enhancement angle estimation module, The reconstruction result of the i-th covariance matrix, Label the i-th self-attention-enhanced covariance matrix reconstruction module. For the i-th DOA estimation result, Label the i-th angle estimation module. The total number of samples, The formula for calculating the loss function of the self-attention enhanced covariance matrix reconstruction module is as follows:

[0016] in, For phase loss weights, For the signal subspace loss component, For the signal subspace loss weights, Weights are assigned to the diagonal elements. For the loss components of diagonal elements, For off-diagonal element loss components, For phase loss components, The formula for calculating the loss function of the channel attention enhancement angle estimation module is as follows:

[0017] in, For the j'-th DOA estimation result, Let j' be the label for the angle estimation module.

[0018] Furthermore, the calculation formula for the signal subspace loss component is as follows:

[0019] in, P s The projection matrix of the signal subspace. N For the number of sensors, B’ The result of covariance matrix reconstruction. Y The self-attention enhanced covariance matrix reconstruction module is tagged with [label]. The formula for calculating the loss component of the diagonal element is as follows:

[0020] in, Let be the element in the i-th row and i-th column of the reconstructed covariance matrix. The element in the i-th row and i-th column of the self-attention enhancement covariance matrix reconstruction module is labeled. The formula for calculating the off-diagonal element line loss component is as follows:

[0021] in, Let be the element in the i-th row and j'-th column of the reconstructed covariance matrix. The element in the i-th row and j'-th column of the self-attention enhancement covariance matrix reconstruction module is the label. The formula for calculating the phase loss component is as follows: .

[0022] Furthermore, the method also includes: The spatial domain is divided into 2G+1 discrete points with a resolution, and a DOA estimation grid G ​​is established. For scenarios with different signal-to-noise ratios, different spatial correlation coefficients of non-Gaussian colored noise, and different texture component parameters, K signal incident angles are selected from the DOA estimation grid.

[0023] Furthermore, the self-attention enhanced covariance matrix reconstruction module includes an encoder, a bottleneck layer, a decoder, and a regularization layer connected in sequence; The channel attention enhancement angle estimation module includes a first channel attention enhancement ConvNeXt layer, a second channel attention enhancement ConvNeXt layer, and a third channel attention enhancement ConvNeXt layer connected in sequence.

[0024] Furthermore, the loss weight of the loss function of the self-attention enhancement covariance matrix reconstruction module is set to 0.45, and the loss weight of the loss function of the channel attention enhancement angle estimation module is set to 0.55.

[0025] Compared with the prior art, the present invention has the following advantages: This invention employs a DOA estimation method for non-Gaussian colored noise interference scenarios. This method recovers the covariance matrix under the Gaussian white noise assumption through a self-attention-enhanced covariance matrix reconstruction module, thus suppressing the impact of non-Gaussian colored noise on DOA estimation. Furthermore, it proposes a channel attention-enhanced angle estimation module based on the ConvNeXt network architecture, extracting angle information from the reconstructed covariance matrix and outputting the corresponding probability distribution. Under low and medium signal-to-noise ratio (SNR) conditions, this method can control the root mean square (RMS) error of DOA estimation within 5°, demonstrating significant stability and accuracy advantages. Moreover, the proposed algorithm also outperforms other comparative algorithms in terms of RMS error under different SNR and snapshot number conditions.

[0026] Based on the above reasons, this invention can be widely applied in fields such as direction of arrival estimation. Attached Figure Description

[0027] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart illustrating a DOA estimation method for non-Gaussian colored noise interference scenarios according to the present invention.

[0029] Figure 2 The diagram shows a comparison of noise amplitude distribution in the embodiments, where (a) Amplitude distribution diagram, (b) Amplitude distribution plot with logarithmic base, (c) Amplitude distribution diagram, (d) Amplitude distribution plot with logarithmic base.

[0030] Figure 3 The following are noise matrix amplitude heatmaps in the embodiments, wherein (a) is a Gaussian white noise matrix amplitude heatmap, and (b) is a Gaussian white noise matrix amplitude heatmap. Time matrix amplitude heatmap, (c) Time matrix amplitude heatmap.

[0031] Figure 4 The following are comparison diagrams of the reconstruction effects of the self-attention enhanced covariance matrix reconstruction module in the embodiments, wherein (a) is a heatmap of the amplitude of the received signal covariance matrix in a non-Gaussian colored scene, (b) is a heatmap of the amplitude of the output matrix of the self-attention enhanced covariance matrix reconstruction network, (c) is a heatmap of the received signal covariance matrix in an ideal Gaussian white noise scene, (d) is the difference between the received signal covariance matrix and the ideal covariance matrix, and (e) is the difference between the reconstructed covariance matrix and the ideal covariance matrix.

[0032] Figure 5 The images show the estimation results of the channel attention enhancement angle estimation module in the embodiment under a snapshot of 500, a signal-to-noise ratio of -10dB, and a Gaussian white noise environment. Among them, (a) is the effect of the channel attention enhancement angle estimation module, (b) is the effect of the DeepCNN network, (c) is the effect of the MUSIC algorithm, (d) is the effect of the Root-MUSIC algorithm, and (e) is the effect of the ESPRIT algorithm.

[0033] Figure 6The images show the error diagrams of the estimated angle and the true angle of the DOA estimation method for non-Gaussian colored noise interference scenarios in the embodiments, under a snapshot size of 500, a signal-to-noise ratio of -10dB, and a non-Gaussian colored noise environment. Among them, (a) is the estimation error diagram of the AEDN network, (b) is the error diagram of the channel attention enhancement angle estimation module, (c) is the estimation error diagram of the DeepCNN network, (d) is the estimation error diagram of the MUSIC algorithm, and (e) is the estimation error diagram of the Root-MUSIC algorithm. Detailed Implementation

[0034] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0035] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0036] like Figure 1 As shown, this invention provides a DOA estimation method for non-Gaussian colored noise interference scenarios, including the following steps: S1. Divide the spatial domain by resolution. Divide the data into 2G+1 discrete points (in degrees) and establish a DOA estimation grid G.

[0037] This invention sets the resolution The estimated spatial domain is discretized into 181 grid points, of which Grid The estimated angular domain range is ,in, G ρ For resolution-based Construct the angle grid points.

[0038] S2. For different signal-to-noise ratios, different spatial correlation coefficients of non-Gaussian colored noise, and different texture component parameters, select K signal incident angles from the DOA estimation grid.

[0039] This invention sets the number of signal sources to 2, generates all possible combinations of incident angle pairs from the grid, totaling 16290 pairs, randomly selects 1000 pairs, and simulates the incident signal under different signal-to-noise ratios (20 / -15 / -10 / -5 / 0dB) and noise conditions (spatial correlation coefficients of 3.0 / 5.0 / 7.0, texture component parameters of 0.1 / 0.3 / 0.6 / 0.9).

[0040] S3. Collect signals affected by non-Gaussian colored noise through an antenna receiving array.

[0041] Specifically, a linear uniform array of 16 antennas is used as the receiving array to receive signals containing non-Gaussian spatial colored noise.

[0042] S4. Using the first antenna in the antenna receiving array as a reference element, establish a narrowband signal model based on the signal to obtain the array received data.

[0043] Assumption A far-field narrowband signal ,from Angled incidence A uniform linear array composed of sensors, and Then the array receives signals affected by noise interference. It can be represented as: ,

[0044] in, To receive signals for an array susceptible to noise interference, For guiding vectors, For the k-th far-field narrowband signal, K For the number of information sources, k As the first serial number, As the guide vector, For noise vectors, For far-field narrowband signals, T is the number of snapshots acquired, and the formula for calculating the steering vector is:

[0045] in, j The imaginary unit, λ The wavelength at the carrier frequency, , c At the speed of light, f For carrier frequency, d For sensor spacing, The angle of signal incidence. N This represents the number of sensors.

[0046] In a non-Gaussian colored noise interference environment, the mathematical model of the array receiving signal can be expressed as: ,

[0047] in, For array to receive signals, To interfere with the amplitude, c ( t () represents the interference component. The formula for calculating the interference component is: ,

[0048] in, Let z(t) be the texture component and z(t) be the speckle component, both having a mean of 0 and a variance of 0. The complex Gaussian distribution. The covariance matrix of colored noise in non-Gaussian space can be expressed as:

[0049] in, This is the spatial correlation coefficient. The index difference between array elements The larger the value, the stronger the correlation between the sensors.

[0050] S5. Based on the array received data, calculate the covariance matrix, and construct the training set, validation set, and labels based on the covariance matrix.

[0051] When the noise has a mean of zero and a variance of... When the additive white Gaussian noise is applied, the covariance matrix of the received signal can be expressed as:

[0052] in, R x The covariance matrix of the received signal. For noise variance, I N It is an N-order identity matrix. R s Let be the signal covariance matrix. , s ( t () represents a far-field narrowband signal.

[0053] Non-Gaussian properties are introduced by amplitude modulation of the speckle component, whose probability follows... Gamma distributed, .in, For shape factor, The smaller the value, the more significant the non-Gaussianity of the noise and the more prominent the heavy-tailed characteristics.

[0054] The formula for calculating the covariance matrix under colored noise in non-Gaussian space is:

[0055] in, Let covariance matrix be the variance matrix. As the guide vector, This represents the index difference between array elements.

[0056] The covariance matrix of the received signal under non-Gaussian colored noise is divided into training and validation sets in a 7:3 ratio, and the covariance matrix under ideal Gaussian white noise and the true incident angle are used as labels.

[0057] S6. Construct the DOA estimation model. The DOA estimation model includes a self-attention-enhanced covariance matrix reconstruction module and a channel attention-enhanced angle estimation module that are connected in sequence. The self-attention-enhanced covariance matrix reconstruction module is constructed based on self-attention enhancement.

[0058] The self-attention-enhanced covariance matrix reconstruction module consists of an encoder, a bottleneck layer, a decoder, and a regularization layer connected in sequence.

[0059] The self-attention-enhanced covariance matrix reconstruction module consists of an encoder, a bottleneck layer, a decoder, and a regularization layer. The encoder employs a progressively downsampled hierarchical convolutional structure. The bottleneck layer introduces a self-attention mechanism module, which directly identifies and models the interference of off-diagonal elements in the noise covariance matrix by calculating the correlation of all position pairs, and generates attention weights to actively suppress these noise-induced pseudo-correlations. The decoder, based on the abstract features extracted by the encoder and the global information processed by the bottleneck layer, progressively reconstructs a covariance matrix representation with the same dimension as the input covariance matrix and possessing Gaussian white noise statistical properties through inverse convolution. The regularization layer is implemented based on matrix theory principles, ensuring the mathematical validity of the output matrix of the self-attention-enhanced covariance matrix reconstruction module through a combination of eigenvalue decomposition and diagonal analysis.

[0060] The channel attention enhancement angle estimation module includes a first channel attention enhancement ConvNeXt layer, a second channel attention enhancement ConvNeXt layer, and a third channel attention enhancement ConvNeXt layer connected in sequence.

[0061] The channel attention-enhanced angle estimation module employs a multi-scale hierarchical design, cascading three channel attention-enhanced ConvNext modules. This module first constructs a wide receptive field using 7×7 depthwise separable convolutions to effectively capture features from the input matrix. Subsequently, it introduces a channel attention mechanism, performing global statistical modeling on each channel to enable the network to adaptively enhance the response of channels critical to the current task stage. Next, the module uses an inverse residual structure for further feature transformation and dimensionality expansion. It enhances nonlinear expressiveness and improves the modeling depth for complex features through 1×1 convolutions followed by dimensionality reduction and GELU activation. Finally, a learnable residual scaling factor is introduced to regulate the strength of skip connections, achieving optimized fusion of deep features while preserving original feature information. Through this process, the module focuses on extracting and integrating phase difference features, local spatial correlation, and global beam direction features at the shallow, middle, and deep layers of the network, respectively, thus systematically constructing a multi-layered angle representation capability.

[0062] S7. Construct the training loss function for the DOA estimation model. The training loss function includes the loss function of the self-attention enhanced covariance matrix reconstruction module and the loss function of the channel attention enhanced angle estimation module. The loss function of the self-attention enhanced covariance matrix reconstruction module includes the signal subspace loss component, the diagonal element loss component, the off-diagonal element line loss component, and the phase loss component.

[0063] The formula for calculating the training loss function is:

[0064] in, To train the loss function, The loss weights for the loss function of the self-attention enhanced covariance matrix reconstruction module. The loss function for the self-attention-enhanced covariance matrix reconstruction module is... The loss weights for the loss function of the channel attention enhancement angle estimation module. The loss function for the channel attention enhancement angle estimation module, The reconstruction result of the i-th covariance matrix, Label the i-th self-attention-enhanced covariance matrix reconstruction module. For the i-th DOA estimation result, Label the i-th angle estimation module. The total number of samples is 0.45. The loss weight of the loss function of the self-attention enhancement covariance matrix reconstruction module is set to 0.45, and the loss weight of the loss function of the channel attention enhancement angle estimation module is set to 0.55.

[0065] The formula for calculating the loss function of the self-attention-enhanced covariance matrix reconstruction module is as follows:

[0066] in, The phase loss weight is set to 0.1 to make a fine adjustment to the phase. For the signal subspace loss component, The signal subspace loss weight is set to 0.8 to ensure the integrity of the signal structure. The weight for the loss of diagonal elements is set to 0.6. For the loss components of diagonal elements, For off-diagonal element loss components, This represents the phase loss component.

[0067] The formula for calculating the signal subspace loss component is:

[0068] in, P s The projection matrix of the signal subspace. N For the number of sensors, B’ The result of covariance matrix reconstruction. Y The self-attention enhanced covariance matrix reconstruction module is tagged with [label]. , The front of the ideal covariance matrix The eigenvector matrix corresponding to the largest eigenvalue.

[0069] The diagonal elements reflect the received power of each channel, and their loss is used to constrain the noise energy of the reconstructed matrix to tend towards whitening. The formula for calculating the diagonal element loss component is as follows:

[0070] in, Let be the element in the i-th row and i-th column of the reconstructed covariance matrix. The element in the i-th row and i-th column of the self-attention enhancement covariance matrix reconstruction module is labeled.

[0071] Off-diagonal element loss reflects the signal correlation between sensors. The formula for calculating the off-diagonal element loss component is as follows:

[0072] in, Let be the element in the i-th row and j'-th column of the reconstructed covariance matrix. The element in the i-th row and j'-th column of the self-attention enhancement covariance matrix reconstruction module label.

[0073] The signal subspace loss and diagonal element loss have constrained the principal phase relationship, so... Setting the value to 0.1 allows for fine-tuning of the phase. The formula for calculating the phase loss component is as follows: .

[0074] In the DOA estimation task, it is modeled as a multi-label classification problem, i.e., determining whether a signal arrives at each angle in a preset angle grid. Each angle grid point corresponds to an independent binary classification task. Therefore, the channel attention-enhanced angle estimation module uses the binary cross-entropy loss function. The formula for calculating the loss function of the channel attention-enhanced angle estimation module is as follows:

[0075] in, For the j'-th DOA estimation result, Let j' be the label for the angle estimation module.

[0076] Experiments have verified that setting the loss weight to [value] is [effective]. The model achieved optimal joint optimization performance on the validation set, demonstrating the best balance between the self-attention-enhanced covariance matrix reconstruction and channel attention-enhanced angle estimation modules. Supervised offline training was employed, simultaneously training the self-attention-enhanced covariance matrix reconstruction and channel attention-enhanced angle estimation modules on the constructed dataset. The parameters of both modules were jointly optimized using backpropagation, with the optimal model selected based on minimizing the validation set loss. The Adam optimizer was used during training to improve convergence efficiency and stability, undergoing 200 training epochs with a batch size of 512. Experiments were conducted on a GPU platform equipped with a 25-core Intel Xeon Platinum 8481C processor and CUDA 12.8 support, with GPU driver version 570.124.04.

[0077] S8. Use supervised offline training to train the DOA estimation model based on the training set, validation set, labels, and training loss function.

[0078] The training set for training the self-attention-enhanced covariance matrix reconstruction module includes the covariance matrix, and the labels include the ideal covariance matrix. The input to the self-attention covariance matrix module learns its mapping relationship to reconstruct the covariance matrix.

[0079] The training set for the channel attention-enhanced angle estimation module includes the reconstructed covariance matrix and labels including the true angle. The module learns the mapping relationship from the reconstructed covariance matrix to the incident angle to achieve the DOA estimation task.

[0080] S9. Based on the trained DOA estimation model, realize the direction of arrival estimation.

[0081] For the DOA estimation task, the incident signal is received under non-Gaussian colored noise environment, the signal is input into the optimal model, and the source incident angle and the source existence probability corresponding to the angle network point are obtained to complete the estimation task.

[0082] This invention also includes a DOA estimation system for non-Gaussian colored noise interference scenarios, implemented based on the aforementioned DOA estimation method for non-Gaussian colored noise interference scenarios. The system includes: Signal receiving module: used to acquire incident signals.

[0083] The self-attention enhancement covariance matrix reconstruction module is used to reconstruct the covariance matrix under non-Gaussian colored noise environment into a covariance matrix under Gaussian white noise environment.

[0084] Channel attention enhancement angle estimation module: used to obtain the source incidence angle from the reconstructed covariance matrix module and output the estimation result.

[0085] The effectiveness of the proposed DOA estimation method for non-Gaussian colored noise interference scenarios was verified by experiments. Figure 2-6 Based on the impact of non-Gaussian colored noise interference on DOA estimation accuracy, the noise reduction capability of the self-attention enhanced covariance matrix reconstruction module, the estimation performance of the channel attention enhanced angle estimation network module in Gaussian white noise environment, and the comprehensive estimation performance of AEDN in low signal-to-noise ratio non-Gaussian colored noise scene are demonstrated.

[0086] Detailed analysis, Figure 2 The blue and red portions represent the amplitude distributions of non-Gaussian colored noise and Gaussian white noise, respectively. The black curve represents the theoretical Rayleigh distribution. The distribution of Gaussian white noise matches the theoretical Rayleigh curve well, consistent with its statistical characteristics. In contrast, non-Gaussian colored noise exhibits different amplitude distributions with varying shape parameters. Under the Gamma distribution, they all exhibit varying degrees of non-Gaussian characteristics with sharp peaks and thick tails. Figure 3 It can be seen that as the spatial correlation coefficient increases, the off-diagonal elements of the matrix show a significant increasing trend, and its off-diagonalization characteristics are gradually enhanced. Figure 4 This study demonstrates that when noise exhibits strong non-Gaussian spatial colored characteristics, the input covariance matrix suffers severe structural distortion, with severely blurred main diagonal edges and energy diffusion, showing a significant difference from the received covariance matrix under ideal Gaussian white noise. However, after reconstruction network processing, the output matrix effectively corrects the structural distortion of the input matrix, the diffusion range of matrix elements converges significantly, the distribution more closely resembles the ideal banded form, and the difference from the ideal covariance matrix is ​​reduced. Figure 5 The channel attention enhancement angle estimation module proposed in this invention provides estimates that all converge near the true value, while other algorithms exhibit significant deviations from the true angle. In non-Gaussian colored noise environments, from Figure 6 It can be seen that the method of this invention has the least error in terms of the actual angle. The comprehensive experimental results show that the method proposed in this invention can reliably estimate DOA in low signal-to-noise ratio, non-Gaussian colored noise environments, and compared with other comparative algorithms, it has the best overall stability and the highest estimation accuracy.

[0087] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0088] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A DOA estimation method for non-Gaussian colored noise interference scenarios, characterized in that, Includes the following steps: Signals affected by non-Gaussian spatial colored noise are collected through an antenna receiving array. Using the first antenna in the antenna receiving array as a reference element, a narrowband signal model is established based on the signal to obtain the array received data; Based on the data received by the array, the covariance matrix is ​​calculated, and a training set, a validation set, and labels are constructed based on the covariance matrix. A DOA estimation model is constructed, which includes a self-attention-enhanced covariance matrix reconstruction module and a channel attention-enhanced angle estimation module connected in sequence. The training loss function for constructing the DOA estimation model includes the loss function of the self-attention enhanced covariance matrix reconstruction module and the loss function of the channel attention enhanced angle estimation module. The loss function of the self-attention enhanced covariance matrix reconstruction module includes signal subspace loss components, diagonal element loss components, off-diagonal element line loss components, and phase loss components. A supervised offline training method is used to train the DOA estimation model based on the training set, validation set, labels, and the training loss function. Based on the trained DOA estimation model, the direction of arrival (DOA) is estimated.

2. The DOA estimation method for non-Gaussian colored noise interference scenarios according to claim 1, characterized in that, The formula for calculating the received signal of the array is as follows: ， in, For array to receive signals, For guiding vectors, For the first K A far-field narrowband signal, K For the number of information sources, k As the first serial number, To interfere with the amplitude, c ( t () represents the interference component, T represents the number of snapshots acquired, and the formula for calculating the guide vector is: in, j The imaginary unit, λ The wavelength at the carrier frequency, , c At the speed of light, f For carrier frequency, d For sensor spacing, The angle of signal incidence. N This represents the number of sensors.

3. The DOA estimation method for non-Gaussian colored noise interference scenarios according to claim 1, characterized in that, The training set for training the self-attention-enhanced covariance matrix reconstruction module includes a covariance matrix, and the labels include an ideal covariance matrix. The training set used to train the channel attention-enhanced angle estimation module includes a reconstructed covariance matrix and labels including the true angle.

4. The DOA estimation method for non-Gaussian colored noise interference scenarios according to claim 1, characterized in that, The formula for calculating the covariance matrix is ​​as follows: in, Let covariance matrix be the variance matrix. As the guide vector, The signal covariance matrix, To interfere with the amplitude, This is the spatial correlation coefficient. This represents the index difference between array elements.

5. The DOA estimation method for non-Gaussian colored noise interference scenarios according to claim 1, characterized in that, The formula for calculating the training loss function is as follows: in, To train the loss function, The loss weights for the loss function of the self-attention enhanced covariance matrix reconstruction module. The loss function for the self-attention-enhanced covariance matrix reconstruction module is... The loss weights for the loss function of the channel attention enhancement angle estimation module. The loss function for the channel attention enhancement angle estimation module, The result of reconstructing the i-th covariance matrix. Label the i-th self-attention-enhanced covariance matrix reconstruction module. For the i-th DOA estimation result, Label the i-th angle estimation module. The total number of samples, The formula for calculating the loss function of the self-attention enhanced covariance matrix reconstruction module is as follows: in, For phase loss weights, For the signal subspace loss component, For the signal subspace loss weights, Weights are assigned to the diagonal elements. For the loss components of diagonal elements, For off-diagonal element loss components, For phase loss components, The formula for calculating the loss function of the channel attention enhancement angle estimation module is as follows: in, For the j'th DOA estimation result, Let j' be the label for the angle estimation module.

6. The DOA estimation method for non-Gaussian colored noise interference scenarios according to claim 5, characterized in that, The formula for calculating the signal subspace loss component is as follows: in, P s The projection matrix of the signal subspace. N For the number of sensors, B’ The result of covariance matrix reconstruction. Y The self-attention enhanced covariance matrix reconstruction module is tagged with [label]. The formula for calculating the loss component of the diagonal element is as follows: in, Let be the element in the i-th row and i-th column of the reconstructed covariance matrix. The element in the i-th row and i-th column of the self-attention enhancement covariance matrix reconstruction module is the label. The formula for calculating the off-diagonal element line loss component is as follows: in, Let be the element in the i-th row and j'-th column of the reconstructed covariance matrix. The element in the i-th row and j'-th column of the self-attention enhancement covariance matrix reconstruction module is the label. The formula for calculating the phase loss component is as follows: 。 7. The DOA estimation method for non-Gaussian colored noise interference scenarios according to claim 1, characterized in that, The method further includes: The spatial domain is divided into 2G+1 discrete points with a resolution, and a DOA estimation grid G ​​is established. For scenarios with different signal-to-noise ratios, different spatial correlation coefficients of non-Gaussian colored noise, and different texture component parameters, K signal incident angles are selected from the DOA estimation grid.

8. The DOA estimation method for non-Gaussian colored noise interference scenarios according to claim 1, characterized in that, The self-attention-enhanced covariance matrix reconstruction module includes an encoder, a bottleneck layer, a decoder, and a regularization layer connected in sequence. The channel attention enhancement angle estimation module includes a first channel attention enhancement ConvNeXt layer, a second channel attention enhancement ConvNeXt layer, and a third channel attention enhancement ConvNeXt layer connected in sequence.

9. The DOA estimation method for non-Gaussian colored noise interference scenarios according to claim 5, characterized in that, The loss weight of the loss function of the self-attention enhancement covariance matrix reconstruction module is set to 0.45, and the loss weight of the loss function of the channel attention enhancement angle estimation module is set to 0.55.

Images