A method and system for DOA estimation based on generative adversarial networks

Abstract

The application discloses a DOA estimation method and system based on a generative adversarial network. The method first acquires time-domain coherent signals received by a uniform linear array, performs preprocessing, and generates a coherent signal covariance matrix and a corresponding incoherent covariance matrix. Secondly, a covariance matrix reconstruction model is constructed, based on the coherent signal covariance matrix and the corresponding incoherent covariance matrix, a generator of the generative adversarial network is used to extract deep features from the preprocessed coherent signal covariance matrix, and the coherent signal covariance matrix is reconstructed. Then, the MUSIC algorithm is used to estimate and verify the DOA based on the reconstructed covariance matrix. The application improves the precision of coherent DOA estimation while not losing the utilization rate of the array aperture. The application has less parameter quantity, lower calculation complexity, faster covariance matrix reconstruction speed and stronger covariance matrix reconstruction capacity.

Description

Technical Field

[0001] This invention relates to the field of Direction of Arrival (DOA) estimation technology, specifically providing a DOA estimation method based on generative adversarial networks. This method utilizes deep learning and is particularly suitable for coherent signal processing in radar, navigation, target localization and tracking, and satellite communications. Background Technology

[0002] DOA estimation plays a crucial role in various fields such as radar, detection, navigation, and wireless communication. In ideal scenarios, signals are typically assumed to be uncorrelated, allowing traditional DOA estimation algorithms to achieve satisfactory performance. However, in real-world electromagnetic environments, multipath effects are prevalent, often resulting in coherent signals incident on the array. This coherence leads to a rank deficiency in the covariance matrix, causing traditional subspace algorithms (such as Multiple Signal Classification (MUSIC) and Rotation Invariant Subspace (ESPRIT)) to fail. To address this issue, traditional methods typically employ spatial smoothing techniques or Toeplitz matrix reconstruction to restore the rank of the covariance matrix. However, these methods have significant limitations: spatial smoothing sacrifices some array aperture, leading to reduced resolution; Toeplitz reconstruction often results in large estimation errors. Furthermore, the performance of both methods degrades significantly under extremely low signal-to-noise ratios and finite snapshot numbers.

[0003] In recent years, the rise of deep learning has opened up promising avenues for coherent signal DOA estimation. Unlike traditional methods that require explicit deconvolution, some deep learning methods attempt to directly process coherent signals or reconstruct subspaces through data-driven approaches. However, existing deep learning methods either perform poorly in harsh environments or require complex network structures and loss functions to improve training performance, thus still exhibiting significant shortcomings. Specifically, existing network designs often struggle to achieve high-accuracy feature extraction and rank recovery while maintaining low computational overhead.

[0004] K. Lee et al. proposed a DOA estimation method combining full-row Toeplitz matrix reconstruction and multilayer perceptron (MLP). Although this method utilizes a preprocessing step to assist network learning, it still has significant shortcomings: the algorithm exhibits poor robustness in low signal-to-noise ratio environments, showing large estimation errors. This means that the simple MLP structure, when combined with traditional reconstruction methods, still struggles to effectively cope with noise interference and cannot meet the high-accuracy requirements in complex electromagnetic environments.

[0005] T. Wang et al. reconstructed the noisy subspace using ResNet and verified orthogonality using a specific loss function. The limitation of this method is that its designed specific loss function imposes a significant computational burden during training, reducing the efficiency of model training. Furthermore, other similar networks often only achieve suboptimal estimation errors when processing coherent signals.

[0006] In summary, while deep learning has shown great potential in signal processing, existing methods for DOA estimation of coherent signals still face challenges: some methods sacrifice array aperture, others suffer from insufficient estimation accuracy under low signal-to-noise ratio and low snapshot conditions, and high-performance deep learning methods often come with prohibitively high computational costs. Therefore, there is an urgent need in this field for a novel method that can directly recover the full-rank covariance matrix without sacrificing array aperture, and possesses efficient feature extraction capabilities and robustness, in order to achieve high-precision coherent DOA estimation under complex conditions. Summary of the Invention

[0007] Purpose of the Invention: To address the limitations of traditional decoherence algorithms in accurately estimating the angle of arrival (DOA) under harsh conditions and the high computational complexity of existing deep learning methods, this invention proposes a DOA estimation method and system based on generative adversarial networks (GANs). This method, without sacrificing array aperture, mines feature information from the coherent signal through a generator within a GAN, further recovers the rank of the coherent signal's covariance matrix, and uses a discriminator to determine whether the generated covariance matrix truly possesses full-rank characteristics.

[0008] Technical solution: In one aspect, this invention provides a DOA estimation method based on generative adversarial networks, comprising the following steps:

[0009] S1: Obtain the time-domain coherent signal received by the uniform linear array, perform preprocessing, and generate the coherent signal covariance matrix and the corresponding incoherent covariance matrix respectively.

[0010] S2: Construct a covariance matrix reconstruction model. Based on the covariance matrix of the coherent signal and the corresponding incoherent covariance matrix, use the generator of the generative adversarial network to extract deep features from the preprocessed covariance matrix of the coherent signal and reconstruct the covariance matrix. Design a dual-constraint loss function composed of implicit adversarial loss and pixel-level reconstruction loss with specific weights.

[0011] S3: DOA estimation is performed using the MUSIC algorithm based on the reconstructed covariance matrix, and the results are verified.

[0012] In another aspect, the present invention also provides a DOA estimation system based on generative adversarial networks (GANs) for implementing the aforementioned DOA estimation method based on GANs, comprising the following modules:

[0013] The data acquisition and processing module is used to acquire the time-domain coherent signal received by the uniform linear array, perform preprocessing, and generate the coherent signal covariance matrix and the corresponding incoherent covariance matrix, respectively.

[0014] The covariance matrix reconstruction module is used to construct a covariance matrix reconstruction model. Based on the covariance matrix of the coherent signal and the corresponding incoherent covariance matrix, it uses the generator of the generative adversarial network to extract deep features from the preprocessed coherent signal covariance matrix and reconstruct the covariance matrix. It also designs a dual-constraint loss function composed of implicit adversarial loss and pixel-level reconstruction loss with specific weights.

[0015] The DOA estimation output module uses the MUSIC algorithm to estimate DOA based on the reconstructed covariance matrix and then verifies it.

[0016] The design principle of this invention is as follows:

[0017] 1. A uniform linear array receives coherent signals from a source, and a generative adversarial network (GAN) extracts deep features and decoheres them. This process relies on a specific dual-constraint loss function (i.e., implicit adversarial loss and pixel-level weights with specific weights). (Combining loss) the network can accurately recover the full-rank characteristic of the original rank-deficient covariance matrix of the coherent signal without introducing complex mathematical constraints, and significantly improve the DOA estimation accuracy in low signal-to-noise ratio and low snapshot environments.

[0018] 2. In the method of the present invention, a uniform linear array is used to receive coherent signal data, and a full-rank covariance matrix is ​​obtained through a generative adversarial network trained by the above-mentioned joint loss function. The DOA spectral peak is obtained from the reconstructed covariance matrix through the MUSIC algorithm, thereby realizing DOA estimation.

[0019] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0020] A uniform linear array receives coherent signals caused by multipath effects. Utilizing the robustness and generalization ability of the network model in this invention, the rank of the coherent signal covariance matrix can be well recovered, improving the accuracy of coherent DOA estimation without sacrificing the utilization rate of the array aperture. In particular, the pixel-level reconstruction loss mechanism introduced in this invention, by imposing strict constraints at the element level, ensures the lossless restoration of sensitive physical characteristics such as the phase and amplitude of the covariance matrix.

[0021] Compared to other deep learning DOA estimation algorithms, this invention employs an efficient joint loss strategy at the loss function level, resulting in fewer parameters, lower computational complexity, faster covariance matrix reconstruction speed, and stronger covariance matrix reconstruction capability, greatly increasing the possibility of deployment in real electromagnetic environments. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the uniform linear array structure involved in this invention;

[0023] Figure 2 This is a schematic diagram of the deep learning model structure involved in this invention;

[0024] Figure 3 This is a performance comparison chart of the method of this invention with other algorithms under different signal-to-noise ratio conditions;

[0025] Figure 4 This is a performance comparison chart of the method of this invention with other algorithms under different snapshot numbers. Detailed Implementation

[0026] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0027] This embodiment is based on the following principle: In one aspect, the present invention provides a DOA estimation method based on generative adversarial networks. The specific method of this embodiment is as follows:

[0028] I. Signal Model and Preprocessing

[0029] Signal model:

[0030] like Figure 1 The uniform linear array antenna structure shown is composed of It consists of several array elements, with the first sensor as the reference element, and the spacing between the array elements is... The wavelength is .

[0031] Assumption The sources are respectively Narrowband incoherent signals are incident on a signal with On a uniform linear array (ULA) with n array elements, then the nth The sensor at the first The complex envelope received in the next snapshot is:

[0032] (1)

[0033] The received signal of equation (1) The summation of different signal sources multiplied by their corresponding phase delays, plus the noise term. For the first The received signals from each sensor Represents the first One signal source, For the number of snapshots, For the first The sensor receives the first signal relative to the reference element. Phase delay of the signal propagation.

[0034] The received data model can be represented as:

[0035] (2)

[0036] The received signal in equation (2) is the product of the array manifold matrix and the signal vector plus the noise component. For the received signal matrix, Let be an array manifold matrix, where As the guide vector, It is a signal vector. It is a noise vector. Then the incoherent covariance matrix... It can be used The calculation yielded:

[0037] (3)

[0038] In the formula, This indicates the conjugate transpose.

[0039] Because of the multipath effect in reality, the received signal is coherent. This is considered as a reference signal. Therefore, the remaining signal can be represented as a scaled version of the reference signal. ,in The complex fading coefficient represents the multipath propagation effect. It is the amplitude attenuation factor. This represents the phase difference.

[0040] Preprocessing:

[0041] From the received signal Obtain the sampling covariance matrix To mitigate the impact of noise terms on the rank of the sample covariance matrix and restore the true rank of the signal subspace, the sampling covariance matrix is... Sequential noise pre-whitening and normalization were performed. The implementation method is as follows:

[0042] (4)

[0043] (5)

[0044] Equation (4) is the pre-whitening operation of the sampling covariance matrix, that is, subtracting the noise term from the original sampling covariance matrix. Equation (5) is the normalization operation. Where, and Let these represent the pre-whitened and normalized covariance matrices, respectively. This represents the estimated noise power. express The identity matrix, This represents the Frobenius norm.

[0045] Next, regarding The real part, imaginary part, and phase are extracted as three input channels, and concatenated along the 0th dimension to form a channel with dimension . The three-dimensional tensor is used to obtain the covariance matrix of the coherent signal.

[0046] II. Reconstruction of Covariance Matrix of Coherent Signal

[0047] like Figure 2 As shown, a deep neural network containing the following modules was constructed:

[0048] 1. Generator Feature Extraction Module

[0049] The entire generator maintains a symmetrical encoder-decoder structure. Input data is first expanded from 3 to 32 channels through a convolutional layer, then enters the generator's feature extraction module. This module consists of cascaded encoder blocks (EBlocks). Each EBlock employs a simple gating mechanism (SG) and simplified channel attention (SCA) instead of traditional activation functions to reduce neuronal information loss, and introduces layer normalization and residual connections to ensure network convergence. Between EBlock combinations, convolutional kernels with a size of [missing information] are used. A two-dimensional convolution operation with a stride of 2 is used for downsampling to extract high-dimensional features of the covariance matrix of the coherent signal. To prevent loss of original information, skip connections are established between each encoder block (EBlock) and its corresponding subsequent decoder module. The mathematical expression for EBlock is...

[0050] (6)

[0051] Equation (6) represents the nonlinear mapping operation of EBlock, used for feature extraction. Where, and These represent the input and output features of EBlock, respectively. This indicates two-dimensional layer normalization. This represents a simplified channel attention mechanism.

[0052] 2. Generator Decoherence Module

[0053] The generator decoherence module consists of a decoder block (DBlock) and maintains structural symmetry with the generator feature extraction module. In DBlock, the input is the sum of the output of the corresponding encoder block EBlock and the output of the previous layer of the decoder block DBlock, then the values ​​are calculated using dilation factors. Dilated convolutions extract multi-scale features. The features from these three branches are fused through element-wise addition, and then sequentially passed through SG, SCA, and a gated multilayer perceptron (MLP) to enhance feature representation and minimize information loss. Before each decoder block (DBlock), pixel shuffle is used for upsampling, enabling the network to reconstruct fine-grained feature details, ultimately outputting a full-rank covariance matrix. The mathematical expression for DBlock is...

[0054] (7)

[0055] (8)

[0056] Equations (7) and (8) represent the nonlinear mapping operations of DBlock, used for decoherence. In the equations, , and These represent the input features of DBlock, the input features of the feedforward network in DBlock, and the output features of DBlock, respectively. This represents a feedforward network. This is obtained by adding the output of the corresponding encoder block EBlock to the output of the previous layer of the decoder block DBlock. Finally, the covariance matrix is ​​reconstructed. The output of the generator decoherence module is expressed as follows:

[0057] (9)

[0058] In the formula, and These represent the generator feature extraction module and the decoherence module, respectively.

[0059] 3. Discriminator

[0060] The discriminator's input is formed by concatenating a full-rank covariance matrix generated by the generator's decoherence module and the true incoherence covariance matrix; its dimension is... The discriminator employs a four-layer convolutional structure, with each layer having a kernel size, stride, and padding of 100%. The convolutional layers have 1, 2, and 1 channels, with 16, 32, 64, and 128 channels respectively. The SiLU activation function is used for non-linearity between each convolutional layer. Finally, the Sigmoid function is used to determine whether the generated covariance matrix can approximate the probability value of the true incoherent covariance matrix. Its mathematical expression is:

[0061] (10)

[0062] Equation (10) represents the discriminator's operation of distinguishing true from false in the generated covariance matrix. Where, This indicates a splicing operation. This represents a four-layer convolution operation. This represents the Sigmoid function.

[0063] III. DOA Estimation Implementation

[0064] Based on the reconstruction of the covariance matrix Eigenvalue decomposition will be performed to extract the noise subspace. The spatial spectrum is used in the calculation of the MUSIC spatial spectrum. The specific calculation of the spatial spectrum is as follows:

[0065] (11)

[0066] Finally, the estimated DOA angle value can be obtained through spectral peak search.

[0067] IV. Dataset Generation and Loss Function

[0068] 1. Dataset Construction

[0069] Use includes A uniform linear array of elements is used to acquire training and validation data. In this setup, the number of snapshots... Maintain at 400, signal-to-noise ratio value is For intervals, the range is from arrive The maximum angle of incidence is limited to Within a given range, information sources were paired to generate 7260 angle combinations. For each angle combination, corresponding coherent and uncorrelated signals were generated, serving as input and true values ​​for the generator. The entire dataset was structured according to... The dataset is divided into training and validation sets. When generating coherent signals, the amplitude attenuation coefficient is fixed at 1, while the phase difference is... It follows a uniform distribution within the range.

[0070] 2. Loss Function

[0071] To ensure stable convergence of generative adversarial networks with low computational complexity, and considering the special physical structure of the covariance matrix of coherent signals, this invention abandons the high computational cost of specific loss function designs in existing technologies (such as complex subspace orthogonality verification) and designs a dual-constraint loss function mechanism. This mechanism effectively solves the significant drawbacks of traditional complex loss functions, which impose huge computational burdens and reduce model training efficiency, by combining implicit rank distribution learning with explicit pixel-level constraints.

[0072] At the beginning of network training, this invention pre-generates a completely one matrix. and all-zero matrix To represent real and fake labels, we first concatenate the coherence covariance matrix. and the true incoherent covariance matrix The discriminator uses real sample pairs for forward propagation, enabling it to learn the relationship between the coherent covariance matrix and the real incoherent covariance matrix, thus obtaining the predicted values ​​of the real sample pairs. Then the coherence covariance matrix and reconstructing the covariance matrix The concatenated samples are used as dummy samples for forward propagation, enabling them to determine the reconstructed covariance matrix. Whether decoherence is completed and the predicted values ​​of fake sample pairs are obtained. Finally, based on the above... , , and To calculate the loss.

[0073] (1) Discriminator loss ( ):

[0074] This loss function aims to guide the network to implicitly learn the true distribution of the incoherent full-rank covariance matrix through a data-driven approach. Unlike traditional methods that require explicit deconvolution, the discriminator loss in this invention calculates the arithmetic mean of the classification losses for real sample pairs and generated sample pairs, enabling the network to autonomously determine whether the generated covariance matrix truly possesses full-rank properties in adversarial games. The specific formula is as follows:

[0075] (12)

[0076] In the formula, This represents the binary cross-entropy loss. This refers to the batch size.

[0077] (2) Generator loss ):

[0078] The generator loss is jointly composed of an adversarial loss term and a pixel-level reconstruction loss term. Given the physical characteristic that the array signal covariance matrix is ​​extremely sensitive to phase and amplitude, relying solely on the adversarial loss is insufficient to guarantee the absolute numerical accuracy of the matrix elements. Therefore, this method introduces a loss term based on a specific penalty coefficient. Pixel-level reconstruction loss forces the network to accurately reconstruct the high-dimensional physical features of the coherent signal at the pixel level (i.e., matrix element level). The specific formula is as follows:

[0079] (13)

[0080] in, express loss function This represents the penalty coefficient.

[0081] In particular, considering the complex environmental requirements of generative adversarial network design and coherent signals, this penalty coefficient... of The loss weight is specifically set to 30. This specific weight setting balances the recovery of the full-rank property of the covariance matrix at the macro level with the accuracy of the numerical values ​​of the matrix elements at the micro level. Through the above loss function design, this model does not need to perform tedious matrix orthogonalization operations during training and forward propagation. This design not only greatly reduces computational complexity but also gives the network a powerful covariance matrix reconstruction capability, enabling it to maintain DOA estimation accuracy far exceeding that of similar deep learning algorithms even under extremely low signal-to-noise ratios and limited snapshot numbers.

[0082] In another aspect, the present invention also provides a DOA estimation system based on generative adversarial networks (GANs) for implementing the aforementioned DOA estimation method based on GANs, comprising the following modules:

[0083] The data acquisition and processing module is used to acquire the time-domain coherent signal received by the uniform linear array, perform preprocessing, and generate the coherent signal covariance matrix and the corresponding incoherent covariance matrix, respectively.

[0084] The covariance matrix reconstruction module is used to construct a covariance matrix reconstruction model. Based on the covariance matrix of the coherent signal and the corresponding incoherent covariance matrix, it uses the generator of the generative adversarial network to extract deep features from the preprocessed coherent signal covariance matrix and reconstruct the covariance matrix. It also designs a dual-constraint loss function composed of implicit adversarial loss and pixel-level reconstruction loss with specific weights.

[0085] The DOA estimation output module uses the MUSIC algorithm to estimate DOA based on the reconstructed covariance matrix and then verifies it.

[0086] Experimental setup and hyperparameters:

[0087] 1. Network Design and Parameters

[0088] (1) Encoder block parameters

[0089] Table 1 shows the specific parameter settings for generating encoder blocks in the adversarial network using this method.

[0090] Table 1

[0091]

[0092] (2) Decoder block parameters

[0093] Table 2 shows the specific parameter settings for generating decoder blocks in the adversarial network using this method.

[0094] Table 2

[0095]

[0096] (3) Upsampling parameters

[0097] Table 3 shows the specific parameters for upsampling in the generative adversarial network of this method.

[0098] Table 3

[0099]

[0100] (4) Downsampling parameters

[0101] Table 4 shows the specific parameters for downsampling in the generative adversarial network of this method.

[0102] Table 4

[0103]

[0104] 2. Training Phase Configuration

[0105] Table 5 shows the specific parameter settings for training using this method.

[0106] Table 5

[0107]

[0108] The model method compared to the present invention (CMR-GAN) is the Residual Network (ResNet), where CMR-GAN-2 and ResNet-2 represent network inputs of two coherent signals, while CMR-GAN-3 and ResNet-3 represent network inputs of three coherent signals. Table 6 shows the specific model index parameters of the present invention model and the comparative model.

[0109] 1. Analysis of Model Parameter Quantity and Computational Complexity

[0110] The formula for calculating the total number of parameters of the generative adversarial network is as follows:

[0111] (14)

[0112] In the formula, For the number of generator parameters, The number of parameters is the number of discriminator parameters. Calculations show that the total number of parameters in the model of this invention is 1.79M, lower than the 6.99M of parameters in the comparative model.

[0113] The computational complexity of the model is measured by the number of floating-point operations (FLOPs) during the forward propagation process. The total model complexity is:

[0114] (15)

[0115] In the formula, The number of floating-point operations performed by the generator. This represents the number of operations performed by the discriminator. Calculations show that the model of this invention performs... The total FLOPs for the input dimension is 21.01M, which is far lower than the 1640.22M of the comparison model.

[0116] Table 6

[0117]

[0118] 2. Experimental Analysis

[0119] To verify the performance of the proposed method, this section designed and conducted a series of experiments. The experimental results and analysis are as follows:

[0120] (1) Experimental performance evaluation indicators

[0121] Signal-to-noise ratio (SNR) is defined as:

[0122] (16)

[0123] in, For signal power, This represents noise power.

[0124] The performance estimation criterion is the Joint Root Mean Square Error (RMSE), defined as follows:

[0125] (17)

[0126] in, For the first The Monte Carlo Trial The precise estimate of DOA for each source. Indicates the number of information sources. This indicates the number of Monte Carlo trials.

[0127] (2) Experimental results diagram

[0128] Figure 3 In signal-to-noise ratio Quick shot number The number of coherent signals were 2 and 3 respectively, and the number of Monte Carlo trials was [number missing]. The RMSE plot of the compared models in the scenario. From Figure 3 As can be seen from this, when the number of coherent signals is 2 and the incident angles are respectively At that time, the RMSE of both CMR-GAN and ResNet can be robustly reduced, and CMR-GAN consistently outperforms ResNet. Furthermore, the gap between the two gradually widens as the signal-to-noise ratio increases, all of which demonstrate that the covariance matrix recovered by CMR-GAN has better stability than that of ResNet. When the number of coherent signals increases to 3, with incident angles of... CMR-GAN can still maintain high estimation accuracy, but ResNet cannot correctly estimate coherent signals.

[0129] Figure 4 In the number of snapshots Signal-to-noise ratio The number of coherent signals were 2 and 3 respectively, and the number of Monte Carlo trials was [number missing]. The RMSE plot of the compared models in the scenario. From Figure 4 As can be seen from this, when there are two coherent signals and the incident angles are respectively At that time, the RMSE of both CMR-GAN and ResNet gradually decreased with the increase of the number of snapshots, and CMR-GAN consistently outperformed ResNet, while maintaining good estimation accuracy even under extreme snapshot counts. However, when the number of coherent signals increased to three, with incident angles of... ResNet also failed to estimate the angle of coherent signals, while CMR-GAN could still maintain a certain level of estimation accuracy. It is also worth noting that when the number of snapshots exceeds 100, CMR-GAN's RMSE is better than ResNet's when estimating three coherent signals, demonstrating that CMR-GAN can better recover the covariance matrix of coherent signals.

Claims

1. A DOA estimation method based on generative adversarial networks, characterized in that, Includes the following steps: S1: Obtain the time-domain coherent signal received by the uniform linear array, perform preprocessing, and generate the coherent signal covariance matrix and the corresponding incoherent covariance matrix respectively. S2: Construct a covariance matrix reconstruction model. Based on the covariance matrix of the coherent signal and the corresponding incoherent covariance matrix, use the generator of the generative adversarial network to extract deep features from the preprocessed covariance matrix of the coherent signal and reconstruct the covariance matrix. The covariance matrix reconstruction model includes a generator feature extraction module, a generator decoherence module, and a discriminator. The specific implementation process of the generator feature extraction module is as follows: The input data is first expanded to include more channels through a convolutional layer, and then enters the generator feature extraction module. This module consists of cascaded encoder blocks EBlock. Each EBlock uses a gating mechanism SG and channel attention SCA to replace the traditional activation function, and introduces layer normalization and residual connections. Between EBlock combinations, downsampling is performed through two-dimensional convolution operations, and skip connections are established between each encoder block EBlock and the corresponding subsequent decoder module. The specific implementation process of the generator decoherence module is as follows: The generator decoherence module consists of a decoder block DBlock and maintains structural symmetry with the generator feature extraction module. In DBlock, the input is the sum of the output of the corresponding encoder block EBlock and the output of the previous layer of the decoder block DBlock. Then, multi-scale features are extracted by dilated convolutions with three different dilation factors. The features of these three branches are fused by element-wise addition. Subsequently, the features are enhanced by SG, SCA, and gated multilayer perceptron and the information loss is minimized. Before different decoder blocks DBlock, pixel rearrangement is used for upsampling, and finally, a full-rank covariance matrix is ​​output. S3: DOA estimation is performed using the MUSIC algorithm based on the reconstructed covariance matrix, and the results are verified.

2. The DOA estimation method based on generative adversarial networks according to claim 1, characterized in that, The preprocessing is specifically implemented as follows: The uniform linear array antenna structure consists of It consists of several array elements, with the first sensor as the reference element, and the spacing between the array elements is... The wavelength is ; Assumption The sources are respectively Narrowband incoherent signals incident on a signal with On a uniform linear array of array elements, then the th array element... The sensor at the first The complex envelope received in the next snapshot is , For receiving signal matrix; using The incoherent covariance matrix was calculated. ; From the received signal Obtain the sampling covariance matrix For the sampling covariance matrix Sequential noise pre-whitening and normalization were performed to obtain the signal. ;right The real part, imaginary part, and phase are extracted as three input channels and concatenated along the 0th dimension to form a three-dimensional tensor, thus obtaining the covariance matrix of the coherent signal.

3. The DOA estimation method based on generative adversarial networks according to claim 2, characterized in that, The specific implementation process of the discriminator is as follows: The discriminator's input is formed by concatenating the full-rank covariance matrix generated by the generator's decoherence module with the true incoherent covariance matrix. The discriminator adopts a convolutional structure, and the SiLU activation function is used to nonlinearize the convolutional layers. Finally, the Sigmoid function is used to determine whether the generated covariance matrix approximates the probability value of the true incoherent covariance matrix.

4. The DOA estimation method based on generative adversarial networks according to claim 3, characterized in that, Step S2 also includes designing a dual-constraint loss function composed of implicit adversarial loss and pixel-level reconstruction loss of weights. The loss function of the covariance matrix reconstruction model is constructed as follows: Discriminator loss The aim is to guide the network to implicitly learn the true distribution of the incoherent full-rank covariance matrix through a data-driven approach. The discriminator loss is calculated by the arithmetic mean of the classification loss of the real sample pairs and the generated sample pairs, which enables the network to autonomously judge whether the generated covariance matrix truly has the full-rank property in adversarial games. Generator loss It consists of a joint adversarial loss term and a pixel-level reconstruction loss term, and introduces a penalty coefficient-based loss term. Pixel-level reconstruction loss.

5. The DOA estimation method based on generative adversarial networks according to claim 4, characterized in that, Step S3 is specifically implemented as follows: the full-rank covariance matrix generated by the generator. Eigenvalue decomposition will be performed to extract the noise subspace. It participates in the calculation of the MUSIC spatial spectrum, and finally obtains the estimated DOA angle value through spectral peak search.

6. A DOA estimation system based on generative adversarial networks, used to implement the DOA estimation method based on generative adversarial networks as described in any one of claims 1 to 5, characterized in that, Includes the following modules: The data acquisition and processing module is used to acquire the time-domain coherent signal received by the uniform linear array, perform preprocessing, and generate the coherent signal covariance matrix and the corresponding incoherent covariance matrix, respectively. The covariance matrix reconstruction module is used to construct a covariance matrix reconstruction model. Based on the covariance matrix of the coherent signal and the corresponding incoherent covariance matrix, it uses the generator of the generative adversarial network to extract deep features from the preprocessed coherent signal covariance matrix and reconstruct the covariance matrix. It also designs a dual-constraint loss function composed of implicit adversarial loss and pixel-level reconstruction loss of weights. The DOA estimation output module uses the MUSIC algorithm to estimate DOA based on the reconstructed covariance matrix and then verifies it.

Images