Deep learning based low snr doa estimation method

Abstract

The application discloses a low signal-to-noise ratio DOA estimation method based on deep learning, comprising the following steps: S1, generating an array signal data set and calculating a covariance matrix; S2, converting the DOA estimation problem into a multi-label classification task, discretizing the angle range into multiple grid points and constructing a multi-label binary vector as a supervised signal; constructing a convolution classification network, the network comprising a multi-scale convolution module, multiple cascaded residual modules with attention mechanisms, a pooling module and an output layer; training the network using the data set; S3, inputting the preprocessed to-be-estimated signal into the trained network to obtain the target existence probability of each angle grid, and obtaining the DOA estimation value through peak value detection. The application strengthens key feature extraction through multi-scale convolution and attention mechanism, maintains high-precision estimation under the condition of low signal-to-noise ratio and low snapshot, and simultaneously realizes lightweight network structure and is suitable for edge deployment.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, specifically relating to a low signal-to-noise ratio (DOA) estimation method based on deep learning. Background Technology

[0002] As an important branch of signal processing, array signal processing has developed rapidly in recent years, with applications spanning numerous military and national economic sectors, including radar, communications, and sonar. Direction-of-arrival (DOA) estimation is an indispensable part of array signal processing research, and its accuracy and robustness directly affect system performance. With increasingly complex application scenarios, especially under non-ideal conditions such as low signal-to-noise ratio (SNR) and low snapshot number, achieving high-precision DOA estimation has become a critical problem urgently needing to be solved in this field.

[0003] Traditional Direction of Arrival (DOA) estimation methods, particularly those based on subspace methods (such as MUSIC and ESPRIT), rely on accurate covariance matrix estimation in low signal-to-noise ratio (SNR) environments. However, at low SNR, noise significantly impacts the estimation of the covariance matrix, making it difficult to distinguish between the signal and noise subspaces and reducing estimation accuracy. Sparse representation-based methods (such as LASSO and BPDN) require precise adjustment of regularization parameters to achieve good performance; however, in practical applications, SNR is often unknown or variable, making parameter selection difficult and affecting estimation results. Furthermore, these methods typically involve solving complex optimization problems, resulting in a heavy computational burden and making them unsuitable for real-time processing.

[0004] In recent years, deep learning has been introduced into the field of DOA estimation due to its powerful feature extraction capabilities. Existing methods typically input the covariance matrix of the array's received signal directly into a neural network, allowing the network to learn a nonlinear mapping from the covariance matrix to the target angle. Compared to traditional methods, deep learning methods exhibit better robustness under low signal-to-noise ratio (SNR) conditions. However, existing deep learning methods still have the following shortcomings: First, the input feature processing methods are relatively simple; most methods directly feed the real and imaginary parts of the complex covariance matrix into the network, failing to fully exploit the structural information of the covariance matrix. Second, network designs often employ general convolutional structures, lacking specific optimizations for the characteristics of array signals, leaving room for improvement in multi-scale feature extraction and key feature enhancement. Third, existing methods often prioritize high accuracy while neglecting computational efficiency, resulting in large network parameters, slow inference speeds, and difficulty in deployment on resource-constrained edge devices. Therefore, how to effectively extract key signal features from noisy array received data under non-ideal conditions such as low SNR and low snapshot speed, and achieve high-precision, low-complexity DOA estimation, has become a core problem urgently needing to be solved in this field. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies and proposes a low signal-to-noise ratio (DOA) estimation method based on deep learning. By transforming the DOA estimation problem into a multi-label classification task, a residual network structure integrating multi-scale convolution and attention mechanisms is designed. This significantly reduces computational complexity while ensuring estimation accuracy, providing an effective solution for real-time DOA estimation in resource-constrained scenarios (such as low snapshot, finite array element, and low signal-to-noise ratio scenarios).

[0006] This invention proposes a deep learning-based method for estimating low signal-to-noise ratio (DOA), comprising the following steps:

[0007] Step S1: Generate an array signal dataset for training and testing, and preprocess the dataset;

[0008] Step S2: Problem Transformation, Network Construction, and Training

[0009] Step S201 constructs a multi-label classification task, transforming the DOA estimation problem into a multi-label classification problem: the angle range is discretized into multiple grid points, each grid point corresponds to a category label, and a multi-label binary vector is constructed based on the true DOA angle as a supervision signal;

[0010] Step S202: Construct a convolutional classifier network, the network comprising:

[0011] ① The multi-scale convolution module uses multiple convolution kernels of different sizes to extract multi-scale features of the input tensor in parallel and then fuses the multi-scale features;

[0012] ② Multiple cascaded residual modules, each including a convolutional layer and an attention mechanism, are used to enhance feature representation;

[0013] ③ Pooling module, used for downsampling feature maps and global feature aggregation;

[0014] ④ Output layer, used to output the probability of target presence for each angle grid;

[0015] Step S203: Use the dataset obtained in step S1 and the multi-label binary vector constructed in step S201 to train the convolutional class network to obtain the trained network;

[0016] Step S3: After processing the array signal to be estimated through step S1 to obtain sample data, input it into the trained convolutional classification network in step S202 to obtain the target existence probability of each discrete angle category, and obtain the DOA estimate through peak detection.

[0017] Preferably, in step S202, the multi-scale convolution module uses four convolution kernels of different sizes to perform parallel convolution, the four sizes being 1×1, 3×3, 5×5, and 7×7, respectively.

[0018] Preferably, in step S202, the feature maps output by the convolutional kernels of each size are first subjected to stepwise cross-scale interaction and fusion through the MSC module, and then spliced ​​and fused before being subjected to channel compression and cross-channel information fusion through a 1×1 convolutional layer.

[0019] Preferably, the MSC module includes multiple window average pooling layers of different sizes and a sparse cross-attention mechanism: firstly, global structural information and local detailed features in the feature map are captured through pooling windows of different sizes, and the pooled features are summed and flattened; then, query Q, key K and value V vectors are generated through linear projection, and attention weights are calculated using sparse cross-attention.

[0020] Preferably, the residual module is an MSPA residual block, and each MSPA residual block includes two convolutional layers and a multi-head self-attention module. The multi-head self-attention module weights the channel dimensions of the feature map and adds the input and output through residual connections.

[0021] Preferably, the multi-head self-attention module includes:

[0022] ① The hierarchical phantom convolution submodule extracts multi-scale spatial information from the input feature map through grouping and hierarchical convolution operations, generating an enhanced multi-scale feature set;

[0023] ② The spatial pyramid recalibration submodule is used to perform cross-dimensional feature interaction on the multi-scale feature set output by the hierarchical phantom convolution submodule and learn channel-level attention weights; the spatial pyramid recalibration submodule adopts a dual-path pooling fusion strategy, including parallel global adaptive average pooling branch and local adaptive max pooling branch, which respectively aggregate global context information and local salient detail features, and after fusion, a preliminary channel attention tensor is generated through pointwise convolution;

[0024] ③ Cross-channel recalibration unit, used to recalibrate the preliminary channel attention tensor through the Softmax activation function, multiply it element-wise with the original multi-scale feature map to achieve feature reweighting, and add it to the input of the MSPA residual block through residual connection.

[0025] Preferably, the MSPA residual blocks are four in number, cascaded sequentially, with each residual block outputting a feature map of constant size and a fixed number of channels of 128.

[0026] The technical solution of the present invention has at least the following technical effects:

[0027] 1. This invention transforms the DOA estimation problem into a multi-label classification task and designs a deep neural network that integrates multi-scale convolution and attention mechanisms, achieving high-precision DOA estimation under low signal-to-noise ratio (SNR) conditions. Example verification: With L=100 snapshots and M=16 array elements, the method of this invention exhibits significantly lower root mean square error (RMSE) than traditional methods such as CBF, MUSIC, and ESPRIT over a wide SNR range of -20dB to 5dB. Particularly at a 5dB SNR, the RMSE of this method drops below 0.3°, entering the error saturation region, while traditional methods remain at a relatively high error level. This effect is attributed to the MSC module's effective suppression of noise interference and enhancement of cross-scale feature correlation through multi-scale pooling and sparse cross-attention mechanisms. Simultaneously, the HPC submodule in the MSPA residual block extracts multi-scale spatial information, and the SPDR submodule generates channel attention weights through dual-path pooling, enabling the network to accurately focus on signal-related features and significantly improve angle estimation accuracy under strong noise backgrounds.

[0028] 2. This invention uses end-to-end training, eliminating the need for precise covariance matrix estimation. It directly learns the mapping from received data to angle probabilities through the network, maintaining stable estimation accuracy even under low snapshot conditions. This solves the technical problem of inaccurate covariance matrix estimation and drastic performance degradation in traditional subspace-based methods when the number of snapshots is small (e.g., L≤100). Experiments show that with L=100 snapshots, this method can still effectively distinguish multiple targets at a low signal-to-noise ratio of -20dB, verifying its robustness in data-constrained scenarios.

[0029] 3. The multi-scale convolution module of this invention uses parallel small convolutional kernels (1×1, 3×3, 5×5, 7×7) to extract features, avoiding the surge in parameters caused by using large-size convolutional kernels; the MSPA residual block enhances features through residual connections and channel attention mechanisms without increasing network depth; the pooling module uses a combination of max pooling and average pooling to achieve efficient dimensionality reduction while preserving salient features. The overall network parameter count is controlled in the millions, enabling millisecond-level inference on edge computing platforms such as the NVIDIA Jetson series, meeting the real-time requirements of scenarios such as drones and autonomous driving.

[0030] 4. This invention does not require complex preprocessing steps. It only requires separating the real and imaginary parts of the covariance matrix before inputting it into the network to achieve end-to-end training. During the training process, the multi-label binary cross-entropy loss function guides the network to adaptively learn the statistical characteristics of signals and noise, avoiding the manual parameter tuning process in traditional methods and improving the practicality and generalization ability of the method.

[0031] 5. This invention naturally supports multi-target DOA estimation through multi-label classification modeling and peak detection post-processing. The multi-scale feature interaction of the MSC module and the channel attention mechanism of the MSPA residual block enable the network to effectively distinguish neighboring targets and accurately distinguish them even when multiple targets exist simultaneously and the angular intervals are small.

[0032] 6. This invention addresses the problem of low signal-to-noise ratio (DOA) estimation in scenarios with low snapshot speed and finite array elements. Through innovative network structure design, it achieves comprehensive advantages in estimation accuracy, computational efficiency, and multi-target resolution, providing an effective technical solution for real-time DOA estimation in resource-constrained environments. Attached Figure Description

[0033] Figure 1 This is a flowchart of the low signal-to-noise ratio (DOA) estimation method based on deep learning according to the present invention;

[0034] Figure 2 This is a network structure diagram of the low signal-to-noise ratio (DOA) estimation method based on deep learning in this invention.

[0035] Figure 3 This invention demonstrates the performance of different DOA estimation methods under low SNR conditions.

[0036] Figure 4 This invention provides the DOA estimation performance of the model under multiple objectives. Detailed Implementation

[0037] This invention provides a low signal-to-noise ratio (DOA) estimation method based on deep learning, addressing the problems of inaccurate covariance matrix estimation, difficulty in multi-target discrimination, and excessive computational complexity of existing deep learning methods in low-shot, finite-element scenarios. The overall technical approach is as follows:

[0038] First, the array signal dataset is generated and the covariance matrix is ​​calculated in step S1 to provide basic data for subsequent processing. Step S201 transforms the DOA estimation problem into a multi-label classification task: the continuous angle space is discretized into multiple grid points, each grid point corresponds to a category, and a multi-label binary vector is constructed based on the real DOA angle as a supervision signal, so that the network can learn multi-target angle information in a classification manner. Step S202 constructs the convolutional classification network MSPDOANet, which achieves feature extraction and enhancement through multi-module collaboration: the multi-scale convolutional module uses four convolutional kernels of different sizes (1×1, 3×3, 5×5, and 7×7) to extract multi-scale features of the input tensor in parallel, capturing signal patterns at different spatial scales; the MSC module performs step-by-step cross-scale interaction and fusion of features at each scale, enhancing the correlation between features; the MSPA residual block uses a multi-head self-attention mechanism to weight the channel dimension of the feature map, strengthening the key feature responses related to the signal, while ensuring the training stability of the deep network through residual connections; in the pooling module, the max pooling layer retains significant local features, and the average pooling layer aggregates global information, achieving feature dimensionality reduction and convergence; the output layer outputs the target existence probability of each angle grid through the sigmoid activation function; Step S203 trains the network using the dataset, making the output probability approximate the supervision signal. Step S3, in the inference stage, the signal to be estimated is preprocessed and input into the trained network, and peak detection is used to obtain the DOA estimates of multiple targets. The embodiments of the present invention achieve multi-objective DOA estimation through the synergistic effect of the above modules: multi-scale convolution provides a rich feature base, the MSC module realizes cross-scale information fusion, the MSPA residual block strengthens key features through the attention mechanism, the pooling module efficiently reduces dimensionality, and finally, peak detection achieves multi-objective DOA estimation.

[0039] The overall technical solution provided by the embodiments of the present invention maintains high-precision estimation under low signal-to-noise ratio and low snapshot conditions, while the network structure is lightweight and suitable for edge deployment.

[0040] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0041] Example 1

[0042] The steps of the deep learning-based low signal-to-noise ratio (DOA) estimation method are as follows:

[0043] Step S1 aims to generate an array signal dataset for training and testing, and preprocess it to obtain a format suitable for deep neural network input. The specific implementation is as follows:

[0044] Traditional signal models for DOA estimation mostly obtain the source incident angle by establishing a model that gives the array output signal vector product array steering matrix. Assume there are K far-field narrowband incoherent target signal sources. and the direction of the incoming wave An incident signal is projected onto an arbitrary M-element uniform linear array with an element spacing of d. It can be represented as:

[0045]

[0046] Where: , indicating that the mean is 0 and the variance is Gaussian white noise; The array incident angle is The steering vector of the incident signal; is the guiding matrix of the array.

[0047]

[0048]

[0049] Where: λ is the signal wavelength; d is the element spacing.

[0050] However, array processing algorithms do not directly estimate the direction of arrival from the actual output of the array elements; instead, they use the covariance matrix. Therefore, the M-row, M-column spatial covariance matrix of the array received signal can be expressed by the following formula:

[0051]

[0052] Where E[·] and [·] H and represent the statistical expectation and Hermitian transpose, respectively. It is the covariance matrix of a K-row, K-column signal. It is the covariance matrix of the M rows and M columns of noise.

[0053] Assume that the noise signals received by different array elements are additive white Gaussian noise, and are independent of s(t). The mean of this noise is 0, and the variance is... Therefore, we can obtain:

[0054]

[0055] Array output The covariance matrix is:

[0056]

[0057] In the formula: E[ ] represents the mathematical expectation; I is the identity matrix.

[0058] In practical applications, since precise values ​​of statistical expectation are often unavailable, finite snapshot data is typically used to estimate the covariance matrix. Let the number of snapshots be L, then the sample covariance matrix... The calculation formula is:

[0059]

[0060] In the formula, For the first The received signal vector at each snapshot moment, This indicates finding its conjugate transpose. In subsequent deep learning model training and inference steps, the sample covariance matrix calculated using this formula will be used. This serves as the basic input data. To reduce computational complexity and preserve complete information, the upper triangular region of the sample covariance matrix is ​​extracted and separated into real and imaginary parts, which are then input into the deep neural network as a two-channel real tensor.

[0061] In this embodiment, the covariance matrix has undergone appropriate preprocessing. As input to the DOA activation network and estimation network.

[0062] Sufficient samples are generated using simulation to meet training requirements: the number of array elements (e.g., 16 elements) and the number of snapshots (e.g., 100) are set, and the number of targets (maximum number can be set, e.g., 5) and their corresponding true DOA angles are randomly selected. Noisy complex array signals are generated within different SNR ranges (e.g., -20dB to 30dB), and the signal covariance matrix is ​​calculated as input to the deep learning model. Figure 3 As shown.

[0063] Step S2: Problem Transformation, Network Construction, and Training

[0064] Step S201 Problem Transformation: Map the DOA to discrete categories for prediction and estimation.

[0065] The core idea of ​​step S201 is to discretize the continuous angle space into a set of fixed categories, with each category corresponding to a predefined angle grid point, thereby mapping the target's arrival angle to a category label. The specific steps are as follows:

[0066] Search the entire angle range [ With fixed step size Divide the grid into C discrete categories at equal intervals to form an angular grid set:

[0067]

[0068] Each category c represents a specific angular range.

[0069] Label construction, for each sample, the actual target angle. Map it to the corresponding category index set That is, select the discrete category that is closest to the true angle.

[0070] Constructing multi-label binary vectors .

[0071]

[0072] Model input and output: The input is the covariance matrix of the array received signals. The complex covariance matrix is ​​split into real and imaginary parts, which are then used as two channels to form a real tensor input to the deep neural network. This preprocessing method fully preserves the magnitude and phase information of the covariance matrix with minimal computational overhead. The neural network outputs a real vector of length C. Each element represents the probability that the angle category is a target.

[0073] During training, the output probability is calculated using a loss function. Optimize network parameters by identifying the difference between the multi-label binary vector y and the supervisory signal.

[0074] Step S202: Construct a convolutional integral class network

[0075] The convolutional class network MSPDOANet employs a multi-scale deep residual network structure, which mainly consists of the following parts:

[0076] Multi-scale Convolutional Module: Input features are first processed through four convolutional layers with different kernel sizes (1×1, 3×3, 5×5, 7×7) to extract multi-scale features. Each branch's features then enter the corresponding multi-scale sparse cross-attention module (MSC). The MSC module's structure and function are as follows: it contains multiple average pooling layers with windows of different sizes and a sparse cross-attention mechanism. First, it captures global structural information and local detail features in the feature map through pooling windows of different sizes, then sums and flattens the pooled features. Subsequently, it generates query (Q), key (K), and value (V) vectors through linear projection, and calculates attention weights using sparse cross-attention. Its function is to effectively suppress interference from irrelevant noise information and achieve adaptive weighting of signal features across receptive fields. After processing by the MSC module, the output features of the four branches are concatenated and fused, and then compressed and adjusted through a 1×1 convolutional layer to achieve deep information fusion across channels.

[0077] MSPA Residual Blocks: The network contains four MSPA residual blocks, each consisting of two convolutional layers followed by an activation layer, and a multi-head self-attention module is introduced into the residual connection structure. The multi-head self-attention module includes hierarchical phantom convolution (HPC) and spatial pyramid recalibration (SPDR) branches, adaptively weighting the channel dimensions of the feature maps. This allows the network to accurately focus on feature channels highly correlated with the real signal, suppressing the responses of purely noisy channels. Simultaneously, by directly adding the input to the weighted output through standard residual connections (i.e., skip connections), the model's feature representation capability is greatly enhanced, and gradient stability during deep network training is ensured. Throughout the cascading process, the spatial size of the feature map output by each MSPA residual block remains unchanged, and the number of feature channels is fixed at 128.

[0078] The internal structure and feature processing flow of the multi-head self-attention module consist of the following three core sub-modules working in tandem:

[0079] (1) Hierarchical phantom convolution (HPC) module:

[0080] The intermediate feature maps input to the MSPA module first enter the HPC module. Unlike the layer-by-layer feature extraction with a single receptive field in traditional convolutional neural networks (CNNs), the HPC module is configured to efficiently extract multi-scale spatial information from the input feature maps in a finer-grained manner. This module generates an enhanced multi-scale feature set containing rich spatial scale cues through grouping and hierarchical convolution operations. In the DOA estimation task, this mechanism can simultaneously capture the local phase differences between adjacent elements in the covariance matrix and the global structural features of the entire array manifold.

[0081] (2) Space Pyramid Recalibration (SPDR) module:

[0082] The multi-scale feature set output from the HPC module then enters the SPDR module to establish cross-dimensional feature interactions and learn channel-level attention weights. To overcome information loss caused by a single pooling method, this module employs a dual-path pooling fusion strategy:

[0083] (2.1) Global branch: Adaptive average pooling (AAP) is applied to aggregate the global context information of the entire feature map to achieve structure regularization;

[0084] (2.2) Local branch: Apply adaptive max pooling (AMP) in parallel to preserve local significant details (such as strong correlation peaks formed by the target signal in the covariance matrix).

[0085] After upsampling and feature fusion, the aforementioned global and local information is used to learn the cross-dimensional correlation between channels through two computationally efficient pointwise convolutional layers (1×1 convolution), thereby generating a preliminary channel attention tensor.

[0086] (3) Cross-channel recalibration and residual connection:

[0087] To further facilitate the adaptive selection of local and global channel attention, the initial attention tensor is recalibrated using the Softmax activation function to establish long-distance dependencies across channels, generating the final adaptive channel-level attention coefficients. These coefficients are then element-wise multiplied with the original multi-scale feature map to reweight the features—that is, the network autonomously learns to focus on high-response channels containing real target signals and significantly attenuates redundant channels dominated by Gaussian white noise. Finally, the attention-weighted features are element-wise added to the original input of the residual block via a skip connection. This residual aggregation mechanism not only endows the network with extremely strong feature representation and multi-target discrimination capabilities but also effectively avoids the gradient vanishing problem during deep network training, ensuring the model's training stability and convergence speed under conditions of low signal-to-noise ratio and few snapshots.

[0088] Pooling module: First, a max pooling layer is used for local feature downsampling and dimensionality reduction, and then an average pooling layer is used for global feature aggregation. The max pooling layer and the average pooling layer respectively play the roles of preserving salient features and aggregating global information, thereby improving the network's discriminative performance and generalization ability.

[0089] Output layer: Finally, a fully connected layer outputs a dimension... The feature vector is a real number corresponding to the preset total number of discrete angle grids, and then processed by the Sigmoid activation function. Since this invention deals with a multi-objective DOA estimation problem (i.e., a multi-label classification task), the Sigmoid function can independently map the features of each angle grid to... The confidence probability between them truly reflects the state where multiple signal sources exist simultaneously in space.

[0090] Step S202 enables the network to simultaneously capture spatial patterns at different scales (through multi-scale convolution) and enhances the response to signal-related features through an attention mechanism. Meanwhile, residual connections ensure the training stability of deep networks, and the output layer uses a fully connected layer to map the extracted high-dimensional features to... The feature vector is then processed by the Sigmoid activation function to output the independent probability of the target's existence on each discrete angle grid.

[0091] Step S203: Train the integral class network

[0092] The generated dataset is divided into training and testing sets in a 10:1 ratio. The convolutional neural network is then trained to obtain the trained training stage. Multi-label binary cross-entropy loss (BCE) is used, with a single-sample loss of:

[0093]

[0094] Based on the loss value calculated above, the network parameters are iteratively updated using an adaptive moment estimator (such as Adam) combined with the backpropagation algorithm to minimize the above loss function until the loss of the model on the validation set no longer decreases or reaches the preset maximum number of iterations. At this point, the model is considered to have converged, and the trained network weights are saved.

[0095] Step S3: Inference and DOA Estimation

[0096] This step uses the trained MSPDOANet to estimate the DOA of the new array signal, as specifically implemented below:

[0097] During the inference phase, the received array signal X to be estimated is processed through step S1 to obtain sample data, and then fed into the trained convolutional classification network in step S202 for prediction. The network output is processed by the sigmoid activation function to obtain the target existence probability for each discrete angle category.

[0098] Finally, a peak detection algorithm is used to identify multiple local maxima in the probability spectrum. The angles corresponding to these peaks are the estimated destinations (DOA) of multiple targets.

[0099] Appendix Figure 3 Performance of different DOA estimation methods under low SNR. For example... Figure 3As shown, simulation results demonstrate the noise robustness of each DOA algorithm. The root mean square error (RMSE) performance under different SNR conditions was compared and analyzed. The proposed method significantly outperforms traditional methods across the entire SNR range. Particularly under low SNR conditions (-20dB to 5dB), the proposed method maintains low error and exhibits strong noise robustness. In contrast, CBF, MUSIC, and ESPRIT show significantly higher errors in this range, with some methods only converging at higher SNR values. The proposed method enters the error saturation region at a SNR of 5dB, with its RMSE dropping below 0.3, far lower than other methods, indicating that the proposed method possesses strong robustness and noise resistance.

[0100] Appendix Figure 4 This describes the DOA estimation performance of the model under multiple objectives in this invention. Figure 4 As shown, the method of this invention can still accurately estimate DOA even under multi-objective conditions. Experimental results demonstrate that the proposed method performs better and is more robust under low signal-to-noise ratio conditions, making it adaptable to various scenarios.

[0101] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A low signal-to-noise ratio (DOA) estimation method based on deep learning, characterized in that, Includes the following steps: Step S1: Generate an array signal dataset for training and testing, and preprocess the dataset; Step S2: Problem Transformation, Network Construction, and Training Step S201 constructs a multi-label classification task, transforming the DOA estimation problem into a multi-label classification problem: the angle range is discretized into multiple grid points, each grid point corresponds to a category label, and a multi-label binary vector is constructed based on the true DOA angle as a supervision signal; Step S202: Construct a convolutional classifier network, the network comprising: ① The multi-scale convolution module uses multiple convolution kernels of different sizes to extract multi-scale features of the input tensor in parallel and then fuses the multi-scale features; ② Multiple cascaded residual modules, each including a convolutional layer and an attention mechanism, are used to enhance feature representation; ③ Pooling module, used for downsampling feature maps and global feature aggregation; ④ Output layer, used to output the probability of target presence for each angle grid; Step S203: Use the dataset obtained in step S1 and the multi-label binary vector constructed in step S201 to train the convolutional class network to obtain the trained network; Step S3: After processing the array signal to be estimated through step S1 to obtain sample data, input it into the trained convolutional classification network in step S202 to obtain the target existence probability of each discrete angle category, and obtain the DOA estimate through peak detection.

2. The low signal-to-noise ratio (DOA) estimation method as described in claim 1, characterized in that, In step S202, the multi-scale convolution module uses four convolution kernels of different sizes to perform parallel convolution, and the four sizes are 1×1, 3×3, 5×5 and 7×7 respectively.

3. The low signal-to-noise ratio (DOA) estimation method as described in claim 2, characterized in that, In step S202, the feature maps output by the convolutional kernels of each size are first subjected to stepwise cross-scale interaction and fusion through the MSC module, and then spliced ​​and fused before being subjected to channel compression and cross-channel information fusion through a 1×1 convolutional layer.

4. The low signal-to-noise ratio (DOA) estimation method as described in claim 3, characterized in that, The MSC module includes multiple window average pooling layers of different sizes and a sparse cross-attention mechanism: First, it captures global structural information and local detailed features in the feature map through pooling windows of different sizes, and then sums and flattens the pooled features; then, it generates query Q, key K and value V vectors through linear projection, and calculates attention weights using sparse cross-attention.

5. The low signal-to-noise ratio (DOA) estimation method as described in claim 1, characterized in that, The residual module is an MSPA residual block. Each MSPA residual block includes two convolutional layers and a multi-head self-attention module. The multi-head self-attention module weights the channel dimensions of the feature map and adds the input and output through residual connections.

6. The low signal-to-noise ratio (DOA) estimation method as described in claim 5, characterized in that, The multi-head self-attention module includes: ① The hierarchical phantom convolution submodule extracts multi-scale spatial information from the input feature map through grouping and hierarchical convolution operations, generating an enhanced multi-scale feature set; ② Spatial pyramid recalibration submodule, used to perform cross-dimensional feature interaction on the multi-scale feature set output by the hierarchical phantom convolution submodule and learn channel-level attention weights; The spatial pyramid recalibration submodule adopts a dual-path pooling fusion strategy, including a parallel global adaptive average pooling branch and a local adaptive max pooling branch, which respectively aggregate global context information and local salient detail features. After fusion, a preliminary channel attention tensor is generated through pointwise convolution. ③ Cross-channel recalibration unit, used to recalibrate the preliminary channel attention tensor through the Softmax activation function, multiply it element-wise with the original multi-scale feature map to achieve feature reweighting, and add it to the input of the MSPA residual block through residual connection.

7. The low signal-to-noise ratio (DOA) estimation method as described in claim 5, characterized in that, The MSPA residual blocks consist of four blocks, which are cascaded sequentially. The output feature map size of each residual block remains unchanged, and the number of channels is fixed at 128.

Images