Direction of arrival determination method and apparatus, electronic device, and storage medium

Abstract

The application provides a direction of arrival determination method, device, electronic equipment and storage medium, wherein the direction of arrival determination method comprises: obtaining a vector covariance matrix of a snapshot data of a to-be-tested signal based on the snapshot data; obtaining a noise residual graph based on the vector covariance matrix; obtaining a de-noised covariance feature based on the noise residual graph and the vector covariance matrix; and determining the direction of arrival of the to-be-tested signal based on the de-noised covariance feature. The application can improve the accuracy of direction of arrival estimation based on a noise residual graph.

Description

Technical Field

[0001] This application relates to the field of array signal processing, and more specifically, to a method, apparatus, electronic device, and storage medium for determining direction of arrival. Background Technology

[0002] In recent years, the continuous advancements in deep learning theory and methods have opened up entirely new research avenues for Direction of Arrival (DOA) estimation. By constructing a mapping relationship between array-received data and signal incident angles, data-driven methods effectively break free from the dependence of traditional algorithms on precise physical models and exhibit stronger adaptability in complex and ever-changing real-world environments. Introducing convolutional neural networks into DOA estimation has significantly improved the accuracy of angle estimation. Some studies have enhanced the robustness of models to covariance features in the presence of interference and array errors by embedding channel attention mechanisms in residual networks; others have attempted to integrate classical spectral estimation ideas with deep learning to further improve performance. Furthermore, carefully designing training samples for specific application scenarios has also proven to effectively improve the model's generalization ability and estimation accuracy. Other methods combine neural networks with the gradient update process of the likelihood function, maintaining good performance even under conditions of limited snapshot numbers; still others employ strategies that combine classification and regression tasks to effectively address the estimation challenges in multi-source scenarios. Although these methods perform well in high signal-to-noise ratio (SNR) environments, their overall performance—especially the accuracy of direction-of-arrival (DOA) estimation—often declines significantly under low SNR conditions, which remains a key issue that urgently needs to be addressed. Summary of the Invention

[0003] The purpose of this application is to provide a method, apparatus, electronic device, and storage medium for determining direction of arrival, so as to improve the above-mentioned problems existing in the prior art.

[0004] In a first aspect, the present invention provides a direction-of-arrival (DOA) determination method, which is applied to a deep learning model. The DOA determination method includes: Based on the snapshot data of the signal under test, the vector covariance matrix of the snapshot data is obtained; Based on the vector covariance matrix, the noise residual map is obtained; Based on the noise residual map and the vector covariance matrix, the denoised covariance features are obtained; and Based on the denoised covariance characteristics, the direction of arrival of the signal under test is determined.

[0005] In the above implementation process, a noise residual map reflecting the covariance distortion caused by finite snapshot data and noise can be obtained. Then, based on the noise residual map, the vector covariance matrix can be processed to obtain the denoised covariance features. Finally, a more accurate direction of arrival can be obtained based on the denoised covariance features.

[0006] In an optional implementation, obtaining the noise residual map based on the vector covariance matrix includes: Convert the vector covariance matrix into a two-dimensional matrix form of the covariance matrix; The covariance matrix in two-dimensional form is convolved to obtain the noise residual map.

[0007] In the above implementation process, before performing convolution processing to obtain the noise residual map, the vector covariance matrix can be converted into a two-dimensional matrix form of covariance matrix, thereby... By preserving the inherent spatial structure information in the vector covariance matrix and avoiding the loss of spatial structure information, we can avoid obtaining a noise residual map based on the covariance matrix after the loss of spatial structure information, thereby improving the calculation accuracy of the noise residual map.

[0008] In an optional implementation, obtaining the denoised covariance features based on the noise residual map and the vector covariance matrix includes: Convert the vector covariance matrix into a two-dimensional matrix form of the covariance matrix; The covariance matrix in two-dimensional form is subtracted from the noise residual map to obtain the denoised covariance features.

[0009] In the above implementation process, by subtracting the covariance matrix in the form of a two-dimensional matrix from the noise residual map, the computational load of the denoising operation can be reduced while obtaining the denoised covariance features.

[0010] In an optional implementation, determining the direction of arrival of the signal under test based on the denoised covariance features includes: The denoised covariance features are flattened to obtain flattened covariance features. The flattened covariance features are added to the vector covariance matrix to obtain the summed covariance features. Based on the covariance characteristics after addition, the direction of arrival of the signal under test is obtained.

[0011] In the above implementation process, adding the denoising result to the original input can supplement the array structure information again, thereby compensating for the destruction of array structure information by excessive denoising during the denoising process and obtaining covariance features with higher fidelity.

[0012] In an optional implementation, obtaining the direction of arrival of the signal under test based on the summed covariance features includes: Based on the summed covariance features, time-domain features and spatial-domain features are extracted. Obtain input condition parameters and determine the gating weight coefficients corresponding to the input condition parameters, wherein the gating weight coefficients characterize the contribution ratio of the time domain features and the spatial domain features to the direction of arrival under the premise of using the input condition parameters; Based on the gating weight coefficients, the time-domain features and the spatial-domain features are weighted and combined to obtain fused features; The direction of arrival of the signal under test is determined based on the fusion features.

[0013] In the above implementation process, after the covariance features are denoised and supplemented, based on the highly correlated covariance features, the gating weight coefficients are further combined to control the contribution of time-domain features and spatial-domain features to the direction of arrival calculation under different input conditions. This achieves adaptive weighted fusion of time-related features and spatial-related features, and ultimately, under different input conditions such as signal-to-noise ratio and data conditions, the relative contribution of time-related features and spatial-domain features can be flexibly adjusted to adapt to different scenarios and improve robustness.

[0014] In an optional implementation, obtaining the input condition parameters and determining the gating weight coefficients corresponding to the input condition parameters includes: The time-domain features and the spatial-domain features are concatenated to obtain the original joint input; The original joint input is subjected to a nonlinear transformation to obtain candidate enhanced feature representations; The gating weight coefficients are obtained based on the original joint input and the candidate enhanced feature representation.

[0015] In the above implementation process, by concatenating the time domain features and the spatial domain features, and performing nonlinear transformation processing on the concatenated original input, a candidate enhancement feature representation can be obtained. Then, based on the original joint input and the candidate enhancement feature representation, the gating weight coefficient can be obtained.

[0016] In an optional implementation, determining the direction of arrival (DOA) of the signal under test based on the fusion features includes: The fused features are subjected to global average pooling to obtain compressed features; Based on the compressed features and weight matrix, a channel weight vector is obtained, wherein the weight matrix is ​​used to characterize the degree of influence of each feature dimension on the direction of arrival estimation; the greater the degree of influence, the greater the corresponding weight. Multiply the channel weight vector by the fused feature to obtain the enhanced feature vector; Based on the enhanced feature vector, the direction of arrival of the signal under test is obtained.

[0017] In the above implementation process, an attention mechanism can be introduced through the channel weight vector, and then the attention recalibration mechanism can be used to highlight the more critical discrimination information for direction of arrival estimation, thereby improving the accuracy of direction of arrival estimation.

[0018] Secondly, the present invention provides a direction-of-arrival (DOA) determination device, the device being applied to a deep learning model, the DOA determination device comprising: The covariance calculation module is used to obtain the vector covariance matrix of the snapshot data based on the snapshot data of the signal under test; The noise residual processing module is used to obtain a noise residual map based on the vector covariance matrix. The denoising module is used to obtain the denoised covariance features based on the noise residual map and the vector covariance matrix; The determination module is used to determine the direction of arrival of the signal under test based on the denoised covariance features.

[0019] In the above implementation process, the device can obtain a noise residual map that reflects the covariance distortion caused by limited snapshot data and noise. Then, based on the noise residual map, the vector covariance matrix is ​​processed to obtain the denoised covariance features. Finally, a more accurate direction of arrival is obtained based on the denoised covariance features.

[0020] Thirdly, the present invention provides an electronic device, comprising: Processor; and The memory is configured to store machine-readable instructions that, when executed by the processor, perform the direction-of-arrival determination method as described in any of the foregoing embodiments.

[0021] In the above implementation process, the processor calls the program instruction in real time to obtain a noise residual map that reflects the covariance distortion caused by finite snapshot data and noise. Then, based on the noise residual map, the vector covariance matrix is ​​processed to obtain the denoised covariance features. Finally, a more accurate direction of arrival is obtained based on the denoised covariance features.

[0022] Fourthly, the present invention provides a storage medium storing a computer program, the computer program being executed by a processor using the direction of arrival determination method as described in any of the foregoing embodiments.

[0023] In the above implementation process, the storage medium completely solidifies the direction of arrival determination method into computer program instructions; when the medium is read and run by the processor, the processor obtains the vector covariance matrix of the snapshot data of the signal under test according to the instructions, obtains the noise residual map based on the vector covariance matrix, obtains the denoised covariance features based on the noise residual map and the vector covariance matrix, and determines the direction of arrival of the signal under test based on the denoised covariance features, thereby realizing the determination of the direction of arrival based on the denoised covariance features and improving the accuracy of the direction of arrival calculation. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating a method for determining direction of arrival provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a direction of arrival determination device provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application; Figure 4 This is a schematic diagram of a DSDE-Ne diagram provided in an embodiment of this application; Figure 5 This is a schematic diagram of a residual denoising network provided in an embodiment of this application; Figure 6 This is a schematic diagram of a dynamic gating fusion network provided in an embodiment of this application; Figure 7a This is a schematic diagram of the direction of arrival estimation result obtained according to the direction of arrival determination method in the embodiments of this application; Figure 7b This is a schematic diagram illustrating the estimation error of the direction of arrival estimation result obtained according to the direction of arrival determination method in the embodiments of this application; Figure 7c This is a schematic diagram of the direction-of-arrival estimation results obtained based on the CNN algorithm in existing technology; Figure 7dThis is a schematic diagram illustrating the estimation error of the direction of arrival estimation result obtained based on the CNN algorithm in the existing technology; Figure 7e This is a schematic diagram of the direction of arrival estimation results obtained based on the ESPRIT algorithm in the existing technology; Figure 7f This is a schematic diagram illustrating the estimation error of the direction of arrival estimation result obtained based on the existing ESPRIT algorithm; Figure 7g This is a schematic diagram of the direction of arrival estimation results obtained based on the ML algorithm in the existing technology; Figure 7h This is a schematic diagram illustrating the estimation error of the direction of arrival estimation result obtained based on the ML algorithm in the existing technology; Figure 8 This is a diagram comparing RMSE performance under different signal-to-noise ratios. Detailed Implementation

[0026] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.

[0027] Please see Figure 1 , Figure 1 This is a flowchart illustrating a direction-of-arrival (DOA) determination method provided in an embodiment of this application, wherein the method is applied to a deep learning model. Figure 1 As shown, the method for determining the direction of arrival includes the following steps: S101. Based on the snapshot data of the signal under test, obtain the vector covariance matrix of the snapshot data; S102. Based on the vector covariance matrix, obtain the noise residual map; S103. Based on the noise residual map and the vector covariance matrix, the denoised covariance features are obtained; S104. Based on the denoised covariance characteristics, determine the direction of arrival of the signal under test.

[0028] In this embodiment, the signal to be measured refers to an electromagnetic wave or acoustic wave signal emitted by one or more radiation sources in space with unknown directions, which travels through a propagation path to the receiving array. Its direction of arrival (DOA), quantity, power and other parameters have not yet been determined and need to be estimated by array signal processing technology. It can be a narrowband far-field plane wave signal or a near-field spherical wave signal.

[0029] In this embodiment, the snapshot data of the signal under test refers to a set of complex signal samples obtained by the receiving array synchronously sampling the signal under test from an unknown direction within a certain observation period. These samples are used for subsequent covariance estimation and parameter inversion. The snapshot data can be a single snapshot, which is the array output vector collected at a certain instant, reflecting the complex amplitude of the signal received by each array element at that moment. Alternatively, it can be multiple snapshots, which are signal matrices collected and stacked at continuous or discrete time points to improve statistical stability and algorithm robustness.

[0030] In this embodiment, the vector covariance matrix refers to the matrix that reflects the spatial correlation of the signal, constructed by autocorrelation operation on the snapshot data.

[0031] In this embodiment, the noise residual map refers to a two-dimensional distribution map of the degree of noise interference at each location estimated from the covariance matrix by a deep learning model. It can be a real-valued residual heatmap based on the output of a convolutional neural network, or a noise confidence map generated by reconstructing errors through an autoencoder.

[0032] In this embodiment, the denoised covariance feature refers to the covariance representation that is closer to the ideal noise-free state obtained by subtracting the noise component estimated by the model from the original covariance matrix. It can be a modified complex covariance matrix obtained by subtracting element by element, or a reconstructed covariance feature that is synthesized after denoising the real and imaginary parts in the complex domain.

[0033] In this embodiment, a uniform linear array (ULA) can be used to receive signals from a far-field narrowband signal source, and 50 snapshots can be acquired in a low signal-to-noise ratio environment to form a 10×50 complex data matrix.

[0034] In this embodiment, the product of the conjugate transpose of the snapshot data of the signal to be tested can be calculated and divided by the number of snapshots to obtain a 10×10 Hermitian sample covariance matrix. In the above implementation process, a noise residual map reflecting the covariance distortion caused by finite snapshot data and noise can be obtained. Then, based on the noise residual map, the vector covariance matrix can be processed to obtain the denoised covariance features. Finally, a more accurate direction of arrival can be obtained based on the denoised covariance features.

[0035] In practical applications of vehicle-mounted radar, due to real-time requirements, systems often can only collect a small number of snapshots (e.g., 10–20), causing the sample covariance matrix to deviate significantly from ideal statistical characteristics. Especially under low signal-to-noise ratio (SNR), noise further distorts the covariance structure, leading to false peaks or angle shifts in direction-of-arrival (DOA) estimation. This embodiment learns from the original vector covariance matrix using a deep neural network and generates a noise residual map. This map accurately depicts the covariance distortion pattern caused by the limited number of snapshots and background noise, such as non-physical high-correlation regions or random fluctuations. Subsequently, the system performs element-wise subtraction between this residual map and the original covariance matrix, effectively removing noise components and obtaining the denoised covariance features. These features are closer to the true covariance structure under noise-free conditions, preserving the spatial correlation of the target signal. Based on this, subsequent DOA estimation algorithms (such as MUSIC) can generate clearer and sharper spatial spectra, accurately distinguishing the true angles of two adjacent vehicles in adjacent lanes, significantly improving the direction-finding accuracy and reliability under limited snapshot numbers and strong noise interference.

[0036] In an optional implementation, a noise residual map is obtained based on the vector covariance matrix, including: Convert the vector covariance matrix into a two-dimensional matrix form of the covariance matrix; The covariance matrix in two-dimensional form is convolved to obtain the noise residual map.

[0037] In this embodiment, the covariance matrix in two-dimensional form refers to the reorganization of the original one-dimensional expanded vector covariance into an image-like matrix with a row-column spatial correspondence. The covariance matrix can be directly regarded as a grayscale image to retain the original index relationship, or the covariance elements can be mapped onto a ring grid in polar coordinates to form a structured image.

[0038] In this embodiment, convolution processing refers to the operation of using a convolution kernel to slide on a two-dimensional covariance image to extract local spatial correlation features. It can be to use a 3×3 standard convolution kernel to capture neighborhood patterns, or to use dilated convolution to expand the receptive field to preserve long-range structural dependencies.

[0039] In this embodiment, one specific way to convert the vector covariance matrix into a two-dimensional form is to directly use the 10×10 covariance matrix as the input of a single-channel image. Normalization can be performed first before conversion to eliminate the influence of energy scale.

[0040] In this embodiment, the process of acquiring snapshot data, constructing a vector covariance matrix, generating a noise residual map, and obtaining the denoised covariance features can be achieved through the backbone denoising sub-network of the DSDE-Net deep learning model. During the training phase, this sub-network uses the ideal noise-free covariance matrix RidealRideal as the supervision label and the noisy covariance matrix RnoisyRnoisy as the input, minimizing the mean squared error (MSE) loss L = ||ΔR (Rnoisy Rideal)∥F2L=∥ΔR (Rnoisy The network learns the noise residual mapping using the Rideal)∥F2 algorithm. During inference, the network directly outputs ΔRΔR without iteration, achieving end-to-end denoising.

[0041] In the above implementation process, before performing convolution processing to obtain the noise residual map, the vector covariance matrix can be converted into a two-dimensional matrix form of the covariance matrix. This preserves the inherent spatial structure information in the vector covariance matrix and avoids the loss of spatial structure information. As a result, the noise residual map is not obtained based on the covariance matrix after the loss of spatial structure information, thereby improving the calculation accuracy of the noise residual map.

[0042] For example, in a vehicle-mounted radar system, the vector covariance matrix essentially contains the spatial phase and amplitude relationships between the channels of the antenna array. This structural information is crucial for accurately modeling noise and signal characteristics. Directly flattening the covariance matrix into a one-dimensional vector before inputting it into a convolutional network would destroy the physical element index relationships corresponding to its rows and columns, leading to the loss of local correlations between adjacent elements. In this embodiment, before convolution processing, the vector covariance matrix is ​​preserved in a two-dimensional matrix form (such as an N×N Hermitian matrix), naturally corresponding to the spatial topology of the array. This allows subsequent convolution operations to effectively perceive local neighborhood patterns; for example, strongly correlated regions near the main diagonal represent the true signal, while random disturbances far from the diagonal are more likely to originate from noise. By preserving this spatial structure, the convolutional neural network can more accurately distinguish between noise-induced distortion and the true signal correlation, thereby generating a high-fidelity noise residual map. For example, when identifying weak targets under low signal-to-noise ratio conditions, this method can avoid misclassifying legitimate spatial correlations determined by array geometry as noise, significantly improve the accuracy of residual maps, and lay a reliable foundation for subsequent denoising and direction-of-arrival estimation.

[0043] In an optional implementation, the denoised covariance features are obtained based on the noise residual map and the vector covariance matrix, including the following sub-steps: Convert the vector covariance matrix into a two-dimensional matrix form of the covariance matrix; The covariance matrix in two-dimensional form is subtracted from the noise residual map to obtain the denoised covariance features.

[0044] In this embodiment, the subtraction operation between the two-dimensional matrix form of the covariance matrix and the noise residual map refers to the element-wise algebraic subtraction of the covariance matrix after spatial structure preservation processing, i.e., the two-dimensional matrix form of the covariance moment, and its corresponding noise residual map at the same position to obtain the denoised covariance representation. This can be done by subtracting the real part and imaginary part of the complex covariance matrix from the real-valued noise residual map respectively, or by expanding the real-valued noise residual map into a complex form (setting the imaginary part to zero) and then directly performing complex subtraction with the original complex covariance matrix.

[0045] In the above implementation process, by subtracting the covariance matrix in two-dimensional matrix form from the noise residual map, the computational cost of denoising can be reduced while obtaining the denoised covariance features.

[0046] In real-time signal processing of automotive millimeter-wave radar, traditional denoising methods (such as covariance reconstruction based on eigenvalue decomposition or iterative optimization) have high computational complexity, making it difficult to meet millisecond-level response requirements, especially under low signal-to-noise ratio conditions where more computation is needed to suppress noise. This embodiment achieves efficient denoising by preserving the spatial structure of the covariance matrix in two-dimensional form and performing element-wise subtraction with the noise residual map generated by the convolutional network. This subtraction operation involves only simple pixel-level addition and subtraction of complex or real numbers, without the need for matrix decomposition, inversion, or iterative convergence, significantly reducing computational overhead. For example, on resource-constrained automotive SoCs (such as the TI AWR2944), traditional subspace denoising may take more than 10 milliseconds, while the subtraction denoising method of this approach can be completed within 2 milliseconds. This not only significantly reduces the processor load, enabling the system to quickly obtain high-quality denoised covariance features under limited snapshot and strong noise conditions, but also provides real-time and reliable input for subsequent high-precision direction-of-arrival estimation, balancing algorithm performance with the engineering feasibility of embedded deployment.

[0047] In an optional implementation, the direction of arrival (DOA) of the signal under test is determined based on the denoised covariance characteristics, including: The denoised covariance features are flattened to obtain the flattened covariance features. The flattened covariance features are added to the vector covariance matrix to obtain the summed covariance features. Based on the covariance characteristics after addition, the direction of arrival of the signal under test is obtained.

[0048] In this embodiment, flattening refers to the operation of expanding the denoised covariance matrix into a one-dimensional vector in a fixed order so that it can be input into a fully connected network or for subsequent fusion. This can be done by concatenating the matrix row by row from the first row into a long vector, or by extracting elements sequentially along the main diagonal and its parallel lines to retain some spatial symmetry structure information.

[0049] In this embodiment, the flattened covariance feature refers to the one-dimensional vector representation of the covariance obtained after flattening. In this embodiment, the summed covariance feature refers to the enhanced covariance representation obtained by adding the flattened and possibly nonlinearly transformed denoised features back to matrix form and then adding them element by element to the original vector covariance matrix.

[0050] In the above implementation process, adding the denoising result to the original input can supplement the array structure information again, thereby compensating for the destruction of array structure information by excessive denoising during the denoising process and obtaining covariance features with higher fidelity.

[0051] For example, in low signal-to-noise ratio scenarios for automotive radar, deep denoising models may misjudge weak but real array phase correlations (such as small phase shifts between adjacent antennas) in the covariance matrix as noise and over-smooth them, leading to the loss of spatial structure information and affecting direction-of-arrival resolution. This embodiment introduces a residual connection mechanism by adding the denoised covariance features to the original vector covariance matrix, effectively recovering the mistakenly deleted structural details. For instance, when distinguishing two adjacent vehicles with an angular interval of less than 1°, the original covariance, although containing noise, still retains key spatial correlation patterns; the addition operation suppresses the main noise while preserving these subtle geometric features, making the final covariance features both clean and high-fidelity, significantly improving the resolution accuracy of algorithms such as MUSIC for dense targets.

[0052] In an optional implementation, the direction of arrival of the signal under test is obtained based on the summed covariance characteristics, including the following steps: Based on the summed covariance features, time-domain features and spatial-domain features are extracted. The input condition parameters are obtained and the corresponding gating weight coefficients are determined. The gating weight coefficients represent the proportion of the contribution of time domain features and spatial domain features to the direction of arrival under the premise of using the input condition parameters. Based on the gating weight coefficient, the time domain features and spatial domain features are weighted and combined to obtain the fused features; The direction of arrival of the signal under test is determined based on the fusion characteristics.

[0053] In this embodiment, the time-domain features refer to the dynamic characteristics extracted from the temporal evolution of the snapshot sequence; these features can be the changing trend of the autocorrelation coefficient between snapshots, or the temporal dependency pattern encoded by the hidden state of the LSTM network.

[0054] In this embodiment, spatial domain features refer to features that reflect the relationship between the array geometry and the signal incident angle. These features can be the spatial spectrum shape corresponding to the principal eigenvector of the covariance matrix, or the differential eigenvector formed by the phase difference between adjacent array elements.

[0055] In this embodiment, the gating weight coefficient is implemented based on a dynamic gating fusion mechanism, which refers to a fusion mechanism that dynamically adjusts the weights of different feature branches according to environmental conditions.

[0056] In this embodiment, the signal-to-noise ratio (SNR) is the ratio of signal power to background noise power, used to measure observation quality. It can be the global average SNR (SNR of the entire snapshot sequence) or the SNR estimate of a local frequency band or angular region.

[0057] In this embodiment, the specific way to extract time-domain features can be to input the snapshot sequence into a bidirectional LSTM and take the final hidden state as the time feature; or to calculate the variance of the snapshot envelope as a simple time-series index.

[0058] In this embodiment, the specific method for extracting spatial domain features can be: inputting the covariance matrix into the CNN and taking the last layer feature map as the spatial feature; or directly using the peak position of the MUSIC spectrum as the manual spatial feature.

[0059] In the above implementation process, after the covariance features are denoised and supplemented, based on the highly correlated covariance features, the gating weight coefficients are further combined to control the contribution of time-domain features and spatial-domain features to the direction of arrival calculation under different input conditions. This achieves adaptive weighted fusion of time-related features and spatial-related features, and ultimately, under different input conditions such as signal-to-noise ratio and data conditions, the relative contribution of time-related features and spatial-domain features can be flexibly adjusted to adapt to different scenarios and improve robustness.

[0060] For example, in actual operation of vehicle-mounted radar, the signal-to-noise ratio (SNR) conditions change significantly: In high SNR scenarios (such as open highways with strong target echoes and little interference), the spatial structure of the covariance matrix is ​​clear, and spatial domain features (such as array phase difference) can accurately reflect the direction of arrival. In this case, the system assigns higher weights to spatial domain features through a gating mechanism to achieve high-precision angle resolution. However, in low SNR scenarios (such as urban canyons or environments with strong electromagnetic interference), the target echo is weak, the covariance matrix is ​​contaminated by noise, and the spatial structure is distorted. But if the target has relative motion, its snapshot sequence may still show a certain time correlation (such as Doppler consistency). In this case, the gating mechanism automatically reduces the spatial domain weights and instead enhances the contribution of time domain features, using the stability of the signal in the time dimension to assist in direction estimation. By combining this adaptive weighted fusion strategy with high-quality covariance after denoising and supplementing structural information, the system can dynamically adjust the fusion ratio of temporal and spatial information according to the signal-to-noise ratio and feature reliability of the input data without pre-setting the scene type. Thus, it can maintain robust and accurate estimation of the direction of arrival under a wide range of conditions from high signal-to-noise ratio to extremely low signal-to-noise ratio.

[0061] In an optional implementation, obtaining the input condition parameters and determining the gating weight coefficients corresponding to the input condition parameters includes: The temporal and spatial features are concatenated to obtain the original joint input; The original joint input is subjected to a nonlinear transformation to obtain candidate enhanced feature representations. Based on the original joint input and candidate enhancement feature representation, the gating weight coefficients are obtained.

[0062] In this embodiment, the original joint input refers to the composite feature vector formed by concatenating the time-domain feature vector and the spatial-domain feature vector.

[0063] In this embodiment, the candidate enhancement feature representation refers to the high-order abstract representation obtained by nonlinear transformation of the original joint input.

[0064] In this embodiment, the gating weight coefficient refers to a learnable parameter used to control the contribution of different feature branches.

[0065] In this embodiment, for splicing time and space features, 64-dimensional time features and 64-dimensional space features can be spliced ​​into a 128-dimensional vector, or feature normalization can be performed first before splicing.

[0066] In this embodiment, the original joint input is subjected to nonlinear transformation to obtain candidate enhancement feature representations, which can be generated by a nonlinear activation function.

[0067] In the above implementation process, by concatenating the time domain features and spatial domain features, and performing nonlinear transformation on the concatenated original input, candidate enhancement feature representations can be obtained. Then, based on the original joint input and candidate enhancement feature representations, gating weight coefficients can be obtained.

[0068] For example, in vehicle-mounted radar systems, temporal features (such as the Doppler variation trend of snapshot sequences) and spatial features (such as the array phase structure reflected by the covariance matrix) each carry complementary information, but their reliability varies significantly under different signal-to-noise ratios. To achieve intelligent fusion, these two types of feature vectors are first concatenated into a raw joint input to form a high-dimensional representation encompassing the entire spatiotemporal landscape. Subsequently, a multilayer perceptron (MLP) is used to perform a nonlinear transformation on this concatenated vector (such as linear projection under ReLU activation) to generate a more discriminative candidate enhanced feature representation. This enhanced feature can capture the implicit interaction between the temporal and spatial dimensions, such as whether "strong Doppler consistency is accompanied by weak spatial coherence". In an optional implementation, determining the direction of arrival (DOA) of the signal under test based on fused features includes: Global average pooling is applied to the fused features to obtain compressed features; Based on the compressed features and weight matrix, the channel weight vector is obtained. The weight matrix is ​​used to characterize the degree of influence of each feature dimension on the direction of arrival estimation. The greater the degree of influence, the greater the corresponding weight. Multiply the channel weight vector by the fused feature to obtain the enhanced feature vector; The direction of arrival (DOA) of the signal under test is obtained based on the enhanced feature vector.

[0069] In this embodiment, the channel weight vector refers to the attention coefficient vector used to reweight the importance of each channel of the fused feature.

[0070] In this embodiment, the enhanced feature vector refers to a new feature representation that highlights key discriminative information after channel weighting.

[0071] In this embodiment, global average pooling refers to averaging all spatial locations of each channel of the feature map and compressing them into spatially independent channel descriptors. This can be either averaging each channel of the two-dimensional feature map to obtain a 1D vector, or averaging a high-dimensional tensor along its spatial dimensions.

[0072] In the above implementation process, an attention mechanism can be introduced through the channel weight vector, and then the attention recalibration mechanism can be used to highlight the more critical discrimination information for direction of arrival estimation, thereby improving the accuracy of direction of arrival estimation.

[0073] For example, in multi-target direction-of-arrival (DOA) estimation using vehicle-mounted millimeter-wave radar, the contributions of different channels in the fused feature vector to angle discrimination are not equal: some channels may primarily respond to strong interference sources or noise, while others are more sensitive to the weak spatial features of the real target. Using all channel information indiscriminately can easily lead to angle estimation bias. This embodiment introduces a channel weight vector to construct an attention recalibration mechanism—first, the fused features are globally averaged and compressed into channel descriptors, then a small fully connected network is used to generate attention weights for each channel; subsequently, these weights are multiplied channel-by-channel with the original features to enhance key information and suppress redundant or noisy channels. For instance, in low signal-to-noise ratio scenarios, the system can automatically reduce the response to channels with high variance and low stability, while amplifying the outputs of channels that maintain a consistent phase relationship across multiple snapshots. This adaptive recalibration makes the peak corresponding to the real target in the DOA spectrum sharper and more prominent, effectively suppressing sidelobes and spurious peaks, thereby significantly improving the accuracy and stability of angle discrimination for adjacent vehicles or vulnerable traffic participants.

[0074] Please see Figure 2 , Figure 2 This is a schematic diagram of a direction-of-arrival (DOA) determination device provided in an embodiment of this application, wherein the device is applied to a deep learning model. Figure 2 As shown, the direction of arrival determination device includes the following functional modules: The covariance calculation module 201 is used to obtain the vector covariance matrix of the snapshot data based on the snapshot data of the signal under test. The noise residual processing module 202 is used to obtain the noise residual map based on the vector covariance matrix; The denoising module 203 is used to obtain the denoised covariance features based on the noise residual map and the vector covariance matrix; The determination module 204 is used to determine the direction of arrival of the signal under test based on the denoised covariance characteristics.

[0075] In the above implementation process, the device can obtain a noise residual map that reflects the covariance distortion caused by limited snapshot data and noise. Then, based on the noise residual map, the vector covariance matrix is ​​processed to obtain the denoised covariance features. Finally, a more accurate direction of arrival is obtained based on the denoised covariance features.

[0076] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. For example... Figure 3 As shown, the electronic device includes: Processor 301; and The memory 302 is configured to store machine-readable instructions that, when executed by the processor 301, perform a direction of arrival determination method as described in any of the foregoing embodiments.

[0077] In the above implementation process, the processor 301 calls the program instruction in real time, which can obtain a noise residual map that reflects the covariance distortion caused by finite snapshot data and noise. Then, based on the noise residual map, the vector covariance matrix is ​​processed to obtain the denoised covariance features. Finally, a more accurate direction of arrival is obtained based on the denoised covariance features.

[0078] Furthermore, embodiments of this application also provide a storage medium storing a computer program, which is executed by a processor as a direction of arrival determination method as described in any of the foregoing embodiments.

[0079] In the above implementation process, the storage medium completely solidifies the direction of arrival determination method into computer program instructions; when the medium is read and run by the processor, the processor obtains the vector covariance matrix of the snapshot data of the signal under test according to the instructions, obtains the noise residual map based on the vector covariance matrix, obtains the denoised covariance features based on the noise residual map and the vector covariance matrix, and determines the direction of arrival of the signal under test based on the denoised covariance features, thereby realizing the determination of the direction of arrival based on the denoised covariance features to improve the accuracy of the direction of arrival calculation.

[0080] As a specific implementation, this approach uses DSDE-Net, a feature enhancement dual-branch network designed for low signal-to-noise ratio environments, as a deep learning model. This model constructs direction-of-arrival estimation as a regression task, directly outputting continuous angle values. Please refer to [link to relevant documentation]. Figure 4 , Figure 4 This is a schematic diagram of a DSDE-Net provided in an embodiment of this application. For example... Figure 4 As shown, DSDE-Net employs a dual-branch architecture, extracting features from the spatial and temporal domains respectively: the spatial branch takes the covariance matrix as input, extracts global features through skip connections, and introduces a residual learning mechanism to predict noise components, achieving effective denoising through subtraction; the temporal branch takes the original signal as input and uses an LSTM network to capture temporal dependencies. To further enhance the discriminative power of the features, a dynamic gating fusion mechanism is designed to adaptively weight and integrate the features from both branches, finally outputting an angle estimate through a multi-layer fully connected mapping.

[0081] Furthermore, for the array signal model, assume a uniform linear array (ULA) consisting of M elements, with an element spacing of half a wavelength. ,in For signal wavelength. (Set) Several incoherent far-field narrowband signal sources are incident on the array at incident angles of... The received signal can then be expressed as: (1) in It's the number of snapshots. To represent noise, the signal source and the array manifold matrix are respectively represented as follows: (2) (3) The array incident angle is The steering vector of the incident signal is specifically expressed as: (4) The covariance matrix of the received signal can then be expressed as: (5) Assuming noise The noise is zero-mean complex white Gaussian noise uncorrelated with the signal, and its noise power is... Then the covariance matrix can be expressed as (6) In the formula Let be the signal covariance matrix. for Identity matrix.

[0082] In practical applications, the covariance matrix is ​​usually estimated using finite snapshot data, and its sample covariance matrix is ​​defined as: (7) Furthermore, the deep learning model introduces a residual denoising network based on two-dimensional convolution into the spatial information extraction network. (See [link to documentation] for details.) Figure 5 , Figure 5 This is a schematic diagram of a residual denoising network provided in an embodiment of this application. This module uses a vectorized covariance matrix... As input, it is then reconstructed into a two-dimensional matrix form to preserve its inherent spatial structure information: (8) Unlike traditional methods that directly learn the denoised signal, this module employs a residual learning strategy. This strategy models the noise components in the covariance matrix using a lightweight, multi-layer convolutional network. (See [link to relevant documentation]). Figure 5 , Figure 5This is a schematic diagram of a residual denoising structure provided in this specific implementation. For example... Figure 5 As shown, its forward propagation process can be represented as: (9) (10) (11) (12) in, This is the predicted noise residual map. Adaptive denoising is achieved through subtraction: (13) To avoid damaging the array structure information due to excessive denoising, the module further introduces a global residual connection, adding the denoising result to the original input, ultimately outputting a covariance feature with higher fidelity: (14) This design enables the network to retain effective array spatial features to the greatest extent while suppressing spatial spectrum distortion caused by limited snapshots and noise, thus providing a more accurate and robust feature representation basis for subsequent DOA estimation.

[0083] Furthermore, to achieve efficient collaborative modeling of time-related and spatially related features, the deep learning model also includes a dynamically gated feature fusion network. (See also...) Figure 6 , Figure 6 This is a schematic diagram of a dynamic gating fusion network provided in an embodiment of this application. Figure 6 As shown, this module takes the concatenated vector of two features as input, denoted as: (15) in and These represent the characteristics of the time domain and the spatial domain, respectively.

[0084] Based on this, the input features are mapped through a nonlinear transformation branch to generate candidate augmentation representations: (16) In the formula This represents a feature enhancement map consisting of multiple fully connected layers and nonlinear activation functions, used for reshaping and compensating the original fused features.

[0085] Subsequently, a dynamic gating branch is constructed to jointly model the original features and enhanced features, and the dimensional gating coefficients are learned through the Sigmoid function: (17) in Indicates a gating network. This is a dimension-wise adaptive weight vector. Based on this gating coefficient, the fused features are obtained through element-wise weighting: (18) This enables the network to adaptively adjust the contribution ratio of temporal and spatial features in each dimension according to different signal-to-noise ratios and other conditions.

[0086] Furthermore, to highlight the discriminative information that is more sensitive to DOA estimation, this specific implementation introduces a vector-level SE attention mechanism into the fusion output to recalibrate the importance of the feature dimensions. This mechanism is implemented through the following three steps: (19) (20) (twenty one) in These are the compressed features after global average pooling. For learnable weight matrix, The compression ratio is... For bias terms, This is the learned channel weight vector. The mechanism first uses global average pooling to weight each sample... The dimensional features are compressed into scalars, and then the weight coefficients of each dimension are learned through two layers of fully connected networks. Finally, the importance of the original features is recalibrated to suppress redundant and noisy features and enhance the response of key dimensions.

[0087] By employing the aforementioned dynamic gating fusion and attention reweighting strategies, the robustness of the fused features is significantly improved while fully leveraging the complementarity of spatiotemporal features, providing a more stable and discriminative feature representation for the subsequent angle regression network.

[0088] Based on the aforementioned deep learning module, this specific implementation uses array element numbers. Quick shot number Experiments were conducted using a uniform linear array, with signal wavelength... The element spacing is 0.5m (half a wavelength). Considering a dual-source scenario, the source incident angle range is... This specific implementation method belongs to gridless DOA estimation, where the signal angle needs to be continuously measured within the search range, resulting in a huge potential sample space. To reduce the complexity of sample generation, this specific implementation method adopts the following data generation strategy: Let the angle interval between the two information sources be . ,Regulation The signal-to-noise ratio (SNR) ranges from [value missing]. Six signal-to-noise ratio (SNR) conditions were included, with 4dB intervals. 500,000 training samples were generated for each SNR condition. This resulted in a dataset containing... One sample.

[0089] All data were randomly divided into training and validation sets in an 8:2 ratio. During training, the Adam optimizer was used with a batch size of 128, along with a cosine annealing learning rate strategy. The initial learning rate was set to 0.0001, and the minimum learning rate was... The learning rate smoothly decays from its initial value to its minimum value according to a cosine function with each training epoch. An early stopping strategy is used, terminating training when the validation set loss shows no improvement over 20 consecutive epochs, while simultaneously saving the best model. Experiments were conducted on a Windows 11 system, using Python 3.12 and PyTorch 2.3.0 (CUDA 13.1) to build the model, and training and testing were completed on an NVIDIA GeForce RTX 4090 GPU.

[0090] The direction-of-arrival (DOA) processing method in this specific embodiment is compared and analyzed with methods used in various existing technologies, including CNN, ESPRIT, and ML algorithms. The training data generation process for the CNN algorithm is detailed in existing technologies. To ensure fairness in the comparative experiments, the CNN training samples are consistent with the method proposed in this specific implementation in terms of angular coverage and signal-to-noise ratio settings, while other training strategies remain unchanged. These settings ensure consistent signal environment conditions for different algorithms during training and performance evaluation, avoiding the influence of differences in external conditions.

[0091] During the testing phase, all DOA estimation algorithms used 10 snapshots of signal and set two fixed intervals. The incident source is [target 1]. The angle of incidence of target 1 is from [target 1]. arrive by With the change in step size, the incident angle of target 2 changes accordingly from Change to A total of 101 test sample pairs were obtained. At a signal-to-noise ratio of... The tests were conducted under the given conditions, and the DOA estimation results for various methods are as follows: Figure 7a , Figure 7b , Figure 7c , Figure 7d , Figure 7e , Figure 7g , Figure 7f , Figure 7g As shown, Figure 7a It includes two fitting curves, one for θ1 and the other for signal θ2. Figure 7a The horizontal axis in the graph represents the signal sampling point, and the vertical axis represents the calculated direction of arrival (DOA) at the corresponding sampling point. Comparing these graphs, it can be seen that the method used in this implementation can effectively distinguish signals θ1 and θ2 at various sampling points. Specifically, the two fitted curves show that the DOA identification results for the two signals do not overlap at different sampling points, while other methods show overlapping DOA estimation results at many sampling points, making it impossible to distinguish the two signals. Figure 7a This is a schematic diagram of the direction-of-arrival estimation result obtained according to the direction-of-arrival determination method of this application. Figure 7b This is a schematic diagram illustrating the estimation error of the direction of arrival (DOA) estimation result obtained according to the DOA determination method in the embodiments of this application. Figure 7c This is a schematic diagram of the direction-of-arrival estimation results obtained based on the existing CNN algorithm. Figure 7d This is a schematic diagram illustrating the estimation error of the direction-of-arrival estimation result obtained based on the existing CNN algorithm. Figure 7e This is a schematic diagram of the direction-of-arrival estimation results obtained based on the existing ESPRIT algorithm. Figure 7f This is a schematic diagram illustrating the estimation error of the direction of arrival estimation result obtained based on the existing ESPRIT algorithm. Figure 7g This is a schematic diagram of the direction-of-arrival estimation results obtained based on the existing ML algorithm. Figure 7h This is a schematic diagram of the estimation error of the direction of arrival estimation result obtained based on the ML algorithm in the existing technology.

[0092] In addition, the maximum and minimum error range, mean absolute error, and overall standard deviation were also statistically analyzed. Detailed results are shown in Table 1.

[0093] Experimental results show that the performance of traditional algorithms degrades significantly under low signal-to-noise ratio conditions. The ESPRIT and ML algorithms struggle to distinguish between two target signals in certain situations, leading to substantial deviations in angle measurement results. In contrast, deep learning-based methods exhibit superior performance, with significantly higher angle measurement accuracy than traditional algorithms. Error analysis reveals that the DSDE-Net implementation in this study achieves correct estimations on all test samples, with the smallest mean absolute error and overall standard deviation, indicating that this network not only boasts the highest estimation accuracy but also the most stable results. The error of the CNN algorithm is... Within this range, the limitations of CNN methods due to grid resolution and low snapshot count result in inaccurate estimation for some samples. Overall, DSDE-Net maintains high estimation accuracy and stability under low signal-to-noise ratio and complex signal environments, demonstrating superior DOA estimation performance.

[0094] Table 1 Error Indicators for Various Algorithms

[0095] To evaluate the statistical performance of the proposed algorithm in DOA (Direction of Arrival) estimation, the Monte Carlo method was used to compare it with traditional ESPRIT, ML, and CNN-based neural network DOA estimation algorithms. The root mean square error (RMSE) was used as the performance evaluation metric. The formula for RMSE is: (twenty two) in, The number of information sources. For the number of test samples, and These are the estimated and true values ​​of the target DOA, respectively. In the simulation experiment, the incident angles of the two signal sources are set as follows: and The sample covariance estimation also uses A quick snapshot. SNR condition set to... There are a total of 5 scenarios. 1,000 independent experiments were conducted for each SNR condition to calculate the RMSE, and the Cramér-Rao lower bound (CRLB) was given as a performance benchmark. The results are as follows: Figure 8 As shown, where, Figure 8 This is a diagram comparing RMSE performance under different signal-to-noise ratios.

[0096] As the signal-to-noise ratio from Gradually increase to Traditional methods have continuously improved their DOA estimation accuracy. However, in the low signal-to-noise ratio (SNR) region, their performance remains significantly limited, with large errors and difficulty in achieving stable and accurate estimation. CNN methods, due to their reliance on discrete angle grids, suffer from limitations when the SNR increases to a certain level. The performance improvement tends to plateau afterward. In contrast, DSDE-Net uses a regression framework for angle estimation, and its estimation accuracy is not limited by the grid. Its estimation accuracy continues to improve with the increase of signal-to-noise ratio, approaching or even exceeding CRLB in some high signal-to-noise ratio regions. Compared with CNN networks, this method not only achieves a significant improvement in estimation accuracy, but also demonstrates adaptability over a wider signal-to-noise ratio range.

[0097] To verify the effectiveness of the proposed module, this section designs an ablation experiment for comparison. The network proposed in this specific implementation includes two key modules: a residual denoising module and a dynamic gating fusion module. Four models were set up for comparison in the experiment: Model 1 (excluding the above two modules), Model 2 (containing only the residual denoising module), Model 3 (containing only the dynamic gating fusion module), and Model 4 (the complete model). At a signal-to-noise ratio of... The angular interval between the two signal sources is The test was conducted under the specified conditions, and the angle estimation error distribution range, root mean square error (RMSE), and mean absolute error (MAE) were used as the core evaluation indicators. The specific experimental results are shown in Table 2.

[0098] Experimental results show that the model has the highest average RMSE (0.6037) without the denoising module and fusion network. Adding the residual denoising module (Model 2) and the dynamic gating fusion module (Model 3) separately significantly improves the RMSE and reduces the average MAE. When all modules are included, both error metrics are minimized, indicating that the combination of modules produces a significant positive synergistic effect, thereby improving the DOA estimation accuracy under low signal-to-noise ratio conditions.

[0099] Table 2 Ablation Experiment Results

[0100] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and there may be other division methods in actual implementation. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0101] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0102] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0103] It should be noted that if a function is implemented as a software module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0104] In the embodiments of this application, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0105] The above are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for determining direction of arrival, characterized in that, The method is applied to a deep learning model, and the direction-of-arrival (DOA) determination method includes: Based on the snapshot data of the signal under test, the vector covariance matrix of the snapshot data is obtained; Based on the vector covariance matrix, the noise residual map is obtained; Based on the noise residual map and the vector covariance matrix, the denoised covariance features are obtained; and Based on the denoised covariance characteristics, the direction of arrival of the signal under test is determined.

2. The method for determining direction of arrival according to claim 1, characterized in that, The step of obtaining the noise residual map based on the vector covariance matrix includes: Convert the vector covariance matrix into a two-dimensional matrix form of the covariance matrix; The covariance matrix in two-dimensional form is convolved to obtain the noise residual map.

3. The method for determining direction of arrival according to claim 1, characterized in that, The step of obtaining the denoised covariance features based on the noise residual map and the vector covariance matrix includes: Convert the vector covariance matrix into a two-dimensional matrix form of the covariance matrix; The covariance matrix in two-dimensional form is subtracted from the noise residual map to obtain the denoised covariance features.

4. The method for determining direction of arrival according to claim 1, characterized in that, Determining the direction of arrival of the signal under test based on the denoised covariance features includes: The denoised covariance features are flattened to obtain flattened covariance features. The flattened covariance features are added to the vector covariance matrix to obtain the summed covariance features. Based on the covariance characteristics after addition, the direction of arrival of the signal under test is obtained.

5. The method for determining direction of arrival according to claim 4, characterized in that, The process of obtaining the direction of arrival of the signal under test based on the summed covariance features includes: Based on the summed covariance features, time-domain features and spatial-domain features are extracted. Obtain input condition parameters and determine the gating weight coefficients corresponding to the input condition parameters; wherein, the gating weight coefficients represent the contribution ratios of the time domain features and the spatial domain features to the direction of arrival under the premise of using the input condition parameters; Based on the gating weight coefficients, the time-domain features and the spatial-domain features are weighted and combined to obtain fused features; Based on the fusion features, the direction of arrival of the signal under test is determined.

6. The method for determining direction of arrival according to claim 5, characterized in that, The step of obtaining input condition parameters and determining the gating weight coefficients corresponding to the input condition parameters includes: The time-domain features and the spatial-domain features are concatenated to obtain the original joint input; The original joint input is subjected to a nonlinear transformation to obtain candidate enhanced feature representations; The gating weight coefficients are obtained based on the original joint input and the candidate enhanced feature representation.

7. The method according to claim 5, characterized in that, Determining the direction of arrival (DOA) of the signal under test based on the fused features includes: The fused features are subjected to global average pooling to obtain compressed features; Based on the compressed features and weight matrix, a channel weight vector is obtained, wherein the weight matrix is ​​used to characterize the degree of influence of each feature dimension on the direction of arrival estimation; the greater the degree of influence, the greater the corresponding weight. Multiply the channel weight vector by the fused feature to obtain the enhanced feature vector; Based on the enhanced feature vector, the direction of arrival of the signal under test is obtained.

8. A direction-of-arrival determination device, characterized in that, The device is applied to a deep learning model, and the direction-of-arrival (DOA) determination device includes: The covariance calculation module is used to obtain the vector covariance matrix of the snapshot data based on the snapshot data of the signal under test; The noise residual processing module is used to obtain a noise residual map based on the vector covariance matrix. The denoising module is used to obtain the denoised covariance features based on the noise residual map and the vector covariance matrix; The determination module is used to determine the direction of arrival of the signal under test based on the denoised covariance features.

9. An electronic device, characterized in that, include: processor; as well as A memory configured to store machine-readable instructions that, when executed by the processor, perform the direction-of-arrival determination method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores a computer program, which is executed by a processor as described in any one of claims 1-7, to determine the direction of arrival.

Images