A DOA estimation method and system based on a deep residual network

Abstract

This invention relates to the interdisciplinary field of array signal processing and deep learning, specifically to a DOA estimation method and system based on a deep residual network. First, temporal snapshot data of spatially incident signals are acquired using a uniform linear array to construct an array-received signal vector model. Based on this model, the sample covariance matrix is ​​calculated, and four-channel real-valued feature extraction is performed. Then, normalization is performed independently on each feature channel. Finally, the normalized feature matrix is ​​input into a pre-trained deep residual network model, and the model inference outputs a high-precision DOA estimation result for the signal incident angle. This invention fully preserves the amplitude and phase information of the covariance matrix through four-channel feature extraction. Combined with the deep feature extraction capability of the deep residual network, it exhibits excellent estimation accuracy and robustness in complex scenarios such as few snapshots, low signal-to-noise ratio, and edge angles. It also features strong physical interpretability, good generalization ability, and high real-time performance.

Description

Technical Field

[0001] This invention relates to the field of array signal processing and deep learning intersection technology, specifically to a DOA estimation method and system based on deep residual networks. Background Technology

[0002] In core application areas such as radar, wireless communication, sonar, and autonomous driving, DOA estimation is a fundamental core technology for array signal processing. Its core objective is to accurately determine the incident direction of a signal source through spatial signals received by an antenna array. It is a prerequisite for realizing advanced functions such as target positioning and tracking, intelligent beamforming, spatial spectrum perception, and autonomous driving environmental perception. Its estimation accuracy and robustness directly determine the performance of subsequent systems.

[0003] Traditional DOA estimation methods are based on subspace algorithms, typically represented by the Multiple Signal Classification (MUSIC) algorithm and the ESPRIT algorithm based on rotation invariant techniques. These algorithms can achieve super-resolution DOA estimation under ideal conditions of high signal-to-noise ratio, high snapshot count, and no array error. However, they have significant drawbacks in complex real-world engineering scenarios: First, their performance is highly dependent on the estimation accuracy of the covariance matrix, requiring a large number of snapshots; in scenarios with few snapshots, the estimation accuracy drops sharply. Second, they have high computational complexity, requiring eigenvalue decomposition and global spectral peak search, making it difficult to meet the requirements of real-time applications. Third, they have poor adaptability to non-Gaussian noise, coherent signals, and array errors, especially in large angle / edge angle regions of 60–90 degrees, where the estimation error increases significantly, failing to meet the accuracy requirements of practical engineering.

[0004] To address the shortcomings of traditional subspace algorithms, researchers have proposed improved fractional low-order moment algorithms and compressed sensing-based DOA estimation methods. However, the former still relies on a large number of snapshots to ensure statistical stability and cannot solve the accuracy problem in scenarios with few snapshots; the latter suffers from grid mismatch, difficulty in manual parameter tuning, and limited generalization ability. In recent years, deep learning technology has brought technological innovation to DOA estimation and has become a research hotspot in this field. However, existing deep learning-based DOA estimation methods still have many shortcomings: the feature learning process of most methods lacks physical interpretability, and the model's generalization ability is insufficient; grid-based deep learning schemes do not fundamentally solve the grid mismatch problem, while gridless schemes suffer from complex network architecture and high training costs; very few methods can simultaneously adapt to multiple practical challenges such as impulse noise, few snapshots, and performance degradation at edge angles, making it difficult to implement in resource-constrained real-time embedded systems. Summary of the Invention

[0005] The purpose of this invention is to propose a DOA estimation method and system based on deep residual networks to achieve high-precision DOA estimation in complex scenarios such as limited snapshots, low signal-to-noise ratio, and edge angles.

[0006] According to a first aspect of the embodiments of this disclosure, a DOA estimation method based on deep residual networks is provided, comprising the following steps: Temporal snapshot data of spatial incident signals are acquired using a uniform linear array, and a vector model of the array-received signal is constructed. The sample covariance matrix is ​​obtained based on the array received signal vector model. Four-channel real-valued feature extraction is performed on the sample covariance matrix to obtain the real part features, imaginary part features, phase sine features and phase cosine features of the matrix elements, and a four-channel real-valued feature matrix is ​​constructed. Normalization is performed on each channel in the four-channel real-valued feature matrix; The normalized feature matrix is ​​input into the trained deep residual network model, and the DOA estimation result of the signal incident angle is output by the deep residual network model.

[0007] In one embodiment, the array received signal vector model is: in, for The array received signal vector at time t; Angle of incidence The corresponding array steering vector characterizes the phase delay of the wavefront among the array elements. For the first The complex amplitude of a source signal; It is an additive noise vector; This represents the number of far-field narrowband signals.

[0008] In one embodiment, the sample covariance matrix for: in, This represents the conjugate transpose operation. This represents the number of snapshots.

[0009] In one embodiment, the phase sine feature is a feature obtained by taking the sine of the phase angle of each element of the sample covariance matrix, and the phase cosine feature is a feature obtained by taking the cosine of the phase angle of each element of the sample covariance matrix.

[0010] In one embodiment, each channel in the four-channel real-valued feature matrix is ​​normalized, specifically as follows: Obtain the mean and standard deviation of the single-channel real-valued feature matrix; If the standard deviation is less than the set threshold, the channel is determined to be a constant channel, and a matrix of all zeros is returned directly; if the standard deviation meets the requirements, the subsequent steps are executed. Perform Z-score normalization on the channel feature: ,in The channel mean. The standard deviation of the channel. To prevent the minimum value of division by zero; The Z-score-normalized features are subjected to tanh smooth compression to achieve smooth saturation constraint of the features; The compressed features are rescaled to map the feature values ​​to a specified interval, thus completing the normalization process for that channel.

[0011] In one embodiment, the deep residual network model is a lightweight deep residual network, whose architecture is as follows: taking the normalized four-channel real-valued feature matrix as input, a feature extraction backbone is constructed through multiple progressively enhanced residual blocks; the backbone network ends by fusing multi-scale features of global average pooling and global max pooling; finally, a fully connected classification head with hierarchical Dropout regularization is used to output the DOA classification probability result corresponding to the angular resolution, so as to obtain the DOA estimation result of the signal incident angle.

[0012] In one embodiment, the deep residual network model employs an edge angle adaptive class weight strategy during training. The adaptive weight coefficients corresponding to the class perspective are: in, This represents the total number of training samples. The total number of angle categories, For the first The number of samples corresponding to the class perspective. This represents the current edge angle value to be weighted.

[0013] According to a second aspect of the present disclosure, a DOA estimation system based on a deep residual network is provided, comprising: The signal acquisition module acquires temporal snapshot data of spatial incident signals through a uniform linear array and constructs a vector model of the array received signals. The covariance matrix construction module obtains the sample covariance matrix based on the array received signal vector model. The multi-channel feature extraction module performs four-channel real-valued feature extraction on the sample covariance matrix, respectively obtaining the real part features, imaginary part features, phase sine features and phase cosine features of the matrix elements, and constructing a four-channel real-valued feature matrix; The feature normalization module performs normalization processing on each channel in the four-channel real-valued feature matrix; The DOA estimation module inputs the normalized feature matrix into the trained deep residual network model, and outputs the DOA estimation result of the signal incident angle through the deep residual network model.

[0014] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the DOA estimation method based on a deep residual network.

[0015] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the aforementioned DOA estimation method based on a deep residual network.

[0016] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: 1. The edge angle adaptive category weight strategy proposed in this invention assigns higher weights to the 60-90 degree edge angle range, which greatly improves the estimation accuracy and robustness of edge angles and makes up for the shortcomings of traditional algorithms in edge angle performance degradation.

[0017] 2. The four-channel features of this invention all have clear physical meanings, are consistent with the principle of electromagnetic field propagation, have physical interpretability, and have strong generalization ability.

[0018] 3. The lightweight deep residual network architecture designed in this invention adopts progressive residual blocks, multi-scale pooling and hierarchical Dropout regularization, which reduces computational complexity while ensuring feature extraction capability. Moreover, the model supports end-to-end inference without the need for complex feature value decomposition and spectral peak search, making it suitable for resource-constrained real-time embedded systems.

[0019] 4. This invention can simultaneously adapt to multiple practical engineering challenges such as limited snapshots, low signal-to-noise ratio, impulse noise, and performance degradation at edge angles, solving the problem of limited applicability of traditional algorithms and existing deep learning methods, and has a wide range of applications. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0021] Figure 1 This is a diagram of the deep residual network structure involved in an embodiment of the present invention; Figure 2 The diagram shows the DOA estimation performance under different SNRs in the embodiments of the present invention. Figure 3 The diagram shows the DOA estimation performance under different snapshot numbers in the embodiments of the present invention. Detailed Implementation

[0022] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0026] Example 1: This embodiment provides a DOA estimation method based on deep residual networks, including the following steps: S1. Obtain temporal snapshot data of spatial incident signals through a uniform linear array, and construct an array received signal vector model; This invention is designed for uniform linear arrays (ULA), wherein the array consists of... It consists of omnidirectional antenna elements, with an adjacent antenna spacing of [missing information]. half wavelength of carrier signal ( To avoid spatial aliasing.

[0027] Assuming that there exists in the space A far-field narrowband signal, from different directions If incident on the array, then The array output signal vector at time t can be modeled as: in, for The array received signal vector at time t; Angle of incidence The corresponding array steering vector characterizes the phase delay of the wavefront among the array elements. For the first The complex amplitude of a source signal; This is an additive noise vector.

[0028] S2. Obtain the sample covariance matrix based on the array received signal vector model; Specifically, the theoretical covariance matrix expression for the array output is: in, It is an array manifold. Represents the source covariance matrix. For noise power, for An identity matrix of order 1. Represents the mathematical expectation. This indicates the conjugate transpose operation.

[0029] In practical engineering, the theoretical covariance matrix needs to be estimated using a finite number of snapshots to obtain the sample covariance matrix: in, This represents the number of snapshots.

[0030] S3. Perform four-channel real-valued feature extraction on the sample covariance matrix to obtain the real part features, imaginary part features, phase sine features and phase cosine features of the matrix elements respectively, and construct a four-channel real-valued feature matrix; Specifically, the obtained sample covariance matrix Four independent feature channels are extracted separately, as follows: ① Real Part Channel: Extracts the real part of each element in the sample covariance matrix; ②Imaginary part channel: Extract the imaginary part of each element in the sample covariance matrix; ③ Phase sine channel: Take the sine value of the phase angle of each element of the sample covariance matrix; ④ Phase cosine channel: Take the cosine value of the phase angle of each element of the sample covariance matrix.

[0031] This four-channel feature representation method simultaneously captures the amplitude and phase information of the covariance matrix, and effectively solves the phase problem by decoupling the phase through sine and cosine components. The fuzzy problem ensures phase continuity; the features of each channel have clear physical meaning, so that the features learned by the network are consistent with the principle of electromagnetic field propagation, which greatly improves the model's generalization ability.

[0032] S4. Normalize each channel in the four-channel real-valued feature matrix; To address the training characteristics of deep residual networks, a dedicated normalization process is designed for each feature channel. The specific steps are as follows: S4.1 Obtain the mean of the single-channel real-valued feature matrix with standard deviation ; S4.2 If the standard deviation Less than If it is determined to be a constant channel, a matrix of all zeros is returned directly; S4.3 If standard deviation If the requirements are met, first perform Z-score standardization: ,in To prevent the minimum value of division by zero; S4.4 performs tanh compression on the standardized features: This achieves smooth saturation constraints on features and avoids gradient explosion; S4.5 performs interval rescaling on the compressed features: Map all feature values ​​to The interval is normalized.

[0033] This normalization process, while preserving the relative relationships of features, perfectly adapts to the training characteristics of deep residual networks, significantly improving training stability and accelerating model convergence.

[0034] S5. Input the normalized feature matrix into the trained deep residual network model, and output the DOA estimation result of the signal incident angle through the deep residual network model.

[0035] This invention specifically designs a lightweight and efficient deep residual network architecture for DOA estimation tasks, such as... Figure 1 As shown. The network is normalized. Using a four-channel feature matrix as input, a feature extraction backbone is constructed using residual blocks with seven progressively enhanced channels. The gradient vanishing problem in deep networks is solved through residual shortcut connections, enabling deep extraction of spatial features of array signals. At the end of the backbone network, multi-scale features of global average pooling and global max pooling are fused to take into account both global statistical properties and significant extreme value features. Finally, a fully connected classification head with hierarchical Dropout regularization is used to output the DOA classification probability results at the corresponding angular resolution. This approach ensures feature extraction capability while also balancing training stability and inference efficiency, and is particularly suitable for feature extraction requirements at edge angles.

[0036] To address the significant performance degradation of traditional algorithms in the 60-90 degree edge angle range, this invention employs a customized design. An adaptive class weighting strategy oriented towards edge angles is used, assigning higher weights to edge angles to improve estimation accuracy and robustness in this range. Adaptive weight coefficients corresponding to the class perspective: in, This represents the total number of training samples. The total number of angle categories, For the first The number of samples corresponding to the class perspective. This represents the current edge angle value to be weighted.

[0037] To fully verify the performance advantages of the proposed method and system, a combination of simulation and experimental tests was conducted to compare it with the traditional MUSIC and ESPRIT algorithms. The core verification indicators were the root mean square error (RMSE) of DOA estimation and the estimation accuracy. The experimental results and analysis are as follows.

[0038] like Figure 2 As shown, with ULA arrays containing 5 and 10 antennas respectively, the RMSE estimated by the present invention for DOA is significantly lower than that of the comparative algorithm as the signal-to-noise ratio (SNR) increases from 0 dB to 10 dB. Under the extremely low SNR condition of 0 dB, the RMSE of the present invention with a 5-antenna configuration can be controlled within 0.1 degrees, while the RMSE of the MUSIC and ESPRIT algorithms exceeds 0.3 degrees; under the condition of SNR of 10 dB, regardless of whether there are 5 or 10 antennas, the RMSE value of the present invention is much lower than that of the comparative algorithm.

[0039] like Figure 3As shown, in the extreme scenario with 50 snapshots, the RMSE of both the MUSIC (5 antennas) and ESPRIT (5 antennas) algorithms is higher than 1.3 degrees, significantly higher than the RMSE of the method of this invention. Even when the number of antennas is increased to 10, the RMSE of the MUSIC and ESPRIT algorithms is still higher than that of the method of this invention. Furthermore, although the RMSE of all algorithms decreases with the increase of the number of snapshots, the DOA estimation performance of the method of this invention still has a significant advantage.

[0040] Example 2: This embodiment provides a DOA estimation system based on deep residual networks, including: The signal acquisition module acquires temporal snapshot data of spatial incident signals through a uniform linear array and constructs a vector model of the array received signals. The covariance matrix construction module obtains the sample covariance matrix based on the array received signal vector model. The multi-channel feature extraction module performs four-channel real-valued feature extraction on the sample covariance matrix, respectively obtaining the real part features, imaginary part features, phase sine features and phase cosine features of the matrix elements, and constructing a four-channel real-valued feature matrix; The feature normalization module performs normalization processing on each channel in the four-channel real-valued feature matrix; The DOA estimation module inputs the normalized feature matrix into the trained deep residual network model, and outputs the DOA estimation result of the signal incident angle through the deep residual network model.

[0041] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.

[0042] Example 3: An electronic device is provided for running the aforementioned "a DOA estimation method based on deep residual networks". The electronic device includes a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S5 of the method described in Embodiment 1, specifically including but not limited to: S1. Obtain temporal snapshot data of spatial incident signals through a uniform linear array, and construct an array received signal vector model; S2. Obtain the sample covariance matrix based on the array received signal vector model; S3. Perform four-channel real-valued feature extraction on the sample covariance matrix to obtain the real part features, imaginary part features, phase sine features and phase cosine features of the matrix elements respectively, and construct a four-channel real-valued feature matrix; S4. Normalize each channel in the four-channel real-valued feature matrix; S5. Input the normalized feature matrix into the trained deep residual network model, and output the DOA estimation result of the signal incident angle through the deep residual network model.

[0043] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage (acquiring observation data and outputting inversion results) through a communication interface. The above hardware configuration does not constitute a limitation of the present invention.

[0044] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to execute the method steps S1 to S5 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.

[0045] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.

[0046] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0047] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A DOA estimation method based on deep residual network, characterized in that, Includes the following steps: Temporal snapshot data of spatial incident signals are acquired using a uniform linear array, and a vector model of the array-received signal is constructed. The sample covariance matrix is ​​obtained based on the array received signal vector model. Four-channel real-valued feature extraction is performed on the sample covariance matrix to obtain the real part features, imaginary part features, phase sine features and phase cosine features of the matrix elements, and a four-channel real-valued feature matrix is ​​constructed. Normalization is performed on each channel in the four-channel real-valued feature matrix; The normalized feature matrix is ​​input into the trained deep residual network model, and the DOA estimation result of the signal incident angle is output by the deep residual network model. 2.The DOA estimation method based on deep residual network according to claim 1, characterized in that, The array received signal vector model is as follows: wherein, is the array received signal vector at time instant is the incident angle is the corresponding array steering vector, representing the phase delay of the wavefront between the array elements; is the complex amplitude of the th source signal; is the additive noise vector; is the number of far-field narrowband signals.

3. The DOA estimation method based on deep residual networks according to claim 2, characterized in that, The sample covariance matrix is: wherein denotes a conjugate transpose operation, is the number of fast taps.

4. The DOA estimation method based on deep residual networks according to claim 1, characterized in that, The phase sine feature is a feature obtained by taking the sine of the phase angle of each element of the sample covariance matrix, and the phase cosine feature is a feature obtained by taking the cosine of the phase angle of each element of the sample covariance matrix.

5. The DOA estimation method based on deep residual networks according to claim 1, characterized in that, Normalization is performed on each channel of the four-channel real-valued feature matrix, specifically as follows: Obtain the mean and standard deviation of the single-channel real-valued feature matrix; If the standard deviation is less than the set threshold, the channel is determined to be a constant channel, and a matrix of all zeros is returned directly; if the standard deviation meets the requirements, the subsequent steps are executed. Perform Z-score normalization on the channel feature: ,in The channel mean. The standard deviation of the channel. To prevent the minimum value of division by zero; The Z-score-normalized features are subjected to tanh smooth compression to achieve smooth saturation constraint of the features; The compressed features are rescaled to map the feature values ​​to a specified interval, thus completing the normalization process for that channel.

6. The DOA estimation method based on deep residual networks according to claim 1, characterized in that, The deep residual network model is a lightweight deep residual network, whose architecture is as follows: the normalized four-channel real-valued feature matrix is ​​used as input, and the feature extraction backbone is constructed through multiple progressively improved residual blocks; The backbone network integrates multi-scale features from global average pooling and global max pooling at its ends; finally, a fully connected classification head with hierarchical Dropout regularization outputs the DOA classification probability results at the corresponding angular resolution to obtain the DOA estimation results for the signal incident angle.

7. The DOA estimation method based on deep residual networks according to claim 1, characterized in that, The deep residual network model employs an edge angle adaptive class weight strategy during training. The adaptive weight coefficients corresponding to the class perspective are: in, This represents the total number of training samples. The total number of angle categories, For the first The number of samples corresponding to the class perspective. This represents the current edge angle value to be weighted.

8. A DOA estimation system based on deep residual networks, characterized in that, include: The signal acquisition module acquires temporal snapshot data of spatial incident signals through a uniform linear array and constructs a vector model of the array received signals. The covariance matrix construction module obtains the sample covariance matrix based on the array received signal vector model. The multi-channel feature extraction module performs four-channel real-valued feature extraction on the sample covariance matrix, respectively obtaining the real part features, imaginary part features, phase sine features and phase cosine features of the matrix elements, and constructing a four-channel real-valued feature matrix; The feature normalization module performs normalization processing on each channel in the four-channel real-valued feature matrix; The DOA estimation module inputs the normalized feature matrix into the trained deep residual network model, and outputs the DOA estimation result of the signal incident angle through the deep residual network model.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the DOA estimation method based on deep residual networks as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements a DOA estimation method based on a deep residual network as described in any one of claims 1-7.

Images