DOA estimation method based on complex convolutional neural network
By employing a DOA estimation method based on complex convolutional neural networks and utilizing depthwise separable convolutional layers and adaptive filtering algorithms, the problems of high computational complexity and accuracy under noisy environments in radar DOA estimation are solved, achieving efficient and accurate DOA estimation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-29
- Publication Date
- 2026-04-14
AI Technical Summary
Existing deep learning models have high computational complexity and insufficient real-time performance in radar DOA estimation, and their estimation accuracy is not high in complex noise environments.
A DOA estimation method based on complex convolutional neural networks is adopted, which replaces the traditional convolutional layers with depthwise separable convolutional layers, combines an adaptive filtering algorithm for noise suppression, and adds random noise and signal amplitude perturbation through training to improve the network's noise resistance.
While reducing the number of parameters and model complexity, it improves the real-time performance and noise immunity of radar DOA estimation, and enhances estimation accuracy.
Smart Images
Figure CN121856923A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar detection technology, and in particular relates to a DOA estimation method based on complex convolutional neural networks. Background Technology
[0002] Direction of Arrival (DOA) estimation refers to estimating the azimuth information of a signal source based on signals received from a radar array. Traditional DOA estimation algorithms, such as MUSIC and ESPRIT, often rely on precise mathematical models and high signal-to-noise ratio (SNR) conditions. However, their accuracy significantly decreases in real-world scenarios with complex noise backgrounds and multipath effects. Deep learning, using convolutional neural networks (CNNs), possesses powerful nonlinear modeling capabilities, automatically learning complex signal features and noise characteristics from large amounts of data and adapting to complex environmental changes. Combining deep learning algorithms with DOA estimation can improve the accuracy and robustness of DOA estimation. Specifically, the DOA estimation problem is transformed into a classification problem in deep learning. Then, a deep learning architecture with a CNN as the front end is constructed and trained to obtain a CNN-based DOA estimation model.
[0003] Currently, various deep learning models have been applied to the field of DOA estimation, showing better performance than traditional algorithms under low signal-to-noise ratio and few snapshots; however, there are still some problems that need to be solved: (1) In practical application scenarios such as radar and sonar that require fast response, the real-time requirements of DOA estimation algorithms are high, while deep learning models usually have high computational complexity and cannot provide accurate and fast signal source localization results in practical scenarios. Therefore, it is necessary to design a robust neural network structure for DOA estimation without increasing network complexity; (2) During radar detection, the signal received by the array will be interfered with by complex noise. Although the neural network can adapt to various complex background noises during training, how to further improve the stability of the algorithm so that the algorithm can maintain high estimation accuracy in complex noise environments is still a problem that needs to be solved. Summary of the Invention
[0004] The purpose of this invention is to provide a DOA estimation method based on complex convolutional neural networks, which reduces the number of parameters and model complexity, and improves the network model's ability to resist noise interference.
[0005] To achieve the objective of this invention, a DOA estimation method based on complex convolutional neural networks is provided, comprising the following steps:
[0006] S1. Preprocess the complex numerical echo signal of the radar receiving array. The preprocessing is to perform noise suppression by applying an adaptive filtering algorithm to the complex numerical echo signal and convert the autocorrelation matrix of the received noise signal into a complex signal tensor.
[0007] S2. Pass the complex signal tensor through the first channel composed of dilated convolutional layers and the second channel composed of standard convolutional layers, and use their complex convolutional kernels to extract the amplitude and phase features of the complex signal tensor to obtain the feature map of the convolution output.
[0008] S3. The feature map is passed through a fully connected layer and a cross-entropy loss function to obtain the angle classification result, and after training, a complex convolutional neural network for DOA estimation is obtained.
[0009] S4. The noise-suppressed complex-valued echo signal is passed through the complex convolutional neural network to obtain the predicted DOA angle range.
[0010] Compared with the prior art, the significant advancement of this invention lies in the fact that it uses depthwise separable convolutional layers instead of traditional convolutional layers in the network structure, which can adapt to the real-time requirements of radar and reduce the number of parameters and model complexity; at the same time, random noise and signal amplitude perturbation are added when training the deep learning model, which improves the network model's ability to resist noise interference.
[0011] To more clearly illustrate the functional characteristics and structural parameters of the present invention, further explanation is provided below in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description
[0012] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:
[0013] Figure 1 This is a schematic diagram of the structure of the two-dimensional planar array of the present invention;
[0014] Figure 2 This is a schematic diagram of the complex convolutional neural network structure used for DOA estimation in this invention. Detailed Implementation
[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0016] This invention provides a DOA estimation method based on complex convolutional neural networks, combining... Figure 1 This includes the following steps:
[0017] S1. Preprocess the complex numerical echo signal of the radar receiving array. The preprocessing is to perform noise suppression by applying an adaptive filtering algorithm to the complex numerical echo signal and convert the autocorrelation matrix of the received noise signal into a complex signal tensor.
[0018] S2. Pass the complex signal tensor through the first channel composed of dilated convolutional layers and the second channel composed of standard convolutional layers, and use their complex convolutional kernels to extract the amplitude and phase features of the complex signal tensor to obtain the feature map of the convolution output.
[0019] S3. The feature map is passed through a fully connected layer and a cross-entropy loss function to obtain the angle classification result. After training, a complex convolutional neural network (C-CNN) is obtained for DOA estimation.
[0020] S4. The noise-suppressed complex-valued echo signal is passed through the complex convolutional neural network to obtain the predicted DOA angle range.
[0021] S1 includes the following steps:
[0022] S11. Based on the complex-valued echo signal from the radar receiving array, an adaptive filtering algorithm is used to suppress noise in the echo signal.
[0023] S12. Set the phase of the incident signal based on the coordinates of the antenna element and the incident angle, and determine the steering vector of the incident signal according to the phase;
[0024] S13. The guide vector matrix of the received signal is obtained by arranging the guide vectors of the incident signal in sequence.
[0025] S14. Based on the steering vector matrix of the received signal, construct an array receiving signal model through matrix operations;
[0026] S15. The autocorrelation matrix of the noise signal is calculated based on the array received signal model.
[0027] S16. Convert the autocorrelation matrix of the noise signal into a complex signal tensor using the tensor concatenation function in pyTorch.
[0028] Suppose there exists an M×N planar two-dimensional array with isotropic antenna elements, and the coordinates of each antenna element are (m,n), where... , Meanwhile, assuming that K plane waves are incident on the array plane from the elevation plane and the azimuth plane respectively, the array receiving signal model of S14 is specifically shown in the following equation:
[0029] ;
[0030] ;
[0031] in, The steering vector matrix for receiving signals. It is the antenna array output vector. This represents the noise vector of additive zero-mean Gaussian noise. This represents the source vector of the signal to be estimated. Representation matrix The transpose of .
[0032] The autocorrelation matrix of S15 specifically includes:
[0033] The phase of the k-th incident signal at the unit with coordinates (m, n) is defined by the following formula:
[0034] ;
[0035] in, Let be the spacing between the array elements in the x-direction. Let be the spacing between the array elements in the y-direction. Let be the angle between the incident signal and the z-axis. Let be the angle between the projection of the incident signal onto the xOy plane and the x-axis. Let x be the x-coordinate of the array element. The y-coordinate of the array element; , The wavelength of the incident signal;
[0036] The steering vector of the kth incident signal is shown in the following formula:
[0037] ;
[0038] in, Let be the steering vector of the k-th incident signal at the unit with coordinates (m, n).
[0039] The steering vector matrix of the received signal The column vectors are guide vectors pointing to K different arrival directions, as shown in the following formula:
[0040] ;
[0041] in, Let be the transpose of the steering vector of the k-th incident signal. Let be the transpose matrix of the steering vector of the Kth incident signal;
[0042] In order to perform DOA estimation, the correlation matrix of the noise signal must be determined.
[0043] Correlation matrix of received noise signal The specific formula is as follows:
[0044] ;
[0045] in, The correlation matrix of the received signal, To calculate the variance of independent noise signals, I is the identity matrix. For matrix The conjugate transpose of , where A is the steering vector matrix of the received signal. Let E be the conjugate transpose of the steering vector matrix of the received signal, and let E be the autocorrelation matrix.
[0046] Divide the range of DOA angles to be estimated into p non-overlapping intervals according to the accuracy requirements. Treat each interval as a category and construct classification labels: .
[0047] S2 includes the following steps:
[0048] S21. A complex convolutional kernel is constructed based on the first channel composed of dilated convolutional layers and the second channel composed of standard convolutional layers. The two channels are connected in the form of pooling layers to obtain the feature extraction part of the complex convolutional neural network.
[0049] S22. Input the complex signal tensor into the feature extraction part of the complex convolutional neural network to obtain the feature map of the convolution output.
[0050] The feature map output by the convolution in S22 The specific formula is as follows:
[0051] ;
[0052] in, and To determine the output feature map size, For complex number multiplication, For complex bias terms; For complex convolution kernels, where and For core size, Number of output channels; Let be a complex signal tensor, where The number of array elements. For the number of snapshots, This is the number of input channels.
[0053] Combination Figure 2 S3 includes the following steps:
[0054] S31. The complex convolutional neural network of the present invention is constructed based on complex convolutional kernels, fully connected layers, and cross-entropy loss function.
[0055] S32. Obtain the angle classification result by passing the feature map through a fully connected layer and a cross-entropy loss function;
[0056] S33. Using MATLAB simulation, source vectors following a Gaussian distribution are generated. After adding noise, an array received signal vector dataset is obtained. This dataset is then divided into a training set and a test set in a 7:3 ratio. During training, the complex signal tensor is used as input, and the Adam optimizer is used to optimize the network parameters to obtain a complex convolutional neural network for DOA estimation.
[0057] The loss function is specifically shown in the following formula:
[0058] ;
[0059] in, The loss value. For the sample size, Let be the real part of the predicted complex value of the nth sample. Let be the real part of the true complex value of the nth sample. Let be the imaginary part of the predicted complex value of the nth sample. Let be the imaginary part of the true complex value of the nth sample. The actual DOA angle of the signal source to be located. These are the predicted values from the complex convolutional neural network model.
[0060] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A DOA estimation method based on complex convolutional neural networks, characterized in that, Includes the following steps: S1. Preprocess the complex numerical echo signal of the radar receiving array. The preprocessing is to perform noise suppression by applying an adaptive filtering algorithm to the complex numerical echo signal and convert the autocorrelation matrix of the received noise signal into a complex signal tensor. S2. Pass the complex signal tensor through the first channel composed of dilated convolutional layers and the second channel composed of standard convolutional layers, and use their complex convolutional kernels to extract the amplitude and phase features of the complex signal tensor to obtain the feature map of the convolution output. S3. The feature map is passed through a fully connected layer and a cross-entropy loss function to obtain the angle classification result, and after training, a complex convolutional neural network for DOA estimation is obtained. S4. The noise-suppressed complex-valued echo signal is passed through the complex convolutional neural network to obtain the predicted DOA angle range.
2. The DOA estimation method based on complex convolutional neural networks according to claim 1, characterized in that, S1 includes the following steps: S11. Based on the complex-valued echo signal from the radar receiving array, an adaptive filtering algorithm is used to suppress noise in the echo signal. S12. Set the phase of the incident signal based on the coordinates of the antenna element and the incident angle, and determine the steering vector of the incident signal according to the phase; S13. The guide vector matrix of the received signal is obtained by arranging the guide vectors of the incident signal in sequence. S14. Based on the steering vector matrix of the received signal, construct an array receiving signal model through matrix operations; S15. The autocorrelation matrix of the noise signal is calculated based on the array received signal model. S16. Convert the autocorrelation matrix of the noise signal into a complex signal tensor using the tensor concatenation function in pyTorch.
3. The DOA estimation method based on complex convolutional neural networks according to claim 2, characterized in that, The array receiving signal model of S14 is specifically shown in the following formula: ； ； in, The steering vector matrix for receiving signals. It is the antenna array output vector. This represents the noise vector of additive zero-mean Gaussian noise. This represents the source vector of the signal to be estimated. Representation matrix The transpose of .
4. The DOA estimation method based on complex convolutional neural networks according to claim 3, characterized in that, The autocorrelation matrix of S15 specifically includes: The phase of the k-th incident signal at the unit with coordinates (m, n) is defined by the following formula: ； in, Let be the spacing between the array elements in the x-direction. Let be the spacing between the array elements in the y-direction. Let be the angle between the incident signal and the z-axis. Let be the angle between the projection of the incident signal onto the xOy plane and the x-axis. Let x be the x-coordinate of the array element. The y-coordinate of the array element; , The wavelength of the incident signal; The steering vector of the kth incident signal is shown in the following formula: ； in, Let be the steering vector of the k-th incident signal at the unit with coordinates (m, n); The steering vector matrix of the received signal The column vectors are guide vectors pointing to K different arrival directions, as shown in the following formula: ； in, Let be the transpose of the steering vector of the k-th incident signal. Let be the transpose matrix of the steering vector of the Kth incident signal; Correlation matrix of received noise signal The specific formula is as follows: ； in, The correlation matrix of the received signal, To calculate the variance of independent noise signals, I is the identity matrix. For matrix The conjugate transpose of , where A is the steering vector matrix of the received signal. Let E be the conjugate transpose of the steering vector matrix of the received signal, and let E be the autocorrelation matrix. Divide the range of DOA angles to be estimated into p non-overlapping intervals according to the accuracy requirements. Treat each interval as a category and construct classification labels: .
5. The DOA estimation method based on complex convolutional neural networks according to claim 1, characterized in that, S2 includes the following steps: S21. A complex convolutional kernel is constructed based on the first channel composed of dilated convolutional layers and the second channel composed of standard convolutional layers. The two channels are connected in the form of pooling layers to obtain the feature extraction part of the complex convolutional neural network. S22. Input the complex signal tensor into the feature extraction part of the complex convolutional neural network to obtain the feature map of the convolution output.
6. The DOA estimation method based on complex convolutional neural networks according to claim 5, characterized in that, The feature map output by the convolution in S22 The specific formula is as follows: ； in, and To determine the output feature map size, For complex number multiplication, For complex bias terms; For complex convolution kernels, where and For core size, Number of output channels; Let be a complex signal tensor, where The number of array elements. For the number of snapshots, This is the number of input channels.
7. The DOA estimation method based on complex convolutional neural networks according to claim 1, characterized in that, S3 includes the following steps: S31. The complex convolutional neural network of the present invention is constructed based on complex convolutional kernels, fully connected layers, and cross-entropy loss function. S32. Obtain the angle classification result by passing the feature map through a fully connected layer and a cross-entropy loss function; S33. Using MATLAB simulation, source vectors following a Gaussian distribution are generated. After adding noise, an array received signal vector dataset is obtained. This dataset is then divided into a training set and a test set in a 7:3 ratio. During training, the complex signal tensor is used as input, and the Adam optimizer is used to optimize the network parameters to obtain a complex convolutional neural network for DOA estimation.
8. The DOA estimation method based on complex convolutional neural networks according to claim 7, characterized in that, The loss function is specifically shown in the following formula: ； in, The loss value. For the sample size, Let be the real part of the predicted complex value of the nth sample. Let be the real part of the true complex value of the nth sample. Let be the imaginary part of the predicted complex value of the nth sample. Let be the imaginary part of the true complex value of the nth sample. The actual DOA angle of the signal source to be located. These are the predicted values from the complex convolutional neural network model.