A two-dimensional DOA estimation method based on convolutional attention network

Abstract

The application relates to a two-dimensional DOA estimation method based on a convolution attention network, which comprises the following steps: in the first step, a target signal is received by a planar antenna array such as a cross array or a uniform circular array, the received signal is sampled and quantized, correlation information is calculated, and a covariance matrix is obtained; in the second step, the covariance matrix is pretreated, and then feature extraction and parameter calculation are performed on the pretreated data through a designed convolution attention neural network to obtain an angle parameter; and in the third step, the angle parameter output by the convolution attention neural network is calculated to estimate the DOA of the signal. The method performs two-dimensional DOA estimation on the signal through the convolution attention neural network, effectively copes with interference factors possibly existing in a receiving system, has robustness, guarantees high DOA estimation precision, and is simple and light in the network model, thereby reducing hardware cost, being deployable on various devices, and being good in compatibility.

Description

Technical Field

[0001] This invention relates to the field of wireless signal direction finding, specifically to a two-dimensional DOA estimation method based on convolutional neural networks and attention mechanisms. Background Technology

[0002] Direction of Arrival (DOA) estimation for wireless signals is a fundamental problem in radio direction finding, with a long history of research and wide applications in military, public security, aviation, navigation, land and water transportation, and disaster relief. Currently, most DOA estimation methods are based on Multi-Signal Classification (MUSIC) algorithms or methods using the Rotation Invariant Technique (ESPRIT) to estimate signal parameters. However, these methods are mostly based on ideal receiving antenna array conditions and have limited ability to fit interference factors under imperfect conditions. When considering multiple interference factors, optimizing the DOA estimation model for the received signal becomes complex. In reality, receiving antenna arrays may exhibit mutual coupling effects or inconsistent gain / phase effects. To address these interference factors, the excellent feature extraction and efficient modeling capabilities of neural networks are used to overcome the impact of interference factors on DOA estimation.

[0003] A search of existing related technologies revealed that Jiangsu Yixin Aerospace Technology Co., Ltd.'s invention [Invention Patent Application Publication No.: CN115877315A] uses low-precision ADC sampling and quantization of the target signal received by the antenna array, and then uses a deep neural network model to recover the quantized data to perform DOA estimation using the MUSIC algorithm. However, this method is aimed at a one-dimensional DOA estimation problem and still requires the use of the MUSIC algorithm to achieve DOA estimation. Ming-Yi You proposed a two-dimensional DOA estimation method for antenna arrays receiving signals with mutual coupling effects and inconsistent gain / phase effects in "A Unified Two-Dimensional Direction Finding Approach for Sensor Arrays Based on Deep Neural Networks With Inhomogeneous Angle and Frequency Partition," in IEEE Sensors Journal, 2022, 7, 22. (i.e., a unified two-dimensional direction finding method for sensor arrays with inhomogeneous angle and frequency partitioning based on deep neural networks). The effectiveness of this method was demonstrated through experiments. However, this method improves the estimation accuracy by dividing the network into classification and regression networks, which expands the network structure and increases the training and deployment overhead. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a two-dimensional DOA estimation method based on convolutional attention networks to solve the two-dimensional DOA estimation problem in real-world environments with interfering factors, while effectively reducing model training and deployment costs.

[0005] The objective of this invention is achieved through the following technical solution: a two-dimensional DOA estimation method based on convolutional attention networks, comprising the following steps:

[0006] The first step is for the receiver to receive the target signal through a planar antenna array such as a cross array or a uniform circular array, and to sample and quantize the received wireless signal, calculate its relevant information, and obtain the covariance matrix.

[0007] The second step is to preprocess the covariance matrix and then use the designed convolutional attention neural network to extract features and calculate parameters from the processed data to obtain the angle parameters.

[0008] The third step is to calculate the angle parameters output by the convolutional attention neural network to estimate the DOA of the target signal.

[0009] Preferably, in the first step, the planar antenna array is a cross array, the number of antennas is N, and the angle between the target signal and the antenna array is θ = [α, β]. T Where α and β represent the azimuth and elevation angles respectively, and T is the transpose symbol. The signal received by the antenna array is:

[0010] x′(t)=as(t),

[0011] Where s(t) is the target signal, Let τ be the steering vector. i (i = 1, ..., N-1) represents the delay time of the signal reaching the i-th antenna relative to its arrival at the reference antenna, where the reference antenna is the 0-th antenna.

[0012] Preferably: the antenna p is calculated from the antenna coordinates. i =[x i y i , z i ] T With reference antenna p0 = [0, 0, 0] T The baseline vector d between i The unit direction vector r = [cosαcosβ, sinαcosβ, sinβ] is calculated from the angle between the target signal and the antenna array. T The delay τ of the signal reaching the i-th antenna relative to its arrival at the reference antenna is... i It can be represented as:

[0013]

[0014] Where c is the wave velocity of the target signal, the phase difference between the signal and the reference antenna at the i-th antenna is:

[0015]

[0016] Preferably, considering the effects of noise, mutual coupling between receiving antennas, and inconsistent gain / phase effects, the signal received by the antenna array is:

[0017] x(t) = BGas(t) + n(t),

[0018] Where n(t) is additive white Gaussian noise, and is a symmetric matrix with all diagonal elements equal to 1. Representing mutual coupling effects, a diagonal matrix This indicates inconsistent gain / phase effects.

[0019] Preferably: The covariance matrix is ​​obtained by sampling and quantizing the signal received by the antenna array and calculating its relevant information.

[0020]

[0021] Where M is the number of sampling points, H is the conjugate transpose sign, and t j This indicates the j-th sampling time.

[0022] Preferably, in the second step, the covariance matrix is ​​preprocessed to obtain the input data for the convolutional attention neural network:

[0023]

[0024] in, and Let C represent the real and imaginary parts of a complex number, respectively. up Let be the vector formed by the upper triangular elements of the covariance matrix.

[0025] Preferably, the convolutional attention neural network is divided into three parts: a convolutional module part, an attention mechanism module part, and a fully connected regression module part. The convolutional module part contains four convolutional modules, each of which contains one 1D convolutional layer, one batch normalization layer, and one ReLU activation function layer; the attention mechanism module part contains one spatial attention module and one channel attention module; and the fully connected regression module part contains two fully connected layers and one ReLU activation function layer.

[0026] Preferably, the convolutional layer parameters of the four convolutional modules are respectively 16 1×3 convolutional kernels, 64 1×3 convolutional kernels, 64 1×3 convolutional kernels, and 128 1×3 convolutional kernels; the number of nodes in the two fully connected layers are 128 and 4 respectively; the convolutional attention neural network uses a weighted extended mean squared error function as the loss function F. loss :

[0027] F loss =μMSE+vEx,

[0028]

[0029]

[0030] Among them, M B Indicates the sample batch size. This represents the k-th true angle parameter. This represents the k-th angle parameter result of the output of the convolutional attention neural network, where μ and v are both weight coefficients.

[0031] Preferably, the convolutional attention neural network model is trained using the network's input data as training data and the weighted extended mean squared error function as the loss function. The parameters set for each training session include: a learning rate of 0.05, a batch size of 512, and 200 training epochs.

[0032] Preferably, in the third step, the angle parameters output by the convolutional attention neural network are calculated to estimate the DOA unit direction vector of the target signal:

[0033]

[0034] In the prediction phase, the input data of the neural network is obtained by sampling, quantizing and processing the received signal of the receiving antenna array. The angle parameters are obtained by feature extraction and parameter calculation. The angle parameters are calculated using the method described above to complete the DOA estimation of the target signal.

[0035] Preferably, the number of antennas in the receiving antenna array is N=5, the azimuth angle α of the target signal ranges from 0 degrees to 358 degrees, the elevation angle β ranges from 30 degrees to 88 degrees, the frequency ranges from 300 MHz to 350 MHz, and the signal-to-noise ratio (SNR) is 10 dB.

[0036] The beneficial effects of this invention are:

[0037] This invention combines convolutional neural networks with attention mechanisms, leveraging their superior feature extraction and efficient modeling capabilities to overcome interference factors present in real-world receiving antenna arrays. It uses sampled, quantized, and processed received signals to form input data for the network model. The network model, trained with a weighted extended mean square error loss function, fully mines and utilizes the features of the data to obtain angle parameters and calculates an estimated two-dimensional DOA. The convolutional attention neural network described in this invention uses a unified network structure to estimate DOA for wireless signals of different frequencies. While ensuring estimation accuracy, it also features a simple and lightweight network structure, reducing hardware costs, allowing for deployment on various devices with good compatibility, and enhancing its application capabilities in different scenarios, thus possessing significant practical value. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the two-dimensional DOA estimation method based on convolutional attention networks of this invention.

[0039] Figure 2 This is a schematic diagram of the cross-shaped receiving antenna array model corresponding to this invention;

[0040] Figure 3 This is the convolutional attention neural network model designed in this invention;

[0041] Figure 4 This is a comparison chart of DOA estimation errors under different numbers of intermediate convolutional modules in the embodiments of the present invention;

[0042] Figure 5 This is a comparison chart of DOA estimation errors under the influence of token and weighted extended mean square error function in embodiments of the present invention;

[0043] Figure 6 This is a comparison chart of the DOA estimation errors of the embodiments of the present invention and the method based on deep neural networks;

[0044] Figure 7 This is a comparison chart of DOA estimation errors under different frequency perturbations in the embodiments of the present invention. Detailed Implementation

[0045] To facilitate understanding of the present invention, a more detailed description of the invention is provided below with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solutions of the present invention, providing detailed implementation methods and specific operating procedures. They represent preferred embodiments of the invention, but the scope of protection of the present invention is not limited to the following embodiments.

[0046] It should be noted that, unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0047] Figure 1 An example of a two-dimensional DOA estimation method based on convolutional attention networks is shown. Figure 1 The process includes the following steps:

[0048] In step S1, the receiving end receives the target signal through a planar antenna array such as a cross array or a uniform circular array, samples and quantizes the received wireless signal, calculates its relevant information, and obtains the covariance matrix.

[0049] The second step, S2, involves preprocessing the covariance matrix and then using the designed convolutional attention neural network to extract features and calculate parameters from the processed data to obtain the angle parameters.

[0050] The third step, S3, calculates the angle parameters output by the convolutional attention neural network to estimate the DOA of the target signal.

[0051] Preferably: In the first step S1, such as Figure 2 As shown, the planar antenna array is a cross array with N=5 antennas. The target signal includes wireless signals emitted from the same transmitting antenna or reflected signals after being reflected by a target. The angle between the target signal arriving at the receiving antenna array and the antenna array is θ=[α, β]. T Where α and β represent the azimuth and elevation angles, respectively, and T is the transpose. Ignoring path loss and time delay, the signal received by the antenna array is:

[0052] x′(t)=as(t),

[0053] Where s(t) is the target signal, Let τ be the steering vector. i (i = 1, ..., N-1) represents the delay time of the signal reaching the i-th antenna relative to its arrival at the reference antenna, where the reference antenna is the 0-th antenna.

[0054] Preferably: the antenna p is calculated from the antenna coordinates. i =[x i y i , z i ] T With reference antenna p0 = [0, 0, 0] TThe baseline vector d between i =[x i y i , z i ] T The unit direction vector r = [cosαcosβ, sinαcosβ, sinβ] is calculated from the angle between the target signal and the antenna array. T The delay τ of the signal reaching the i-th antenna relative to its arrival at the reference antenna is... i It can be represented as:

[0055]

[0056] Where c = 3e8 is the wave velocity of the target signal, the phase difference between the signal and the reference antenna at the i-th antenna is:

[0057]

[0058] Preferably, considering the effects of noise, mutual coupling between receiving antennas, and inconsistent gain / phase effects, the signal received by the antenna array is:

[0059] x(t) = BGas(t) + n(t),

[0060] Where n(t) is additive white Gaussian noise, and is a symmetric matrix with all diagonal elements equal to 1. Representing mutual coupling effects, a diagonal matrix Indicates inconsistent gain / phase effects, with diagonal elements as follows: Among them, g i and ψ i Let be the gain and phase of the i-th antenna relative to the reference antenna, respectively. Then the phase difference of the signal between the i-th antenna and the reference antenna is:

[0061] φ i '=φ i +ψ i .

[0062] Preferably: The covariance matrix is ​​obtained by sampling and quantizing the signal received by the antenna array and calculating its relevant information.

[0063]

[0064] Where M = 1000 is the number of sampling points, H is the conjugate transpose sign, and t j This indicates the j-th sampling time.

[0065] Preferably, in the second step S2, the covariance matrix is ​​preprocessed to obtain the input data for the convolutional attention neural network:

[0066]

[0067] in, and Let C represent the real and imaginary parts of a complex number, respectively. up =[C 1，2 C 1，3 C 2，3 C 1，N C N-1，N ] T Let C be a vector formed by the upper triangular elements of the covariance matrix, with a length of N(N-1) / 2, where C m，n This represents the element in the m-th row and n-th column of the covariance matrix C.

[0068] Preferably: such as Figure 3 As shown, the convolutional attention neural network is divided into three parts: a convolutional module part, an attention mechanism module part, and a fully connected regression module part. The convolutional module part contains four convolutional modules, each containing one 1D convolutional layer, one batch normalization layer, and one ReLU activation function layer. The attention mechanism module part contains one spatial attention module and one channel attention module. The fully connected regression module part contains two fully connected layers and one ReLU activation function layer. In the convolutional module part of this convolutional attention neural network, there are multiple intermediate convolutional modules with identical parameter settings between the first and last convolutional modules. These convolutional modules are interconnected to form the network's convolutional module part, which is used for feature extraction. Figure 4 As shown, the number of intermediate convolutional modules has different effects on the final DOA estimation. Experiments verified that selecting two intermediate convolutional modules achieves the optimal effect; therefore, the convolutional module section contains four convolutional modules. The attention mechanism module constructs an attention map in the feature space through spatial attention and channel attention modules, thereby overcoming interference factors. In the fully connected regression module, the ReLU activation function layer is located between two fully connected layers. This module performs regression fitting on the final features to obtain the angle parameters of the network output.

[0069] Preferably: such as Figure 3 As shown, the convolutional layer parameters of the four convolutional modules are respectively 16 1×3 convolutional kernels, 64 1×3 convolutional kernels, 64 1×3 convolutional kernels, and 128 1×3 convolutional kernels; the number of nodes in the two fully connected layers are 128 and 4 respectively; the convolutional attention neural network uses a weighted extended mean squared error function as the loss function F. loss :

[0070] F loss =μMSE+vEx,

[0071]

[0072]

[0073] Among them, M B Indicates the sample batch size. This represents the k-th true angle parameter. This represents the k-th angle parameter result output by the convolutional attention neural network, where μ and v are both weight coefficients. In the convolutional module section, the first convolutional module is the input layer, using 16 convolutional kernels, and the middle two convolutional modules have a total of 64 kernels. In the fully connected regression module, the input data of the first fully connected layer needs to include the frequency of the target signal as a token. The weights of the loss function are preferably μ = ν = 1. Figure 5 As shown, the DOA estimation error curves of the proposed method are given with and without tokens and with a weighted extended mean squared error loss function. Preferably, frequency tokens and a weighted extended mean squared error loss function are used.

[0074] Preferably: The convolutional attention neural network model described in step S2 is trained using the network's input data as training data and the weighted extended mean squared error function as the loss function. During training, the learning rate of the neural network model is 0.05, the batch size is 512, and the number of training epochs is 200 per training iteration.

[0075] Preferably: In the third step S3, the angle parameters output by the convolutional attention neural network are calculated to estimate the DOA unit direction vector of the target signal.

[0076]

[0077] In the prediction phase, the input data of the neural network is obtained by sampling, quantizing and processing the received signal of the receiving antenna array. The angle parameters are obtained by feature extraction and parameter calculation. The angle parameters are calculated using the method described above to complete the DOA estimation of the target signal.

[0078] Preferably, the number of antennas in the receiving antenna array is N=5, the azimuth angle α of the target signal ranges from 0 degrees to 358 degrees, the elevation angle β ranges from 30 degrees to 88 degrees, the frequency ranges from 300 MHz to 350 MHz, and the signal-to-noise ratio (SNR) is 10 dB.

[0079] Figure 6The paper presents DOA estimation error curves for embodiments of the proposed method and a deep neural network-based method under different SNRs. As can be seen from the figures, the proposed method outperforms the deep neural network-based method under various SNRs. Furthermore, the network structure of the proposed method is simpler and lighter than that of the deep neural network-based method, reducing hardware costs, allowing for deployment on various devices, and exhibiting good compatibility.

[0080] Figure 7 The DOA estimation error curves of the proposed method under various frequency perturbations are presented. As can be seen from the figures, the proposed method maintains high DOA estimation accuracy under various frequency perturbations and also overcomes the impact of different SNRs on performance. The proposed method exhibits good robustness.

[0081] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments and fall within the scope of protection of the technical solution of this application.

Claims

1. A two-dimensional DOA estimation method based on convolutional attention networks, characterized in that, Includes the following steps: The first step is for the receiver to receive the target signal through a planar antenna array of cross or uniform circular array, and to sample, quantize and calculate the received signal to obtain the covariance matrix. The second step is to preprocess the covariance matrix and then use the designed convolutional attention neural network to extract features and calculate parameters from the processed data to obtain the angle parameters. The third step is to calculate the angle parameters output by the convolutional attention neural network to estimate the DOA of the received target signal; The convolutional attention neural network is divided into three parts: a convolutional module, an attention mechanism module, and a fully connected regression module. The convolutional module section contains four convolutional modules, each of which contains one 1D convolutional layer, one batch normalization layer, and one ReLU activation function layer. The four convolutional modules each use 16 parameters for their convolutional layers. Convolution kernel, 64 Convolution kernel, 64 Convolution kernel and 128 The convolutional kernel; the two fully connected layers have 128 and 4 nodes respectively; the convolutional attention neural network uses a weighted extended mean squared error function as the loss function. : in, Indicates the sample batch size. Indicates the first A true angle parameter The first value represents the output of the convolutional attention neural network. Results of each angle parameter. and All are weighting coefficients. , The first and second parts of the output of the convolutional attention neural network are respectively represented by the first part. Azimuth and elevation angles.

2. The two-dimensional DOA estimation method based on convolutional attention networks according to claim 1, characterized in that, In the first step, the planar antenna array is a cross array, and the number of antennas is... The angle between the target signal and the antenna array is ,in and These represent the azimuth and elevation angles, respectively. Using the transpose symbol, the signal received by the antenna array is: in, For target signal, Let be the steering vector; where For frequency, For the signal to arrive at the The time delay of the 0th antenna relative to the arrival time at the reference antenna. .

3. The two-dimensional DOA estimation method based on convolutional attention networks according to claim 2, characterized in that, Calculate each antenna from its coordinates With reference antenna Baseline vectors between The angle between the target signal and the antenna array Calculate the unit direction vector Then the signal reaches the first The time delay of the root antenna relative to the arrival time of the reference antenna Represented as: in, The wave speed of the target signal.

4. The two-dimensional DOA estimation method based on convolutional attention networks according to claim 3, characterized in that, Considering the effects of noise, mutual coupling between receiving antennas, and inconsistent gain / phase effects, the signal received by the antenna array is: in, It is a symmetric matrix with all diagonal elements equal to 1, and is additive white Gaussian noise. Representing mutual coupling effects, a diagonal matrix This indicates inconsistent gain / phase effects.

5. A two-dimensional DOA estimation method based on a convolutional attention network according to claim 4, characterized in that, By sampling and quantizing the signal received by the antenna array, its relevant information is calculated to obtain the covariance matrix: in, It is the number of sampling points. The sign for conjugate transpose. Indicates the first Each sampling time.

6. A two-dimensional DOA estimation method based on a convolutional attention network according to claim 5, characterized in that, In the second step, the covariance matrix is ​​preprocessed to obtain the input data for the convolutional attention neural network: in, and Let them represent the real and imaginary parts of a complex number, respectively. Let be the vector formed by the upper triangular elements of the covariance matrix. Represents the real number field.

7. A two-dimensional DOA estimation method based on a convolutional attention network according to claim 6, characterized in that, The attention mechanism module includes one spatial attention module and one channel attention module; The fully connected regression module consists of two fully connected layers and one ReLU activation function layer.

8. A two-dimensional DOA estimation method based on a convolutional attention network according to claim 1, characterized in that, The convolutional attention neural network model is trained using the input data of the convolutional attention network as training data and the weighted extended mean squared error function as the loss function. The parameters set for each training session include: a learning rate of 0.05, a batch size of 512, and a training epoch of 200.

9. A two-dimensional DOA estimation method based on convolutional attention networks according to claim 1, characterized in that, In the third step, the angle parameters output by the convolutional attention neural network are calculated to estimate the DOA unit direction vector of the target signal: in, , These represent the azimuth and elevation angles output by the convolutional attention neural network, respectively. In the prediction phase, the input data of the convolutional attention neural network is obtained by sampling, quantizing, and processing the received signal from the receiving antenna array. The angle parameters are obtained through feature extraction and parameter calculation. The angle parameters are calculated using the above formula to complete the DOA estimation of the target signal.

Images