DOA estimation method based on SE and CNN

Abstract

The invention relates to the technical field of DOA estimation, in particular to a DOA estimation method based on SE and CNN, which comprises the following steps: constructing a signal model by using sensor array configuration, acquiring an array receiving signal, and calculating a sampling estimation value of a covariance matrix of the array receiving signal; zero setting is carried out on diagonal elements of a sampling estimation value of the array receiving signal covariance matrix to obtain a sampling estimation covariance matrix, and a real part, an imaginary part and a phase of the sampling estimation covariance matrix are obtained; and constructing an SE-CNN model, and training the SE-CNN model by taking a real part, an imaginary part and a phase of the sampling estimation covariance matrix as input and taking an angle range of the discretized DOA as a label. According to the invention, the precision and robustness of DOA estimation are improved by using the deep learning technology in a complex environment.

Description

Technical Field

[0001] The present invention relates to the technical field of DOA estimation, and in particular to a DOA estimation method based on SE and CNN. Background Art

[0002] DOA estimation, as an important research direction in array signal processing, plays a key role in many fields such as radar, sonar, navigation and wireless communication.

[0003] Traditional DOA estimation methods, such as minimum variance distortionless response (MVDR) and MUSIC, usually rely on assumed signal characteristics and array geometry configurations, but these methods often face problems of unstable accuracy and high computational complexity in practical environments.

[0004] At present, deep learning methods have the following problems in DOA estimation:

[0005] 1. Existing deep learning models are highly dependent on data quality. Especially in low signal-to-noise ratios, the estimation accuracy often drops significantly, and it is difficult to effectively cope with changes in the correlation between signal sources;

[0006] 2. Existing methods usually assume that the array position accuracy is perfect, but in practical applications, the error of array element position and signal-to-noise ratio mismatch will affect the stability of DOA estimation;

[0007] 3. The deep learning model does not adequately model the relationship between channels, resulting in the model’s inability to fully capture the mutual influence between different signal sources in complex scenarios, which in turn affects the accuracy of the estimation. Summary of the invention

[0008] In view of the shortcomings of the existing methods, the present invention uses deep learning technology to improve the accuracy and robustness of DOA estimation in complex environments.

[0009] The technical solution adopted by the present invention is: a DOA estimation method based on SE and CNN comprises the following steps:

[0010] Step 1: construct a signal model using the sensor array configuration, obtain the array received signal, and calculate the sampled estimate of the array received signal covariance matrix;

[0011] As a preferred embodiment of the present invention, step one specifically includes:

[0012] Assume that there are K far-field narrowband signals s(t) = [s1(t),…,s k (t)] T With the incident angle θ=[θ1,…,θ K ] T It is incident on the sensor array, and the array receiving signal at time t is:

[0013]

[0014] Among them, A(θ)=[a(θ1),…,a(θ K )]∈C N×K is the array manifold matrix; n(t)∈C N is Gaussian white noise; is the steering vector of the Kth incident signal, λ=c / f is the wavelength at the carrier frequency f, and c is the speed of sound;

[0015] Calculate the sampled estimate of the array received signal covariance matrix The formula is:

[0016]

[0017] Where T represents the number of snapshots, x(t)=[x1(t),…,x N (t)] T ∈C N is the array received signal, and H is the conjugate transpose.

[0018] Step 2: setting the diagonal elements of the sampled estimated values ​​of the array received signal covariance matrix to zero to obtain the sampled estimated covariance matrix, and calculating the real part, imaginary part and phase of the sampled estimated covariance matrix;

[0019] As a preferred embodiment of the present invention, step 2 specifically includes:

[0020] Construct the sampling estimation covariance matrix, the formula is:

[0021]

[0022] in, is a sampled estimate of the array received signal;

[0023] Estimate the covariance matrix for the sample Perform three-channel division and obtain and They are The real part, imaginary part and phase of .

[0024] Step 3: Build the SE-CNN model, take the real part, imaginary part and phase of the sampled estimated covariance matrix as input, and discretize the angle range of DOA as label to train the SE-CNN model;

[0025] As a preferred embodiment of the present invention, the SE-CNN model includes: a first convolution block, an SE module, a second, third, and fourth convolution blocks, a first fully connected layer, a second fully connected layer, and a third fully connected layer cascaded, and the output layer adopts a Sigmoid activation function; wherein the first convolution block includes: a convolution layer with a filter number of 256, a kernel size of 3×3, and a stride of 2×2, a batch normalization layer, and a ReLU activation function;

[0026] The second, third, and fourth convolution blocks all use convolutional layers with 256 filters, a kernel size of 3×3, a stride of 1×1, a batch normalization layer, and a ReLU activation function;

[0027] The number of neurons in the first to third fully connected layers are 4096, 2048, and 1024, respectively.

[0028] As a preferred embodiment of the present invention, the formula for discretizing the angle range of DOA is:

[0029] The discretized grid is defined as:

[0030] φ={-G min ,…,-G ρ ,0°,G ρ ,…,G max} (7)

[0031] Among them, φ represents the set of signal arrival directions;

[0032] Considering DOA estimation as a classification task, the number of categories is:

[0033]

[0034] Among them, G max Indicates the maximum reaching direction; -G min Indicates the minimum reaching direction; G ρ is the resolution of the direction estimate.

[0035] As a preferred embodiment of the present invention, step three also includes: in the test phase, the real part, the imaginary part and the phase of the sampled estimated covariance matrix are used as the input of the SE-CNN model, the confidence probability of each discrete angle is output, and the direction estimation is completed by selecting the angle with the largest probability. The formula is:

[0036]

[0037] Among them, φ i is the predicted DOA of the ith signal source, Pr(G i =q|X) means that when the observation data X is given, the signal source G i The probability of being in direction q.

[0038] As a preferred embodiment of the present invention, binary cross entropy loss is used as the optimization target of the SE-CNN model.

[0039] As a preferred embodiment of the present invention, a DOA estimation system based on SE and CNN includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement a DOA estimation method based on SE and CNN.

[0040] As a preferred embodiment of the present invention, a computer readable medium stores a computer program code, and when the computer program code is executed by a processor, a DOA estimation method based on SE and CNN is implemented.

[0041] Beneficial effects of the present invention:

[0042] 1. By setting the diagonal elements of the array covariance matrix sampling estimate to zero and using its real part, imaginary part and phase as three-channel input, the time-frequency characteristics of the signal are fully utilized to effectively reduce noise interference;

[0043] 2. In terms of network design, the SE attention mechanism is combined with the CNN convolutional neural network to improve the sensitivity to important signal features, and local and global features are extracted through multi-layer convolution operations to enhance the ability to capture spatial consistency;

[0044] 3. Through training in a low signal-to-noise ratio environment, it is verified that the model can still accurately predict DOA under high signal-to-noise ratio, showing good adaptability and robustness. Comprehensive evaluation results show that the present invention is significantly better than existing methods under complex noise conditions, with high accuracy and strong generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the network structure in the present invention;

[0046] Figure 2 It is a schematic diagram of the SE attention mechanism module in the present invention;

[0047] Figure 3 It is MUSIC, R-MUSIC, ESPRIT, UnESPRIT, l 2,1 - Comparison of DOA estimation performance of SVD, CNN, ResNet, and SENet under low signal-to-noise ratio;

[0048] Figure 4 It is MUSIC, R-MUSIC, ESPRIT, UnESPRIT, l 2,1 - Comparison of DOA estimation errors of SVD, CNN, ResNet, and SENet under low signal-to-noise ratio;

[0049] Figure 5 It is MUSIC, R-MUSIC, ESPRIT, UnESPRIT, l 2,1 - DOA estimation of SVD, CNN, ResNet, and SENet under different signal-to-noise ratios;

[0050] Figure 6 It is MUSIC, R-MUSIC, ESPRIT, UnESPRIT, l 2,1 - DOA estimation performance of SVD, CNN, ResNet, and SENet under different snapshots. DETAILED DESCRIPTION

[0051] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.

[0052] like Figure 1 As shown, a DOA estimation method based on SE and CNN includes the following steps:

[0053] Step 1: construct a signal model using a sensor array configuration; the sensor array is referred to as an array;

[0054] Building a signal model includes:

[0055] Assume that there are K far-field narrowband signals s(t) = [s1(t),…,s k (t)] T With the incident angle θ=[θ1,…,θ K ] T It is incident on the sensor array, and the array receiving signal at time t is:

[0056]

[0057] where x(t)=[x1(t),…,x N (t)] T ∈C N The array receives the signal; A(θ)=[a(θ1),…,a(θ K )]∈C N×K is the array manifold matrix; n(t)∈C N is Gaussian white noise, T represents the number of snapshots; is the steering vector of the Kth incident signal, λ=c / f is the wavelength at the carrier frequency f, and c is the speed of sound.

[0058] When the array receives T snapshots, the array receiving signal can be expressed in the following vector form:

[0059] X=AS+N (2)

[0060] Where X = [x(1),…,x(T)]∈C N×T The array receives the signal, S = [s(1), ..., s(T)] ∈ C K×T is the incoming signal waveform, N=[n(1),…,n(T)]∈C N×T Receive noise for the array.

[0061] Formula (1) is the representation of a single moment (single snapshot), and formula (2) is the representation after integrating the received signals at multiple moments into a matrix; the expression method is expanded from "one sampling point" to a collection of "multiple sampling points".

[0062] The covariance matrix of the array output can be given by:

[0063] R = E[x(t)x H (t)]

[0064] =AE[s(t)s H (t)]A H +E[v(t)v H (t)]

[0065] =AR s A N +R n

[0066] =AR s A N +σ 2 I N (3)

[0067] in, and R n =σ 2 I N denote the signal covariance matrix and the noise covariance matrix respectively. diag[·] denotes generating a diagonal matrix with the given sequence as the diagonal; [·] H represents the conjugate transpose of the matrix; E[·]Table I N Show expectations. represents the power of the Kth signal; σ 2 Represents the noise power; I N is the N×N identity matrix.

[0068] In practical applications, the covariance matrix of the array is unknown, and the sampled estimate of the covariance matrix of the array received signal is usually used. To express, It can be calculated based on T snapshots:

[0069]

[0070] Step 2: Design a training data generation method, use the real part, imaginary part and phase information of the covariance matrix after the diagonal is zeroed as input features, and combine the direction of arrival (DOA) label under the set signal-to-noise ratio (SNR) to generate a training data set;

[0071] Set the diagonal elements of the covariance matrix to zero and define a matrix whose diagonal elements are the same as the diagonal elements of The same, the other elements are zero, that is:

[0072]

[0073] Among them, diag() means generating a diagonal matrix with the given sequence as the diagonal.

[0074] Then, the diagonal elements of the sampled estimate of the covariance matrix are set to zero to obtain the sampled estimate covariance matrix

[0075]

[0076] in, The diagonal elements are all zero.

[0077] In DOA estimation, the diagonal elements of the array covariance matrix are often affected by noise and mainly reflect the autocorrelation information of the signal rather than the effective cross-correlation relationship. In order to solve this problem, the present invention innovatively adopts formulas (5) and (6) to remove noise interference by setting the diagonal elements to zero, thereby enhancing the expression ability of the matrix non-diagonal elements for the true cross-correlation between signals. This strategy is particularly suitable for multi-signal source scenarios and can significantly improve the accuracy of DOA estimation. At the same time, this processing method also enhances the robustness of the algorithm in complex noise environments, so that it can still maintain stable and reliable performance under low signal-to-noise ratio or strong interference conditions.

[0078] The input data is a matrix with the diagonal of the covariance matrix set to zero, with a size of N×N×3, and the third dimension represents different "channels"; the first and second channels are the real and imaginary parts of the covariance matrix, that is, and The third channel corresponds to the phase, i.e.

[0079] Each channel carries different signal characteristics. The real part mainly reflects the amplitude information of the signal, while the imaginary part carries the phase information of the signal. The phase information is crucial to the estimation of the spatial distribution and arrival angle of the signal source. By inputting these three signal characteristics separately, the model can better capture the multi-dimensional relationship between signals and avoid the information loss and mutual interference caused by traditional single feature input.

[0080] Step 3: Convert the direction estimation problem into a multi-label classification task, discretize the angle range of DOA and perform label encoding on the training data;

[0081] The discretized grid is defined as:

[0082] φ={-G min ,...,-G ρ ,0°,G ρ ,...,G max} (7)

[0083] Where φ represents the set of all possible signal arrival directions; G max Indicates the maximum reaching direction; -G min Indicates the minimum reaching direction; G ρ is the resolution of the direction estimate (in degrees);

[0084] Through the modeling of formula (7), DOA estimation is regarded as a classification task, and the number of categories is:

[0085]

[0086] A deep learning model (SE-CNN network) based on the Squeeze-and-Excitation (SE) attention mechanism and convolutional neural network (CNN) was constructed. The expressiveness of important features was enhanced by adaptively recalibrating the weights of each feature channel, optimizing hyperparameters and evaluating the performance of the model on the validation set to ensure the robustness and generalization ability of the model.

[0087] like Figure 1 , 2 As shown in the figure, in terms of network architecture, the model uses a combination of multi-layer convolution, batch normalization (BatchNormalization), activation function (ReLU), Squeeze-and-Excitation (SE) modules and fully connected layers; the input layer accepts a tensor of shape [N, M, channels]; where N and M are the dimensions of the covariance matrix, and channels represents the number of channels (including real part, imaginary part and phase).

[0088] The first convolutional block contains a convolutional layer with 256 filters, a kernel size of 3×3, and a stride of 2×2, with valid padding. Subsequently, a batch normalization layer normalizes the convolution output, and then a ReLU activation function is used to add nonlinearity.

[0089] The SE module is introduced; the feature map is compressed into the global features of each channel through global average pooling, and then the channel weights are adaptively recalibrated through two fully connected layers, and finally multiplied channel by channel with the original convolution output to enhance important features;

[0090] The second, third, and fourth convolution blocks all use convolutional layers with 256 filters, a kernel size of 3×3, and a stride of 1×1 to reduce the spatial dimensions of the feature map in turn; each convolutional layer is followed by a batch normalization layer and a ReLU activation function to accelerate the training process, stabilize model performance, and alleviate the gradient vanishing problem.

[0091] After four convolution blocks, the feature map is flattened into a one-dimensional vector and enters the fully connected layer.

[0092] The first fully connected layer contains 4096 neurons and 30% Dropout is applied to prevent overfitting;

[0093] The second fully connected layer contains 2048 neurons and is also equipped with 30% Dropout;

[0094] The third fully connected layer contains 1024 neurons and continues to use 30% Dropout.

[0095] Finally, the output layer uses the Sigmoid activation function, and the number of neurons is 121, corresponding to the number of categories, to achieve multi-label binary classification tasks, and can predict the existence probability of multiple directions at the same time.

[0096] In this embodiment, the angle range is set to -60 to 60 degrees, and the angle resolution is 1, so there are 121 directions.

[0097] During the model training process, the Adam optimizer was used, the initial learning rate was set to 0.001, the loss function was binary cross entropy (binary_crossentropy), and the evaluation indicator was binary accuracy (BinaryAccuracy); in order to optimize the learning rate, the ReduceLROnPlateau callback function was used, and when the validation loss did not decrease within 10 cycles, the learning rate was reduced to 0.7 times the original. The model was trained in 200 cycles, the batch size was 32, the data was randomly shuffled before each cycle, and the validation set accounted for 10% of the total data.

[0098] Figure 1 The parameter settings of the key layers are shown in Table 1.

[0099] Table 1 Parameter settings

[0100]

[0101] The design of the SE-CNN network of the present invention has many advantages. The introduced SE module enhances the model's ability to focus on important features and improves the effectiveness of feature expression by adaptively recalibrating channel weights. The batch normalization and ReLU activation function after each convolutional layer not only accelerate the training process, but also stabilize the model performance and alleviate the problem of gradient disappearance. The deep fully connected network effectively learns complex feature relationships through multi-layer nonlinear transformations combined with Dropout, while preventing overfitting. The dynamic learning rate adjustment strategy ensures that the model can perform more detailed parameter optimization in the later stage of training, further improving the performance of the model. Finally, the output layer activated by Sigmoid enables the model to adapt to multi-label binary classification tasks, which is suitable for direction estimation requirements in complex signal processing scenarios.

[0102] In the test phase, the real part, imaginary part, and phase of the sampled estimated covariance matrix are used as input, and the confidence probability of each discrete angle is output. The direction estimation is completed by selecting the angle with the largest probability.

[0103] Selecting the angle with the highest probability to complete the direction estimation is formally expressed as:

[0104]

[0105] Among them, φ i is the predicted DOA of the ith signal source, Pr(G i =q|X) means that when the observation data X is given, the signal source G i The probability of being in direction q.

[0106] In the present invention, the binary cross entropy loss function is used as the optimization target of the model. The binary cross entropy loss function guides the optimizer to gradually adjust the model parameters by measuring the difference between the model prediction value and the actual label, which can effectively improve the classification performance of the model. The formula is as follows:

[0107]

[0108] in, is the predicted value of the model, Z (i) (n) is the actual label value.

[0109] For each label of each sample, the logarithmic difference between the predicted probability and the actual label value is calculated and the negative value is taken. Then, the loss values ​​for all labels are summed and averaged.

[0110] The performance of the model is compared with that of traditional methods (such as MUSIC, ESPRIT, etc.) under various conditions such as different signal-to-noise ratios, different numbers of snapshots, changes in signal source correlation, and array element position errors, and the improvement effect of the deep learning method in terms of angular resolution and robustness is analyzed.

[0111] In all experiments, the signal-to-noise ratio (SNR) is defined by the following formula, which ensures that in the case of multiple signal sources, the overall signal-to-noise ratio of the system depends on the strength of the weakest signal source. It is defined as:

[0112]

[0113] in, represents the signal power of the ith signal source, Represents the noise power.

[0114] RMSE is used to quantify the error of the SE-CNN neural network in DOA estimation; a low RMSE value indicates a smaller error between the model prediction and the actual value, and better performance. Compared with other error indicators (such as mean absolute error MAE), RMSE is more sensitive to large errors and can more effectively reflect the performance of the model in extreme cases. It is defined as:

[0115]

[0116] in, represents the kth actual angle of arrival (DOA, ground truth) in the mth test sample, represents the k-th predicted arrival angle in the m-th test sample, D te is the total number of test samples in each experiment, and K is the number of DOAs for each sample.

[0117] Figure 3 , 4 It is MUSIC, R-MUSIC, ESPRIT, UnESPRIT, l 2,1 -Comparison of DOA estimation results of SVD, CNN, ResNet, and SENet under low signal-to-noise ratio.

[0118] The experimental results show that the RMSE of SENet is 0.2563, that of ResNet is 0.37278, that of CNN is 0.31239, that of MUSIC is 5.0138, that of R-MUSIC is 0.34934, that of ESPRIT is 0.35265, that of UnESPRIT is 0.27154, and that of 2,1-SVD is 1.0022. The estimated value of the MUSIC algorithm deviates from the true value between samples 1 to 20 and 100-116, with an error of up to ±15°; most data points of R-MUSIC, ESPRIT and UnESPRIT are closely around the true value, but the performance is weak at the edge of the grid angle, with large error fluctuations; they show poor performance and high errors; deep learning methods are significantly better than traditional methods in DOA estimation. The error of ResNet is usually within ±2 degrees, and the error of CNN is usually within ±1 degree, with small error fluctuations. The error of SENet on most samples is controlled within ±0.5 degrees, and the error distribution is extremely uniform, showing extremely high estimation accuracy and stability.

[0119] Figure 5 It is MUSIC, R-MUSIC, ESPRIT, UnESPRIT, l 2,1 - DOA estimation results of SVD, CNN, ResNet, and SENet under different signal-to-noise ratios.

[0120] Experimental results show that the root mean square error (RMSE) of most algorithms gradually decreases with the increase of signal-to-noise ratio (SNR). Among them, R-MUSIC, ESPRIT and UnESPRIT perform particularly well in the high SNR area, and their RMSE values ​​are close to the Cramer-Rao lower bound (CRLB uncr ), indicating that its estimation performance is close to the theoretical optimum; in the entire signal-to-noise ratio range, l 2,1 -SVD's RMSE fluctuates greatly, especially under low SNR conditions, where the RMSE increases significantly. Under extremely low SNR conditions, SENet shows lower RMSE than two deep learning-based methods, ResNet and CNN. As SNR increases, SENet still shows good robustness and maintains a low RMSE level. It is worth noting that even without training for high SNR conditions, SENet can still accurately estimate angles without any parameter adjustments.

[0121] Figure 6 It is MUSIC, R-MUSIC, ESPRIT, UnESPRIT, l 2,1 - DOA estimation results of SVD, CNN, ResNet, and SENet under different snapshots.

[0122] The traditional MUSIC and R-MUSIC methods have high RMSE when the number of snapshots is small (T < 500), and the MUSIC method has the largest error. When the number of snapshots increases, the RMSE decreases significantly, but is still higher than other methods. The ESPRIT and UnESPRIT methods perform well in high SNR areas, and their RMSE is close to the Cramer-Rao lower bound (CRLB uncr ); however, when the number of snapshots is small, the RMSE of the ESPRIT method fluctuates greatly, showing a certain instability. 2,1 -SVD method has large fluctuations in the whole range of snapshot numbers, especially when the number of snapshots is small, the error is high.

[0123] Deep learning methods (CNN, ResNet, and SENet) perform well under low snapshot numbers, especially SENet, which has low RMSE over the entire snapshot number range; SENet has the lowest RMSE under extremely low snapshot numbers (T = 100), showing superior robustness and adaptability. As the number of snapshots increases, the RMSE of SENet continues to decrease, and it can provide accurate DOA estimates without retraining the model under different snapshot numbers.

[0124] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A DOA estimation method based on SE and CNN, characterized in that: The following steps are involved: Step 1: construct a signal model using the sensor array configuration, obtain the array received signal, and calculate the sampling estimate of the covariance matrix of the array received signal; Step 2: Set the diagonal elements in the sampled estimated values ​​of the array received signal covariance matrix to zero to obtain the sampled estimated covariance matrix, and calculate the real part, imaginary part and phase of the sampled estimated covariance matrix; Step 3: Construct the SE-CNN model, take the real part, imaginary part and phase of the sampled estimated covariance matrix as input, and the angular range of the discretized DOA as the label, and train the SE-CNN model.

2. The DOA estimation method based on SE and CNN according to claim 1, characterized in that: Step 1 specifically includes: Assume that there are K far-field narrowband signals s(t)=[s1(t),...,s k (t)] T With the incident angle θ=[θ1,...,θ K ] T It is incident on the sensor array, and the array receiving signal at time t is: Among them, A(θ)=[a(θ1),...,a(θ K )]∈C N×K is the array manifold matrix; n(t)∈C N is Gaussian white noise; is the steering vector of the Kth incident signal, λ=c / f is the wavelength at the carrier frequency f, and c is the speed of sound; Calculate the sampled estimate of the array received signal covariance matrix The formula is: Where T represents the number of snapshots, x(t)=[x1(t),...,x N (t)] T ∈C N is the array received signal, and H is the conjugate transpose.

3. The DOA estimation method based on SE and CNN according to claim 1, characterized in that: Step 2 specifically includes: Construct the sampling estimation covariance matrix, the formula is: in, is a sampled estimate of the array received signal; Estimate the covariance matrix for the sample Perform three-channel division and obtain and They are The real part, imaginary part and phase of .

4. The DOA estimation method based on SE and CNN according to claim 1, characterized in that: The SE-CNN model includes: the first convolution block, SE module, second, third, and fourth convolution blocks, the first fully connected layer, the second fully connected layer, and the third fully connected layer in cascade, and the output layer uses the Sigmoid activation function; among them, the first convolution block includes: a convolution layer with 256 filters, a kernel size of 3×3, and a stride of 2×2, a batch normalization layer, and a ReLU activation function; The second, third, and fourth convolution blocks all use convolutional layers with 256 filters, a kernel size of 3×3, a stride of 1×1, a batch normalization layer, and a ReLU activation function; The number of neurons in the first to third fully connected layers are 4096, 2048, and 1024, respectively.

5. The DOA estimation method based on SE and CNN according to claim 1, characterized in that: The formula for discretizing the angular range of DOA is: The discretized grid is defined as: φ={-G min ,…,-G ρ ,0°,G ρ ,…,G max } (7) Among them, φ represents the set of signal arrival directions; Considering DOA estimation as a classification task, the number of categories is: Among them, G max Indicates the maximum reaching direction; -G min Indicates the minimum reaching direction; G ρ is the resolution of the direction estimate.

6. The DOA estimation method based on SE and CNN according to claim 1, characterized in that: Step 3 also includes: in the test phase, the real part, imaginary part and phase of the sampled estimated covariance matrix are used as the three-channel input of the SE-CNN model, the confidence probability of each discrete angle is output, and the angle with the largest probability is selected to complete the direction estimation. The formula is: Among them, φ i is the predicted DOA of the ith signal source, Pr(G i =q|X) means that when the observation data X is given, the signal source G i The probability of being in direction q.

7. The DOA estimation method based on SE and CNN according to claim 1, characterized in that: Binary cross entropy loss is used as the optimization objective of the SE-CNN model.

8. The DOA estimation system based on SE and CNN is characterized by: include: a memory for storing instructions executable by a processor; A processor, configured to execute instructions to implement the DOA estimation method based on SE and CNN as described in any one of claims 1 to 7.

9. A computer readable medium storing a computer program code, characterized in that: When the computer program code is executed by a processor, the DOA estimation method based on SE and CNN is implemented as described in any one of claims 1 to 7.

Images