360-degree whole-airspace direction-of-arrival estimation method based on space sectorization neural network

Abstract

The invention discloses a 360-degree whole-airspace direction-of-arrival estimation method based on a space sectorization neural network, which mainly solves the problems of poor accuracy, high model training complexity and the like in the whole-airspace interval direction-of-arrival estimation in the prior art, and comprises the following implementation steps of: modeling a whole-airspace received signal and constructing neural network input; dividing the airspace into sectors and calculating sector indexes; mapping the direction of arrival of the whole airspace to a reference sector by using a sector index; designing a space sectorization neural network; training and reasoning a space sectorization neural network; and recovering the direction of arrival of the whole airspace by using the obtained sector information and the mapped angle value. According to the method, mapping and estimation of the full-airspace direction of arrival in a reference sector which is a high-performance interval are achieved by introducing space sector information, the method has the advantages of being high in direction of arrival estimation accuracy, high in robustness, easy to train and the like, and the method has wide prospects in practical application such as a new-generation wireless communication system, passive positioning and target detection.

Description

Technical Field

[0001] The present invention belongs to the field of signal processing technology, and in particular relates to the estimation of the direction of arrival of radar signals, acoustic signals and electromagnetic signals in a full airspace scenario. Specifically, it is a 360° full airspace direction of arrival estimation method based on a spatial sectorized neural network, which can be used for passive positioning, target detection and a new generation of wireless communication systems. Background Art

[0002] Direction-of-Arrival Estimation (DOA) uses array antennas to receive spatial signals and processes the received signals through statistical signal processing techniques and various optimization methods to recover the DOA information contained in the signals. It is one of the basic problems in the field of array signal processing and has a wide range of applications in radar, sonar, speech, radio astronomy, seismology, wireless communications, medical imaging and other fields.

[0003] Model-driven methods, such as MUSIC and ESPRIT algorithms, are classic DOA estimation methods. In order to ensure the immediacy and accuracy of the estimation, such methods usually presuppose that the signal source arrives from a limited angle range. Within this range, the model-driven method can obtain a higher estimation accuracy with an acceptable computational complexity. The present invention refers to such an interval as a high-performance interval, and typical high-performance intervals include 0 to 45 degrees, -60 to 60 degrees, etc. However, in actual application scenarios such as target direction finding and passive positioning represented by the new generation of vehicle networking, array antennas need to receive omnidirectional and full-airspace signals. When the model-driven method faces a 360° full-airspace DOA estimation that exceeds the high-performance interval, some signal sources located near the boundary of the angle period will experience a sudden drop in the accuracy of the DOA estimation due to estimation errors.

[0004] With the increasing complexity of the communication environment, the DOA estimation method based on neural network has shown strong adaptability in terms of estimation accuracy, robustness and inference speed. Its data-driven characteristics enable such methods to handle DOA estimation in various angle intervals and are no longer limited by high-performance intervals. However, since the 360° full-space DOA estimation belongs to a two-dimensional DOA estimation with a large angle range, the traditional DOA estimation method based on neural network still faces challenges in this scenario. On the one hand, the classification neural network assumes that the signal arrives from certain specific directions and classifies these directions as DOA; in order to maintain good estimation accuracy, the interval between adjacent specific directions is usually small, which makes the number of categories required for the classification network very large under the full-space angle condition, and the training cost is extremely high. On the other hand, although the recurrent neural network can directly estimate the AOA, due to the 360° periodicity of the AOA in the full-space scenario, the recurrent network will cause serious estimation errors at the boundary of the angle value period range. Therefore, how to design a neural network structure for DOA estimation in full airspace scenarios to solve the performance degradation of existing technologies in 360° full airspace DOA estimation is an important issue that needs to be solved urgently. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and propose a 360° full-spatial-domain direction of arrival estimation method based on a spatial sectorized neural network. By modeling antenna array signals, a signal covariance matrix for training the neural network is obtained; then, the range of the angle of arrival in the full airspace is effectively compressed through spatial sectorization and angle mapping; finally, a spatial sectorized neural network is designed to estimate the direction of arrival, thereby avoiding the problem of too many categories required by the classification network in the full-spatial-domain scenario and the estimation error of the regression network at the angle boundary, thereby realizing the direction of arrival estimation in the full-spatial-domain scenario, and providing technical support for its application in the fields of new generation wireless communication systems, passive positioning, target detection, etc.

[0006] The object of the present invention is to achieve the following technical solution: a 360° full airspace direction of arrival estimation method based on a spatial sectorized neural network, comprising the following steps:

[0007] (1) Modeling of the full spatial domain received signal and construction of neural network input: Assume that there are K signals from {(θ k ,φ k ), k = 1, 2, ..., K} angular direction, where the azimuth angle θ k ∈[0°, 360°], pitch angle φ k∈[-90°, 90°], indicating that the signal arrives from the entire 360° airspace; the azimuth angle is the angle between the signal arrival direction and the x-axis after being projected onto the antenna array plane, and the elevation angle is the angle between the signal arrival direction and the antenna array plane; at the receiving end, an antenna array is formed using M×N physical antenna elements for signal reception; the signal received by the array is modeled as:

[0008]

[0009] in, represents the received signal waveform matrix of K sources, [·]T represents the transposition operation, Contains the waveform of the kth source, T is the number of sampling snapshots; is the array steering matrix, where a k is the steering vector corresponding to the kth source, expressed as:

[0010]

[0011] Among them, (u i , v i ), i = 1, 2, ..., MN, represents the actual position coordinates of the i-th array element in the plane array, and u1 = v1 = 0; μ k = sinφ k cosθ k , v k = sinφ k sinθ k , represents the imaginary unit; N represents additive white Gaussian noise; using the received signal matrix X, the covariance matrix of the signal is calculated As the neural network model input, [·]H represents the conjugate transpose operation;

[0012] (2) Divide the airspace into sectors and calculate the sector index: Substitute the azimuth angle θ k and the pitch angle φ k The angle range is divided into P and Q regions of equal size, that is, the entire airspace is divided into P×Q sectors; the arrival angle of the kth source belongs to a specific sector, and its index h k Calculate as follows:

[0013]

[0014] in, Respectively represent the range of azimuth and elevation angles in the sector, Indicates a round-down operation; for example: the kth source from {θ k =120°,φ k=60°}, assuming that the space is divided into P=8, Q=2, that is, A θ =45°, A φ =90° sector, then this source belongs to h k = Sector 10;

[0015] (3) Mapping the full-space arrival angle to the reference sector using the sector index: The existing neural network-based method processes the full-space arrival angle θ = [θ1, θ2…, θ K ],φ=[φ1,φ2…,φ K ], there is a performance degradation due to the increase in model complexity, and it is necessary to use the sector index to map the wave arrival angle value range to the reference sector to compress the angle value range; here, the reference sector is zero (h k =0) sector; the numerical relationship between the mapped wave angle, sector index and the full airspace wave angle is expressed by the following formula:

[0016]

[0017]

[0018] in, and They represent the azimuth and elevation angles of the k-th source after the arrival angle is mapped in the reference sector, and mod(·) represents the modulo operation, h=[h1,h2,…,h K ] represents the index of the sector corresponding to the source; after angle mapping, the full spatial wave arrival angle {θ, φ} is a triplet with a compressed value range The equivalent representation is: if this triplet can be estimated from the input signal, the full spatial domain angle of arrival can also be recovered;

[0019] (4) Spatial sectorization neural network design: After the spatial sectorization and angle mapping process, the original full-space AOA is equivalent to a triplet Therefore, it is necessary to design a neural network architecture for estimating this triplet; the spatial sectorization neural network architecture consists of two sub-networks: the classifier and the estimator, which are used to estimate the sector index h and the mapped wave arrival angle information respectively. The classifier is a classification neural network, and the estimator is a regression neural network, both using the signal covariance matrix As input; the classifier uses a convolutional neural network and a multi-layer perceptron cascade structure, and the output is K probability distributions corresponding to each source. The index corresponding to the maximum value in the kth probability distribution represents the spatial sector index h corresponding to the source's arrival angle k, k = 1, 2, ..., K; the estimator uses a convolutional neural network as the network structure, and the output is 2K values, corresponding to the azimuth and elevation angles of the K sources in the reference sector after angle mapping. This network model is designed to estimate triples Designed for;

[0020] (5) Spatial sectorized neural network training and inference: In the training phase, the classifier and estimator are trained independently. Both the classifier and the estimator use the gradient descent method with a learning rate of α for training, and the training objective is to minimize the loss function. In the inference phase, the signal covariance matrix is ​​input into the model, and the output of the classifier and the estimator are the estimated values ​​of the sector index and the mapped angle.

[0021] (6) Use the obtained sector information and the mapped angle value to restore the full spatial domain arrival angle: When the spatial sectorization neural network estimates the sector index h and the angle mapped arrival angle After that, the full airspace bode angle is restored through the equivalent relationship between this triplet and the full airspace bode angle, that is:

[0022]

[0023]

[0024] in, and They represent the sector index estimated by the spatial sectorization neural network and the angle of arrival after angle mapping respectively; This is the direction of arrival estimation result in the full airspace scenario.

[0025] Furthermore, the antenna array described in step (1) is laid out using a uniform rectangular array: two integers M and N are selected, M rows of array elements are set up on the x-axis, N array elements are set up in each row along the y-axis, and the array element spacing d is half the wavelength of the incident narrowband signal; for the array element in the i-th row and j-th column in the array, its position coordinate is (u i , v j ),u i =(i-1)d,v j =(j-1)d; wherein i=1,2,...,M, j=1,2,...,N; thereby an antenna array for signal reception is obtained.

[0026] Furthermore, the number of sectors described in step (2) needs to refer to the following performance balance principle: when selecting the number of sectors, an increase in the number of sectors will increase the number of categories that the classifier needs to identify and reduce performance; but at the same time, it will further reduce the sector angle range and increase the accuracy of the estimator; when selecting the number of sectors, do not allow excessive classification errors to overwhelm the performance gain brought about by the increase in estimator accuracy.

[0027] Furthermore, the number of sectors described in step (2) needs to be determined comprehensively with reference to the following principles: the sector size needs to be appropriate to ensure the accuracy of the estimator; the number of sectors should not be too many or too few; the performance balance principle of the classifier and the estimator; taking the above principles into consideration, the number of sectors is selected as P=8, Q=4, that is, the azimuth and elevation angle ranges of each sector are both 45°.

[0028] Furthermore, the loss function in step (5) is selected in the following manner: the classifier uses a binary cross entropy function as the loss function, namely:

[0029]

[0030] Where D = KPQ is the number of classifier output neurons, p = [p1, p2, ..., p D ]∈{0,1} and are the probability distributions of the true and classifier-estimated corresponding sectors, respectively;

[0031] The estimator uses the Huber function as the loss function, namely:

[0032]

[0033] in, are the real and estimated angles of arrival after angle mapping, respectively; ||·||1 and ||·||2 represent the Manhattan norm and the Euclidean norm, respectively; η is a threshold used to control the choice of the Huber function between the mean square error (MSE) and the mean absolute error (MAE). This threshold control makes the Huber function more robust than the separate MSE or MAE function.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] (1) The present invention divides the airspace into sectors and angle mapping, thereby achieving effective compression of the value range of the full airspace wave angle, and equates the full airspace wave angle with a triplet with a limited value range, thereby achieving accurate estimation of the full airspace wave direction.

[0036] (2) The present invention combines the advantages of classification neural network and recurrent neural network, and can achieve relatively stable 360° full-space DOA estimation under low signal-to-noise ratio conditions, with the advantage of strong robustness. At the same time, the classifier category is limited by the number of sectors, and the training complexity of the spatial sectorized neural network is lower than that of the existing DOA estimation technology based on classification network. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is the overall flow chart of the present invention.

[0038] Figure 2 It is a schematic diagram of the spatial sector segmentation and angle mapping process in the present invention.

[0039] Figure 3 It is a schematic diagram of the spatial sectorized neural network framework in the present invention.

[0040] Figure 4 It is a performance comparison diagram of the root mean square error of the method proposed in the present invention as the signal-to-noise ratio changes.

[0041] Figure 5 It is a performance comparison diagram of the root mean square error of the method proposed in the present invention as the azimuth angle range changes.

[0042] Figure 6 It is a performance comparison diagram of the root mean square error of the method proposed in the present invention as the pitch angle range changes. DETAILED DESCRIPTION

[0043] The technical solutions and effects of the present invention are further described in detail below with reference to the accompanying drawings.

[0044] Regarding the application of DOA estimation technology in practical systems such as the new generation of connected vehicles, full-airspace target direction finding, and passive positioning, existing DOA estimation methods based on neural networks face challenges such as complex classification network models and poor estimation accuracy of regression networks. In order to address the above challenges, the present invention introduces spatial sector information to achieve the mapping of full-airspace DOA in the high-performance interval of the reference sector, and combines the advantages of the classification network and the regression network to propose a 360° full-airspace DOA estimation method based on spatial sectorized neural networks. Figure 1 , the implementation steps of the present invention are as follows:

[0045] Step 1: Modeling of the full spatial domain received signal and construction of neural network input. Assume that there are K signals from {(θ k ,φ k ), k = 1, 2, ..., K} angular direction, where the azimuth angle θ k ∈[0°, 360°], pitch angle φ k∈[-90°, 90°], indicating that the signal arrives from the entire 360° airspace; the azimuth angle is the angle between the signal arrival direction and the x-axis after being projected onto the antenna array plane, and the elevation angle is the angle between the signal arrival direction and the antenna array plane. At the receiving end, M×N physical antenna array elements are used to form a uniform rectangular array (Uniform Rectangular Array) for signal reception, and the array element spacing is half the wavelength of the incident narrowband signal. The signal received by the array can be modeled as:

[0046]

[0047] in, represents the signal matrix of K sources, [·] T represents the transpose operation, is the waveform of the kth signal source, T is the number of sampling snapshots; is the array steering matrix, where

[0048]

[0049] is the steering vector of the kth source, where μ k = sinφ k cosθ k , v k = sinφ k sinθ k ,and Represents the imaginary unit. N represents additive Gaussian white noise. Using the received signal matrix, the covariance matrix of the signal can be estimated The neural network uses this covariance matrix as input. Since most existing deep learning frameworks only support real number operations, it is necessary to stack the real and imaginary parts of the covariance matrix into a three-dimensional tensor as the real input of the network model, where [·] H Represents the conjugate transpose operation.

[0050] Step 2: Divide the airspace into sectors and calculate the sector index. Figure 2 , the azimuth angle θ k and the pitch angle φ k The angle range of is divided into P and Q regions of equal size, that is, the entire airspace is divided into P×Q sectors. The arrival angle of the kth source belongs to a specific sector, and its index h k It can be calculated as follows:

[0051]

[0052] in, Respectively represent the value range of azimuth angle and elevation angle in the sector, and Indicates the rounding operation. For example: The kth source is from {θ k =120°,φ k =60°}, assuming that the space is divided into P=8, Q=2, that is, A θ =45°, A φ =90° sector, then this source belongs to h k =Sector No. 10.

[0053] Step 3: Use the sector index to map the full airspace arrival angle to the reference sector. The full airspace arrival angle θ = [θ1, θ2…, θ K ],φ=[φ1,φ2…,φ K ] does not belong to the high-performance interval, and the existing method has performance degradation. It is necessary to use the sector index to map the wave angle value range to the reference sector to compress the angle value range. Here, the reference sector is the zero (h=0) sector, because it belongs to the high-performance interval, and the estimation accuracy in this sector is high. The numerical relationship between the mapped wave angle, sector index and full airspace wave angle can be expressed by the following formula:

[0054]

[0055]

[0056] in, and They represent the azimuth and elevation angles of the k-th source after the arrival angle is mapped in the reference sector, and mod(·) represents the modulo operation, h=[h1,h2,…,h K ] represents the index of the sector corresponding to the source. After angle mapping, the full spatial domain arrival direction angle {θ, φ} can be expressed as a triplet with a compressed value range. The equivalent representation is , and the value range of this triplet has been significantly compressed. If this triplet can be estimated from the input signal, the full spatial direction of arrival angle can also be recovered.

[0057] Step 4: Spatial sectorization neural network design. After the spatial sectorization and angle mapping process, the original full-space AOA is equivalent to a triplet Therefore, it is necessary to design a neural network architecture for estimating this triplet from the input signal. The spatial sectorization neural network architecture consists of two sub-networks: the classifier and the estimator, which are used to estimate the sector index and the mapped wave angle information respectively. Figure 3Both the classifier and the estimator use the signal covariance matrix as the network input. The classifier is a classification neural network that uses a convolutional neural network and a multi-layer perceptron cascade structure. The output is K probability distributions corresponding to each source. The index corresponding to the maximum value in the kth probability distribution represents the spatial sector index h corresponding to the source's arrival angle. k The estimator uses a convolutional neural network as the network structure, and the output is 2K values, corresponding to the azimuth and elevation angles of the K sources in the reference sector after angle mapping. This network model is designed to estimate triples To avoid the model being too complex, the network depth is kept within a small range, so the pooling layer is not necessary in the part of the spatial sectorized neural network involving the convolutional neural network.

[0058] Step 5: Spatial sectorization neural network training and reasoning, estimating the sector index and the angle value after mapping. In the training phase, the classifier and the estimator are trained independently. Both the classifier and the estimator use the gradient descent method and are trained with a learning rate of α. The training goal is to minimize the loss function. The classifier uses the binary cross entropy function as the loss function, that is:

[0059]

[0060] Where D = KPQ is the number of classifier output neurons, p = [p1, p2, ..., p D ]∈{0,1} and are the true and classifier estimated probability distributions of the corresponding sectors, respectively.

[0061] The estimator uses the Huber function as the loss function, namely:

[0062]

[0063] in, are the real and estimated angles of arrival after mapping, respectively, ||·||1 and ||·||2 represent the Manhattan norm and the Euclidean norm, respectively. η is a threshold used to control the choice of the Huber function between the mean square error (MSE) and the mean absolute error (MAE). This threshold control makes the Huber function more robust than the separate MSE or MAE function. Here η=1.

[0064] In the inference phase, the signal covariance matrix is ​​input into the model, and the output of the classifier and estimator are the estimated values ​​of the sector index and the mapped angle.

[0065] Step 6: Use the obtained sector information and the mapped angle value to restore the full spatial domain arrival angle. When the spatial sectorization neural network estimates the sector index h and the angle mapped arrival angle After that, the full airspace bode angle can be restored through the equivalent relationship between this triplet and the full airspace bode angle, that is:

[0066]

[0067]

[0068] in, and They represent the sector index estimated by the full-space neural network and the angle of arrival after angle mapping. The calculation result of the above formula is the direction of arrival estimation result in the full-space scenario.

[0069] The present invention divides the airspace into sectors and angle mappings, designs a spatial sectorized neural network, realizes the wave direction estimation of the entire airspace, avoids the problem of too many categories required for the classification network in the full airspace scenario and the estimation error of the regression network at the angle boundary, and has the advantages of high estimation accuracy, easy training, and strong robustness.

[0070] The effect of the method proposed in the present invention is further described below in conjunction with a simulation example.

[0071] The number of elements of the antenna array is selected as M=N=7, that is, the dimension of the signal covariance tensor of the input network is 49×49×2. The number of sources K=1, and the number of sampling snapshots is selected as 250. When selecting the number of sectors, the increase in the number of sectors will increase the number of categories that the classifier needs to identify and the performance will decrease; but at the same time, it will further reduce the sector angle range and increase the accuracy of the estimator. Too much classification error will overwhelm the increase in the accuracy of the estimator. Considering the performance trade-off here, and taking into account the principle that the reference sector must be in the high-performance range and the number of sectors should not be too many or too few, the number of sectors is selected as P=8, Q=4, that is, the azimuth and elevation angle ranges of each sector are 45°. The spatial sectorized neural network optimizer uses the Adam optimizer.

[0072] Simulation Example 1: Under the condition that the azimuth angle and elevation angle {θ, φ} are randomly selected between [0°, 360°] and [-90°, 90°] respectively, the performance comparison diagram of the root mean square error of the proposed method with the existing ESPRIT method (model-driven method) and convolutional neural network method (data-driven method) versus signal-to-noise ratio is shown in the figure. Figure 4As shown. For each SNR simulation setting, the root mean square error is obtained by averaging 10,000 Monte Carlo experiments. From the comparison results shown in the figure, it can be seen that since the classifier has high classification accuracy under the appropriate number of sectors and the estimator is in the high performance range, the root mean square error of the method proposed in the present invention is better than the existing method under each SNR condition, and has the advantages of high accuracy and strong robustness.

[0073] Simulation Example 2: Under the condition of a signal-to-noise ratio of 5 dB, the performance comparison of the root mean square error of the proposed method with the existing ESPRIT method and convolutional neural network method as the azimuth and elevation angles vary is shown in the figure below: Figure 5 and Figure 6 For each azimuth or elevation angle range, the root mean square error is obtained by averaging 10,000 Monte Carlo experiments. Figure 5 It can be seen from the comparison results that, since the spatial sector segmentation maps the airspace angles into the reference sector, the performance of the method proposed in the present invention can remain relatively stable when the azimuth and elevation angle value ranges are expanded, showing its unique advantages in the full airspace scenario.

[0074] In summary, the present invention compresses the value range of the wave arrival angle by introducing spatial sector information, avoids the excessive number of categories required by the classification network in the full airspace scenario and the estimation error of the regression network at the angle boundary, has the advantages of high accuracy, strong robustness and easy training, and has broad prospects for new generation wireless communication systems and practical applications such as passive positioning and target detection.

Claims

1. A 360° full-space direction of arrival estimation method based on spatial sectorized neural network, characterized in that: The following steps are involved: (1) Modeling of the full spatial domain received signal and construction of neural network input: Assume that there are K signals from {(θ k ,φ k ),k=1,2,…,K}, where the azimuth angle θ k ∈[0°,360°], pitch angle φ k ∈[-90°,90°], indicating that the signal arrives from the entire 360° airspace; the azimuth angle is the angle between the signal arrival direction and the x-axis after being projected onto the antenna array plane, and the elevation angle is the angle between the signal arrival direction and the antenna array plane; at the receiving end, an antenna array is formed using M×N physical antenna elements for signal reception; the signal received by the array is modeled as: in, represents the received signal waveform matrix of K sources, [·] T represents the transpose operation, Contains the waveform of the kth source, T is the number of sampling snapshots; is the array steering matrix, where a k is the steering vector corresponding to the kth source, expressed as: Among them, (u i ,v i ), i = 1, 2, ..., MN, represents the actual position coordinates of the i-th array element in the plane array, and u1 = v1 = 0; μ k = sinφ k cosθ k , ν k = sinφ k sinθ k , J represents the imaginary unit; N represents additive Gaussian white noise; using the received signal matrix X, calculate the signal covariance matrix As the neural network model input, where [·] H represents the conjugate transpose operation; (2) Divide the airspace into sectors and calculate the sector index: Substitute the azimuth angle θ k and the pitch angle φ k The angle range is divided into P and Q regions of equal size, that is, the entire airspace is divided into P×Q sectors; the arrival angle of the kth source belongs to a specific sector, and its index h k Calculate as follows: in, Respectively represent the range of azimuth and elevation angles in the sector, Indicates floor operation; (3) Mapping the full-space arrival angle to the reference sector using the sector index: The existing neural network-based method processes the full-space arrival angle θ = [θ1, θ2…, θ K ],φ=[φ1,φ2…,φ K ], there is a performance degradation due to the increase in model complexity, and it is necessary to use the sector index to map the wave arrival angle value range to the reference sector to compress the angle value range; here, the reference sector is zero (h k =0) sector; the numerical relationship between the mapped wave angle, sector index and the full airspace wave angle is expressed by the following formula: in, and They represent the azimuth and elevation angles of the k-th source after the arrival angle is mapped in the reference sector, and mod(·) represents the modulo operation, h=[h1,h2,…,h K ] represents the index of the sector corresponding to the source; after angle mapping, the full spatial wave arrival angle {θ, φ} is a triplet with a compressed value range. The equivalent representation; (4) Spatial sectorization neural network design: After the spatial sectorization and angle mapping process, the original full-space AOA is equivalent to a triplet Therefore, it is necessary to design a neural network architecture for estimating this triplet; the spatial sectorization neural network architecture consists of two sub-networks: the classifier and the estimator, which are used to estimate the sector index h and the mapped wave arrival angle information respectively. The classifier is a classification neural network, and the estimator is a regression neural network, both using the signal covariance matrix As input; the classifier uses a convolutional neural network and a multi-layer perceptron cascade structure, and the output is K probability distributions corresponding to each source. The index corresponding to the maximum value in the kth probability distribution represents the spatial sector index h corresponding to the source's arrival angle k , k = 1, 2, ..., K; the estimator uses a convolutional neural network as the network structure, and the output is 2K values, corresponding to the azimuth and elevation angles of the K sources in the reference sector after angle mapping. (5) Spatial sectorized neural network training and inference: In the training phase, the classifier and estimator are trained independently. Both the classifier and the estimator use the gradient descent method with a learning rate of α for training, and the training objective is to minimize the loss function. In the inference phase, the signal covariance matrix is ​​input into the model, and the output of the classifier and the estimator are the estimated values ​​of the sector index and the mapped angle. (6) Use the obtained sector information and the mapped angle value to restore the full spatial domain arrival angle: When the spatial sectorization neural network estimates the sector index h and the angle mapped arrival angle After that, the full airspace bode angle is restored through the equivalent relationship between this triplet and the full airspace bode angle, that is: in, and They represent the sector index estimated by the spatial sectorization neural network and the angle of arrival after angle mapping respectively; This is the direction of arrival estimation result in the full airspace scenario.

2. The 360° full-spatial-domain direction-of-arrival estimation method based on spatial sectorized neural network according to claim 1, characterized in that: The antenna array described in step (1) is laid out using a uniform rectangular array: two integers M and N are selected, M rows of array elements are set up on the x-axis, N array elements are set up in each row along the y-axis, and the array element spacing d is half the wavelength of the incident narrowband signal; for the array element in the i-th row and j-th column in the array, its position coordinate is (u i ,v j ),u i =(i-1)d,v j =(j-1)d; wherein i=1,2,…,M, j=1,2,…,N; thereby an antenna array for signal reception is obtained.

3. The 360° full-spatial-domain direction-of-arrival estimation method based on spatial sectorized neural network according to claim 1, characterized in that: The number of sectors described in step (2) needs to refer to the following performance balance principle: when selecting the number of sectors, an increase in the number of sectors will increase the number of categories that the classifier needs to identify and reduce performance; but at the same time, it will further reduce the sector angle range and increase the accuracy of the estimator; when selecting the number of sectors, do not allow excessive classification errors to overwhelm the performance gain brought by the increase in estimator accuracy.

4. The 360° full-spatial-domain direction-of-arrival estimation method based on spatial sectorized neural network according to claim 1, characterized in that: The number of sectors described in step (2) needs to be determined comprehensively with reference to the following principles: the sector size needs to be appropriate to ensure the accuracy of the estimator; the number of sectors should not be too many or too few; the performance balance principle between the classifier and the estimator; taking the above principles into consideration, the number of sectors is selected as P = 8, Q = 4, that is, the azimuth and elevation angle ranges of each sector are both 45°.

5. The 360° full-spatial-domain direction-of-arrival estimation method based on spatial sectorized neural network according to claim 1, characterized in that: The loss function described in step (5) is selected as follows: the classifier uses a binary cross entropy function as the loss function, namely: Where D = KPQ is the number of classifier output neurons, p = [p1, p2, ..., p D ]∈{0,1} and are the probability distributions of the true and classifier-estimated corresponding sectors, respectively; The estimator uses the Huber function as the loss function, namely: in, are the real and estimated angles of arrival after angle mapping, ||·||1 and ||·||2 represent the Manhattan norm and Euclidean norm, respectively; η is a threshold used to control the choice of the Huber function between the mean square error (MSE) and the mean absolute error (MAE).

Images