Collaborative learning assisted DOA estimation method and system of dual hybrid structure

Abstract

The invention discloses a DOA (direction of arrival) estimation method and a DOA estimation system for a dual hybrid structure assisted by collaborative learning. The DOA estimation method specifically comprises the following steps: constructing a dual MIMO (multiple input multiple output) receiver H2AD-FD system model in which an all-digital FD sub-array and a heterogeneous hybrid H2AD structure are combined; generating a plurality of candidate angle sets by performing DOA estimation on each group in the H2AD structure; for the FD sub-array, a rough estimation value of the radiation source direction is calculated by adopting a multilayer deep learning algorithm or a Root-MUSIC method; taking the rough estimation value of the FD sub-array as a true solution class initial sample point, and establishing a relationship between the FD mode and the H2AD mode so as to deduce a true solution set of the H2AD mode; and an optimal weighted fusion algorithm is designed based on a minimum mean square error principle, two parts of true solutions are synthesized, and a final DOA estimated value is obtained. According to the invention, low-time-delay and high-time-efficiency accurate direction finding is realized.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technologies, and particularly relates to a method and system for DOA estimation with a dual hybrid structure assisted by collaborative learning. Background Art

[0002] Direction of Arrival (DOA) estimation has important research significance in wireless communication networks and radar communications, and has currently been widely applied in many fields such as radar, navigation, UAV communication, intelligent transportation, millimeter-wave communication, etc. Different from traditional active direction-finding technologies, passive DOA estimation is a passive direction-finding method that measures the direction by processing the phase difference of the received electromagnetic waves between different antennas.

[0003] Fully digital (FD) large-scale multiple-input multiple-output (MIMO) antenna arrays have been widely applied in data transmission, direction of arrival (DOA) measurement, etc. due to their excellent rate and resolution performance. However, it faces two main challenges: high computational complexity and circuit cost. These two problems can be well solved by a hybrid analog-digital (HAD) structure. However, HAD has the problem of phase ambiguity, resulting in low efficiency or high time delay. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for DOA estimation with a dual hybrid structure assisted by collaborative learning, which has low cost, low complexity, high time efficiency, and can quickly eliminate phase ambiguity.

[0005] The technical solution for achieving the purpose of the present invention is: a method for DOA estimation with a dual hybrid structure assisted by collaborative learning, comprising the following steps:

[0006] Step 1, construct a dual MIMO receiver HAD-FD system model combining a fully digital FD subarray and a heterogeneous hybrid HAD structure; 2 AD structure; 2 AD-FD system model;

[0007] Step 2, perform DOA estimation on each group in the HAD structure to generate multiple candidate angle sets; 2 AD structure to generate multiple candidate angle sets;

[0008] Step 3, for the FD subarray, use a multi-layer deep learning algorithm or the Root-MUSIC method to calculate a rough estimated value of the radiation source direction;

[0009] Step 4, use the rough estimated value of the FD subarray as the initial sample point of the true solution class, and establish the relationship between the FD and HAD modalities to infer the true solution set of the HAD modality; 2 AD modalities to infer the true solution set of the HAD modality; 2 AD modality;

[0010] Step 5: Design an optimal weighted fusion algorithm based on the principle of minimum mean square error, synthesize the two parts of the true solution, and obtain the final DOA estimation value.

[0011] Further, in Step 1, the all-digital FD subarray and the heterogeneous hybrid H 2 AD structure are combined to form a dual MIMO receiver H 2 The basic parameters of the AD-FD system are specifically as follows:

[0012] A far-field narrowband signal where x(t) and f c are the baseband signal and the carrier frequency respectively, and t represents time. Considering an FD with M antennas, the output signal is expressed as:

[0013] y FD (t) = a FD (θ 0 )x(t) + w(t) (1.1)

[0014] where is the additive white Gaussian noise (AWGN) vector, is the noise variance, I is the identity matrix, and a FD represents the array manifold vector, which is expressed as:

[0015]

[0016] where d = λ / 2 represents the antenna element spacing, i.e., half of the wavelength λ; in the H 2 AD structure, the uniform linear array with M antennas is divided into H groups, each group contains N h antennas, and each group is further divided into K h subarrays, each subarray contains M h antennas, i.e., In this array, M 1 ≠ M 2 ≠ … ≠ M H and M 1 , M 2 , …, M H values are preferably chosen as prime numbers;

[0017] Considering that ψ h,k,m represents the phase corresponding to analog beamforming, the output of the kth subarray is:

[0018]

[0019] where represents the noise in the kth subarray of the hth group, and τ h,k,m is the propagation delay established by the direction of the signal source relative to the array, which is expressed as:

[0020]

[0021] Among them, τ 0 represents the propagation delay from the radiation source to the reference point of the antenna array, and c represents the speed of light;

[0022] Stack the outputs y h of all the K h,k th sub-arrays. The baseband signal vector of the hth group is expressed as:

[0023]

[0024] Among them, y h (t) is a vector formed by stacking the output signals y h,k (t) of the kth sub-array of the hth subgroup,

[0025] In addition, is an additive white Gaussian noise complex vector with a dimension of K h ×1, M h is the number of antennas included in the sub-array of the hth group, is the noise at the M h th antenna, and (·) H represents the conjugate transpose operation, is the array manifold vector, which is expressed as:

[0026]

[0027] And Υ A,h represents a block diagonal matrix:

[0028]

[0029] Among them, γ A,h,k is the kth block diagonal element:

[0030]

[0031] Among them, represents the phase shift of the kth sub-array in the hth subgroup. Through the analog-to-digital converter ADC, y h (t) is converted to:

[0032]

[0033] where n = 1, 2,..., L, L is the number of snapshots, θ 0 is the ideal incident angle, and a h (θ 0 ) represents the array manifold vector of the ideal incident angle θ 0 , and w(n) is the noise.

[0034] Further, in step 2, by performing DOA estimation on each group in the H 2 AD structure, multiple candidate angle sets are generated. The specific steps are as follows:

[0035] Each subarray in the h-th group is regarded as a virtual antenna. Assuming the analog beamforming vector Then K h The output vector of the subarray is:

[0036]

[0037] where y 1 (n), y 2 (n), …, y Kq (n) represent a series of received signals in the h-th group, where represents a series of noises in the h-th group, is the array manifold of each virtual subarray, expressed as:

[0038]

[0039] and b h (θ 0 ) is a constant obtained by summing all elements of each subarray, defined as:

[0040]

[0041] The covariance matrix of the output vector y h (n) is:

[0042]

[0043] where denotes taking the expectation, is the signal-to-noise ratio of the received signal, is the noise variance; R h The eigenvalue decomposition EVD of is:

[0044] R h = [U S U N Σ[U S U N H (1.14)

[0045] where U N is the K h ×(K h - 1) noise subspace, U S is the K h ×1 signal subspace,​

[0046] is K h ×K h diagonal matrix;

[0047] Then, the pseudo-spectrum corresponding to the virtual antenna array is:

[0048]

[0049] where θ represents the arrival angle of the initial input, and the direction corresponding to the peak of P H2AD (θ) is the final DOA estimate; in addition, the optimal transmitting source direction is obtained by Root-MUSIC, and the polynomial equation is defined as:

[0050]

[0051] where Q ij represents the element located in the i-th row and j-th column of matrix Q;

[0052]

[0053] where and f(z) has 2K h -2 roots, that is, Z RM ={z i , i ∈ {1, 2, …, 2K h -2}}, where z i represents the i-th root of the polynomial equation (1.16), and the DOA estimate set is obtained as follows:

[0054]

[0055] where

[0056]

[0057] Then, the DOA estimate of the h-th array group selects the root closest to the unit circle:

[0058]

[0059] is a function of with a period of 2π Therefore, a feasible solution containing M h solutions is:

[0060]

[0061] where

[0062]

[0063] Combining all H groups gives:

[0064]

[0065] where j h ∈ {1, 2, …, M h}, M H is the number of subarray antennas in the H-th group, H represents the number of groups, θ H,jH represents the set of candidate solutions for the H-th group, represents the estimated value of the H-th group, j H is the imaginary unit of the H-th group, representing the cyclic period index in

[0066]

[0067] The total candidate set includes solutions, where each candidate set has one true solution and M h - 1 false solutions.

[0068] Furthermore, in step 3, for the FD subarray, a multi-layer deep learning algorithm or the Root-MUSIC method is used to calculate a rough estimate of the radiation source direction. The specific steps are as follows:

[0069] (1) Estimate the initial rough DOA value through the Root-MUSIC method:

[0070] The covariance matrix of y h (n) is:

[0071]

[0072] where represents the signal-to-noise ratio of the received signal,

[0073] Then, the corresponding spatial spectrum function is as follows:

[0074]

[0075] where a FD (θ) represents the array manifold vector. By constructing a 2(M - 1) polynomial equation through Root-MUSIC, where M represents the number of antennas of the FD subarray, the root closest to the unit circle is the DOA estimate i.e., the initial mean of the true solution class:

[0076]

[0077] (2) Estimate the initial rough DOA value through the CNN method:

[0078] 1) Dataset and its labels: The input data of the CNN model is a real-valued matrix, expressed as:

[0079]

[0080] where represent the real part, imaginary part, and phase part respectively; assuming S represents the number of training sets, the input samples of the CNN are defined as I′ = {I 1 , I 2 , …, I S}; the angles of the training data are selected from p grid points in the range [-θ, θ], where θ ∈ [1°, 90°] and p is the interval unit; therefore, the desired grid is defined as Δ = {-θ, …, -p, 0°, p, …, θ}, and the binary vector corresponding to the training angle is used as the label; Δ = {-45°, …, -1°, 0°, 1°, …, 45°} has |Δ| = 91 grid points when θ = 45° and p = 1°; in addition, the angle -45° is converted into a binary vector

[0081] Therefore, the training dataset and the corresponding label set of the CNN are:

[0082] I′ = {I 1 , I 2 , …, I S} (1.29)

[0083] V = {ν 1 , ν 2 , …, ν S} (1.30)

[0084] 2) CNN structure: The goal of DOA estimation based on CNN is f: The mapping from the input to the output of the CNN network is defined as:

[0085]

[0086] where f Ci (i = 1, …, 5) is the i-th convolutional layer, and f Fj (j = 1, …, 4) is the j-th fully connected layer;

[0087] 3) Loss function:

[0088]

[0089] where and represent the true label and the predicted value respectively, μ represents the weight decay coefficient, and W

[0090] are the weighted parameters of the FC layer, and the model is trained by minimizing the loss function L.

[0091] Furthermore, in step 4, the rough estimate of the FD sub-array is used as the initial sample point of the true solution class, and the relationship between the FD and H 2 AD modalities is established to infer the true solution set of the H 2 AD modality. The specific steps are as follows:

[0092] The true solution of the h-th group is expressed as:

[0093]

[0094] Therefore, the true solution set of the H 2 AD modality is:

[0095]

[0096] Furthermore, in step 5, an optimal weighted fusion algorithm is designed based on the principle of minimum mean square error to synthesize the two parts of the true solution and obtain the final DOA estimation value. The specific steps are as follows:

[0097] Fusing the two parts of the true solution to obtain the optimal DOA estimation:

[0098]

[0099] where

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106] where L represents the number of snapshots, SNR represents the signal-to-noise ratio, represents the real part of b h 2 (θ 0 ).

[0107] The present invention also provides a DOA estimation system with a dual hybrid structure for collaborative learning assistance, which is used to implement the DOA estimation method with the dual hybrid structure for collaborative learning assistance. The system includes a first module to a fifth module, and the specific functions of each module are as follows:

[0108] The first module constructs a dual MIMO receiver H 2 AD structure combined with a full-digital FD subarray to form a dual MIMO receiver H 2 AD-FD system model;

[0109] The second module generates multiple candidate angle sets by performing DOA estimation on each group in the H 2 AD structure;

[0110] The third module calculates a rough estimate of the radiation source direction for the FD subarray using a multi-layer deep learning algorithm or the Root-MUSIC method;

[0111] The fourth module uses the rough estimate of the FD subarray as the initial sample points of the true solution class, and establishes the relationship between the FD and H 2 AD modes to infer the true solution set of the H 2 AD mode;

[0112] The fifth module designs an optimal weighted fusion algorithm based on the minimum mean square error principle to synthesize the two parts of the true solutions and obtain the final DOA estimate value.

[0113] The present invention also provides a mobile terminal, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the DOA estimation method with the dual hybrid structure for collaborative learning assistance.

[0114] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the steps in the DOA estimation method with the dual hybrid structure for collaborative learning assistance.

[0115] Compared with the prior art, the present invention has the following significant advantages: (1) The DOA estimation framework with the dual hybrid structure based on collaborative learning and multi-modal deep learning proposed by the present invention can achieve low-latency and high-timeliness phase ambiguity elimination; (2) The method proposed by the present invention realizes ultra-low latency target direction finding. A complete direction finding can be completed in a single sampling time slot, and the direction finding speed is twice that of the traditional hybrid structure, and the computational complexity and circuit cost are much lower than those of the full-digital MIMO structure; (3) The collaborative learning proposed by the present invention can significantly improve the true solution clustering process. Description of the Drawings

[0116] Figure 1Flowchart of the DOA estimation method with a dual hybrid structure for collaborative learning assistance in the present invention.

[0117] Figure 2 Schematic diagram of the overall structure of the DOA estimation system with a dual hybrid structure for collaborative learning assistance in the present invention.

[0118] Figure 3 DOA estimation model diagram of the FD array based on CNN provided by the present invention.

[0119] Figure 4 Relationship curve diagram of the mean square error and signal-to-noise ratio of the MD-Root-MUSIC method of the present invention.

[0120] Figure 5 Relationship curve diagram of the mean square error and signal-to-noise ratio of the MDDL method of the present invention. Detailed implementation manners

[0121] In order to make the objectives, technical solutions and advantages of the present invention more obvious, exemplary embodiments according to the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments of the present invention. It should be understood that the present invention is not limited by the exemplary embodiments described herein. Based on the embodiments of the present invention described herein, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present invention.

[0122] In the following description, numerous specific details are given to provide a more thorough understanding of the present invention. However, it is obvious to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, in order to avoid confusion with the present invention, some technical features well known to those skilled in the art are not described.

[0123] It should be understood that the present invention can be implemented in different forms and should not be construed as limited to the embodiments presented herein. On the contrary, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the present invention to those skilled in the art.

[0124] The present invention relates to the technical research on eliminating phase ambiguity of the DOA estimation (CoMDDL and CoMD-RootMUSIC) methods with a dual hybrid structure of multi-modal deep learning based on collaborative learning.

[0125] The present invention provides a DOA estimation method and system of a dual hybrid structure of multimodal deep learning based on collaborative learning, which is composed of a full digital structure mode, a heterogeneous hybrid structure mode and a multimodal fusion device. The basic implementation steps of the method are as follows: the heterogeneous hybrid structure uses the Root-MUSIC method to generate multiple candidate solution sets; the full digital subarray constructs a multi-layer deep learning algorithm to calculate a rough estimate of the direction of the radiation source, and uses it as the initial clustering center of the true solution class in the candidate solution set, which can significantly accelerate the rapid clustering of positive and false solution classes and remove pseudo solutions in the candidate solution set; finally, based on the principle of minimum mean square error, an optimal weighted fusion algorithm is designed to synthesize the two parts of the true solution to obtain an improved DOA estimate. The present invention realizes accurate direction finding with low latency and high timeliness.

[0126] Combination Figures 1 to 3 The present invention provides a DOA estimation method of a dual hybrid structure of multimodal deep learning based on collaborative learning, the method comprising the following steps:

[0127] Step 1: Construct the fully digital (FD) sub-array and the heterogeneous hybrid (H 2 A dual MIMO receiver (H AD) structure combined with 2 AD-FD) system model;

[0128] Step 2: By 2 Each group in the AD structure performs DOA estimation to generate multiple candidate angle sets;

[0129] Step 3: For the FD subarray, a rough estimate of the radiation source direction is calculated using a multi-layer deep learning algorithm or a Root-MUSIC method;

[0130] Step 4: Use the rough estimate of the FD subarray as the initial sample point of the true solution class to establish FD and H 2 The relationship between AD modalities to infer H 2 The set of true solutions of AD mode;

[0131] Step 5: Design the optimal weighted fusion algorithm based on the minimum mean square error principle, synthesize the two parts of the true solution, and obtain the final DOA estimation value.

[0132] As a specific example, the basic parameters of the dual MIMO receiver system combining the FD subarray and the H2AD structure described in step 1 are specifically:

[0133] A far-field narrowband signal where x(t) and f c are the baseband signal and carrier frequency respectively, and t represents time. Considering the FD with M antennas, the output signal is expressed as:

[0134] y FDx(t) = a FD x(θ 0 )x(t) + w(t), (2.1)

[0135] where is the additive white Gaussian noise (AWGN) vector, is the noise variance, I is the identity matrix, and a FD represents that the array manifold vector can be expressed as:

[0136]

[0137] where d = λ / 2 represents the antenna element spacing, i.e., half of the wavelength λ; in the H 2 AD structure, a uniform linear array with M antennas is divided into H groups, each group containing N h antennas, and each group is further divided into K h sub - arrays, each sub - array containing M h antennas, i.e., In this array, M 1 ≠ M 2 ≠ … ≠ M H and preferably select the values of M 1 , M 2 , …, M H as prime numbers.

[0138] Considering that ψ h,k,m represents the phase corresponding to analog beamforming, the output of the k - th sub - array is:

[0139]

[0140] where represents the noise in the k - th sub - array of the h - th group. τ h,k,m is the propagation delay established by the direction of the signal source relative to the array, expressed as:

[0141]

[0142] where τ 0 represents the propagation delay from the radiation source to the reference point of the antenna array, and c represents the speed of light.

[0143] Adding up the outputs y h (t) of all the K h,k sub - arrays, the baseband signal vector of the h - th group is represented as:

[0144]

[0145] where y h (t) is the output signal y h,k(t) is the vector formed by stacking. In addition, is a complex additive white Gaussian noise vector of dimension K h ×1, M h is the number of antennas included in the subarray of the h-th group, is the noise at the M h th antenna, (·) H denotes the conjugate transpose operation, is the array manifold vector, expressed as:

[0146]

[0147] and Υ A,h denotes a block diagonal matrix:

[0148]

[0149] where γ A,h,k is the k-th block diagonal element:

[0150]

[0151] where represents the phase shift of the k-th subarray in the h-th subgroup. Through the analog-to-digital converter (ADC), y h (t) is transformed into:

[0152]

[0153] where n = 1, 2, …, L, and L is the number of snapshots. θ 0 is the ideal angle of incidence, a h (θ 0 ) represents the array manifold vector of the ideal angle of incidence θ 0 , and w(n) is the noise.

[0154] As a specific example, in step 2, multiple candidate angle sets are generated by performing DOA estimation on each subgroup in the H 2 AD structure. The specific steps are as follows:

[0155] Each subarray in the h-th group can be regarded as a virtual antenna. Assuming the analog beamforming vector then the output vector of the K h subarrays is:

[0156]

[0157] where represents a series of received signals in the h-th subgroup, where represents a series of noises in the h-th subgroup, is the array manifold of each virtual subarray, expressed as:

[0158]

[0159] and b h (θ 0 ) is a constant obtained by summing all elements of each subarray, defined as:

[0160]

[0161] The covariance matrix of the output vector y h (n) is:

[0162]

[0163] where denotes the expectation, is the signal-to-noise ratio of the received signal, is the noise variance. R h The eigenvalue decomposition (EVD) of is:

[0164] R h =[U S U N Σ[U S U N H ,(2.14)

[0165] where U N is the K h ×(K h -1) noise subspace, and U S is the K h ×1 signal subspace. is the K h ×K h diagonal matrix.

[0166] Then, the pseudospectrum corresponding to the virtual antenna array is:

[0167]

[0168] where θ represents the arrival angle of the initial input, and the direction corresponding to the peak is the final DOA estimate. In addition, the optimal transmitter direction is obtained by Root-MUSIC. The polynomial equation can be defined as:

[0169]

[0170] where Q ij represents the element located in the i-th row and j-th column of matrix Q.​

[0171]

[0172] in And f(z) has 2Kh-2 roots, namely Z RM ={z i ,i∈{1,2,…,2K h -2}}. Among them, z i represents the i-th root of the polynomial equation (16). The DOA estimation set is obtained as follows:

[0173]

[0174] in

[0175]

[0176] Then, the DOA estimate of the hth array group is You can choose the root closest to the unit circle:

[0177]

[0178] About Function with period 2π So a containing M h The feasible solution of a solution is:

[0179]

[0180] in

[0181]

[0182] Combining all H groups gives:

[0183]

[0184] where j h ∈{1,2,…,M h},M H is the number of subarray antennas in the Hth group, where H represents the number of groups. represents the candidate solution set of group H, represents the estimated value of group H, j H is the imaginary unit of group H, indicating The cyclic period index in has a period of 2π. Therefore, the candidate angle solution set is:

[0185]

[0186] Total candidate set include solution, where each candidate set has one true solution and M h - 1 false solutions. Therefore, finding the true solution is a challenging task.

[0187] As a specific example, in step 3, for the FD sub - array, a multi - layer deep learning algorithm or the Root - MUSIC method is used to calculate the rough estimate of the radiation source direction. The specific steps are as follows:

[0188] Estimate the initial rough DOA value through the Root - MUSIC method:

[0189] y h (n)'s covariance matrix is:

[0190]

[0191] where represents the signal - to - noise ratio of the received signal,

[0192] Then, the corresponding spatial spectrum function is as follows:

[0193]

[0194] where a FD (θ) represents the array manifold vector. By constructing a 2(M - 1) polynomial equation through Root - MUSIC, where M represents the number of antennas of the FD sub - array, the root closest to the unit circle is the DOA estimate that is, the initial mean of the true solution class:

[0195]

[0196] Estimate the initial rough DOA value through the CNN method:

[0197] 1) Dataset and its labels: The input data of the CNN model is a real - valued matrix, denoted as:

[0198]

[0199] where represent the real part, imaginary part, and phase part respectively. Assuming S represents the number of training sets, the input samples of the CNN can be defined as I′ = {I 1 ,I 2 ,…,I S}。The training data angles are selected from p grid points in the range [-θ, θ], where θ ∈ [1°, 90°] and p is the interval unit. Thus, the desired grid is defined as Δ = {-θ, …, -p, 0°, p, …, θ}, and the binary vector corresponding to the training angle is used as the label. For example, Δ = {-45°, …, -1°, 0°, 1°, …, 45°} has |Δ| = 91 grid points when θ = 45° and p = 1°. Additionally, the angle -45° is converted into a binary vector

[0200] Therefore, the training data set and the corresponding label set for the CNN are:

[0201] I′ = {I 1 , I 2 , …, I S} (2.29)

[0202] V = {ν 1 , ν 2 , …, ν S} (2.30)

[0203] 2) CNN structure: The objective of CNN-based DOA estimation is f: The mapping from the input to the output of this network is defined as:

[0204]

[0205] where is the i-th convolutional layer, is the j-th fully connected layer.

[0206] 3) Loss function:

[0207]

[0208] where and represent the true label and the predicted value respectively, and μ represents the weight decay coefficient. W is the weighted parameter of the FC layer. The model is trained by minimizing the loss function L.

[0209] As a specific example, in step 4, the rough estimation value of the FD subarray is used as the initial sample point of the true solution class, and the relationship between the FD and H 2 AD modalities is established to infer the set of true solutions of the H 2 AD modality. The specific steps are as follows:

[0210] The true solution of the h-th group can be expressed as.

[0211]

[0212] Therefore, H2 The true solution set of the AD mode is as follows:

[0213]

[0214] As a specific example, in step 5, an optimal weighted fusion algorithm is designed based on the least mean square error principle to synthesize two parts of true solutions and obtain the final DOA estimation value. The specific steps are as follows:

[0215] Fuse two parts of true solutions to obtain the optimal DOA estimation:

[0216]

[0217] where

[0218]

[0219]

[0220]

[0221]

[0222]

[0223]

[0224] where L represents the number of snapshots and SNR represents the signal-to-noise ratio, denotes the real part of.

[0225] Figure 4 and Figure 5 Plots the RMSE of the proposed method versus the signal-to-noise ratio with the corresponding Cramer-Rao lower bound (CRLB) as the performance benchmark. It can be seen from the figure that the proposed method can quickly select the true solution in the group and approach or even reach the corresponding Cramer-Rao lower bound, achieving good DOA estimation performance.

[0226] The present invention also provides a DOA estimation system with a collaborative learning-assisted dual hybrid structure, which is used to implement the DOA estimation method with the collaborative learning-assisted dual hybrid structure. The system includes a first module to a fifth module, and the specific functions of each module are as follows:

[0227] The first module constructs a dual MIMO receiver H 2 AD-FD system model combining a full-digital FD subarray and a heterogeneous hybrid H 2 AD structure;

[0228] The second module, by performing H 2Each group in the AD structure performs DOA estimation to generate multiple candidate angle sets;

[0229] The third module, for the FD subarray, uses a multi-layer deep learning algorithm or the Root-MUSIC method to calculate a rough estimate of the radiation source direction;

[0230] The fourth module uses the rough estimate of the FD subarray as the initial sample points of the true solution class, and establishes the relationship between the FD and H 2 AD modes to infer the true solution set of the H 2 AD modes;

[0231] The fifth module designs an optimal weighted fusion algorithm based on the principle of minimum mean square error, synthesizes the two parts of the true solutions, and obtains the final DOA estimate value.

[0232] The present invention also provides a mobile terminal, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the DOA estimation method of the collaborative learning-assisted dual hybrid structure described above.

[0233] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the steps in the DOA estimation method of the collaborative learning-assisted dual hybrid structure described above.

[0234] As mentioned above, only the preferred specific embodiments of the present invention are described, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

[0235] It should be understood that in order to streamline the present invention and help those skilled in the art understand various aspects of the present invention, in the above description of the exemplary embodiments of the present invention, various features of the present invention are sometimes described in a single embodiment or with reference to a single figure. However, the present invention should not be construed as including the features included in the exemplary embodiments as essential technical features of the patent claims of the present invention.

Claims

1. A collaborative learning-assisted dual hybrid structure DOA estimation method, characterized in that: The following steps are involved: Step 1: Construct all-digital FD subarray and heterogeneous hybrid H 2 Dual MIMO receiver combined with AD structure 2 AD-FD system model; Step 2: By 2 Each group in the AD structure performs DOA estimation to generate multiple candidate angle sets; Step 3: For the FD subarray, a rough estimate of the radiation source direction is calculated using a multi-layer deep learning algorithm or a Root-MUSIC method; Step 4: Use the rough estimate of the FD subarray as the initial sample point of the true solution class to establish FD and H 2 The relationship between AD modalities to infer H 2 The set of true solutions of AD mode; Step 5: Design the optimal weighted fusion algorithm based on the minimum mean square error principle, synthesize the two parts of the true solution, and obtain the final DOA estimation value.

2. The DOA estimation method of the dual hybrid structure assisted by collaborative learning according to claim 1, characterized in that: In step 1, the all-digital FD subarray is combined with the heterogeneous hybrid H 2 Dual MIMO receiver combined with AD structure 2 The basic parameters of the AD-FD system are as follows: A far-field narrowband signal where x(t) and f c are the baseband signal and carrier frequency respectively, t represents time, considering the FD with M antennas, the output signal is expressed as: y FD (t)=a FD (θ0)x(t)+w(t)(1.1) in is the additive white Gaussian noise (AWGN) vector, is the noise variance, I is the identity matrix, a FD The vector representation of the array manifold is: Where d = λ / 2 represents the antenna element spacing, which is half of the wavelength λ; 2 In the AD structure, a uniform linear array containing M antennas is divided into H groups, each containing N h antennas, and each group is divided into K h subarrays, each containing M h Antenna, that is In this array, M1≠M2≠…≠M H And M1, M2, ..., M are preferred. H The value of is considered a prime number; Considering ψ h,k,m represents the phase corresponding to the simulated beamforming, then the output of the kth subarray is: in represents the noise of the kth subarray in the hth group, τ h,k,m is the propagation delay established by the direction of the signal source relative to the array and is expressed as: Where τ0 represents the propagation delay from the radiation source to the reference point of the antenna array, and c represents the speed of light; All K h The output y of the subarray h,k (t) superposition, the baseband signal vector of the hth group is expressed as: Among them, y h (t) is the output signal y of the kth subarray of the hth group h,k (t) is a vector formed by stacking. In addition, The dimension is K h ×1 additive Gaussian white noise complex vector, M h is the number of antennas contained in the h-th subarray, It is the Mth h Noise at the root antenna, (·) H represents the conjugate transpose operation, is an array manifold vector, represented as: And A,h Represents a block diagonal matrix: Among them, γ A,h,k is the kth block diagonal element: in, represents the phase shift of the kth subarray in the hth group, through the analog-to-digital converter ADC, y h (t) is transformed into: Where n = 1, 2, ..., L, L is the number of snapshots, θ0 is the ideal angle of incidence, a h (θ0) represents the array manifold vector of the ideal incident angle θ0, and w(n) is the noise.

3. The DOA estimation method of the dual hybrid structure assisted by collaborative learning according to claim 2, characterized in that: In step 2, by 2 Each group in the AD structure performs DOA estimation to generate multiple candidate angle sets. The specific steps are as follows: Each subarray in the hth group is considered as a virtual antenna, assuming the simulated beamforming vector K h The output vector of the subarray is: in, represents a series of received signals in the hth group, where represents a series of noises in the h-th group, is the array manifold of each virtual sub-array, expressed as: And b h (θ0) is a constant obtained by adding all elements of each subarray and is defined as: Output vector y h The covariance matrix of (n) is: in Expressing hope, is the signal-to-noise ratio of the received signal, is the noise variance; R h The eigenvalue decomposition EVD is: R h =[U S IN N ]Σ[U S IN N ] H (1.14) Among them U N It's K h ×(K h -1) Noise subspace, U S It's K h ×1 signal subspace, It's K h ×K h diagonal matrix; Then, the pseudo-spectrum corresponding to the virtual antenna array is: Where θ represents the arrival angle of the initial input, and The direction corresponding to the peak is the final DOA estimate; in addition, Root-MUSIC is used to obtain the optimal source direction, and the polynomial equation is defined as: in Q ij represents the element located in the i-th row and j-th column of the matrix Q; in And f(z) has 2K h -2 roots, namely Z RM ={z i ,i∈{1,2,…,2K h -2}}, where z i represents the i-th root of the polynomial equation (1.16), from which the DOA estimation set is obtained as follows: in Then, the DOA estimate of the hth array group is Choose the root closest to the unit circle: About Function with period 2π So a containing M h The feasible solution of a solution is: in Combining all H groups gives: where j h ∈{1,2,…,M h },M H is the number of subarray antennas in the Hth group, where H represents the number of groups. represents the candidate solution set of group H, represents the estimated value of group H, j H is the imaginary unit of group H, indicating The cyclic period index in is 2π; therefore, the candidate angle solution set is: Total candidate set include Solution, where each candidate set There is a true solution and M h -1 pseudo solution.

4. The DOA estimation method of the dual hybrid structure assisted by collaborative learning according to claim 3, characterized in that: In step 3, for the FD subarray, a multi-layer deep learning algorithm or Root-MUSIC method is used to calculate a rough estimate of the radiation source direction. The specific steps are as follows: (1) Estimate the initial rough DOA value by Root-MUSIC method: y h The covariance matrix of (n) is: in represents the signal-to-noise ratio of the received signal, Then, the corresponding spatial spectrum function is as follows: where a FD (θ) represents the array popular vector, and 2(M-1) polynomial equations are constructed by Root-MUSIC, where M represents the number of antennas in the FD subarray, so that the root closest to the unit circle is the DOA estimate That is, the initial mean of the true solution class: (2) Estimate the initial rough DOA value by CNN method: 1) Dataset and its labels: Input data for CNN model is a real-valued matrix, represented as: in denote the real part, imaginary part and phase part respectively; assuming S denotes the number of training sets, the input sample of CNN is defined as I′={I1,I2,…,I S }; the angles of the training data are selected from p grid points in [-θ,θ], where θ∈[1°,90°] and p is an interval unit; therefore, the desired grid is defined as Δ={-θ,…,-p,0°,p,…,θ}, and the binary vectors corresponding to the training angles are used as labels; Δ={-45°,…,-1°,0°,1°,…,45°} has |Δ|=91 grid points when θ=45° and p=1°; in addition, the angle -45° is converted into a binary vector Therefore, the training data set and corresponding label set of CNN are: I′={I1,I2,…,I S } (1.29) V={ν1,ν2,…,ν S } (1.30) 2) CNN structure: The goal of CNN-based DOA estimation is The mapping from CNN network input to output is defined as: in, is the i-th convolutional layer, is the jth fully connected layer; 3) Loss function: in and Represent the true label and predicted value respectively, μ represents the weight decay coefficient, W is the weight parameter of the FC layer, and the model is trained by minimizing the loss function L.

5. The DOA estimation method of the dual hybrid structure assisted by collaborative learning according to claim 4, characterized in that: In step 4, the rough estimate of the FD subarray is used as the initial sample point of the true solution class to establish FD and H 2 The relationship between AD modalities to infer H 2 The true solution set of AD mode, the specific steps are: The true solution of the hth group is expressed as: Therefore, H 2 The true solution set of AD mode is:

6. The DOA estimation method of the dual hybrid structure assisted by collaborative learning according to claim 5, characterized in that: In step 5, the optimal weighted fusion algorithm is designed based on the minimum mean square error principle to synthesize the two parts of the true solution and obtain the final DOA estimation value. The specific steps are as follows: The optimal DOA estimation is obtained by combining the two true solutions: in Ξ=M h +K h SNR||b h (θ0)|| 2 (1.41) Where L represents the number of snapshots, SNR represents the signal-to-noise ratio, Indicates b h 2 (θ0)The real part of μ.

7. A collaborative learning-assisted dual hybrid structure DOA estimation system, characterized in that: The system is used to implement the collaborative learning-assisted dual hybrid structure DOA estimation method according to any one of claims 1 to 6. The system includes a first module to a fifth module. The specific functions of each module are as follows: The first module constructs a fully digital FD subarray and a heterogeneous hybrid H 2 Dual MIMO receiver combined with AD structure 2 AD-FD system model; The second module is to 2 Each group in the AD structure performs DOA estimation to generate multiple candidate angle sets; The third module uses a multi-layer deep learning algorithm or Root-MUSIC method to calculate a rough estimate of the radiation source direction for the FD subarray; The fourth module uses the rough estimate of the FD subarray as the initial sample point of the true solution class to establish FD and H 2 The relationship between AD modalities to infer H 2 The set of true solutions of AD mode; In the fifth module, the optimal weighted fusion algorithm is designed based on the minimum mean square error principle to synthesize the two parts of the true solution and obtain the final DOA estimation value.

8. A mobile terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the DOA estimation method of the dual hybrid structure assisted by collaborative learning is implemented as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the DOA estimation method of a dual hybrid structure assisted by collaborative learning are implemented.

Images