DOA estimation method and device based on multi-modal depth fusion

Abstract

The invention discloses a DOA (direction of arrival) estimation method and device based on multi-modal depth fusion. The method specifically comprises the following steps: constructing a dual MIMO (multiple input multiple output) receiver H2AD-FD system model in which a full-digital FD sub-array and a heterogeneous mixed H2AD structure are combined; aiming at the FD sub-array, calculating a rough estimation value of a radiation source direction; generating a plurality of candidate angle sets by performing DOA estimation on each group in the H2AD structure; based on the rough estimation value of the radiation source direction as an initial clustering center point, inferring a true solution corresponding to each group from the candidate angle set by adopting a global maximum similarity and global minimum distance clustering method; and constructing a fusion network based on a three-layer full-connection network FCNN, and fusing true solutions deduced by all the sub-array groups of the H2AD structure and a radiation source direction rough estimation value calculated by the FD sub-array to obtain a final DOA estimation value. The method has the direction finding performance of high energy efficiency, low cost and low complexity, and can quickly eliminate phase ambiguity.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technologies, and particularly to a DOA estimation method and device based on multi-modal deep fusion. Background Art

[0002] Massive multiple-input multiple-output (MIMO) has high spatial angular resolution and spectral efficiency, and is regarded as one of the key technologies for future green wireless communication networks. The direction-of-arrival (DOA) estimation of radiation sources plays a crucial role in target positioning, covert communication, communication system performance improvement, etc., and its combination with massive multiple-input multiple-output (MIMO) technology provides ultra-high direction-finding accuracy and resolution. However, the large number of antennas also causes problems such as high computational complexity of traditional algorithms based on all-digital structures and high circuit costs; while the hybrid analog and digital structures face the problem of how to quickly eliminate phase ambiguity. In addition, most of the existing DOA estimations are studied based on the ideal incident angle of the signal source, but the ideal incident angle of the signal source does not exist in the actual communication system. Summary of the Invention

[0003] The purpose of the present invention is to provide a DOA estimation method and device based on multi-modal deep fusion, which can quickly eliminate phase ambiguity and have high energy efficiency, low cost, and low-complexity direction-finding performance.

[0004] The technical solution for achieving the purpose of the present invention is: a DOA estimation method based on multi-modal deep fusion, including the following steps:

[0005] Step 1, construct a dual MIMO receiver H 2 AD structure combined with an all-digital FD subarray H 2 AD-FD system model;

[0006] Step 2, for the FD subarray, use the Root-MUSIC method to calculate the rough estimated value of the radiation source direction;

[0007] Step 3, generate multiple candidate angle sets by performing DOA estimation on each group in the H 2 AD structure;

[0008] Step 4, based on the rough estimated value of the radiation source direction provided by the FD subarray as the initial clustering center point, use the global maximum similarity (GMaxCS) and global minimum distance (GMinD) clustering methods to infer the true solution corresponding to each group from the candidate angle sets;

[0009] Step 5, synthesize the true solutions of the two parts through a deep fusion algorithm: construct a fusion network fusionNet based on a three-layer fully connected network (FCNN), for all H 2The true solution inferred by the sub-array group of the AD structure and the rough estimated value of the radiation source direction calculated by the FD sub-array are fused to obtain the final DOA estimation value.

[0010] Further, in step 1, the dual MIMO receiver H 2 The parameters of the AD-FD system model are specifically:

[0011] A far-field narrowband signal e(t)e j2πfct , where e(t) and f c are the baseband signal and the carrier frequency respectively, t represents time, where H 2 The AD consists of P groups and each group is divided into K p sub-arrays, and each sub-array contains M p antennas, that is In this array, M 1 ≠M 2 ≠…≠M P and preferably select the values of M 1 , M 2 ,…, M P as prime numbers;

[0012] Assume there is a transmitted signal incident from the θ direction, then the received signal vector y F (n) and the received signal vector y p (n) of the p-th group are respectively expressed as:

[0013] y F (n) = a F (θ 0 )e(n)+v F (n),(1.1)

[0014]

[0015] where e(n) is the baseband signal, and are additive white Gaussian noise (AWGN) vectors, represents the noise variance, I is the identity matrix, represents a complex vector of dimension K p ×1, represents a series of noises in the q-th subgroup; and represent the array manifold vectors:

[0016]

[0017]

[0018] where

[0019]

[0020]

[0021] Among them, M represents the number of antennas of the FD subarray, N p represents the number of antennas of the subarray in the q-th group, d represents the spacing between adjacent array elements in the array antenna, and λ represents the wavelength;

[0022] The block diagonal matrix Ξ A,p is expressed as:

[0023]

[0024] Among them, ε A,p,k is expressed as:

[0025]

[0026] Among them, Υ p,k,m is the analog beam phase shift matrix, ε A,p,k represents the analog beamforming vector of the k-th subarray in each group, and M p represents that each subarray contains M p antenna elements.

[0027] Furthermore, in step 2, for the FD subarray, the Root-MUSIC method is used to calculate the rough estimate of the radiation source direction. The specific steps are as follows:

[0028] The spatial spectrum corresponding to the FD subarray is expressed as:

[0029]

[0030] represents the conjugate transpose of the array manifold vector of the FD subarray, θ represents the initial input arrival angle, and U N represents the signal noise subspace;

[0031] S F (θ)'s denominator can be expressed as:

[0032]

[0033] z F is a function containing θ, and U N represents the signal noise subspace, and s F (z F ) represents a 2(M F -1) degree polynomial;

[0034] The polynomial equation s(z F ) has 2(M F-1) roots, so the root closest to the unit circle represents the expected DOA direction:

[0035]

[0036] Further, in step 3, 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:

[0037] For the H 2 AD array, the spatial spectrum calculation formula of the virtual antenna array is:

[0038]

[0039] where

[0040]

[0041] where, r p (θ) is a constant obtained by summing all elements of each subarray, represents the array manifold vector of M q root virtual antennas, and E N represents the noise subspace of the signal;

[0042] Define Then:

[0043]

[0044] is the element in the i 1 -th row and i 2 -th column of B;

[0045] Through Root-MUSIC, the direction corresponding to the peak is the final DOA estimate, the polynomial equation in

[0046]

[0047] s(z) represents a 2K p - 2-degree polynomial, z is a function containing θ, and s(η) represents a 2K p - 2-degree polynomial;

[0048] where

[0049]

[0050]

[0051] The equation s(θ) has 2K p - 2 roots, that is,

[0052] Z R ={z l ,l∈[1,2K p -2]}(1.18)

[0053] η p is a function of θ, M p represents that each sub-array contains M p antennas;

[0054] DOA estimation set is expressed as:

[0055]

[0056] where

[0057]

[0058] z l represents the l-th root of the polynomial equation (1.15);

[0059] Therefore, the estimated value of the p-th group is:

[0060]

[0061] represents the estimation of η p , represents the estimated value of the p-th group;

[0062] Considering the function with a period of 2π, the feasible set of the M P solutions of the p-th group is extended to:

[0063]

[0064] where

[0065]

[0066] Integrating all groups gives:

[0067]

[0068] where q p ∈{1,2,…,M p} is the ambiguity coefficient. Therefore, the candidate set of H 2 AD is expressed as:

[0069]

[0070] Candidate set contains Solution, each candidate set has a true solution and M p - 1 false solutions. In the case of noise ignored, the true solution class and the false solution class are respectively expressed as

[0071]

[0072]

[0073] where

[0074]

[0075] χ p,m represents a constant determined by M p and m;

[0076] Then, the true solution class is expressed as

[0077]

[0078] where is the predicted true solution of the p-th group of the HAD structure. 2

[0079] Furthermore, in step 4, based on the rough estimated value of the radiation source direction provided by the FD subarray as the initial clustering center point, the global maximum similarity (i.e., GMaxCS) and the global minimum distance (i.e., GMinD) clustering methods are adopted to infer the true solution corresponding to each group from the candidate angle set. The specific steps are as follows:

[0080] When the initial DOA value and the candidate set of the p-th group are obtained, a vector is defined as

[0081]

[0082] indicating the q-th p candidate solution of the p-th group;

[0083] Then, the cosine similarity is defined as

[0084]

[0085] Therefore, by searching for the direction with the maximum similarity, the true angle

[0086]

[0087] Furthermore, in step 5, the two parts of the true solutions are synthesized through a deep fusion algorithm. The specific steps are as follows:

[0088] ​The adaptive fusionNet processes H through three - layer FCNN fusion 2 the true solutions of the AD structure and the FD sub - array, and obtains the desired DOA value:

[0089]

[0090] where is a new vector formed by splicing the true solution of the H 2 AD structure and the rough estimate of the FD sub - array, and ω is the learning parameter of the fusionNet;

[0091] According to the characteristics of the FCNN layer, the above formula is transformed into:

[0092]

[0093] where W is the weight matrix learned during the FCNN training process, and b is the bias vector.

[0094] The present invention also provides a DOA estimation device based on multi - modal deep fusion, which is used to implement the above - mentioned DOA estimation method based on multi - modal deep fusion. The device includes a first module to a fifth module, and the functions of each module are as follows:

[0095] The first module is used to construct a dual MIMO receiver H 2 AD - FD system model by combining a fully digital FD sub - array and a heterogeneous hybrid H 2 AD structure;

[0096] The second module, for the FD sub - array, uses the Root - MUSIC method to calculate the rough estimate value of the radiation source direction;

[0097] The third module generates multiple candidate angle sets by performing DOA estimation on each group in the H 2 AD structure;

[0098] The fourth module, based on the rough estimate value of the radiation source direction provided by the FD sub - array as the initial clustering center point, uses the global maximum similarity (i.e., GMaxCS) and global minimum distance (i.e., GMinD) clustering methods to infer the true solution corresponding to each group from the candidate angle sets;

[0099] The fifth module synthesizes the true solutions of the two parts through a deep fusion algorithm: constructs a fusion network fusionNet based on a three - layer fully connected network FCNN, and fuses the true solutions inferred from all the sub - array groups of the H 2 AD structure and the rough estimate value of the radiation source direction calculated by the FD sub - array to obtain the final DOA estimate value.

[0100] 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, the DOA estimation method based on multi-modal depth fusion as described above is implemented.

[0101] 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, the steps in the DOA estimation method based on multi-modal depth fusion as described above are implemented.

[0102] Compared with the prior art, the significant advantages of the present invention are: (1) The present invention uses a heterogeneous hybrid H 2 AD structure, which has lower requirements for circuit hardware, reduces hardware costs, and can quickly eliminate phase ambiguity; (2) Compared with the all-digital MIMO structure, the multi-modal depth fusion framework of the present invention has higher energy efficiency, lower cost, and lower complexity while having the same low latency. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] Figure 1 is a flowchart of the DOA estimation method based on multi-modal depth fusion provided by the present invention.

[0104] Figure 2 is an overall structural schematic diagram of the DOA estimation device based on multi-modal depth fusion of the present invention.

[0105] Figure 3 is a relationship curve diagram of the mean square error and the signal-to-noise ratio of the MM-fusionNet-GMaxCS method of the present invention.

[0106] Figure 4 is a relationship curve diagram of the mean square error and the signal-to-noise ratio of the MM-fusionNet-GMinD method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0107] In order to make the objectives, technical solutions, and advantages of the present invention more apparent, 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 the 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.

[0108] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the present invention. However, it will be apparent to one of ordinary skill in the art that the present invention may be practiced without one or more of these specific details. In other instances, well-known features have not been described in order to avoid obscuring the invention.

[0109] It should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. On the contrary, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0110] The present invention provides a DOA estimation method based on multi-modal depth fusion. This method constructs a fusion network (fusionNet) based on a three-layer fully connected neural network (FCNN) to fuse the true solutions inferred from all sub-array groups of the hybrid structure and the rough DOA angles estimated by the FD sub-arrays. This method achieves fast elimination of phase ambiguity and has the advantages of high energy efficiency, low cost, and low complexity in direction finding performance.

[0111] Referring to Figure 1 and Figure 2 , the present invention provides an enhanced DOA estimation method based on a multi-modal depth fusion framework for fast eliminating direction finding methods. The method includes the following steps:

[0112] Step 1: Construct a dual MIMO receiver (H 2 AD-FD) system model combining a fully digital (FD) sub-array with a heterogeneous hybrid (H 2 AD) structure;

[0113] Step 2: For the FD sub-array, use the Root-MUSIC method to calculate a rough estimate of the radiation source direction;

[0114] Step 3: Generate multiple candidate angle sets by performing DOA estimation on each group in the H 2 AD structure;

[0115] Step 4: Based on the true solution-like initial clustering center points provided by the FD sub-array, use the global maximum similarity (GMaxCS) and global minimum distance (GMinD) clustering methods to infer the true solutions corresponding to each group in the candidate angle sets;

[0116] Step 5: Synthesize the two parts of the true solutions through a depth fusion algorithm to obtain the final DOA estimate value.

[0117] A far-field narrowband signal where e(t) and f c are the baseband signal and the carrier frequency respectively, and t represents time. Among them, H2 AD consists of P groups, and each group is divided into K p sub - arrays, and each sub - array contains M p antennas, that is in this array, M 1 ≠M 2 ≠…≠M P and M 1 , M 2 ,…, M P values are preferably selected as prime numbers.

[0118] Assume there is a transmitted signal incident from the θ direction. Then the received signal vectors of the FD sub - array and the p - th group are respectively expressed as:

[0119] y F (n)=a F (θ 0 )e(n)+v F (n),(2.1)

[0120]

[0121] where e(n) is the base - band signal, and are additive white Gaussian noise (AWGN) vectors, represents the noise variance, I is the identity matrix, represents a complex vector of dimension K p ×1, represents a series of noises in the q - th group; and represent the array manifold vectors:

[0122]

[0123]

[0124] where

[0125]

[0126]

[0127] where M represents the number of antennas in the FD sub - array, N p represents the number of antennas in the sub - array of the q - th group, d represents the spacing between adjacent array elements in the array, and λ represents the wavelength.

[0128] The block - diagonal matrix Ξ A,p is expressed as:

[0129]

[0130] Among them, ε A,p,k is expressed as:

[0131]

[0132] Among them, Υ p,k,m is the analog beam phase shift matrix. ε A,p,k represents the analog beamforming vector of the k-th subarray in each group, M p represents that each subarray contains M p antenna elements.

[0133] In step 2, for the FD subarray, the Root-MUSIC method is used to calculate the rough estimate of the radiation source direction.

[0134] The specific steps are as follows:

[0135] The spatial spectrum corresponding to the FD subarray is expressed as:

[0136]

[0137] represents the conjugate transpose of the array manifold vector of the FD subarray, θ represents the initial input arrival angle, U N represents the signal noise subspace.

[0138] S F (θ)'s denominator can be expressed as:

[0139]

[0140] z F is a function containing θ, U N represents the signal noise subspace, s F (z F ) represents a 2(M F -1) degree polynomial.

[0141] The polynomial equation s(z F ) has 2(M F -1) roots, so the root closest to the unit circle represents the expected DOA direction:

[0142]

[0143] In step 3, multiple candidate angle sets are generated by performing DOA estimation on each group in the H 2 AD structure. The specific steps are as follows:

[0144] For the H 2 AD array, the spatial spectrum calculation formula of the virtual antenna array is:

[0145] Among them

[0146]

[0147]

[0148] where r p (θ) is a constant obtained by summing all elements of each subarray, denotes the array manifold vector of M q root virtual antennas, and E N denotes the noise subspace of the signal.

[0149] Define Then:

[0150]

[0151] is the element in the i 1 -th row and i 2 -th column of B.

[0152] Through Root-MUSIC, the direction corresponding to the peak is the final DOA estimate. The polynomial equation in

[0153]

[0154] s(z) represents a 2K p - 2-degree polynomial, z is a function containing θ, and s(η) represents a 2K p - 2-degree polynomial.

[0155] where

[0156]

[0157]

[0158] The equation s(θ) has 2K p - 2 roots, that is,

[0159] Z R ={z l , l ∈ [1, 2K p - 2]}(2.18) η p is a function containing θ, and M p represents that each subarray contains M p root antennas.

[0160] . The DOA estimation set can be expressed as:

[0161]

[0162] wherein

[0163]

[0164] z l represents the l-th root of the polynomial equation (15).

[0165] Therefore, the estimated value of the p-th group is:

[0166]

[0167] represents the estimation of η p , representing the estimated value of the p-th group.

[0168] Considering the function with a period of 2π, the feasible set of the M P -th solution of the p-th group can be generalized as:

[0169]

[0170] wherein

[0171]

[0172] Integrating all groups gives:

[0173]

[0174] where q p ∈ {1, 2, …, M p} is the fuzzy coefficient. Therefore, the candidate set of H2AD can be expressed as:

[0175]

[0176] From the above analysis, it can be seen that the candidate set contains solutions, and each candidate set has one true solution and M p - 1 false solutions. Therefore, eliminating false solutions is a crucial issue. In the case where the noise is negligible, the true solution class and the false solution class can be respectively expressed as

[0177]

[0178]

[0179] wherein

[0180]

[0181] χp,m Denote the constant determined by M p and m.

[0182] Then, the true solution class is expressed as

[0183]

[0184] where is the predicted true solution of the p-th group.

[0185] In step 4, based on the initial clustering center points of the true solution class provided by the FD subarray, the true solution corresponding to each subgroup is inferred from the candidate angle set by using the global maximum similarity (GMaxCS) and global minimum distance (GMinD) clustering methods. The specific steps are as follows:

[0186] When the initial DOA value and the candidate set of the p-th group are obtained. Define the vector:

[0187]

[0188] Denote the q p -th candidate solution of the p-th group.

[0189] Then, the cosine similarity can be defined as

[0190]

[0191] Therefore, by searching for the direction with the maximum similarity, the true angle

[0192]

[0193] In step 5, the two parts of the true solutions are fused by the network fusionNe to obtain the final DOA estimation value. The specific steps are as follows:

[0194] The adaptive fusionNet fuses and processes the true solutions of the H 2 AD array and the FD subarray through three-layer FCNN. The desired DOA value is obtained:

[0195]

[0196] where is the new vector formed by splicing the true solution of H 2 AD and the rough estimation of the FD subarray. ω is the learning parameter of the fusionNet.

[0197] According to the characteristics of the FCNN layer, the above formula is transformed into:

[0198]

[0199] Where \(W\) is the weight matrix learned during the FCNN training process, and \(b\) is the bias vector.

[0200] Figure 3 and Figure 4 The curves of the RMSE of the proposed method versus the signal-to-noise ratio are plotted 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.

[0201] The present invention also provides a DOA estimation device based on multi-modal depth fusion, which is used to implement the DOA estimation method based on multi-modal depth fusion. The device includes a first module to a fifth module, and the functions of each module are as follows:

[0202] The first module is used to construct a dual MIMO receiver \(H_{AD - FD}\) system model combining a fully digital FD subarray and a heterogeneous hybrid \(H_{AD}\) structure; 2 \(H_{AD - FD}\) 2 system model;

[0203] The second module is used to calculate a rough estimate of the radiation source direction by using the Root-MUSIC method for the FD subarray;

[0204] The third module is used to generate multiple candidate angle sets by performing DOA estimation on each group in the \(H_{AD}\) structure; 2 AD

[0205] The fourth module is used to infer the true solution corresponding to each group from the candidate angle sets by using the global maximum similarity (i.e., GMaxCS) and global minimum distance (i.e., GMinD) clustering methods, based on the rough estimate of the radiation source direction provided by the FD subarray as the initial clustering center point;

[0206] The fifth module is used to synthesize the true solutions of the two parts through a depth fusion algorithm: construct a fusion network fusionNet based on a three-layer fully connected network FCNN, and fuse the true solutions inferred from all the subarray groups of the \(H_{AD}\) structure and the rough estimate of the radiation source direction calculated by the FD subarray to obtain the final DOA estimate value. 2 AD

[0207] 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, the DOA estimation method based on multi-modal depth fusion is implemented.

[0208] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps in the DOA estimation method based on multi-modal depth fusion are implemented.

[0209] As described above, only the preferred specific embodiments of the present invention are provided, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

[0210] 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 meaning that the features included in the exemplary embodiments are all essential technical features of the patent claims.

Claims

1. A DOA estimation method based on multimodal deep fusion, 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: For the FD subarray, use the Root-MUSIC method to calculate a rough estimate of the radiation source direction; Step 3: By 2 Each group in the AD structure performs DOA estimation to generate multiple candidate angle sets; Step 4: Based on the rough estimate of the radiation source direction provided by the FD subarray as the initial cluster center point, the global maximum similarity (GMaxCS) and the global minimum distance (GMinD) clustering method are used to infer the true solution corresponding to each group from the candidate angle set; Step 5: Synthesize the two true solutions through a deep fusion algorithm: Construct a fusion network fusionNet based on a three-layer fully connected network FCNN, and 2 The true solution inferred by the subarray group of the AD structure and the rough estimate of the radiation source direction calculated by the FD subarray are fused to obtain the final DOA estimate.

2. The DOA estimation method based on multimodal deep fusion according to claim 1, characterized in that: In step 1, the dual MIMO receiver H 2 The parameters of the AD-FD system model are as follows: A far-field narrowband signal Where e(t) and f c are the baseband signal and carrier frequency respectively, t represents time, where H 2 AD consists of P groups and each group is divided into K p subarrays, each containing M p Antenna, that is In this array, M1≠M2≠…≠M P And M1, M2, ..., M are preferred. P The value of is considered a prime number; Assuming there is a transmitted signal incident from the direction of θ, the received signal vector y of the FD subarray is F (n) and the received signal vector y of the pth group p (n) are respectively expressed as: the F (n)=a F (θ0)and(n)+v F (n), (1.1) Where e(n) is the baseband signal, and is the additive white Gaussian noise (AWGN) vector, represents the noise variance, I is the unit matrix, Indicates that the dimension is K p ×1 complex vector, v1(n), v2(n),…, represents a series of noises in the qth group; and Represents an array manifold vector: in Where M represents the number of antennas in the FD subarray, N p represents the number of antennas in the subarray in the qth group, d represents the spacing between adjacent array elements in the array antenna, and λ represents the wavelength; Block diagonal matrix Ξ A,p It is expressed as: Among them, ε A,p,k It is expressed as: Among them p,k,m is the simulated beam phase shift matrix, ε A,p,k represents the simulated beamforming vector of the kth subarray in each group, M p Indicates that each subarray contains M p Antenna.

3. The DOA estimation method based on multimodal deep fusion according to claim 2, characterized in that: In step 2, for the FD subarray, the Root-MUSIC method is used to calculate a rough estimate of the radiation source direction. The specific steps are: The spatial spectrum corresponding to the FD subarray is expressed as: represents the conjugate transpose of the array manifold vector of the FD subarray, θ represents the arrival angle of the initial input, and U N represents the noise subspace of the signal; S F The denominator of (θ) can be expressed as: z F is a function containing θ, U N represents the noise subspace of the signal, s F (z F ) means 2(M F -1) polynomial; Polynomial equation s(z F ) has 2(M F -1) root, so the root closest to the unit circle represents the expected DOA direction:

4. The DOA estimation method based on multimodal deep fusion according to claim 3 is characterized in that: In step 3, by 2 Each group in the AD structure performs DOA estimation to generate multiple candidate angle sets. The specific steps are as follows: For H 2 AD array, the spatial spectrum calculation formula of the virtual antenna array is: in Among them, r p (θ) is a constant obtained by adding all elements of each subarray, Indicates M q The array manifold vector of the root virtual antenna, E N represents the noise subspace of the signal; definition but: is the element in row i1 and column i2 of B; Through Root-MUSIC, The direction corresponding to the peak is the final DOA estimate. The polynomial equation in is defined as s(z) represents 2K p -2 degree polynomial, z is a function including θ, s(η) represents 2K p -2nd degree polynomial; where The equation s(θ) has 2K p -2 roots, i.e. WITH R ={z l ,l∈[1,2K p -2]}(1.18) η p is a function containing θ, M p Indicates that each subarray contains M p Antenna; DOA estimation set It is expressed as: in z l represents the lth root of the polynomial equation (1.15); Therefore, the estimated value of group p for: Represents η p The estimate, represents the estimated value of group p; Consider the function The period is 2π, then M of the pth group P The feasible set of solutions is generalized to: in The groups that integrate all of them are: where q p ∈{1,2,…,M p } is the fuzzy coefficient, so H 2 The candidate set of AD is expressed as: Candidate Set Include Solution, each candidate set There is a true solution and M p -1 pseudo solution. When the noise is ignored, the true solution class and the false solution class are expressed as in χ p,m Indicated by M p A constant determined by and m; Then, the true solution class is expressed as in, H 2 The predicted true solution of the pth group of AD structures.

5. The DOA estimation method based on multimodal deep fusion according to claim 4 is characterized in that: In step 4, based on the rough estimate of the radiation source direction provided by the FD subarray as the initial cluster center point, the global maximum similarity, i.e., GMaxCS, and the global minimum distance, i.e., GMinD clustering method, are used to infer the true solution corresponding to each group from the candidate angle set. The specific steps are as follows: When the initial DOA value is obtained and the candidate set of group p When , define the vector: It means the qth p Candidate solutions; Then, the cosine similarity is defined as Therefore, by searching for the direction with the greatest similarity, the true angle 6. The DOA estimation method based on multimodal deep fusion according to claim 5, characterized in that: In step 5, the two parts of the true solution are synthesized through the deep fusion algorithm. The specific steps are as follows: Adaptive fusionNet processes H through three-layer FCNN fusion 2 True solution of AD structure and FD subarray to obtain the expected DOA value: in It is made by H 2 The new vector is formed by concatenating the true solution of the AD structure and the rough estimate of the FD subarray. ω is the learning parameter of fusionNet. According to the characteristics of the FCNN layer, the above formula is transformed into: Where W is the weight matrix learned during FCNN training, and b is the bias vector.

7. A DOA estimation device based on multimodal deep fusion, characterized in that: The device is used to implement the DOA estimation method based on multimodal deep fusion according to any one of claims 1 to 6, and the device includes a first module to a fifth module, wherein the functions of each module are as follows: The first module is used to construct a fully digital FD subarray and a heterogeneous hybrid H 2 Dual MIMO receiver combined with AD structure 2 AD-FD system model; In the second module, for the FD subarray, the Root-MUSIC method is used to calculate a rough estimate of the radiation source direction; The third module is to 2 Each group in the AD structure performs DOA estimation to generate multiple candidate angle sets; The fourth module uses the rough estimate of the radiation source direction provided by the FD subarray as the initial cluster center point, and uses the global maximum similarity (GMaxCS) and global minimum distance (GMinD) clustering methods to infer the true solution corresponding to each group from the candidate angle set; The fifth module synthesizes the two parts of the true solution through a deep fusion algorithm: construct a fusion network fusionNet based on a three-layer fully connected network FCNN, and 2 The true solution inferred by the subarray group of the AD structure and the rough estimate of the radiation source direction calculated by the FD subarray are fused to obtain the final DOA estimate.

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 based on multimodal deep fusion as described in any one of claims 1 to 6 is implemented.

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 in the DOA estimation method based on multimodal deep fusion as described in any one of claims 1 to 6 are implemented.

Images