CN-116933001-B - DOA estimation method based on deep learning

CN116933001BCN 116933001 BCN116933001 BCN 116933001BCN-116933001-B

Abstract

The invention discloses a DOA estimation method based on deep learning, which comprises the following steps of S1, constructing an array data model by utilizing received array data, obtaining an output signal X by the array data model, S2, constructing a fourth-order cumulant of the output signal X, obtaining a fourth-order cumulant matrix C Y of the output signal X, S3, carrying out vectorization processing on C Y , carrying out de-duplication on a processing result, then calculating by combining a space smoothing algorithm to obtain a covariance matrix, and S4, inputting the calculated covariance matrix into a trained CNN model to obtain a DOA estimation result. The invention locates the interference signal by combining the sparse array, the space smoothing algorithm and the deep neural network model. The DOA estimation method effectively improves the accuracy of the DOA estimation algorithm, reduces the calculated amount and reduces the hardware cost and the operation burden.

Inventors

LI JIANZHONG
LIN CANXIN
LIANG ZEXIAO
FU ZHE

Assignees

广东工业大学

Dates

Publication Date: 20260505
Application Date: 20230718

Claims (8)

1. The DOA estimation method based on deep learning is characterized by comprising the following steps of: s1, constructing an array data model by utilizing received array data, and obtaining output signals from the array data model ; S2, constructing a fourth-order cumulant of the output signal X to obtain a fourth-order cumulant matrix of the output signal X ; S3, pairing Performing vectorization processing, de-duplicating a processing result, and then calculating by combining a space smoothing algorithm to obtain a covariance matrix; s4, inputting the calculated covariance matrix into a trained CNN model to obtain a DOA estimation result; Step S3 is performed on After vectorization processing, de-duplicating the processing result, and then calculating by combining a space smoothing algorithm to obtain a covariance matrix, wherein the specific process is as follows: Will be Vector processing to obtain vector : Wherein, the An equivalent direction vector matrix composed of virtual elements in the fourth-order cumulants, As a vector of the equivalent signal, Representing the fourth order cumulative amount of the source; , ; Processing is performed according to the position relation of the distributed array to remove repeated signals, wherein the position relation is shown as the following formula: Wherein, the Representing the real number field, D is the set of positions of the virtual antennas in the virtual array, in effect The combination of the index part of (a) and the above part is just that De-duplication and ordering constructs as new vectors The expression is as follows: wherein V is the virtual array element position set of the receiving array element, For the total number of virtual array elements, G is the selection matrix of the virtual array elements receiving the actual signals, q refers to the q-th element of V, Is that The dimension matrix is used to determine the dimensions of the matrix, The parameter p refers to the p-th element of D, By constructing the screening matrix G Structure of the device T represents the transpose of the matrix; Will be Divided into Multiple overlapping subarrays , Each subarray has A virtual element, wherein The sensors of the subarray are located Wherein And then will be The covariance matrix of the subarray is defined as After calculating the covariance matrix of each sub-array, Dimensional spatially smoothed covariance matrix Defined as the average of all subarray covariance matrices: Wherein the method comprises the steps of Is one A diagonal matrix of dimensions is used, Is that An equivalent vector matrix of directions of the dimensions, ; Step S4, inputting the calculated covariance matrix into a CNN model to obtain a DOA estimation result, wherein the specific process is as follows: covariance matrix Converted into amplitude and phase form as follows: Wherein, is called As the magnitude factor of the covariance matrix, Designing a characteristic matrix for phase factor The upper triangle is The imaginary part of the upper triangle phase factor, the lower triangle is The real part of the upper triangular phase factor, the diagonal element is Normalized values of diagonal elements in order to input valid information into model training or prediction are as follows: Wherein, the The function representation takes the real part of the complex number, The representation takes the imaginary part of the complex number, Vectorizing it to obtain feature vector There is The following is shown: Finally, will And inputting the obtained DOA estimation value into a trained CNN model to obtain a final DOA estimation value, wherein the DOA estimation value is the arrival angle information of the interference signal.
2. The DOA estimation method based on deep learning as claimed in claim 1, wherein in step S1, an array data model is constructed by using the received array data, and an output signal is obtained from the array data model The specific process is as follows: Setting use The array elements form an extended Kantol array, and The distribution rule is as follows: Wherein, the Is a non-negative integer, meaning that the number of elements of the extended Contole array can only be 2 The power of the power, i.e., the value of M can only be 1,2,4,8,16, S is a conventional set of contuo array position distributions, E is an extended set of contuo array position distributions employed, To extend the distribution of a Contoll array (E-CA), a Contoll array Array of data Composition, array By passing through The offset is added to form A variant of (C), and The elements of (a) are always non-negative integers, Is the first The steering vector of the array distribution is reached by the source, where j refers to the imaginary number i in the mathematical concept complex number a + bi, , , Is the first of the array The number of the array elements is one, Is the actual array element distribution position set, d is ; Represent the first Actual distribution positions of the array elements; The setting space is provided with A plurality of far-field narrow-band signals respectively Is incident on the receiving array, and the incident signal is processed by the receiving array After the sub-reception of the samples, the output signal of the array The method comprises the following steps: Wherein X represents a signal received by an array element, A represents a direction vector matrix, S represents an incident signal, N represents array noise, Is that Matrix array Is that The matrix is formed by a matrix of, Is that N is A matrix; ; Representing the signal vector received by M array elements for Proceeding with Point sampling, the problem to be processed is converted into a signal to be output Is { of sampling } The direction of arrival angle of the signal source is estimated by M, which is the number of array elements, so that the array signal can be regarded as a superposition of several spatial harmonics of noise interference naturally, and the direction of arrival estimation problem is linked with the spectrum estimation.
3. The DOA estimation method as claimed in claim 1, wherein the constructing the fourth order cumulant of the output signal X in the step S2 obtains a fourth order cumulant matrix of the output signal X The specific process is as follows: Processing the output signal X by the following formula to obtain a fourth-order cumulant matrix of the output signal X : Wherein the method comprises the steps of Representative mathematics it is desirable that the first and second heat exchangers, Represents the Kronecker product Representing the conjugate transpose of the matrix, t representing the t-th sampling point, at which time the fourth order cumulative amount matrix can be decomposed into: To the direction matrix which is extended after using the fourth order cumulant Columns may be represented as , For the fourth order cumulative amount of the incident signal S, The following formula is shown: 。
4. the DOA estimation method based on deep learning as claimed in claim 1, wherein, And training the CNN model by using a ReLU function as an activation function of each layer, wherein the expression is as follows: The mean square error loss function MSE is adopted as the loss function, and the expression is as follows: Wherein the method comprises the steps of Is a predicted value of the angle of the DOA, Is the actual angle value of the DOA.
5. The DOA estimation method based on deep learning as claimed in claim 1, wherein, In the training of the CNN model, the momentum SGD algorithm is combined to avoid the problems that the calculated amount of the back propagation update parameters of the deep neural network is too large, the operation speed is greatly reduced and memory overflow is even caused, and the momentum SGD algorithm update rule is as follows: (1) Initializing learning rate Momentum parameter Initial parameters Initial velocity ; (2) Selecting m samples from the training set, and calculating gradient: ; Wherein, the Is assigned the first Input samples Can be regarded as Is used to update the update amplitude of (a), The larger the size of the container, The smaller the value of the change in (c) is, Relative to The larger the previous gradient, the greater the effect on the current direction, L ()' is a loss function Refers to an objective function, representing input data, The mapping relation with the predicted value is that, Is a gradient calculation, pair (-) solution Is a bias guide of (2); (3) Speed update Parameter update ; (4) Repeating the steps (2) and (3) until No longer changes.
6. The DOA estimation system based on the deep learning is characterized by comprising a memory and a processor, wherein the memory comprises a DOA estimation method program based on the deep learning, and the DOA estimation method program based on the deep learning realizes the following steps when being executed by the processor: s1, constructing an array data model by utilizing received array data, and obtaining output signals from the array data model ; S2, constructing a fourth-order cumulant of the output signal X to obtain a fourth-order cumulant matrix of the output signal X ; S3, pairing Performing vectorization processing, de-duplicating a processing result, and then calculating by combining a space smoothing algorithm to obtain a covariance matrix; s4, inputting the calculated covariance matrix into a trained CNN model to obtain a DOA estimation result; Step S3 is performed on After vectorization processing, de-duplicating the processing result, and then calculating by combining a space smoothing algorithm to obtain a covariance matrix, wherein the specific process is as follows: Will be Vector processing to obtain vector : Wherein, the An equivalent direction vector matrix composed of virtual elements in the fourth-order cumulants, As a vector of the equivalent signal, Representing the fourth order cumulative amount of the source; , ; Processing is performed according to the position relation of the distributed array to remove repeated signals, wherein the position relation is shown as the following formula: Wherein, the Representing the real number field, D is the set of positions of the virtual antennas in the virtual array, in effect The combination of the index part of (a) and the above part is just that De-duplication and ordering constructs as new vectors The expression is as follows: wherein V is the virtual array element position set of the receiving array element, For the total number of virtual array elements, G is the selection matrix of the virtual array elements receiving the actual signals, q refers to the q-th element of V, Is that The dimension matrix is used to determine the dimensions of the matrix, The parameter p refers to the p-th element of D, By constructing the screening matrix G Structure of the device T represents the transpose of the matrix; Will be Divided into Multiple overlapping subarrays , Each subarray has A virtual element, wherein The sensors of the subarray are located Wherein And then will be The covariance matrix of the subarray is defined as After calculating the covariance matrix of each sub-array, Dimensional spatially smoothed covariance matrix Defined as the average of all subarray covariance matrices: Wherein the method comprises the steps of Is one A diagonal matrix of dimensions is used, Is that An equivalent vector matrix of directions of the dimensions, ; Step S4, inputting the calculated covariance matrix into a CNN model to obtain a DOA estimation result, wherein the specific process is as follows: covariance matrix Converted into amplitude and phase form as follows: Wherein, is called As the magnitude factor of the covariance matrix, Designing a characteristic matrix for phase factor The upper triangle is The imaginary part of the upper triangle phase factor, the lower triangle is The real part of the upper triangular phase factor, the diagonal element is Normalized values of diagonal elements in order to input valid information into model training or prediction are as follows: Wherein, the The function representation takes the real part of the complex number, The representation takes the imaginary part of the complex number, Vectorizing it to obtain feature vector There is The following is shown: Finally, will And inputting the obtained DOA estimation value into a trained CNN model to obtain a final DOA estimation value, wherein the DOA estimation value is the arrival angle information of the interference signal.
7. A DOA estimation system based on deep learning as claimed in claim 6, wherein, Step S1, constructing an array data model by using the received array data, and obtaining output signals from the array data model The specific process is as follows: Setting use The array elements form an extended Kantol array, and The distribution rule is as follows: Wherein, the Is a non-negative integer, meaning that the number of elements of the extended Contole array can only be 2 The power of the power, i.e., the value of M can only be 1,2,4,8,16, S is a conventional set of contuo array position distributions, E is an extended set of contuo array position distributions employed, To extend the distribution of a Contoll array (E-CA), a Contoll array Array of data Composition, array By passing through The offset is added to form A variant of (C), and The elements of (a) are always non-negative integers, Is the first The steering vector of the array distribution is reached by the source, where j refers to the imaginary number i in the mathematical concept complex number a + bi, , , Is the first of the array The number of the array elements is one, Is the actual array element distribution position set, d is ; Represent the first Actual distribution positions of the array elements; The setting space is provided with A plurality of far-field narrow-band signals respectively Is incident on the receiving array, and the incident signal is processed by the receiving array After the sub-reception of the samples, the output signal of the array The method comprises the following steps: Wherein X represents a signal received by an array element, A represents a direction vector matrix, S represents an incident signal, N represents array noise, Is that Matrix array Is that The matrix is formed by a matrix of, Is that N is A matrix; ; Representing the signal vector received by M array elements for Proceeding with Point sampling, the problem to be processed is converted into a signal to be output Is { of sampling } The direction of arrival angle of the signal source is estimated by M, which is the number of array elements, so that the array signal can be regarded as a superposition of several spatial harmonics of noise interference naturally, and the direction of arrival estimation problem is linked with the spectrum estimation.
8. A computer readable storage medium, characterized in that a depth learning based DOA estimation method program is included in the computer readable storage medium, which, when executed by a processor, implements the steps of a depth learning based DOA estimation method according to any one of claims 1 to 5.

Description

DOA estimation method based on deep learning Technical Field The present invention relates to the field of signal processing, and in particular, to a DOA estimation method, system and storage medium based on deep learning. Background In recent years, along with the development of Beidou satellite navigation systems, the application of the Beidou satellite navigation systems relates to a plurality of fields, and along with the gradual rising of the integration of artificial intelligence and the Internet of things in the world, the application of the Beidou satellite navigation systems further goes into various aspects of daily life and social production. However, due to the vulnerability of the Beidou satellite signals, most of the satellite signals are affected by active or passive interference in the satellite signal propagation process. Hundreds of satellite communication interference events are reported to occur every year, and these interference events show an increasing trend with the development of modern society. From the practical application of satellite systems in China, jamming events also occur. Most of the interference signals are short-lived and can be eliminated by conventional satellite monitoring measures, but a small number of interference signals with long duration can cause the whole system to not work normally. Therefore, how to quickly and effectively reduce and even shield the influence of interference signals is a new research topic of the safety protection of the Beidou satellite navigation system. Among the problems in this respect, how to locate the source of interference is a critical issue, which involves locating the incoming direction of arrival (DOA, direction ofArrival) estimation of the spatial signal. The positioning of signal sources is an extremely important research topic in the field of array signal processing. In recent 30 years, the technology for the subject, namely the spatial spectrum estimation technology, is rapidly developed, wherein the most typical meaning is a multiple signal classification algorithm (MUSIC) proposed by Schmidt R O et al, the main idea of the algorithm is to divide a covariance matrix of array received data into a signal subspace and a noise subspace orthogonal with the signal subspace according to a matrix spectrum decomposition theory, and an angle estimation value of the signal is obtained by solving a peak value of a spectrum function established by utilizing the mutually orthogonal characteristics of the two subspaces, so that the performance of the algorithm in terms of resolution and the like is improved, and meanwhile, the physical aperture limit-Rayleigh limit of the array estimation on the direction of arrival (DOA) in the traditional spatial spectrum estimation algorithm is broken through, and the development of the research about subspace algorithm is promoted. The root-finding MUSIC algorithm, the beam space MUSIC algorithm, the MNM and the decorrelation algorithm are all based on the derivation of the MUSIC algorithm. Because the MUSIC algorithm needs to perform high-calculation-amount spectrum peak search, a signal subspace class algorithm typified by the ESPRIT algorithm is generated, the class algorithm can perform DOA estimation without performing spectrum peak search, the DOA estimation speed is greatly improved, but the class algorithm needs specific array distribution to be used, and the application range is relatively limited. Subsequently, subspace fitting algorithms typified by maximum likelihood algorithms, weighted subspace fitting algorithms, and the like, which perform better than the subspace decomposition type algorithms described above, begin to emerge. The performance of the algorithm is obviously better than that of a subspace decomposition algorithm under the conditions of smaller signal-to-noise ratio and smaller array received data sampling number, but the algorithm needs to calculate the optimal solution of the direction estimation likelihood function with nonlinear characteristics through multidimensional search so as to obtain the DOA estimated value algorithm, and compared with the DOA estimated value algorithm, the DOA estimated value algorithm has larger operation amount. Disclosure of Invention The invention provides a DOA estimation method based on deep learning, which aims to overcome the defects that the DOA estimation algorithm in the prior art is difficult to process non-Gaussian signals, the second moment cannot fully describe the statistical characteristics of the signals, the performance of the algorithm is greatly reduced, and the influence of interference signals is basically required to be efficiently reduced or shielded. The primary purpose of the invention is to solve the technical problems, and the technical scheme of the invention is as follows: The first aspect of the invention provides a DOA estimation method based on deep learning, comprising the following steps: S1, constructing a