MIMO radar intelligent DOA estimation method

Abstract

The application discloses a kind of MIMO radar intelligent DOA estimation methods, first, the centralized MIMO radar echo model under sparse scene is established, the iteration process and single-layer network structure of depth unfolding network are constructed, then the denoising auto-encoder based on deep neural network is constructed, the reconstructed spatial spectrum is obtained by training depth unfolding network, finally, the reconstructed spatial spectrum is searched for spectral peak, and DOA estimation is realized.The method of the application learns the implicit features in the data by training the deep neural network to enhance the robustness of DOA estimation, and gives the network model interpretability to improve the generalization ability of the model, which not only ensures the estimation accuracy and robustness of DOA estimation under low signal-to-noise ratio and single-snapshot sampling, but also has significant effect in beam sharpening and sidelobe suppression.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing, specifically relating to a smart DOA estimation method for MIMO radar. Background Technology

[0002] As a novel radar system, MIMO radar has received increasing attention for its parameter estimation, anti-jamming capabilities, and remote sensing. Compared to other existing array antenna radars, MIMO radar achieves waveform diversity through multiple transmit and receive antennas, enabling more accurate azimuth measurements with fewer physical antennas. However, existing MIMO radar beamforming methods based on matched filtering suffer from wide main lobes and high side lobes, resulting in low angular resolution and hindering practical applications.

[0003] To further improve the target azimuth angular resolution of MIMO radar, the paper "T.Yardibi, J.Li, P.Stoica, M.Xue and ABBaggeroer, et al. Source Localization and Sensing: A Nonparametric Iterative Adaptive Approach Based on Weighted Least Squares. IEEE Transactions on Aerospace and Electronic Systems. 2010, 46(1): 425-443" proposes an iterative adaptive approach (IAA). This method adaptively updates the autocorrelation matrix through iteration, which greatly improves the angular resolution of radar direction finding. However, the IAA algorithm suffers from high computational complexity, making it unsuitable for real-time processing. With the rise of artificial intelligence, the paper "L.Wu, Z.Liu and Z.Huang. Deep Convolution Network for Direction of Arrival Estimation With Sparse Prior. IEEE Signal Processing Letters, 2019, 26(11): 1688-1692" proposes an array DOA (Direction of Arrival) estimation method based on convolutional neural networks. This method utilizes the sparsity of the spatial domain to establish a mapping relationship between the array output covariance vector and the spatial spectrum, achieving high-precision spatial spectrum reconstruction. However, this method uses an end-to-end "black box" model, lacking interpretability and thus weakening its generalization ability. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes an intelligent DOA estimation method for MIMO radar. By designing an interpretable deep unfolded network model, the method estimates the DOA of targets detected by MIMO radar in sparse scenes. The deep unfolded network constructed by this method not only ensures the estimation accuracy and robustness of DOA estimation in single snapshot sampling, but also has significant effects on beam sharpening and sidelobe suppression.

[0005] The technical method employed in this invention is: a smart DOA estimation method for MIMO radar, the specific steps of which are as follows:

[0006] Step 1: Establish a centralized MIMO radar echo model in a sparse scene;

[0007] Employ a device with M t One transmitting antenna and M r A MIMO radar system model with K receiving antennas is proposed, assuming the MIMO radar detects K independent far-field targets, and the direction of the targets can be represented as... The target signal can be represented as T represents the transpose of the matrix, and t represents the distance-time variable.

[0008] When the target source is sparse relative to the entire airspace, let Let L represent the discrete set of sampled angles in the entire spatial domain, where L represents the number of sampled points and the sampling interval is 1. In a sparse background, the overcomplete form y(t) of the receiving antenna array after matched filtering is expressed as:

[0009]

[0010] in, Represents the array steering matrix. M represents t M r A complex column vector of ×L dimensions, with symbols Let represent the Kronecker product, and n(t) represent additive noise that follows a Gaussian distribution. Indicates the transmission array steering vector. The receiver array steering vector is expressed as follows:

[0011]

[0012]

[0013] Where λ represents the carrier wavelength, d t and d r Let represent the element spacing between the transmitting antenna and the receiving antenna, respectively, and satisfy d r =dt / M r =λ / 2.

[0014] s(t)=[s1(t),s2(t),...,s L (t)] T Represents the original target signal vector Mapping from θ to The extended form satisfies:

[0015]

[0016] Among them, s l (t) represents the l-th element in s(t). express The k-th element.

[0017] When the output of the receiving array is sampled in a single snapshot, equation (1) is expressed as:

[0018] y = As + n (5)

[0019] in, and These represent the observed signal vector, the target signal vector, and the noise vector, respectively. This represents the guidance matrix.

[0020] Step 2: The iterative process of constructing a deep unfolded network;

[0021] The iterative process is represented as:

[0022]

[0023]

[0024]

[0025] Among them, s (n) and v (n) Let h represent the reconstructed signal and residual in the nth iteration, respectively; β represent the update step size; func(·) represents sparse signal reconstruction; h (n) H represents the input of func(·), and H represents the conjugate transpose.

[0026] Step 3: Construct a single-layer network structure for the deep unfolded network;

[0027] Combining equations (6), (7), and (8), the sparse signal reconstructed in the nth iteration is represented as:

[0028]

[0029] Where W and B represent pre-defined learnable parameters, and W = I - βA H A, B = βA H , where I represents the identity matrix.

[0030] For an N-layer deep unfolded network, the parameters W and B are iteratively unfolded into a special feedforward neural network structure. This structure is optimized using the backpropagation algorithm, and the parameters to be optimized are... and The single-layer network structure can then be represented as:

[0031]

[0032] Step 4: Construct a denoising autoencoder based on a deep neural network;

[0033] Reconsider equation (7), let:

[0034]

[0035] Where, δ (n) In the nth iteration The error between the actual target signal vector s0 and the target signal vector s0.

[0036] When the algorithm converges, the sparse recovery signal It will infinitely approach the true target signal vector s0, that is:

[0037]

[0038] Based on equation (12), a noise reduction autoencoder is constructed as a specific implementation of func(·).

[0039] Step 5: Train the deep unfolded network to obtain the reconstructed spatial spectrum;

[0040] The dataset required for model training is generated based on the MIMO radar-related parameters mentioned in step one. The specific training process is as follows:

[0041] First, the observed signal from the receiving array after matched filtering is input into the deep unfolded network for forward propagation, and the loss function value is calculated, specifically using the mean squared error (MSE) function as the loss function. Then, the adaptive moment estimation Adam optimizer is used to update the network parameters through backpropagation. Finally, after several iterations until the loss function converges, the trained deep unfolded network model is obtained.

[0042] right and Perform a special initialization, namely:

[0043]

[0044] and Initialize to a zero vector.

[0045] Step 6: Perform peak search on the reconstructed spatial spectrum to achieve DOA estimation.

[0046] The beneficial effects of this invention are as follows: The method of this invention first establishes a centralized MIMO radar echo model in a sparse scene, constructs an iterative process and single-layer network structure for a deep unfolded network, then constructs a noise reduction autoencoder based on a deep neural network. By training the deep unfolded network, a reconstructed spatial spectrum is obtained. Finally, spectral peak search is performed on the reconstructed spatial spectrum to achieve DOA estimation. This method enhances the robustness of DOA estimation by training a deep neural network to learn latent features in the data and provides interpretability to the network model, thereby improving its generalization ability. It not only ensures the estimation accuracy and robustness of DOA estimation under low signal-to-noise ratio and single-shot sampling conditions, but also has significant effects on beam sharpening and sidelobe suppression. Attached Figure Description

[0047] Figure 1 This is a flowchart of a MIMO radar intelligent DOA estimation method according to the present invention.

[0048] Figure 2 This is a diagram of the MIMO radar array configuration used in an embodiment of the present invention.

[0049] Figure 3 This is a schematic diagram of the deep unfolded network structure in an embodiment of the present invention.

[0050] Figure 4 The diagram shows the spatial spectrum results of DOA estimation using different methods provided in the embodiments of the present invention.

[0051] Figure 5 This is a comparison chart of the statistical performance of the method of the present invention with other methods under different signal-to-noise ratios in the embodiments of the present invention. Detailed Implementation

[0052] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0053] like Figure 1 As shown, a smart DOA estimation method for MIMO radar according to the present invention includes the following steps:

[0054] Step 1: Establish a centralized MIMO radar echo model in a sparse scene;

[0055] like Figure 2 As shown, in this embodiment, a device equipped with M is used. t =4 transmitting antennas and M r=8 received antennas in a MIMO radar system. Assume the MIMO radar detects K = 2 independent far-field targets with equal power, and the direction of the targets can be expressed as... The target signal can be represented as T represents the transpose of the matrix, and t represents the distance-time variable.

[0056] When the target source is sparse relative to the entire airspace, let Let L represent the discrete set of sampled angles in the entire spatial domain, where L represents the number of sampled points and the sampling interval is 1. In this embodiment, the considered spatial range is [-90°, 90°], and the sampling interval is... The spatial domain is then sampled into L = 181 sampling points. Therefore, in a sparse background, the overcomplete form of the receiving antenna array after matched filtering can be expressed as:

[0057]

[0058] in, Represents the array steering matrix. M represents t M r A complex column vector of ×L dimensions, with symbols Let represent the Kronecker product, and n(t) represent additive noise that follows a Gaussian distribution. Indicates the transmission array steering vector. The receiver array steering vector is expressed as follows:

[0059]

[0060]

[0061] Where λ represents the carrier wavelength, d t and d r Let d represent the element spacing between the transmitting antenna and the receiving antenna, respectively. The element spacing between the transmitting antenna and the receiving antenna satisfies d r =d t / M r =λ / 2.

[0062] s(t)=[s1(t),s2(t),...,s L (t)] T Represents the original target signal vector Mapping from θ to The extended form satisfies:

[0063]

[0064] Among them, s l(t) represents the l-th element in s(t). express The k-th element in the array. And s(t) is... This is an approximate representation and contains a certain amount of quantization error.

[0065] When the output of the receiving array is sampled in a single snapshot, equation (1) is expressed as:

[0066] y = As + n (5)

[0067] in, and These represent the observed signal vector, the target signal vector, and the noise vector, respectively. In this embodiment, when M... t M r << L, i.e., the guidance matrix It is also an underdetermined matrix.

[0068] Step 2: The iterative process of constructing a deep unfolded network;

[0069] Deep unfolded networks are a method of constructing deep neural networks by unfolding an iterative process. In essence, a single-layer deep unfolded network structure is a single-round iterative update process.

[0070] In this embodiment, the proposed iterative process is expressed as follows:

[0071]

[0072]

[0073]

[0074] Among them, s (n) and v (n) Let h represent the reconstructed signal and residual in the nth iteration, respectively; β represent the update step size; func(·) represents sparse signal reconstruction; h (n) H represents the input of func(·), and H represents the conjugate transpose.

[0075] Step 3: Construct a single-layer network structure for the deep unfolded network;

[0076] Combining equations (6), (7), and (8), the sparse signal reconstructed in the nth iteration can be expressed as:

[0077]

[0078] Where W and B represent pre-defined learnable parameters, and W = I - βA H A, B = βA H , where I represents the identity matrix.

[0079] Based on equation (9), each iteration of the deep unfolded network proposed in this embodiment performs one linear operation and one nonlinear operation on the observed signal vector y. For an N-layer deep unfolded network, when parameters W and B are considered as learnable parameters, its iterative process can be unfolded into a special feedforward neural network structure. This structure can use the backpropagation algorithm to optimize the network parameters, and the parameters to be optimized are... and The single-layer network structure in this embodiment can be represented as follows:

[0080]

[0081] Step 4: Construct a denoising autoencoder based on a deep neural network;

[0082] Reconsider equation (7), let:

[0083]

[0084] Where, δ (n) In the nth iteration The error between the actual target signal vector s0 and the target signal vector s0 can also be considered as a noise component.

[0085] Rearrange equation (11) as follows: Then the input of func(·) can be equivalent to the real target signal vector s0 and the noise component δ. (n) The sum of . Therefore, when the algorithm converges, the sparsely recovered signal. It will infinitely approach the true target signal vector s0, that is:

[0086]

[0087] As can be seen from equation (12), the function of func(·) is equivalent to a denoiser. Leveraging the powerful learning capabilities of deep neural networks, this embodiment constructs a denoising autoencoder (DAE) as the specific implementation of func(·). By learning the latent features in a large amount of observation data, the denoising autoencoder can improve the denoising capability during signal recovery and achieve robust DOA estimation.

[0088] Specifically, the denoising autoencoder consists of fully connected layers. Except for the last layer, each layer is followed by a rectified linear unit (ReLU) as a non-linear activation function. In this embodiment, the specific structure of the denoising autoencoder is shown in Table 1.

[0089] Table 1

[0090] Layer Category size Activation function Fully connected layer 1200 ReLU Fully connected layer 1000 ReLU Fully connected layer 800 ReLU Fully connected layer 300 ReLU Fully connected layer 100 ReLU Fully connected layer 300 ReLU Fully connected layer 800 ReLU Fully connected layer 1000 ReLU Fully connected layer 1200 ReLU Fully connected layer L none

[0091] Based on this, embedding a denoising autoencoder into a single-layer network structure yields a deep unfolded network model. This model can both suppress noise by learning latent features from the data and enhance generalization ability by providing interpretability. A schematic diagram of the deep unfolded network structure proposed in this embodiment is shown below. Figure 3 As shown.

[0092] Step 5: Train the deep unfolded network to obtain the reconstructed spatial spectrum;

[0093] In this embodiment, the dataset required for model training is generated based on the MIMO radar-related parameters mentioned in step one. The specific training process is as follows:

[0094] First, the observed signal from the receiving array after matched filtering is input into the deep unfolded network proposed in this embodiment for forward propagation, and the loss function value is calculated. Specifically, the mean square error (MSE) function is used as the loss function. Then, the adaptive moment estimation (Adam) optimizer is used to update the network parameters through backpropagation. Finally, after several iterations until the loss function converges, the trained deep unfolded network model is obtained.

[0095] Furthermore, to prevent the model from getting trapped in local minima during training, this embodiment... and Perform a dedicated initialization, that is

[0096]

[0097] and Initialize to a zero vector.

[0098] Step 6: Perform peak search on the reconstructed spatial spectrum to achieve DOA estimation.

[0099] In summary, the proposed method can be used to estimate the DOA of MIMO radar in sparse scenes. This invention compares the proposed method with iterative adaptive methods and iterative soft thresholding algorithms to verify the DOA estimation performance of the proposed method. Figure 4 The reconstructed spatial spectra obtained by the three methods after 100 Monte Carlo experiments are shown when the signal-to-noise ratio is 2dB and single-shot sampling is used. Figure 4 (a) shows the result of the iterative adaptive method. Figure 4 (b) shows the result of the iterative soft thresholding algorithm. Figure 4 (c) is the result of the method proposed in this invention. Figure 5The statistical performance of the three methods in DOA estimation under different signal-to-noise ratios is presented. The root mean square error (RMSE) is used as the metric for statistical performance. Figure 4 , Figure 5 The results show that, in the case of a single snapshot, the method of the present invention not only improves the angular resolution and robustness of DOA estimation, but also has significant effects on beam sharpening and sidelobe suppression.

[0100] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A smart DOA estimation method for MIMO radar, the specific steps of which are as follows: Step 1: Establish a centralized MIMO radar echo model in a sparse scene; Use a device with One transmitting antenna and A MIMO radar system model with one receiving antenna is provided, assuming the MIMO radar detects... There are three independent far-field targets, and the direction of the targets is represented as... The target signal is represented as , To represent the transpose of a matrix, Represents distance as a time variable; When the target source is sparse relative to the entire airspace, let Represents the discrete sampling angle set of the entire spatial domain. This indicates the number of sampling points, with a sampling interval of 1. In a sparse background, the overcomplete form of the receiving antenna array after matched filtering... Represented as: (1)； in, Represents the array steering matrix. express A complex-valued column vector of dimension, with symbol Represents the Kronecker product. This represents additive noise that follows a Gaussian distribution. Indicates the transmission array steering vector. The receiver array steering vector is expressed as follows: (2)； (3)； in, Indicates the carrier wavelength. and Let represent the element spacing between transmitting antennas and between receiving antennas, respectively, and satisfy . ; Represents the original target signal vector from Mapped to The extended form satisfies: (4)； in, express The Middle One element, express The Middle One element; When the output of the receiving array is sampled in a single snapshot, equation (1) is expressed as: (5)； in, , and These represent the observed signal vector, the target signal vector, and the noise vector, respectively. Represents the guidance matrix; Step 2: The iterative process of constructing a deep unfolded network; The iterative process is represented as: (6)； (7)； (8)； in, and Let these represent the reconstructed signal and residual in the nth iteration, respectively. Indicates the update step size. This indicates that sparse signal reconstruction is being performed. express Input, Indicates conjugate transpose; Step 3: Construct a single-layer network structure for the deep unfolded network; Combining equations (6), (7), and (8), the sparse signal reconstructed in the nth iteration is represented as: (9)； in, and This represents the pre-defined learnable parameters, and , , Represents the identity matrix; For an N-layer deep unfolded network, the parameters and The iterative process unfolds into a special feedforward neural network structure, which optimizes the network parameters using the backpropagation algorithm. The parameters to be optimized are... and Then, the single-layer network structure is represented as: (10)； Step 4: Construct a denoising autoencoder based on a deep neural network; Reconsider equation (7), let: (11)； in, In the nth iteration With the real target signal vector The error between; When the algorithm converges, the sparse recovery signal It will infinitely approach the real target signal vector ,Right now: (12)； Based on equation (12), a noise reduction autoencoder is constructed as... The specific implementation; Step 5: Train the deep unfolded network to obtain the reconstructed spatial spectrum; The dataset required for model training is generated based on the MIMO radar-related parameters mentioned in step one. The specific training process is as follows: First, the observed signal from the receiving array after matched filtering is input into the deep unfolded network for forward propagation, and the loss function value is calculated, specifically using the mean squared error (MSE) function as the loss function. Then, the adaptive moment estimation Adam optimizer is used to update the network parameters through backpropagation. Finally, after several iterations until the loss function converges, the trained deep unfolded network model is obtained. right and Perform a special initialization, namely: (13)； and Initialize to a zero vector; Step 6: Perform peak search on the reconstructed spatial spectrum to achieve DOA estimation.

Images