A method for estimating the DOA and distance parameters of a near-field target positioning

Abstract

The application belongs to the technical field of millimeter wave radar signal processing, and provides a DOA and distance parameter estimation method for near-field target positioning, comprising: receiving near-field target signals by using an array of linearly distributed array elements; wherein the number M of linearly distributed array elements in the array is greater than the number K of near-field target signals to be measured; establishing a near-field model to generate a training data set and a test data set; constructing an NFLnet network structure; the NFLnet network structure is obtained by training and testing the training data set and the test data set; obtaining an output matrix G based on the NFLnet network structure; obtaining a distance-angle two-dimensional spectrum based on the output matrix G, and determining a loss function; and minimizing the loss function to train the NFLnet network structure. By using the technical scheme of the application, the problem of solving the distance-angle coupling term parameters can be effectively solved, the high-complexity two-dimensional spectrum peak search can be avoided, and the advantages of high robustness and low complexity are achieved.

Description

Technical Field

[0001] This invention relates to the field of millimeter-wave radar signal processing technology, specifically to a method for estimating the DOA and range parameters for near-field target localization. Background Technology

[0002] Target localization has been extensively studied in the field of array signal processing. Based on the distance from the source to the array antenna, it can be divided into near-field source localization and far-field source localization. When the distance from the source to the array antenna is much larger than the array aperture, the source is located in the far-field region of the array, and the signal propagates as a plane wave. When the distance from the source to the array antenna is small, within the Fresnel zone, the source is located in the near-field region, and the signal propagates as a spherical wave, with the wavefront shape changing with the array element position. Therefore, far-field source localization mainly involves estimating the direction of arrival (DOA), while near-field source localization requires estimating not only the DOA but also the distance parameter from the source to the array.

[0003] Traditional DOA estimation methods, such as the MUSIC (Multiple Signal Classification) algorithm, the ESPRIT (Estimating Signal Parameters via Rotational Invariance Techniques) algorithm, and a series of compressed sensing-based algorithms such as the OMP (Orthogonal Matching Pursuit) algorithm and the ANM (Atomic Norm Minimization) algorithm, are all based on the plane wave assumption of far-field signals and cannot estimate the range dimension parameters of near-field sources. However, with the widespread application of high-frequency bands and large-aperture array antennas, signals in reality often do not satisfy the far-field assumption but are located in the near field. Therefore, the localization problem of near-field sources has attracted researchers' attention. Currently, near-field source localization mainly faces two problems: firstly, the algorithm complexity is high. For example, the classic TS-MUSIC (TwoStage MUSIC) algorithm can estimate not only the angle and range parameters of the near-field source but also the DOA of the mixed far-field source. However, this type of method based on spectral peak search is computationally complex and computationally intensive. Another issue is that, due to the spherical wave transmission state, the signal amplitude received by each element of the array antenna is not consistent and is related to distance and angle, but this is often ignored in traditional near-field source models, affecting parameter estimation performance.

[0004] Therefore, establishing a more accurate near-field target signal model and proposing a new near-field target localization method to achieve high-precision, low-complexity joint estimation of distance and angle has significant research and application value. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for estimating the DOA and distance parameters of near-field target localization, thereby solving the problems that existing target localization methods cannot simultaneously estimate the angle and distance parameters of near-field signals, and ignore the influence of signal amplitude.

[0006] This invention provides a method for estimating the DOA and range parameters of near-field target localization, comprising:

[0007] An array of linearly distributed array elements is used to receive near-field target signals; wherein, the number M of linearly distributed array elements in the array is greater than the number K of near-field target signals to be measured;

[0008] Build a near-field model and generate training and testing datasets;

[0009] Construct the NFLnet network structure; the NFLnet network structure is obtained by training and testing the training dataset and the test dataset.

[0010] The output matrix G is obtained based on the NFLnet network structure;

[0011] The distance-angle two-dimensional spectrum is obtained based on the output matrix G, and the loss function is determined; the loss function is minimized to train the NFLnet network structure.

[0012] As can be seen from the above technical solution, the DOA and distance parameter estimation method for near-field target localization provided by the present invention can obtain the DOA and distance parameters of near-field signals by establishing a near-field model and training the NFLnet network structure accordingly, thus ensuring low complexity.

[0013] Optionally, the near-field model is

[0014]

[0015] in,

[0016]

[0017] y m (t) represents the signal received by the m-th array element, and t is the sampling time, s. k (t) represents the k-th near-field target signal, n m (t) represents the noise signal received by the m-th array element;

[0018] The relative amplitude γ of the signal received by the m-th element k,m Phase τ k,m for

[0019]

[0020]

[0021] d m θ is the element spacing. k ,r k,0 (k=0,1,…,K-1) are the azimuth angle and distance from the near-field target signal k to the reference array element, respectively, and λ is the operating wavelength.

[0022] Optionally, the data input into the NFLnet network structure to obtain the output matrix G is the raw signal data received by the array.

[0023] Optionally, the NFLnet network structure includes two FC layers and five CBR fusion layers, wherein the CBR fusion layers include a fusion of convolutional layers, batch normalization layers, and activation function layers.

[0024] Optionally, the output X of the CBR fusion layer out =max(0,W F X+B F ), where W F B is the weight matrix after fusing the convolutional layer and the batch planning layer. F This is the bias parameter matrix after fusing the convolutional layer and the batch planning layer.

[0025] Optionally, the distance-angle two-dimensional spectrum includes the estimated spectrum f. sp (θ,r) and reference spectrum f ref (θ,r); the estimated spectrum f sp (θ,r) is derived from the output M B ×M θ M r Reshaping the dimensional matrix into M B ×M θ ×M r The reference spectrum is obtained.

[0026]

[0027] Where, σ G The standard deviation determines the resolution of the spectral peaks.

[0028] Optionally, based on the estimated spectrum f sp (θ,r) and the reference spectrum f ref (θ,r), obtain the loss function

[0029]

[0030] Among them, M θ and M r It is the number of grids used to divide the angular and distance dimensions.

[0031] By adopting the above technical solution, this application has the following beneficial effects:

[0032] Since the near-field signal is transmitted as a spherical wave, the wavefront shape changes with the position of the array elements, and the amplitude of the received signal of each array element is no longer consistent and depends on the angle and distance parameters. By deriving the amplitude and phase expressions of the received signal of each channel, a more accurate near-field model of the signal received by the receiving array can be established.

[0033] To address the problem of solving the distance-angle coupling term parameters, an NFLnet network structure based on a low-complexity neural network was constructed. The NFLnet network structure employs two fully connected FC layers and five CBR fusion layers. It takes the raw echo signal data as input and the angle-distance two-dimensional spectrum as output, directly and effectively performing joint estimation of the target angle and distance. This not only effectively solves the problem of solving the distance-angle coupling term parameters but also avoids the high-complexity two-dimensional spectrum peak search, exhibiting advantages such as high robustness and low complexity. Attached Figure Description

[0034] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0035] Figure 1 The flowchart illustrates a method for estimating the DOA and distance parameters of near-field target localization according to an embodiment of the present invention.

[0036] Figure 2 A schematic diagram of the near-field model provided in an embodiment of the present invention is shown;

[0037] Figure 3 A schematic diagram of the NFLnet network structure provided in an embodiment of the present invention is shown;

[0038] Figure 4 A schematic diagram illustrating the simulation results of locating multiple targets according to an embodiment of the present invention is shown;

[0039] Figure 5 A schematic diagram of a near-field target localization test scenario constructed as a verification example of the present invention is shown;

[0040] Figure 6 A schematic diagram of the near-field signal target localization results provided in the verification example of the present invention is shown. Detailed Implementation

[0041] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore merely examples, and should not be construed as limiting the scope of protection of the present invention.

[0042] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0043] Because near-field signals propagate as spherical waves, the wavefront shape changes with the array element positions, resulting in inconsistent amplitudes of the received signals at each element, and these amplitudes depend on angle and range parameters. Existing methods suffer from two drawbacks: one is that they can only estimate the DOA (Direction of Area) but not the range parameter; the other is that while they can estimate both DOA and range parameters simultaneously, the complexity is high and they cannot be applied to the spherical wave state of near-field source signals. To address these issues, this invention proposes a method for estimating DOA and range parameters for near-field target localization. Figure 1 As shown in the figure, an embodiment of the present invention provides a method for estimating the DOA and distance parameters for near-field target localization, including:

[0044] S1. An array of linearly distributed array elements is used to receive near-field target signals; wherein, the number M of linearly distributed array elements in the array is greater than the number K of the near-field target signals to be measured.

[0045] This step is based on Figure 2 For example, the first array element is used as the reference array element, and the array element spacing is d. m (m=0,1,…,M-1), assuming K information sources are located at (θ0,r) 0,m ),(θ1,r 1,m ),…,(θ K-1 ,r K-1,m ), where θ k ,r k,0 (k = 0, 1, ..., K-1) represent the azimuth and distance from the near-field target signal k to the reference array element, respectively. k,m It is the distance between the near-field target signal k and the m-th array element.

[0046] When r k,m satisfy When the target is in the Fresnel zone, it is a near-field signal. At this time, the electromagnetic wave propagates in the form of a spherical wave, and the wavefront shape changes with the position of the array elements. Here, D is the aperture of the receiving antenna array, and λ is the operating wavelength.

[0047] S2. Build a near-field model and generate training and test datasets;

[0048] Considering that the signal received in step S1 is a near-field signal, the transmission state of electromagnetic spherical waves means that the signal amplitude received by each element of the array antenna is inconsistent and depends on angle and distance parameters. Based on this, the relative amplitude γ of the signal received by the m-th receiving element relative to the reference element is... k,m Phase τ k,m It can be represented as

[0049]

[0050]

[0051] The signal received by the m-th array element can be represented as

[0052]

[0053] Where t is the sampling time, s k () represents the signal transmitted by the k-th source, n m () represents the noise signal received by the m-th array element. The near-field model received by the receiving array at this time is...

[0054]

[0055] Based on the established near-field model, training and testing datasets are generated.

[0056] S3. Construct the NFLnet network structure.

[0057] The purpose of this step is to construct an NFLnet network structure based on low-complexity neural networks, such as... Figure 3 As shown, the NFLnet network structure illustrated in this embodiment includes two fully connected (FC) layers and five CBR fusion layers, wherein the CBR fusion layers include the fusion of convolutional layers (Conv), batch normalization layers (BN), and activation function layers (ReLU).

[0058] It should be noted that this embodiment uses five CBR fusion layers to simultaneously satisfy sufficiently high network structure complexity and computational accuracy, while ensuring a certain computational load. Using too few layers might result in ineffective output. In practical applications, the number of CBR fusion layers can be adjusted based on considerations of complexity and computational load.

[0059] In one possible implementation, the raw data of the array-received signal is used as input to the NFLnet network. The raw data contains more information than the covariance matrix. The input data is then transmitted to the Conv layer after passing through a fully connected (FC) layer.

[0060] Taking any CBR fusion layer as an example:

[0061] Assume the input to the Conv layer is X, and the weight matrix is ​​W. Conv Then the output of the Conv layer is

[0062] X Conv =W Conv X

[0063] The output X of the Conv layer Conv As input to the BN layer, the output of the BN layer is

[0064]

[0065] Introduced To ensure stability, W BN It is the weight matrix of the BN layer, B BN It is the bias parameter matrix of the BN layer.

[0066] To reduce the computational complexity of the NFLnet network structure, the Conv layer and the BN layer are fused. Then, when the input is X, a new weight matrix and output matrix of the fused layer can be obtained.

[0067]

[0068] in, W is the output matrix of the fusion layer. F This is the weight matrix of the fusion layer.

[0069] Based on the above steps, the Batch Normalization (BN) layer is essentially omitted, and the convolutional kernel is changed without increasing the computational cost of the convolutional layer. The resulting CBR fusion layer output... As input to the ReLU layer, the signal passes through the ReLU nonlinear activation function f(x) = max(0,x) to obtain the output X. out =max(0,W F X+B F Finally, the output of the neural network is obtained through a fully connected (FC) layer.

[0070] S4. Obtain the output matrix G based on the NFLnet network structure.

[0071] Assume the number of samples input to the NFLnet network structure is M. B The input data is then represented as

[0072]

[0073] in, This involves splitting and reassembling the real and imaginary parts of the received data. The output matrix is ​​obtained after training with the NFLnet network architecture.

[0074]

[0075] Where C represents the set of complex numbers, M θ and M r It is the number of grids used to divide the angular and distance dimensions.

[0076] S5. Obtain the distance-angle two-dimensional spectrum based on the output matrix G and determine the loss function; minimize the loss function to train the NFLnet network structure.

[0077] First, the output matrix G is reshaped to obtain the estimated spectrum f of the distance-angle two-dimensional spectrum. sp (θ,r). The output M B ×M θ M r Reshaping the dimensional matrix into M B ×M θ ×M r M per group θ ×M r The matrix corresponds to a set of estimated distance-angle two-dimensional spectra f sp (θ,r).

[0078] Secondly, obtain the reference spectrum f of the distance-angle two-dimensional spectrum. rrf (θ,r).

[0079] During network training, the reference spectrum of the output distance-angle spectrum is given by a Gaussian function.

[0080]

[0081] Where, σ G The standard deviation determines the resolution of the spectral peaks and needs to be selected appropriately.

[0082] After obtaining the estimated spectrum and the reference spectrum, define the loss function.

[0083]

[0084] The NFLnet network structure is trained based on the criterion of minimizing the loss function.

[0085] The following is a specific verification example of the present invention:

[0086] (1) Given parameters:

[0087] Assuming the number of receiving array elements is 16, the element spacing is half a wavelength, the number of targets is 3, and the operating frequency is 2.45 GHz, the number of targets, angles, distances, and signal-to-noise ratios are randomly selected within the range of angle dimension (-90°, 90°), distance dimension (0.1, 10), and signal-to-noise ratio (-20 dB, 20 dB). According to the near-field signal model established in step two, 10,000 sets of raw received signal data y() are generated. At the same time, the corresponding reference spectrum is generated using a Gaussian function. Among them, 8,000 sets are used as training datasets and 2,000 sets are used as test datasets for training and testing the NFLnet network.

[0088] (2) Set network parameters:

[0089] The convolutional layer uses two 32×32 filters, the kernel size is 3, the standard deviation of the Gaussian function is 5, and the learning rate is 0.005. 8000 training sets are input into the NFLnet network and the network is trained.

[0090] (3) Input 2000 test sets into the trained NFLnet network and output the test results:

[0091] Taking three targets located at (-19.45°, 1.95m), (25.21°, 1.95m), and (30.01°, 3.75m) as examples, the angle-distance spectrum obtained by NFLnet testing is as follows: Figure 4 As shown, the estimated angle and distance values ​​are (-18.95°, 1.48m), (26.01°, 5.87m), and (29.82°, 4.22m), respectively. The mean square error of the angle estimation is approximately 0.33°, and the mean square error of the distance estimation is approximately 0.44m. Simulation results show that NFLnet can effectively and accurately perform joint angle-distance estimation for near-field targets.

[0092] (4) Construct a prototype under microwave anechoic chamber conditions, such as Figure 5 As shown:

[0093] A horn antenna is used as the transmitter to transmit a 2.45 GHz signal, and a 4-element uniform linear array is placed as the receiver at near-field distance. The signal received by the receiving antenna is input into the NFLnet network, and the output estimated spectrum is obtained as follows. Figure 6 As shown, the estimated angle and distance parameters are approximately (-44.6°, 0.92m), which matches the actual values ​​(-45°, 1.2m).

[0094] The above embodiments are only used to provide a detailed description of the technical solutions of this application. However, the descriptions of the above embodiments are only for the purpose of helping to understand the methods of the embodiments of the present invention and should not be construed as limiting the embodiments of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art should be covered within the protection scope of the embodiments of the present invention.

Claims

1. A method for estimating DOA and distance parameters for near-field target localization, characterized in that, include: An array of linearly distributed array elements is used to receive near-field target signals; wherein, the number M of linearly distributed array elements in the array is greater than the number K of near-field target signals to be measured; A near-field model is built, generating training and testing datasets; the near-field model generates multiple sets of raw received signal data. Meanwhile, the Gaussian function is used to generate the corresponding reference spectrum, which serves as the training and testing datasets. Construct the NFLnet network structure; the NFLnet network structure is obtained by training and testing the training dataset and the test dataset. Obtaining the output matrix based on the NFLnet network structure ; Based on the output matrix Obtain the distance-angle two-dimensional spectrum and determine the loss function; minimize the loss function to train the NFLnet network structure.

2. The method according to claim 1, characterized in that, The near-field model is in, Let t be the signal received by the m-th array element, and t be the sampling time. For the k-th near-field target signal, The noise signal received by the m-th array element; The relative amplitude of the signal received by the m-th array element Phase for For the spacing between array elements, These are the azimuth and distance from the near-field target signal k to the reference array element, respectively. This is the operating wavelength.

3. The method according to claim 1, characterized in that, The input is fed into the NFLnet network structure to obtain the output matrix. The data is the raw signal data received by the array.

4. The method according to claim 1, characterized in that, The NFLnet network structure includes two FC layers and five CBR fusion layers. The CBR fusion layers are fusions of convolutional layers, batch normalization layers, and activation function layers.

5. The method according to claim 4, characterized in that, The output of the CBR fusion layer ),in, This is the weight matrix after fusing the convolutional layer and the batch planning layer. The bias parameter matrix is ​​the result of fusing the convolutional layer and the batch planning layer. This is the input to the Conv layer.

6. The method according to claim 2, characterized in that, The distance-angle two-dimensional spectrum includes the estimated spectrum. and reference spectrum The estimated spectrum From the output 3D matrix reshaping to The reference spectrum is obtained. in, The standard deviation determines the resolution of the spectral peaks. The number of samples for inputting the NFLnet network structure.

7. The method according to claim 6, characterized in that, According to the estimated spectrum and the reference spectrum Obtain the loss function in, and It is the number of grids used to divide the angular and distance dimensions.

Images