CN110471026B - A Phase-Enhanced Method for Estimating Low-Elevation DOA for Meter-Wave Radar Targets - Google Patents

Abstract

The invention discloses a phase-enhanced metric wave radar target low-elevation DOA estimation method. The method includes normalizing an input data set; inputting the normalized input data set into a neural network model to obtain a network output data set ; Construct the objective function according to the network output data set and label data set, transfer the objective function in reverse, and update the network parameter set of the neural network model; according to the network parameter set, the test data set is enhanced to obtain a new data covariance matrix, A new test data set is obtained according to the new data covariance matrix; DOA estimation is performed on the new test data set to obtain the elevation angle of the target. In the DOA estimation method provided by the present invention, the neural network model is constructed according to the actual scene, which effectively solves the problem of mismatch between the actual signal model and the ideal far-field plane wave model existing in the physical-driven DOA estimation, thereby making the DOA estimation accuracy more accurate. high.

Description

A Phase-Enhanced Method for Estimating Low-Elevation DOA for Meter-Wave Radar Targets

technical field

The invention belongs to the technical field of radar, in particular to a phase-enhanced metric wave radar target low-elevation DOA estimation method.

Background technique

At present, the DOA estimation problem of low-elevation-angle targets is an important problem to be solved urgently in the field of meter-wave radar. This is because the beam of meter-wave radar is wide, and when detecting low-elevation-angle targets, there is the phenomenon of beam "hitting" and the multipath reflected by the ground. The direct signal reflected by the signal and the target is received by the radar from the main lobe direction, which leads to the direct signal characteristic no longer satisfying the simple far-field plane wave model characteristics, but a far-field plane wave model with amplitude and phase distortion. Amplitude and phase distortion is caused by multipath signals.

In response to this problem, a large number of researchers have explored some effective physics-driven DOA estimation methods in recent years. In the early days, the far-field plane wave model with amplitude and phase distortion was modeled as a two-point coherent source model, that is, the direct signal and the multipath signal were modeled as a pair of strongly coherent point source signals. In 1985, Shan et al. published a spatial smoothing-based decoherent multiple signal classification (Spatial smoothing multiple signal classification, referred to as SSMUSIC) method in the 34th issue of IEEE Transactions on Antennas and Propagation, pp. 806-811. , the method first achieves the decoherence processing of the coherent source through the sub-array smoothing method, and finally performs the super-resolution estimation. In 1998, Ziskind et al. published in IEEE Transaction on Acoustics Speech, and Signal Processing, Volume 36, Issue 10, pages 1553 to 1560, under the condition that the noise distribution characteristics are known a priori, that is, the noise distribution conforms to Gaussian white noise. A maximum-likelihood estimation method based on alternating projection is proposed. Through the optimization process of alternating projection, the DOA estimation of coherent sources is realized. In recent years, with the rapid development of deep learning technology, deep learning technology has not only been applied in fields such as image processing, but also gradually applied to the field of signal processing. In 1997, Zooghby et al. proposed a DOA estimation method using radial-basis-function network (RBFN) in IEEE Transactions on Antennas and Propagation, Vol. 45, No. 11, pages 1611 to 1617. The DOA estimation problem is simplified to a complex mapping relationship between covariance matrix data and output. By training RBFNN, the relationship between data covariance matrix data and DOA information is learned, and DOA estimation for coherent and incoherent sources is performed. The driven method is less computationally intensive and more accurate. In January 2019, Xiang et al. published a DOA estimation method based on unsupervised learning in the IETRadar Sonar & Navigation Journal, Volume 13, Issue 1. By learning the airspace data characteristics of the received data, the DOA information was inverted. The performance under multipath conditions is more accurate than the existing SSMUSIC method, and the calculation amount is smaller than that of the SSMUSIC method. In April 2019, Wang et al. proposed a DOA estimation method based on Support Vector Regression (SVR) in IEEE Signal Processing Letters, Vol. 26, No. 4, pages 642 to 646. The complex mapping relationship between the real part and imaginary part of the data and the DOA information is used to realize the DOA estimation, and the performance of the method is higher than that of the MUSIC of the rotation signal subspace.

Although the above-mentioned physical-driven super-resolution DOA estimation methods are the most commonly used DOA estimation methods at present, in the actual array environment, the multipath signal distribution is more complicated. The received multipath signal may not satisfy the far-field plane wave model, etc., resulting in low DOA estimation accuracy; the above-mentioned DOA estimation methods based on deep learning can be regarded as an end-to-end learning method, that is, to directly learn a certain part of the received data. The complex mapping relationship between these features and DOA is used to realize DOA estimation. Because the network model is simple, and there are no effective features and redundant features for the DOA estimation analysis data, the redundant features are used as the input of the network. On the one hand, it brings the network The complexity of training, on the other hand, the accuracy of DOA estimation is not high.

SUMMARY OF THE INVENTION

In order to solve the above problems existing in the prior art, the present invention provides a phase-enhanced metric wave radar target low-elevation DOA estimation method, the method comprising:

Obtain the phase feature of the input dataset and the phase feature of the label dataset;

Normalizing the phase feature of the input data set to obtain the normalized phase feature of the input data set;

Inputting the phase feature of the normalized input data set into the neural network model to obtain the phase feature of the network output data set;

According to the phase feature of the network output data set and the phase feature of the label data set, an objective function is constructed, the objective function is reversely transmitted by the first preset method, and the second preset method is used to update the objective function. The network parameter set of the neural network model;

The test data set is enhanced according to the network parameter set to obtain a new data covariance matrix, and a new test data set is obtained according to the new data covariance matrix;

Using a preset DOA estimation model to perform DOA estimation on the new test data set to obtain the elevation angle of the meter-wave radar target.

In an embodiment of the present invention, acquiring the phase feature of the input dataset and the phase feature of the label dataset includes:

constructing a receiving array, and obtaining an array receiving signal and an array steering vector according to the receiving array;

Obtain an input data set and a label data set according to the array received signal and the array steering vector, respectively;

respectively dividing the input data set and the label data set to obtain a divided input data set and a divided label data set;

The phase features of the upper triangular elements of the divided input data set and the divided label data set are respectively extracted to obtain the phase features of the input data set and the phase features of the label data set.

In an embodiment of the present invention, the input data set and the label data set are respectively divided to obtain a divided input data set and a divided label data set, including:

Obtain the input data of the current frame and the number of data divisions m, where m is a positive odd number;

Taking the current frame input data as the central frame, extracting (m-1)/2 frames of input data from the input data set forward and backward respectively to obtain forward input data and backward input data;

Obtain the divided input data set according to the current frame input data, the forward input data and the backward input data;

The divided label data set is obtained according to the input data of the current frame.

In an embodiment of the present invention, normalizing the phase feature of the input data set to obtain the normalized phase feature of the input data set includes:

obtaining the mean and standard deviation of the phase characteristics of the input data set;

Gaussian normalization is performed on the phase feature of the input data set according to the mean value and the standard deviation to obtain the phase feature of the normalized input data set.

In an embodiment of the present invention, the phase feature of the network output dataset is obtained by training the neural network model according to the phase feature of the normalized input dataset, including:

Build a four-layer deep neural network model;

The phase feature of the normalized input data set is input into the four-layer deep neural network model, and the phase feature of the network output data set is obtained by training.

In an embodiment of the present invention, the constructed objective function is:

表示网络输出数据集的相位特征，

表示标签数据集的相位特征。in,

represents the phase feature of the network output dataset,

Represents the phase features of the labeled dataset.

In an embodiment of the present invention, the first preset method includes a backpropagation method, and the second preset method includes an adaptive moment estimation method.

In an embodiment of the present invention, the network parameter set of the neural network model includes a network weight W and a network bias value b, wherein,

The network weight W is:

The network bias value b is:

分别为网络权值W的中间变量，

分别为网络偏置值b的中间变量。where α represents the learning rate, ∈ represents a fixed constant,

are the intermediate variables of the network weight W, respectively,

are the intermediate variables of the network bias value b, respectively.

In an embodiment of the present invention, the test data set is enhanced according to the network parameter set to obtain a new data covariance matrix, and a new test data set is obtained according to the new data covariance matrix, including:

Get the test dataset;

Dividing the test data set to obtain a divided test data set;

Extracting the phase feature of the upper triangular element of the divided test data set to obtain the phase feature of the test data set;

inputting the normalized phase feature of the test data set into the phase feature of the test data set enhanced by the four-layer deep neural network model;

Obtain the new data covariance matrix according to the phase feature of the enhanced test data set;

A new test data set is obtained according to the new data covariance matrix.

In an embodiment of the present invention, the preset DOA estimation model is:

表示估计的米波雷达目标的仰角。where R′ _i represents the new data covariance matrix in the new test dataset, a(θ _i ) represents the array steering vector, a ^H (θ _i ) represents the conjugate transpose of a(θ _i ),

Represents the estimated elevation angle of the meter-wave radar target.

Compared with the prior art, the beneficial effects of the present invention:

The phase-enhanced metric wave radar target low-elevation-angle DOA estimation method provided by the present invention, the neural network model of which is constructed according to the actual scene, effectively solves the problem between the actual signal model and the ideal far-field plane wave model existing in the physical-driven DOA estimation. Therefore, the DOA estimation accuracy of the metric wave radar target is higher.

The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

Description of drawings

1 is a schematic flowchart of a phase-enhanced metric-wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention;

2 is a schematic structural diagram of a four-layer deep neural network in a phase-enhanced metric wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention;

3 is a schematic diagram of a comparison result of angle measurement errors under different signal-to-noise ratio conditions of a phase-enhanced metric wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention;

4 is a schematic diagram of a comparison result of the feature fitting goodness of a phase-enhanced metric-wave radar target low-elevation-angle DOA estimation method provided by an embodiment of the present invention under different signal-to-noise ratio conditions;

5 is a schematic diagram of a comparison result of angle measurement errors under a conditional mismatch of signal-to-noise ratios with a phase-enhanced metric-wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention;

6 is a schematic diagram of a comparison result of the feature fitting goodness of a phase-enhanced metric-wave radar target low-elevation-angle DOA estimation method provided in an embodiment of the present invention under the conditional mismatch of signal-to-noise ratio;

7 is a schematic diagram of a comparison result of angle measurement errors under different phase error conditions of a phase-enhanced metric wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention;

8 is a schematic diagram of a comparison result of feature fitting goodness under different phase error conditions of a phase-enhanced metric wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention;

9 is a schematic diagram of a comparison result of angle measurement errors under a phase error condition mismatch of a phase-enhanced metric wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention;

10 is a schematic diagram of a comparison result of the feature fitting goodness of a phase-enhanced metric-wave radar target low-elevation-angle DOA estimation method provided by an embodiment of the present invention under the phase error condition mismatch;

11 is a schematic diagram of the measured data track in a phase-enhanced metric wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention;

12 is a schematic diagram of the angle measurement error of the two traditional methods provided by the embodiment of the present invention for processing the measured data;

13 is a schematic diagram of the angle measurement error of the measured data processing by a phase-enhanced metric wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention.

Detailed ways

The present invention will be described in further detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

Example 1

At present, radar DOA estimation methods include physics-driven super-resolution DOA estimation methods and deep learning-based DOA estimation methods. Among them, the physical-driven super-resolution DOA estimation method, with the complexity of the multipath signal distribution in the actual array environment, makes the multipath signals received by the array may not satisfy the far-field plane wave model, resulting in low DOA estimation accuracy. DOA estimation method, because the existing network model is simple, and there are no effective features and redundant features for DOA estimation analysis data, resulting in redundant features as the input of the network, on the one hand, it brings the complexity of network training, on the other hand DOA estimation accuracy is not high.

For the above problems, please refer to FIG. 1. FIG. 1 is a schematic flowchart of a phase-enhanced metric wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention. This embodiment provides a phase-enhanced metric wave radar target low-elevation DOA estimation method, the method comprising:

Step 1. Obtain the phase feature of the input dataset and the phase feature of the label dataset;

Step 2, normalizing the phase feature of the input data set to obtain the normalized phase feature of the input data;

Step 3, input the phase feature of the normalized input data set into the neural network model to obtain the phase feature of the network output data set;

Step 4, constructing an objective function according to the phase characteristics of the network output data set and the phase characteristics of the label data set, using the first preset method to perform reverse transmission processing on the objective function, and using the second preset method to update the network of the neural network model parameter set;

Step 5, performing enhancement processing on the test data set according to the phase characteristics of the network parameter set to obtain a new data covariance matrix, and obtaining a new test data set according to the new data covariance matrix;

Step 6, using the preset DOA estimation model to perform DOA estimation on the new test data set to obtain the elevation angle of the meter-wave radar target.

Specifically, in this embodiment, a neural network model suitable for this scene is obtained according to the phase feature of the input data set, and the objective function is constructed by training the phase feature of the network output data set of the neural network model and the phase feature of the label data set, The back-propagation processing of the objective function makes the phase characteristics of the network output data set approach the phase characteristics of the label data set. At this time, the optimal network parameter set corresponding to the neural network model is obtained at the same time, and then through the set neural network with the optimal network parameters. The model is trained on any test data set to obtain an enhanced new test data set, thereby obtaining the accurate elevation angle of the meter wave radar target from the enhanced new test data set.

The phase-enhanced metric wave radar target low-elevation DOA estimation method provided by this embodiment, the neural network model of which is constructed according to the actual scene, effectively solves the actual signal model and the ideal far-field plane wave model existing in the physical-driven DOA estimation. The mismatch problem makes the DOA estimation accuracy of the metric wave radar target higher.

Further, step 1 obtains the phase feature of the input dataset and the phase feature of the label dataset, including step 1.1, step 1.2, step 1.3, and step 1.4:

Step 1.1, construct a receiving array, and obtain array received signals and array steering vectors according to the receiving array.

Specifically, the receiving array constructed in the actual scene of this embodiment is a uniform linear array of K array elements, and the number of snapshots (number of samples) is L, then the array steering vector of this embodiment is designed as:

Among them, d represents the array element spacing, λ represents the signal wavelength, and θ represents the elevation angle of the metric wave radar target.

The multipath signal in the scene of this embodiment can be simplified as the amplitude and phase disturbance superimposed on the direct signal. If the amplitude and phase disturbance is τ, the array received signal in this embodiment is designed as:

x(t)=τ⊙a(θ)s(t)+n(t), t=1,...,L (2)

where n(t)=[n ₁ (t),n ₂ (t),...,n _K (t)] ^T represents the noise data vector, s(t) represents the signal source data, and a(θ) represents the above array Steering vector, ⊙ means dot product.

Step 1.2, according to the received signal of the array and the steering vector of the array, the input data set and the label data set are obtained respectively.

R_i表示多径阵列接收输入数据协方差矩阵，

表示多径阵列接收标签数据协方差矩阵，输入数据协方差矩阵R_i、标签数据协方差矩阵

均包括幅度特征和相位特征，其中，Specifically, for the meter-wave radar target in the flight state, the real elevation angle of the target is in a state of continuous change. In this embodiment, the meter-wave radar target is continuously detected N times, where N is an integer greater than 0, and the input data set {R ₁ ,R ₂ ,…,R _N } and label dataset

R _i represents the covariance matrix of the input data received by the multipath array,

Indicates that the multipath array receives the label data covariance matrix, the input data covariance matrix R _i , the label data covariance matrix

Both include amplitude characteristics and phase characteristics, where,

The input data covariance matrix R _i in each input dataset is designed as:

设计为：Label data covariance matrix in each label dataset

Designed to:

Step 1.3. Divide the input data set and the label data set respectively to obtain the divided input data set and the divided label data set.

Specifically, in order to obtain a more accurate DOA estimation in this embodiment, step 1.3 specifically includes step 1.3.1, step 1.3.2, step 1.3.3, and step 1.3.4:

Step 1.3.1. Obtain the input data of the current frame and the number of data divisions m, where m is a positive odd number.

中获取当前帧标签数据，比如前帧标签数据可以为

也可以为

同时确认数据划分数目m，该数据划分数据m决定了进行多少帧相位增强，本实施例数据划分数目m取值为正奇数，即m取值为1、3、5、……。Specifically, this embodiment obtains the current frame input data from the input data set {R ₁ , R ₂ ,...,R _N }, for example, the current frame input data may be R ₁ or R ₂ ; from the label data set

Get the current frame label data from , for example, the previous frame label data can be

can also be

At the same time, confirm the data division number m, which determines how many frames of phase enhancement are performed.

Step 1.3.2. Taking the input data of the current frame as the central frame, extract (m-1)/2 frames of input data from the input data set continuously forward and backward respectively to obtain forward input data and backward input data.

Specifically, in this embodiment, the number m of data divisions is taken as an example, and the input data of the current frame is taken as the central frame, and (m-1) is continuously extracted from the input data set {R ₁ , R ₂ ,...,R _N }. /2 frames of input data are used as forward input data, and (m-1)/2 frames of input data are continuously extracted from the input data set {R ₁ , R ₂ ,..., R _N } in the backward direction to ensure the division The subsequent input data is m frames, so as to divide the entire input data set {R ₁ , R ₂ ,...,R _N }. Specifically, for example, the number of data divisions m is 3. In the input data set {R ₁ , R ₂ ,..., R _N }, with R ₂ as the input data of the current frame, one frame of input data is continuously extracted forward, that is, the forward extracted The input data is R ₁ , and R ₁ is used as the forward input data, and then 1 frame of input data is continuously extracted backward, that is, the input data extracted backward is R ₃ , and R ₃ is used as the backward input data. For example, the number of data divisions m is 5, in the input data set {R ₁ , R ₂ ,...,R _N }, with R ₃ as the current frame input data, then continuously extract 2 frames of input data forward, that is, the forward extracted input data is R ₁ and R ₂ , take R ₁ and R ₂ as the forward input data, and continuously extract 2 frames of input data backward, that is, the backward extracted input data are R ₄ and R ₅ , and use R ₄ and R ₅ as the backward input data .

Step 1.3.3. Obtain a divided input data set according to the current frame input data, the forward input data, and the backward input data.

Specifically, in this embodiment, the input data of the current frame, the forward input data, and the backward input data obtained in step 1.3.2 are packaged to obtain the divided input data. For the entire input data set {R ₁ , R ₂ ,..., R _N } is divided to obtain the divided input data set {Φ ₁ ,Φ ₂ ,...,Φ _M }, where Φ _i indicates that the input data of a current frame, the forward input data and the backward input data are packaged to obtain the division subsequent input data. Specifically, for example, the number of data divisions m is 3, and R ₂ is the current frame input data, the obtained forward input data is R ₁ , the backward input data is R ₃ , the current frame input data R ₂ and the forward input data R ₁ are obtained. , backward input data R ₃ is packaged to obtain the divided input data Φ ₁ , specifically Φ ₁ ={R ₁ , R ₂ , R ₃ }, for example, the number of data divisions m is 5, in the input data set {R ₁ ,R ₂ ,...,R _N } with R ₃ as the current frame input data, the obtained forward input data is {R ₁ ,R ₂ }, the backward input data is {R ₄ ,R ₅ }, and the current frame input data R ₃ , the forward input data {R ₁ , R ₂ } and the backward input data {R ₄ , R ₅ } are packaged to obtain the divided input data Φ ₂ , specifically Φ ₂ ={R ₁ , R ₂ , R ₃ , R ₄ , R ₅ }. Among them, M≤N.

It should be noted that Φ ₁ or Φ ₂ obtained when the number of data divisions m is 3 or 5 in this embodiment is only for illustration, and the specific Φ ₁ or Φ ₂ is determined by the actual input data set.

Step 1.3.4. Obtain a divided frame label data set according to the current frame input data.

进行划分得到划分后的标签数据集

具体比如数据划分数目m为3，以R₂为当前帧输入数据，则该当前帧输入数据R₂对应的标签数据为

将标签数据

作为划分后的标签数据

再比如数据划分数目m为5，以R₃为当前帧输入数据，则该当前帧输入数据R₂对应的标签数据为

将标签数据

作为划分后的标签数据

其中，M≤N。Specifically, in this embodiment, the label data corresponding to the above-mentioned central frame is used as the divided label data, and the entire label data set is

Divide to get the divided label dataset

Specifically, for example, the number of data divisions m is 3, and R ₂ is the current frame input data, then the label data corresponding to the current frame input data R ₂ is:

label data

as divided label data

Another example is that the number of data divisions m is 5, and R ₃ is the current frame input data, then the label data corresponding to the current frame input data R ₂ is

label data

as divided label data

Among them, M≤N.

或

只是举例说明，具体

或

由实际标签数据集决定。It should be noted that, in this embodiment, when the number of data divisions m is 3 or 5, the

or

Just an example, specific

or

Determined by the actual labeled dataset.

Step 1.4: Extract the phase features of the upper triangular elements of the divided input data set and the divided label data set, respectively, to obtain the phase feature of the input data set and the phase feature of the label data set.

的上三角元素的相位特征，得到标签数据集的相位特征

Specifically, this embodiment extracts the phase features of the upper triangular elements of the divided input dataset {Φ ₁ ,Φ ₂ ,...,Φ _M } to obtain the phase features of the input dataset {Φ ₁ ,Φ ₂ ,..., φ _M }; in the same way, extract the divided label dataset

The phase features of the upper triangular elements of , get the phase features of the label dataset

The phase feature of the input data set obtained in this embodiment not only utilizes the feature of the input data of the current frame, but also fully utilizes the feature of the correlation between the input data of multiple frames, so that the DOA estimation accuracy is higher.

Further, step 2 normalizes the phase feature of the input data set to obtain the normalized phase feature of the input data set, including:

Obtain the mean and standard deviation of the phase features of the input dataset;

Gaussian normalization is performed on the phase features of the input data set according to the mean and standard deviation to obtain the normalized phase features of the input data set.

Specifically, in order to ensure the non-linearity in the subsequent neural network training process and retain the distribution characteristics of the input data set, the phase features {φ ₁ , φ ₂ ,...,φ _M } of the input data set are determined in the feature dimension in this embodiment. After Gaussian normalization, the mean of the feature dimension of the input data set is u ₁ and the standard deviation is σ ₁ , then the normalized phase feature corresponding to the input data set is designed as:

Normalize the phase feature of each input data in the phase feature {φ ₁ ,φ ₂ ,…,φ _M } of the input dataset to obtain the phase feature of the normalized input dataset

Further, step 3 trains the neural network model according to the phase characteristics of the normalized input data set to obtain the phase characteristics of the network output data set.

Specifically, this embodiment trains the neural network model according to the phase feature set of the normalized input data set obtained in step 2, and step 3 specifically includes steps 3.1 and 3.2:

Step 3.1. Build a four-layer deep neural network model.

Specifically, please refer to FIG. 2, which is a schematic structural diagram of a four-layer deep neural network in a phase-enhanced metric wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention. This embodiment adopts a four-layer deep neural network model. The specific structure of the network model is shown in Figure 2. The number of hidden layer nodes in the network is 1024, and the activation function in the network adopts the ReLU activation function. The specific ReLU activation function is expressed as f( x)=max(x,0).

Step 3.2: Input the phase feature of the normalized input data set into the four-layer deep neural network model, and train to obtain the phase feature of the network output data set.

为了使得训练过程中网络输出数据集的相位特征

和标签数据集的相位特征

相匹配，本实施例四层深度神经网络模型中的最后一层采用线性映射。Specifically, in this embodiment, the phase feature of the normalized input dataset is input into the four-layer deep neural network model constructed in step 3.1, and the phase feature of the network output dataset is obtained by training.

In order to make the phase characteristics of the network output data set during the training process

and phase features of the label dataset

Correspondingly, the last layer in the four-layer deep neural network model in this embodiment adopts linear mapping.

Further, step 4 constructs an objective function according to the phase characteristics of the network output data set and the phase characteristics of the label data set, uses the first preset method to perform reverse transmission processing on the objective function, and uses the second preset method to update the neural network model. set of network parameters.

Specifically, the objective function constructed in this embodiment is:

表示网络输出数据集的相位特征，

表示标签数据集的相位特征。in,

represents the phase feature of the network output dataset,

Represents the phase features of the labeled dataset.

标签数据集的相位特征

相匹配，同时利用第二预设方法更新四层深度神经网络模型的网络参数集。This embodiment uses the first preset method to perform reverse transmission processing on the objective function of formula (6), so that the network outputs the data set

Phase Features of Labeled Datasets

match, and at the same time update the network parameter set of the four-layer deep neural network model by using the second preset method.

The network parameter set of the four-layer deep neural network model in this embodiment includes a network weight W and a network bias value b, wherein,

The network weight W is designed as:

The network bias value b is designed as:

表示网络参数集求解过程中的中间变量，具体

的设计如下：where α represents the learning rate, ∈ represents a fixed constant,

Represents the intermediate variables in the process of solving the network parameter set, specifically

is designed as follows:

Among them, β ₁ is the exponential decay rate of the first-order moment of the network weight W and the network bias value, β ₂ is the exponential decay rate of the second-order moment of the network weight W and the network bias value, v _dW represents the network weight value The first-order moment of W, s _dW represents the second-order moment of the network weight W, v _db represents the first-order moment of the network bias value b, s _db represents the second-order moment of the network bias value b, v _dW , v _db , The initial values of s _dW and s _db are both 0.

By repeatedly training the four-layer deep neural network until the objective function Loss converges, the corresponding training network parameter set is the optimal network parameter of the four-layer deep neural network.

Preferably, the first preset method is a backpropagation (Backpropagation, BP for short) method, and the second preset method is an adaptive moment estimation (Adaptive moment estimation, Adam for short) method.

Further, step 5 performs enhancement processing on the test data set according to the network parameter set to obtain a new data covariance matrix, and obtains a new test data set according to the new data covariance matrix.

Specifically, step 5 of this embodiment includes step 5.1, step 5.2, step 5.3, step 5.4, step 5.5, and step 5.6:

Step 5.1. Obtain the test data set.

Specifically, in this embodiment, a test data set is obtained from an actual scene, and the construction of the test data set is the same as the construction of the input data set in formula (3).

Step 5.2: Divide the test data set to obtain a divided test data set.

Specifically, this embodiment divides the test data set according to the division method of the input data set in the above step 1.3 to obtain the divided test data set.

Step 5.3: Extract the phase feature of the upper triangular element of the divided test data set to obtain the phase feature of the test data set.

Specifically, in this embodiment, according to the extraction method of the divided input data set and the divided label data set in the above step 1.3, the phase characteristics of the upper triangular elements of the divided test data set are extracted to obtain the phase characteristics of the test data set. .

Step 5.4: Input the phase feature of the normalized test data set into the four-layer deep neural network model, and train the phase feature of the enhanced test data set.

Specifically, in this embodiment, before the phase feature of the test data set is input into the four-layer deep neural network model, the phase feature of the test data set is first normalized. The normalized processing method, the phase feature of the normalized test data set is designed as:

表示步骤5.3得到的测试数据集的相位特征，u₂表示通过统计标签数据集的相位特征的均值，σ₂表示通过统计标签数据集的相位特征的标准差。in,

represents the phase feature of the test data set obtained in step 5.3, u ₂ represents the mean value of the phase feature of the statistical label dataset, and σ ₂ represents the standard deviation of the phase feature of the statistical label dataset.

Step 5.5, obtaining a new data covariance matrix according to the phase feature of the enhanced test data set.

作为四层深度神经网络的输入，经四层深度神经网络训练后得到测试数据集增强的相位特征φ_i′，将原测试数据集中的幅度信息ρ_i、对角元素Γ_i和增强的相位特征φ′_i，重组得到新的数据协方差矩阵R′_i。Specifically, this example is obtained in step 5.4

As the input of the four-layer deep neural network _, the enhanced phase feature φ _i ′ of the test data set is obtained after the training of the four-layer deep neural network _. φ′ _i , recombine to obtain a new data covariance matrix R′ _i .

Step 5.6: Obtain a new test data set according to the new data covariance matrix.

Specifically, in this embodiment, the phase characteristics of the normalized test data set are respectively subjected to the training process in step 5.5 above to obtain a new data covariance matrix R′ _i corresponding to each test data. The variance matrix R′ _i is constructed to obtain a new test data set {R′ ₁ , R′ ₂ ,...,R′ _N }.

Further, step 6 uses a preset DOA estimation model to perform DOA estimation on the new test data set to obtain the elevation angle of the meter-wave radar target.

Specifically, in this embodiment, DOA estimation is performed on each new data covariance matrix in the new test data set. Specifically, DOA estimation is performed on each new data covariance matrix by using a preset DOA estimation model. This embodiment The preset DOA estimation model is designed as:

where R′ _i represents the new data covariance matrix in the new test dataset, a(θ _i ) represents the array steering vector, and a ^H (θ) represents the conjugate transpose of a(θ _i ).

In this embodiment, the elevation angle of each meter-wave radar target in the test data set is obtained by formula (12).

To sum up, compared with the physically driven DOA estimation method, the DOA estimation method provided in this embodiment effectively solves the mismatch between the actual signal model and the ideal far-field plane wave model, so that the DOA estimation accuracy is higher; For the existing data-driven DOA estimation method (such as the DOA estimation method based on SVR), the DOA estimation method provided in this embodiment not only utilizes the characteristics of the input data of the current frame, but also makes full use of the difference between the input data of multiple frames. The characteristics of the correlation make DOA estimation more accurate.

In order to verify the effectiveness of the phase-enhanced metric wave radar target low-elevation DOA estimation method provided by this application, this embodiment is further described by the following simulation experiments:

Simulation environment

The data generation and processing of the experiment was completed on MATLAB2017a, and the training of the four-layer deep neural network model was completed on Python 3.5. This embodiment verifies the effectiveness of the method through four simulation scenarios.

Experimental Scenario 1

The array structure in the scene is a uniform linear array with 21 array elements, the wavelength λ is 1 meter, the array element spacing d is half wavelength, the number of snapshots is 42, the signal-to-noise ratio range is -10dB to -2dB, and the signal-to-noise ratio sampling interval is 2dB, the target elevation angle ranges from 1° to 5°, and the phase error in the array is 20°. Please refer to FIG. 3 and FIG. 4. FIG. 3 is a schematic diagram of the comparison results of angle measurement errors under different signal-to-noise ratio conditions of a phase-enhanced metric wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention. A schematic diagram of the comparison results of the feature fitting goodness of a phase-enhanced metric-wave radar target low-elevation DOA estimation method under different signal-to-noise ratio conditions provided by the embodiment of the invention, the abscissa of FIG. 3 represents the signal-to-noise ratio, and the ordinate represents the measurement. Angle error, the abscissa in Figure 4 represents the signal-to-noise ratio, and the ordinate represents the feature fit goodness. The angle measurement error in this embodiment is the root mean square error. It can be seen from Figures 3 and 4 that the higher the signal-to-noise ratio, the higher The larger the number of frames (number of data divisions), the better the feature fit and the smaller the angle measurement error.

Experimental Scenario 2

The array structure in the scene is a uniform linear array with 21 array elements, the wavelength λ is 1 meter, the array element spacing d is half wavelength, the number of snapshots is 42, the signal-to-noise ratio of the tag data set ranges from -10dB to 0dB, and the test data set The signal-to-noise ratio ranges from -11dB to -1dB, the signal-to-noise ratio sampling interval is 2dB, the target elevation angle ranges from 1° to 5°, and the phase error existing in the array is 20°. Please refer to FIG. 5 and FIG. 6 . FIG. 5 is a schematic diagram of the comparison result of the angle measurement error under the mismatch of the signal-to-noise ratio condition of a phase-enhanced metric wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention, and FIG. 6 is the present invention. A schematic diagram of the comparison results of the feature fitting goodness of a phase-enhanced metric wave radar target low-elevation DOA estimation method under the conditional mismatch of the signal-to-noise ratio provided by the embodiment of the invention, the abscissa of FIG. 5 represents the signal-to-noise ratio, and the ordinate represents the measurement. Angle error, the abscissa in Figure 6 represents the signal-to-noise ratio, and the ordinate represents the feature fit goodness. In this embodiment, the angle measurement error is the root mean square error. It can be seen from Figures 5 and 6 that the higher the signal-to-noise ratio, the The larger the number of frames, the better the feature fit, and the smaller the angle measurement error. At the same time, it can be seen from FIG. 6 that when the number of frames increases to a certain extent, because the correlation between frames decreases, the network feature fitting performance tends to be saturated. Strong generalization ability.

Experimental Scenario 3

The array structure in the scene is a uniform linear array with 21 array elements, the wavelength λ is 1 meter, the array element spacing d is half wavelength, the number of snapshots is 42, the signal-to-noise ratio is 0dB, the target elevation angle range is 1°~5°, the array The range of the phase error that exists in 5% to 20% is 5%, and the error sampling interval is 5%. Please refer to FIG. 7 and FIG. 8. FIG. 7 is a schematic diagram of the comparison results of angle measurement errors under different phase error conditions of a phase-enhanced metric wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention, and FIG. 8 is the present invention. A schematic diagram of the comparison results of the feature fit goodness of a phase-enhanced metric wave radar target low-elevation DOA estimation method under different phase error conditions provided by the embodiment, the abscissa of FIG. 7 represents the phase error, and the ordinate represents the angle measurement error, The abscissa in Fig. 8 represents the phase error, and the ordinate represents the feature goodness of fit. The angle measurement error in this embodiment is the root mean square error. It can be seen from Fig. 7 that the larger the number of frames, the smaller the angle measurement error. Compared with the existing method, the angle measurement accuracy of the present application has been effectively improved. It can be seen from FIG. 8 that the larger the number of frames, the better the feature fitting goodness. Compared with the existing method, the present application network feature fitting performance is better. on existing methods.

Experimental Scenario 4

The array structure in the scene is a uniform linear array with 21 array elements, the wavelength λ is 1 meter, the array element spacing d is half wavelength, the number of snapshots is 42, the signal-to-noise ratio is 0dB, and the phase error range of the tag data set is 5% to 20 %, the phase error range of the test data set is 3% to 23%, the sampling interval of the phase error is 5%, and the target elevation angle range is 1° to 5°. FIG. 9 is a schematic diagram of a comparison result of angle measurement errors under a phase error condition mismatch of a phase-enhanced metric wave radar target low-elevation DOA estimation method provided by an embodiment of the present invention, and FIG. 10 is a phase-enhanced method provided by an embodiment of the present invention. Schematic diagram of the comparison results of the characteristic goodness-of-fit of the metric wave radar target low-elevation DOA estimation method under the mismatch of the phase error condition, the abscissa of Fig. 9 represents the phase error, the ordinate of Fig. , the ordinate represents the feature goodness of fit, and the angle measurement error in this embodiment is the root mean square error. It can be seen from FIG. 9 that the larger the number of frames, the smaller the angle measurement error. Compared with the existing method, this application measures The angular accuracy has been effectively improved. It can be seen from Figure 10 that the larger the number of frames, the better the feature fitting goodness. Compared with the existing method, the network feature fitting performance of the present application is better than the existing method. 10 It can be seen that when the number of frames increases to a certain extent, because the correlation between frames decreases, the network fitting performance tends to be saturated. performance.

In order to verify the practicability of the DOA estimation method provided in this application, this embodiment processes the measured data of a meter wave radar at a certain position. In the actual scene, the 3dB beam width of the radar is about 5°. The position environment where the target is located is very harsh, and there are many objects such as trees and hills. Please refer to FIG. 11. FIG. Schematic diagram of the measured data track in the low-elevation DOA estimation method of the meter-wave radar target.

Please refer to FIG. 12 and FIG. 13 . FIG. 12 is a schematic diagram of the angle measurement results of the two traditional methods provided by the embodiment of the present invention processing the measured data. Schematic diagram of the angle measurement error of the elevation DOA estimation method for the processing of the measured data. For the two traditional methods in Figure 12, the DBF method and the SSMUSIC method are used to process a certain route data offline, as can be seen from Figure 12 , both the DBF method and the SSMUSIC method have failed, because after the smoothing of these two methods, all feature vectors are added to the original data as noise vectors and multipath signals, so that the test data set and the label data set do not match, which in turn leads to The angle measurement error is large. For the present application, after the phase feature enhancement is performed by the four-layer deep neural network, it can be seen from Figure 13 that the angle measurement error is almost uniformly distributed within the range of ±0.5°, and the measurement performance is significantly improved. Taking the angle measurement error of 0.3° as the evaluation standard of the effective point trace, the effective point trace after the enhancement of the four-layer deep neural network is increased to 95.6%, and the performance is very significant. This application has high reliability.

The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be considered that the specific implementation of the present invention is limited to these descriptions. For those of ordinary skill in the technical field of the present invention, without departing from the concept of the present invention, some simple deductions or substitutions can be made, which should be regarded as belonging to the protection scope of the present invention.

Claims (8)

1. A phase-enhanced meter-wave radar target low elevation DOA estimation method is characterized by comprising the following steps:

acquiring phase characteristics of an input data set and phase characteristics of a tag data set;

normalizing the phase characteristics of the input data set to obtain the phase characteristics of the normalized input data set;

inputting the phase characteristics of the normalized input data set into a neural network model to obtain the phase characteristics of a network output data set;

constructing an objective function according to the phase characteristics of the network output data set and the label data set, performing reverse transmission processing on the objective function by using a first preset method, and updating a network parameter set of the neural network model by using a second preset method;

according to the network parameter set, carrying out enhancement processing on a test data set to obtain a new data covariance matrix, and according to the new data covariance matrix, obtaining a new test data set;

performing DOA estimation on the new test data set by using a preset DOA estimation model to obtain an elevation angle of the meter-wave radar target;

wherein obtaining phase characteristics of the input data set and the tag data set comprises:

constructing a receiving array, and obtaining an array receiving signal and an array guide vector according to the receiving array;

respectively obtaining an input data set and a tag data set according to the array receiving signals and the array steering vectors;

dividing the input data set and the label data set respectively to obtain a divided input data set and a divided label data set;

respectively extracting the phase characteristics of the upper triangular elements of the divided input data set and the divided label data set to obtain the phase characteristics of the input data set and the phase characteristics of the label data set;

wherein, divide said input data set and label data set separately and get the input data set after dividing, label data set after dividing, including:

acquiring input data of a current frame and the data division number m, wherein m is a positive odd number;

respectively and continuously extracting (m-1)/2 frames of input data from the input data set forwards and backwards by taking the current frame of input data as a central frame to obtain forward input data and backward input data;

obtaining the divided input data set according to the current frame input data, the forward input data and the backward input data;

and obtaining the divided label data set according to the current frame input data.

2. The method of phase-enhanced meter wave radar target low elevation DOA estimation according to claim 1, wherein normalizing the phase signature of the input data set to obtain a normalized phase signature of the input data set comprises:

obtaining a mean and a standard deviation of phase characteristics of the input data set;

and carrying out Gaussian normalization processing on the phase characteristics of the input data set according to the mean value and the standard deviation to obtain the phase characteristics of the normalized input data set.

3. The phase-enhanced meter wave radar target low elevation DOA estimation method according to claim 1, wherein inputting the phase characteristics of the normalized input data set to a neural network model to obtain the phase characteristics of a network output data set comprises:

constructing a four-layer deep neural network model;

and inputting the phase characteristics of the normalized input data set into the four-layer deep neural network model, and training to obtain the phase characteristics of the network output data set.

4. The phase-enhanced meter-wave radar target low elevation DOA estimation method according to claim 1, wherein the constructed objective function is:

wherein,

representing the phase characteristics of the network output data set,

representing the phase characteristics of the tag data set.

5. The phase-enhanced meter-wave radar target low elevation DOA estimation method according to claim 1, wherein the first preset method comprises a back propagation method and the second preset method comprises an adaptive moment estimation method.

6. The phase-enhanced meter-wave radar target low elevation DOA estimation method according to claim 1, wherein the set of network parameters of the neural network model includes a network weight W and a network bias value b, wherein,

the network weight W is:

the network bias value b is:

where α represents a learning rate, ∈ represents a fixed constant,

respectively, are intermediate variables of the network weight W,

respectively, intermediate variables of the network bias value b.

7. The method of phase-enhanced metric wave radar target low elevation DOA estimation according to claim 3, wherein the enhancing the test dataset according to the network parameter set to obtain a new data covariance matrix, and the obtaining the new test dataset according to the new data covariance matrix comprises:

acquiring a test data set;

dividing the test data set to obtain a divided test data set;

extracting the phase characteristics of the upper triangular elements of the divided test data set to obtain the phase characteristics of the test data set;

inputting the phase characteristics of the test data set subjected to normalization processing into the four-layer deep neural network model to obtain the phase characteristics of the enhanced test data set;

obtaining the new data covariance matrix according to the phase characteristics of the enhanced test data set;

and obtaining a new test data set according to the new data covariance matrix.

8. The phase-enhanced meter-wave radar target low elevation angle DOA estimation method according to claim 1, wherein the preset DOA estimation model is:

wherein R'_iRepresents a new data covariance matrix, a (θ), in a new test data set_i) Representing array steering vectors, a^H(θ_i) Denotes a (theta)_i) The conjugate transpose of (a) is performed,

representing the estimated elevation angle, theta, of a meter-wave radar target_iRepresenting the true elevation angle of the meter-wave radar target.

Images