Cross-platform phase-enhanced super-resolution DOA estimation method

Abstract

The invention discloses a cross-platform phase-enhanced super-resolution DOA (direction of arrival) estimation method, which belongs to the field of direction of arrival estimation, and comprises the following steps: constructing a DNN (deep neural network) supervised deep neural network architecture, and extracting phase difference distribution characteristics; building a radar system array of the training platform; the method comprises the following steps: training a DNN supervised deep neural network on a training platform, using an original noise sampling covariance matrix and a corresponding clean sampling covariance matrix, extracting phase distribution characteristics without wavelength errors through a compensation phase label, and inputting the phase distribution characteristics into a defined DNN for training; and placing the trained DNN supervised deep neural network on different platforms to carry out DOA estimation, and completing cross-platform phase-enhanced super-resolution DOA estimation through a result output by the DNN supervised deep neural network and a cross-platform transformation formula. According to the cross-platform phase enhancement super-resolution DOA estimation method provided by the invention, the negative influence of the phase disturbance of a multipath signal on DOA estimation is effectively reduced, and the estimation problem of cross-platform frequency change is effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of direction of arrival estimation, and in particular to a cross-platform phase-enhanced super-resolution DOA estimation method. Background Art

[0002] Deep learning is a branch of machine learning that analyzes and learns data by mimicking the structure and function of neural networks in the human brain. Deep learning models consist of multiple layers of neural networks, each containing many neurons, each connected to neurons in the previous and next layers. Data is passed between layers through these connections, and each layer is responsible for extracting increasingly advanced features from the input data. Through the back-propagation algorithm, deep learning models can automatically adjust weights and biases to minimize the error between predicted output and actual output, thereby achieving efficient processing and prediction of complex data.

[0003] In the field of radar, the application of deep learning has made significant progress. The application of deep learning in the field of radar is mainly reflected in target detection and tracking, signal processing and waveform recognition. By utilizing the powerful learning ability of deep neural networks, radar systems can achieve more efficient and accurate target recognition and positioning, and improve the accuracy and stability of target tracking. In addition, deep learning can also optimize radar signal processing methods, such as automatically suppressing clutter and extracting effective target information, thereby improving radar performance. In terms of waveform recognition, deep learning can help radar systems automatically identify the characteristics of complex targets, improve the ability to distinguish different targets, and further enhance the radar's perception and decision-making capabilities.

[0004] In harsh terrain environments, when radar tracks low-angle targets, the distribution characteristics of the received signal depend not only on the direct signal, but also on the multipath signals including mirror reflection signals and diffuse reflection signals. In general, multipath signals will cause amplitude and phase disturbances to the direct signal. In a multipath environment, due to the ambiguity of the far-field plane wave characteristics of the direct signal, the performance of existing super-resolution methods based on far-field plane wave basis (such as digital beamforming (DBF), multiple signal classification (MUSIC) and maximum likelihood (ML)) is reduced, and they are very sensitive to the size of unknown disturbances. Therefore, restoring the far-field plane wave characteristics of the direct signal is the only way to improve the performance of the physical drive method. In order to solve this problem, the present invention introduces a phase enhancement technology, which is based on a supervised deep neural network (DNN) method to improve the accuracy of the direction of arrival (DOA) estimation. This technology also focuses on proposing a cross-platform phase enhancement super-resolution DOA estimation technology to cope with the DOA estimation problem under different platform environments. Summary of the invention

[0005] The purpose of the present invention is to provide a cross-platform phase-enhanced super-resolution DOA estimation method to solve the problems mentioned in the above background technology.

[0006] To achieve the above object, the present invention provides a cross-platform phase-enhanced super-resolution DOA estimation method, comprising the following steps:

[0007] S1. Build a DNN supervised deep neural network architecture to extract phase difference distribution features. Through the training of the neural network, the phase enhancement result of the input phase disturbance signal can be obtained to achieve the de-disturbance effect;

[0008] S2, build radar system array for training platform;

[0009] S3, training the DNN supervised deep neural network on the training platform, using the original noise sampling covariance matrix and the corresponding clean sampling covariance matrix, extracting the phase distribution characteristics without wavelength error by compensating the phase label, and inputting it into the defined DNN for training;

[0010] S4. Place the trained DNN supervised deep neural network on different platforms for DOA estimation, and complete the cross-platform phase-enhanced super-resolution DOA estimation through the output results of the DNN supervised deep neural network and the cross-platform transformation formula.

[0011] Preferably, the DNN supervised deep neural network architecture in S1 includes 4 layers.

[0012] Preferably, the specific content of S2 is as follows:

[0013] For a uniform linear array with M isotropic elements, where the spacing d between each element is equal to half the wavelength, and the array contains L snapshots, the received vector is:

[0014]

[0015] The steering vector a(θ,λ) is:

[0016]

[0017] where θ d represents the direction of arrival (DOA) of the direct signal; ρ i represents the attenuation factor corresponding to the i-th multipath; n represents the number of unknown multipath signals; s(t) represents the complex echo signal; n(t) has a mean of 0 and a variance of σ 2 Gaussian white noise; in formula (1) When, in fact For a(θ d ,λ) causes amplitude and phase disturbances; therefore, the signal model is described as follows:

[0018] y(t)=Γ⊙a(θ d ,λ)s(t)+n(t) (3)

[0019] The disturbance term Γ is:

[0020] Γ=[τ1,τ2,…,τ M ] T (4).

[0021] Preferably, the training process of S3 is as follows:

[0022] S31, initialize network parameters W k and b k ;

[0023] S32, obtaining a phase tag for compensating for wavelength error

[0024] Assume that the phase information of the labeled training dataset is represented as Among them and and Representing the input data and the corresponding known data, the phase label is compensated according to the formula:

[0025]

[0026] S33, normalize the training data set to obtain parameters μ1 and σ1;

[0027] For the phase information of the compensated training data set where Φ1 and Representing the input data and the corresponding known output data, the normalized data set is based on the formula:

[0028]

[0029] S34, will Randomly divide into several batches and perform forward propagation;

[0030] S35. According to the error of the output layer, the BP strategy and AdaM method are used to adjust the network parameters W. k and b k Make fine adjustments;

[0031] The optimization formula is as follows:

[0032]

[0033] Where φ′1 is the output data, For known output data

[0034] S36. Repeat S33 to S35 by learning and adjusting the network architecture until the error of the objective function converges.

[0035] Preferably, in S4, the wavelength error caused by the cross-platform condition is corrected by the following method: the structure and parameters of the DNN are fixed during the cross-platform test phase, and the phase of the input noise covariance matrix is The corresponding enhanced phase φ′2 of the network is obtained; the enhanced phase is corrected:

[0036]

[0037] The sampling covariance matrix is ​​reconstructed according to the corrected enhanced phase Φ information, and DOA estimation is achieved using DBF, MUSIC and ML.

[0038] Preferably, the correction principle in S4 is as follows:

[0039] S411, Φ(λ1) represents the phase information of the training data at wavelength λ1. In the training stage, the wavelength is λ1, and the Φ(λ1)×λ1 operation is performed:

[0040]

[0041] There are M elements, and the mth element represents the phase information of the signal received by the mth array element in the array. represents the phase disturbance of the received signal of the mth array element in the array; represents the phase disturbance of the received signal of the second array element in the array;

[0042] S412. In the test phase, the corresponding enhanced phase φ′2 of the DNN supervised deep neural network is obtained, and the wavelength is λ2:

[0043]

[0044] S413, perform phase correction according to the enhanced phase φ′2:

[0045]

[0046] The network enhancement data φ2′ is adapted to the wavelength λ2 to obtain Φ, where Φ represents the final enhanced phase. The phase information Φ obtained at this time is the cross-platform compensation result for suppressing multipath. A high-resolution estimation result is obtained by signal reconstruction and using traditional DOA estimation methods.

[0047] Therefore, the present invention adopts the above-mentioned cross-platform phase-enhanced super-resolution DOA estimation method, which has the following beneficial effects:

[0048] (1) Only the phase feature is trained to improve the DOA estimation accuracy. That is, in the present invention, only the phase feature is trained, instead of training the amplitude feature and the phase feature together, or training the real part feature and the imaginary part feature together;

[0049] (2) A phase enhancement framework based on feature-to-feature learning is established to reduce the negative impact of features;

[0050] (3) The cross-platform phase-enhanced DOA estimation problem is solved. Even if the wavelength of the training platform is inconsistent with the wavelength of the test platform, the present invention can still obtain phase-enhanced super-resolution DOA estimation.

[0051] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a schematic diagram of a flow chart of an embodiment of the present invention;

[0053] Figure 2 A block diagram of a method implementation of an embodiment of the present invention;

[0054] Figure 3 A schematic diagram of a DNN network structure constructed for an embodiment of the present invention;

[0055] Figure 4 A schematic diagram of the technical simulation experiment steps of an embodiment of the present invention;

[0056] Figure 5 It is a schematic diagram of comparison results of phase angle details of signal reconstruction applied to each array element by the cross-platform phase enhancement signal technology in scenario 1 of an embodiment of the present invention;

[0057] Figure 6 It is a schematic diagram of the comparison results of the angle measurement error of the cross-platform phase-enhanced super-resolution DOA estimation method under different estimation methods in Scenario 1 of an embodiment of the present invention;

[0058] Figure 7 It is a schematic diagram of comparison results of signal phase angle detail features of each array element when the cross-platform phase enhancement signal technology is applied in Scenario 2 of an embodiment of the present invention;

[0059] Figure 8 It is a schematic diagram of the comparison results of the angle measurement error of the cross-platform phase-enhanced super-resolution DOA estimation method under different estimation methods in Scenario 2 of an embodiment of the present invention;

[0060] Fig. 9 It is a schematic diagram of comparison results of signal phase angle detail features of each array element when the cross-platform phase enhancement signal technology is applied in Scenario 3 of an embodiment of the present invention;

[0061] Fig.10It is a schematic diagram showing the comparison results of angle measurement errors of the cross-platform phase-enhanced super-resolution DOA estimation method under different estimation methods in Scenario 3 of an embodiment of the present invention;

[0062] Fig.11 It is a schematic diagram of comparison results of signal phase angle detail features of each array element when the cross-platform phase enhancement signal technology is applied in Scenario 4 of an embodiment of the present invention;

[0063] Fig.12 It is a schematic diagram of the comparison results of the angle measurement error of the cross-platform phase-enhanced super-resolution DOA estimation method under different estimation methods in Scenario 4 of an embodiment of the present invention;

[0064] Fig.13 4 is a comparison result of the simulation experiment of the embodiment of the present invention. DETAILED DESCRIPTION

[0065] Example

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] Reference Figure 1-Figure 4 The present invention discloses a cross-platform phase-enhanced super-resolution DOA estimation method, comprising the following steps:

[0068] S1. Build a DNN supervised deep neural network architecture to extract the phase difference distribution characteristics. Through the training of this neural network, the phase enhancement result of the input phase disturbance signal can be obtained to achieve the de-disturbance effect.

[0069] S2. Build a radar system array for the training platform.

[0070] Consider a uniform linear array (ULA) with M isotropic elements, where the spacing d between each element is equal to half the wavelength, and the array contains L snapshots. The received vector is:

[0071]

[0072] The steering vector a(θ,λ) is:

[0073]

[0074] where θ d Indicates the direction of arrival (DOA) of the direct signal.i is the attenuation factor corresponding to the i-th multipath. n represents the number of unknown multipath signals. s(t) is the complex echo signal. n(t) has a mean of 0 and a variance of σ 2 Gaussian white noise. In the analysis formula (1) When, in fact will be a(θ d ,λ) causes amplitude and phase disturbances. Therefore, the signal model can be re-described as follows:

[0075] y(t)=Γ⊙a(θ d ,λ)s(t)+n(t) (3)

[0076] The disturbance term Γ is:

[0077] Γ=[τ1,τ2,…,τ M ] T (4)

[0078] Due to the existence of the perturbation term Γ, the amplitude and phase distribution of a(θ) are distorted. In order to alleviate the negative impact of Γ on DOA estimation, a new phase enhancement method using DNN is proposed to learn the mapping relationship between input data and DOA, and the echo signal feature a(θ) may be distorted due to the λ switching between different platforms. d ,λ), which affects the direction of arrival angle (DOA) estimation, and a corresponding solution is proposed.

[0079] The data is integrated as follows: Since the signal type considered in this embodiment is a far-field plane wave, the array elements in each column maintain the same phase difference distribution. Therefore, the data of the synthetic size represented by y(t) can be expressed as:

[0080]

[0081] n p are independent of each other and therefore cannot be integrated. Therefore, the composite amplitude y(t) is k times stronger than the single column x(t). We can conclude that the phase of the received data determines the amplitude of the composite size data, that is, the signal-to-noise ratio. The estimated expression of the DBF method is as follows:

[0082]

[0083] According to formula (16), it can be observed in this embodiment that for the DBF algorithm, λ is a known quantity, the steering vector a(θ,λ) is determined by the angle θ, and θ determines the phase of a(θ,λ). Therefore, X or The phase of must be accurate so that DOA can be accurately estimated. Therefore, this embodiment will obtain a more accurate phase through the DNN method.

[0084] S3. Train the DNN supervised deep neural network on the training platform, use the original noise sampling covariance matrix and the corresponding clean sampling covariance matrix, extract the phase distribution characteristics without wavelength error by compensating the phase label, and input it into the defined DNN for training.

[0085] In the training phase, the original noise sampling covariance matrix and the corresponding clean sampling covariance matrix are used. The phase difference distribution features are extracted and input into the defined DNN for training. This technology designs a 4-layer DNN architecture. Considering that the mapping relationship between input data and output data is unknown, it is necessary to optimize the architecture of the DNN. p and b p (p∈1,2,3,4) represents the parameters of the DNN, i.e., the weights and biases of the pth layer. After DNN learning, the phase of the noise sampling covariance matrix is ​​enhanced. The output of the DNN, i.e., the enhanced phase, is used to reconstruct a new sampling covariance matrix with the original amplitude and DOA estimation.

[0086] S31, initialize network parameters W k and b k ;

[0087] S32, obtaining a phase tag for compensating for wavelength error

[0088] Assume that the phase information of the labeled training dataset is represented as Among them and and Representing the input data and the corresponding known data, the phase label is compensated according to the formula:

[0089]

[0090] S33, normalize the training data set to obtain parameters μ1 and σ1;

[0091] For the phase information of the compensated training data set where Φ1 and Representing the input data and the corresponding known output data, the normalized data set is based on the formula:

[0092]

[0093] S34, will Randomly divide into several batches and perform forward propagation;

[0094] Assume that p th The input is x p , then the output h p It can be expressed as follows:

[0095]

[0096] Where f(·) is an activation function. In order to prevent the gradient vanishing problem and maintain the sparsity of neurons, we use the Rectified Linear Unit (ReLU) function. The function can be expressed as follows:

[0097] f(z)=max(z,0) (15)

[0098] Where z represents the independent variable of the function. In order to match the data distribution between the output of the third hidden layer and the unnormalized label data, a linear mapping layer is used. The output vector φ′1 can be formulated as follows:

[0099] φ′1=W 4 x4+b 4 (16)

[0100] S35. According to the error of the output layer, the BP strategy and AdaM method are used to adjust the network parameters W. k and b k Make fine adjustments;

[0101] The optimization formula is as follows:

[0102]

[0103] Where φ′1 is the output data, is the known output data. In order to accelerate the convergence of the training phase, the entire training data set is randomly divided into several batches with a batch size of D. Formula (7) can be rewritten as:

[0104]

[0105] S36, repeating S33 to S35 by learning and adjusting the network architecture until the error of the objective function converges;

[0106] DOA estimation is achieved using DBF (digital beam forming, DBF), MUSIC (multiple signal classification, MUSIC) and ML (maximum likelihood, ML).

[0107] S4. Place the trained DNN supervised deep neural network on different platforms for DOA estimation, and complete the cross-platform phase-enhanced super-resolution DOA estimation through the output results of the DNN supervised deep neural network and the cross-platform transformation formula.

[0108] To correct the wavelength error caused by cross-platform conditions, the following method is used: the structure and parameters of DNN are fixed during the cross-platform test phase, and the phase of the input noise covariance matrix is The corresponding enhanced phase φ′2 of the network is obtained; the enhanced phase is corrected:

[0109]

[0110] The sampling covariance matrix is ​​reconstructed according to the corrected enhanced phase Φ information, and DOA estimation is achieved using DBF, MUSIC and ML.

[0111] The correction principle is as follows:

[0112] S411, Φ(λ1) represents the phase information of the training data at wavelength λ1. In the training stage, the wavelength is λ1, and the Φ(λ1)×λ1 operation is performed:

[0113]

[0114] There are M elements, and the mth element represents the phase information of the signal received by the mth array element in the array. represents the phase disturbance of the received signal of the mth array element in the array; It represents the phase disturbance of the signal received by the second element in the array. The processed data features do not have the wavelength factor, and the final training of the network is robust to the wavelength.

[0115] S412. In the test phase, the corresponding enhanced phase φ′2 of the DNN supervised deep neural network is obtained, and the wavelength is λ2:

[0116] φ2′=[0,-2πdsinθ d ×1,...,-2πdsinθ d ×(M-1)] (10)

[0117] S413, perform phase correction according to the enhanced phase φ′2:

[0118]

[0119] The network enhancement data φ2′ is adapted to the wavelength λ2 to obtain Φ, where Φ represents the final enhanced phase. The phase information Φ obtained at this time is the cross-platform compensation result for suppressing multipath. A high-resolution estimation result is obtained by signal reconstruction and using traditional DOA estimation methods.

[0120] It is only necessary to perform DNN network training and DOA estimation on the echo data through the original platform, and then correct it according to the transformation formula to complete accurate phase-enhanced super-resolution DOA estimation.

[0121] The following is a simulation experiment:

[0122] The data generation and processing of the experiment were completed on MATLAB2021a, and the four-layer deep neural network model training was completed on Python3.5. This embodiment verifies the effectiveness of this method through four simulation scenarios. The wavelength of the training platform is λ = 1 meter, and the test platform uses a wavelength of λ = 1.2 meters.

[0123] Scenario 1: Impact of SNR Matching on DOA Estimation Performance

[0124] The array structure of this scenario is a uniform linear array with M=24 elements, wavelength λ=1.2 meters, element spacing d is half a wavelength, and the fixed snapshot number is 1. The channel error of the training and test samples is 50%, the signal-to-noise ratio range of the training and test samples is 0dB~20dB, and the signal-to-noise ratio sampling interval is 2dB. During the experiment, we will generate received signal data under different signal-to-noise ratios and target azimuths, and use three direction estimation methods (DBF, MUSIC and ML) for direction estimation. Please refer to Figure 5 The schematic diagram shows the comparison results of the phase angle details of the signal reconstruction under each array element using the cross-platform phase enhancement signal technology provided by the embodiment of the present invention, where the horizontal axis represents the array element number and the vertical axis represents the angle. By adopting the DNN processing method, we can reduce the error of the phase angle feature of the signal. Figure 6 The cross-platform phase-enhanced super-resolution DOA estimation method provided by the embodiment of the present invention shows a schematic diagram of the comparison results of the angle measurement error under different estimation methods, where the horizontal axis represents the signal-to-noise ratio and the vertical axis represents the angle measurement error. By adopting the method provided by this technology, we significantly reduce the angle measurement error of various estimation methods, which proves the superiority of this method in improving the accuracy of direction estimation under the condition of signal-to-noise ratio matching.

[0125] Scenario 2: Impact of SNR mismatch on DOA estimation performance

[0126] The array structure built in this scenario is a uniform linear array with M=24 elements, wavelength λ=1.2 meters, element spacing d is half a wavelength, and the fixed snapshot number is 1. The channel error between training and test samples is 50%, the signal-to-noise ratio range of training samples is 0dB~20dB, the signal-to-noise ratio range of test samples is -3dB~23dB, and the signal-to-noise ratio sampling interval is 2dB. During the experiment, we will generate received signal data at different signal-to-noise ratios and target azimuths, and use three direction estimation methods (DBF, MUSIC, and ML) for direction estimation. Please refer to Figure 7The schematic diagram shows the comparison results of the signal phase angle detail features under each array element using the cross-platform phase enhancement signal technology provided by the embodiment of the present invention, where the horizontal axis represents the array element number and the vertical axis represents the angle. By using the DNN processing method, we found that the error of the signal phase angle feature can still be reduced under the condition of mismatched signal-to-noise ratio. Figure 8 A schematic diagram showing the comparison results of the angle measurement error of the cross-platform phase-enhanced super-resolution DOA estimation method provided by an embodiment of the present invention under different estimation methods is shown, wherein the horizontal axis represents the signal-to-noise ratio and the vertical axis represents the angle measurement error. By adopting the method provided by this technology, we have significantly reduced the angle measurement errors of various estimation methods. The results of this figure show that, through the cross-platform phase-enhanced super-resolution DOA estimation method provided by the present invention, more accurate direction estimation results can be obtained regardless of low or high signal-to-noise ratios. Compared with traditional estimation methods, our method can still maintain a high degree of accuracy when the signal-to-noise ratio does not match, proving that it has strong generalization and practicality in practical applications.

[0127] Scenario 3: Impact of channel error matching on DOA estimation performance

[0128] The array structure built in this scenario is a uniform linear array with M=24 elements, wavelength λ=1.2 meters, element spacing d is half a wavelength, and the fixed snapshot number is 1. The channel error of the training and test samples is 10% to 50%, the channel error sampling interval is 10%, and the signal-to-noise ratio range of the training and test samples is 10dB. During the experiment, we will generate received signal data under different signal-to-noise ratios and target azimuths, and use three direction estimation methods (DBF, MUSIC, and ML) for direction estimation. Please refer to Fig. 9 The schematic diagram of the comparison results of the signal phase angle detail features under each array element using the cross-platform phase enhancement signal technology provided by the embodiment of the present invention is shown, where the horizontal axis represents the array element number and the vertical axis represents the angle. By adopting the DNN processing method, we can reduce the error of the phase angle feature of the signal. Fig.10 The cross-platform phase-enhanced super-resolution DOA estimation method provided by the embodiment of the present invention shows a schematic diagram of the comparison results of the angle measurement error under different estimation methods, where the horizontal axis represents the channel error and the vertical axis represents the angle measurement error. By adopting the method provided by this technology, we significantly reduce the angle measurement error of various estimation methods, which proves the superiority of this method in improving the accuracy of direction estimation under the condition of channel error matching.

[0129] Scenario 4: Impact of channel error mismatch on DOA estimation performance

[0130] The array structure built in this scenario is a uniform linear array with M=24 elements, wavelength λ=1.2 meters, element spacing d is half a wavelength, and the fixed snapshot number is 1. The channel error of the training sample is 10%~50%, the channel error of the test sample is 5%~65%, the channel error sampling interval is 10%, and the signal-to-noise ratio range of the training and test samples is 10dB. During the experiment, we will generate received signal data under different signal-to-noise ratios and target azimuths, and use three direction estimation methods (DBF, MUSIC and ML) for direction estimation. Please refer to Fig.11 The schematic diagram shows the comparison results of the signal phase angle detail features under each array element using the cross-platform phase enhancement signal technology provided by the embodiment of the present invention, where the horizontal axis represents the array element number and the vertical axis represents the angle. By adopting the DNN processing method, we can still reduce the error of the signal phase angle feature under the condition of channel error mismatch. Fig.12 A schematic diagram showing the comparison results of the angle measurement error of the cross-platform phase-enhanced super-resolution DOA estimation method provided by an embodiment of the present invention under different estimation methods is shown, wherein the horizontal axis represents the channel error and the vertical axis represents the angle measurement error. By adopting the method provided by this technology, we still significantly reduce the angle measurement errors of various estimation methods under the condition of channel error mismatch. This experiment demonstrates the superiority of the cross-platform phase-enhanced signal technology in terms of the root mean square error of the real and imaginary parts of the signal and the accuracy of direction estimation, especially under the condition of channel error mismatch. It still shows good performance. This is of great significance for improving the generalization and accuracy of direction estimation.

[0131] Scenario 5: Performance comparison with other data-driven algorithms

[0132] This scenario compares this technology with other angle measurement schemes based on machine learning, and compares three other data-driven angle measurement schemes: SVR-based phase learning scheme, DNN super-resolution angle measurement method based on real and imaginary part learning, and DNN-super-resolution method based on phase learning. Among them, the jointly used super-resolution method adopts ML estimation. The experiment will be carried out in two scenarios, namely the angle measurement scenario under channel error and the angle measurement scenario under SNR simulation. The data training phase uses λ=1 meter data for training, and the test phase uses λ=1.2 meter data for error statistics. The array structure is a uniform linear array with array elements M=24, the array element spacing d is half wavelength, and the fixed snapshot number is 1. For the angle measurement scenario under the channel error scenario, the channel error is set from 10% to 50% with a step size of 10% in the training and testing phases. For the angle measurement environment simulated by SNR, the signal-to-noise ratio is set from 0 to 20dB with a step size of 2dB in both the training and testing phases. The root mean square error under each signal-to-noise ratio or the average of the root mean square error under each signal-to-noise ratio is taken as the final statistical result.

[0133] Fig.13The comparison results of the simulation experiment are shown, where the dark bar graph represents the comparison of the angle measurement error between other machine learning methods and the cross-platform DNN phase enhancement super-resolution method proposed in this paper in the channel error scenario, and the light bar graph represents the performance comparison of various machine learning methods in the signal-to-noise ratio simulation scenario. The boxed bar graph represents the method mentioned in this paper. It can be seen that the method mentioned in this paper is superior to other data-driven methods in terms of angle measurement error, including the angle measurement scheme based on SVR learning, the DNN super-resolution method based on real and imaginary part learning, and the DNN super-resolution method based on phase learning. This experiment proves that the method mentioned in this paper is superior to many intelligent angle measurement schemes.

[0134] Therefore, the present invention adopts the above-mentioned cross-platform phase-enhanced super-resolution DOA estimation method, introduces a method based on a DNN neural network into the array signal DOA estimation method, and realizes cross-platform phase-enhanced super-resolution DOA estimation. It effectively reduces the negative impact of the phase disturbance of multipath signals on DOA estimation, and effectively solves the estimation problem of cross-platform frequency changes.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A cross-platform phase-enhanced super-resolution DOA estimation method, characterized in that: The following steps are involved: S1. Build a DNN supervised deep neural network architecture to extract phase difference distribution features; S2, build radar system array for training platform; S3, training the DNN supervised deep neural network on the training platform, using the original noise sampling covariance matrix and the corresponding clean sampling covariance matrix, extracting the phase distribution characteristics without wavelength error by compensating the phase label, and inputting it into the defined DNN for training; S4. Place the trained DNN supervised deep neural network on different platforms for DOA estimation, and complete the cross-platform phase-enhanced super-resolution DOA estimation through the output results of the DNN supervised deep neural network and the cross-platform transformation formula.

2. The cross-platform phase-enhanced super-resolution DOA estimation method according to claim 1, characterized in that: The DNN supervised deep neural network architecture in S1 consists of 4 layers.

3. The cross-platform phase-enhanced super-resolution DOA estimation method according to claim 1, characterized in that: The specific contents of S2 are as follows: For a uniform linear array with M isotropic elements, where the spacing d between each element is equal to half the wavelength, and the array contains L snapshots, the received vector is: The steering vector a(θ,λ) is: where θ d represents the direction of arrival (DOA) of the direct signal; ρ i represents the attenuation factor corresponding to the i-th multipath; n represents the number of unknown multipath signals; s(t) represents the complex echo signal; n(t) has a mean of 0 and a variance of σ 2 Gaussian white noise; in formula (1) hour, For a(θ d ,λ) causes amplitude and phase disturbances; therefore, the signal model is described as follows: y(t)=Γ⊙a(θ d ,λ)s(t)+n(t) (3) The disturbance term Γ is: C=[τ1,τ2,…,τ M ] T (4)。 4. The cross-platform phase-enhanced super-resolution DOA estimation method according to claim 1, characterized in that: The training process of S3 is as follows: S31, initialize network parameters W k and b k ; S32, obtaining a phase tag for compensating for wavelength error Assume that the phase information of the labeled training dataset is represented as Among them and and Representing the input data and the corresponding known data, the phase label is compensated according to the formula: S33, normalize the training data set to obtain parameters μ1 and σ1; For the phase information of the compensated training data set where Φ1 and Representing the input data and the corresponding known output data, the normalized data set is based on the formula: S34, will Randomly divide into several batches and perform forward propagation; S35. According to the error of the output layer, the BP strategy and AdaM method are used to adjust the network parameters W. k and b k Make fine adjustments; The optimization formula is as follows: Where φ1′ is the output data, is the known output data; S36. Repeat S33 to S35 by learning and adjusting the network architecture until the error of the objective function converges.

5. The cross-platform phase-enhanced super-resolution DOA estimation method according to claim 4, characterized in that: Cross-platform in S4 will cause wavelength error. The following method is used to correct the wavelength error caused by cross-platform conditions: S41. Fix the structure and parameters of the DNN during the cross-platform test phase and the phase of the input noise covariance matrix The enhanced phase φ2′ corresponding to the DNN supervised deep neural network is obtained; the enhanced phase is corrected: S42, and reconstruct the sampling covariance matrix according to the corrected enhanced phase Φ information, and use DBF, MUSIC and ML to achieve DOA estimation.

6. The cross-platform phase-enhanced super-resolution DOA estimation method according to claim 5, characterized in that: The correction principle in S41 is as follows: S411, Φ(λ1) represents the phase information of the training data at wavelength λ1. In the training stage, the wavelength is λ1, and the Φ(λ1)×λ1 operation is performed: in represents the phase disturbance of the received signal of the mth array element in the array; represents the phase disturbance of the received signal of the second array element in the array; S412. In the test phase, the corresponding enhanced phase φ′2 of the DNN supervised deep neural network is obtained, and the wavelength is λ2: φ′2=[0,-2πd sinθ d ×1,...,-2πd sinθ d ×(M-1)] (10) S413, perform phase correction according to the enhanced phase φ′2: The network enhancement data φ′2 is adapted to the wavelength λ2 to obtain Φ, where Φ represents the final enhanced phase. The phase information Φ obtained at this time is the cross-platform compensation result for suppressing multipath. A high-resolution estimation result is obtained by signal reconstruction and using traditional DOA estimation methods.

Images