A method for simultaneously estimating the number of sources and DOA based on deep neural network and clustering

Abstract

The application discloses a kind of based on deep neural network and clustering's simultaneous estimation method of source number and DOA, first by array covariance matrix is structured, then the training data-label set of automatic encoder and the training data-label set of parallel classifier, next using training data-label set trains deep neural network, data is input into the deep neural network after training, and output spatial spectrum vector is obtained;Obtain the large spectral peak value of spatial spectrum vector by spectral peak search, form spectral peak data set and corresponding angle value;Spectral peak data set is extended, and the expanded spectral peak data set and corresponding angle value are obtained;Finally, the expanded spectral peak data set is clustered using clustering algorithm, and the source number estimation value and the corresponding DOA estimation value are obtained.The application can realize accurate source number estimation and high-precision angle estimation under uniform or non-uniform noise conditions.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, specifically relating to a method for simultaneous estimation of source number and DOA based on deep neural networks and clustering. Background Technology

[0002] Direction of Arrival (DOA) estimation is a crucial research area in array signal processing, with the MUSIC (Multiple Signal Classification) algorithm being widely applied. The MUSIC algorithm divides the eigenvector space into signal and noise subspaces based on the eigenvalues ​​of the covariance matrix to obtain the target DOA estimate. However, MUSIC requires known source numbers, necessitating source number estimation methods when the number of sources is unknown. Source number estimation typically employs statistical methods such as AIC (Akaike Information Criterion) and MDL (Minimum Description Length). AIC and MDL use likelihood function values ​​and coding length, respectively, to select the most suitable model. Under uniform noise conditions, source number estimation is accurate, but significant estimation errors may occur under low signal-to-noise ratios and with limited observation snapshots. For non-uniform noise, Aouada et al. proposed the Non-Uniform Noise MDL (NU-MDL) method, but prediction accuracy decreases under high noise power ratios. Recently, Wax et al. proposed the Signal Subspace Matching (SSM) method, a source number detection method based on signal subspace matching. However, the SSM method is suitable for medium and high signal-to-noise ratio conditions.

[0003] In summary, the two-step strategy of first estimating the number of sources and then estimating the DOA has an error accumulation problem. That is, the error in estimating the number of sources will affect the subsequent DOA estimation results, leading to an increase in the root mean square error. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a method for simultaneous estimation of source number and DOA based on deep neural networks and clustering. First, a covariance matrix is ​​constructed using an array. Then, training data-label sets for an autoencoder and parallel classifiers are used. Next, a deep neural network is trained using these training data-label sets. The data is input into the trained deep neural network to obtain the output spatial spectral vector. Peak search is used to obtain the largest spectral peaks in the spatial spectral vector, forming a peak dataset and corresponding angle values. The peak dataset is then expanded to obtain an extended peak dataset and corresponding angle values. Finally, a clustering algorithm is used to cluster the extended peak dataset to obtain source number estimates and corresponding DOA estimates. This invention can achieve accurate source number estimation and high-precision angle estimation under both uniform and non-uniform noise conditions.

[0005] The technical solution adopted by this invention to solve its technical problem is as follows:

[0006] Step 1: The array consists of M elements, and K far-field, narrowband target signals are incident on the array; the received signal of the m-th element is r. m (t); The observation data vector of M array elements is r(t)=[r1(t),r2(t),…,r M (t)] T Construct the covariance matrix in,(·) T This indicates the transpose operation, (·) H This represents the conjugate transpose operation, where N is the number of samples;

[0007] Step 2: According to Obtain the training data-label set {(x1,y1),(x2,y2),…(x...}} of the autoencoder, consisting of F sets of data-label pairs. i ,y i ),…,(x F ,y F )}, where x i Let y represent the i-th input vector. i This represents the corresponding label vector; and the training data-label set {(x1,Z1),(x2,Z2),…(x...} of the parallel classifier consisting of S sets of data-label pairs. i Z i ),…,(x S Z S )}, where x i Let Z represent the i-th input vector. i This represents the corresponding label vector;

[0008] Step 3: Utilize the training data - label set {(x1,y1),(x2,y2),…,(xF ,y F {(x1,Z1),(x2,Z2),…,(x)} and {(x1,Z1),(x2,Z2),…,(x)} S Z S Training deep neural networks;

[0009] Step 4: Testing phase, obtaining input data not in the training dataset. The input is fed into the deep neural network trained in step 3 to obtain the output spatial spectral vector p;

[0010] Step 5: Obtain the M-1 large spectral peaks of the spatial spectral vector p through spectral peak search, which form the spectral peak dataset {peak}. i The values ​​of i = 1, 2, ..., M-1 and their corresponding angles.

[0011] Step 6: Expand the spectral peak dataset to obtain the expanded spectral peak dataset {peak}. i {i = 1, 2, ..., M} and their corresponding angle values

[0012] Step 7: Use a clustering algorithm to analyze the expanded spectral peak dataset {peak} i Clustering is performed on the group i = 1, 2, ..., M to obtain an estimate of the number of sources. and the corresponding DOA estimate

[0013] Further, step 1 specifically includes:

[0014] Step 1-1: Let the angles of the K received far-field signals be θ1, θ2, ..., θ K s(t) is a vector consisting of K information sources, s(t) = [s1(t), ... s2(t)]. K (t)] T , A=[a(θ1),a(θ2),…,a(θ K )],a(θ i ) is the steering vector at the i-th angle. Where d m λ is the distance between the m-th array element and the first array element, λ is the wavelength of the signal, j is the imaginary unit, and the noise vector n(t) = [n1(t), ..., n M (t)] T The array receives the signal vector as: r(t) = As(t) + n(t);

[0015] Step 1-2: The estimated value of the covariance matrix is

[0016] Furthermore, step 2 specifically includes:

[0017] Step 2-1: Take The upper right triangle elements are arranged to form a A vector b cor =[R 1,2 ,R 1,3 ,…,R 1,M ,R 2,3 ,…,R 2,M ,…,R M-1,M] T b cor It is a complex vector containing real and imaginary parts. Decomposing and concatenating its real and imaginary parts yields an M(M-1) dimensional vector b = [Real{b cor T},Imag{b cor T}], where Real(·) is the real part operation and Imag(·) is the imaginary part operation; then b is normalized to obtain the vector. Where mean(·) represents taking the mean, and ||·||2 represents the 2-norm;

[0018] Step 2-2: The data label corresponding to the autoencoder is an M×(M-1)×NUM dimensional vector y. i Where NUM is the number of parallel classifiers, thus obtaining a data-label pair as (x i ,y i ), where the subscript i is the sequence number; the data label corresponding to the parallel classifier is A vector of dimension Z i Where grid is the grid spacing of the deep neural network, and a data-label pair is further obtained as (x i Z i ), where the subscript i is the sequence number; Z i It is a sparse vector in which most elements are equal to 0, and its non-zero positions are the target DOA;

[0019] Steps 2-3: Obtain observation data r(t) under different scenarios, and obtain the corresponding results based on Step 1. Finally, the training data-label set of the autoencoder, consisting of F sets of data-label pairs, is obtained: {(x1,y1),(x2,y2),…,(x...} F ,y F )}, and the training data-label set {(x1,Z1),(x2,Z2),…,(x)} of a parallel classifier consisting of S sets of data-label pairs. S Z S )}.

[0020] Furthermore, step 3 specifically includes:

[0021] Step 3-1: Utilize the training data - label set {(x1,y1),(x2,y2),…,(x F ,y F Training the autoencoder layer: Let the number of layers in the autoencoder network be L, excluding the input layer, using net... l This represents the output of layer l, net. l =W l,l-1 h l-1 +b l l = 1, 2, ..., L, where W l,l-1 b represents the weight matrix between layer (l-1) and layer l. l Indicates the deviation of the l-th layer;

[0022] Step 3-2: Utilize the training data - label set {(x1,Z1),(x2,Z2),…,(x s Z S Train parallel classifier layers, assuming the total number of layers in the classifier's neural network is . Without an input layer, use net l This represents the output of the l-th layer. in This represents the weight matrix between the (l-1)th layer and the lth layer. G represents the deviation of the l-th layer. l (·) is the activation function; for the output layer in Indicates the first Layer and First Weight matrix between layers Indicates the first Layer deviation.

[0023] Furthermore, step 6 specifically includes:

[0024] Step 6-1: At the angle At this point, virtual spectral peak data is added to the spectral peak dataset, and its amplitude is set to amp1, thereby obtaining the Mth spectral peak. M =amp1, and its corresponding angle is θ M =θ vs ;

[0025] Step 6-2: Obtain the extended spectral peak dataset {peak} i {i = 1, 2, ..., M} and their corresponding angle values

[0026] Furthermore, step 7 specifically includes:

[0027] Step 7-1: Cluster the expanded spectral peak dataset {peak} i Perform binary clustering on the {i = 1, 2, ..., M};

[0028] Step 7-2: The clustering results are divided into two categories: signal and noise. Obtain the angle values ​​corresponding to the data in the signal category and filter out the virtual angle θ. vs Finally, an estimate of the number of information sources was obtained. and the corresponding DOA estimate

[0029] The beneficial effects of this invention are as follows:

[0030] This invention proposes a method for simultaneous estimation of the number of sources and DOA based on deep neural networks and clustering. It can achieve accurate source number estimation and high-precision angle estimation under both uniform and non-uniform noise conditions. This invention is mainly applied to target detection under conditions of unknown source number. Attached Figure Description

[0031] Figure 1 Flowchart of the method of this invention;

[0032] Figure 2 The curve showing the change in the probability of detection (PD) as a function of the number of snapshots (N) when the number of sources is 0.

[0033] Figure 3 When the number of sources is 1, the curves of the source detection success rate PD and the estimated RMSE of DOA as a function of SNR are shown. (a) is the curve of the source number estimated PD as a function of SNR, and (b) is the curve of the estimated RMSE of DOA as a function of SNR.

[0034] Figure 4 When the number of sources is 1, the curves of the source number detection success rate PD and the estimated RMSE of DOA change with the source DOA are shown. (a) is the curve of the source number estimated PD change with the source DOA, and (b) is the curve of the estimated RMSE of DOA change with the source DOA.

[0035] Figure 5 When the number of sources is 1, the curves of the source number detection success rate PD and the estimated RMSE of DOA with WNPR are shown. (a) is the curve of the source number estimated PD with WNPR, and (b) is the curve of the estimated RMSE of DOA with WNPR.

[0036] Figure 6When the number of sources is 2, the curves showing the changes in the source number detection success rate (PD) and the estimated RMSE of DOA as a function of the dual-target angular interval are shown in (a) and (b).

[0037] Figure 7 When the number of sources is 2, the curves of the source detection success rate (PD) and the estimated RMSE of DOA as a function of WNPR are shown. (a) is the curve of the source number estimated PD as a function of WNPR, and (b) is the curve of the estimated RMSE of DOA as a function of WNPR. Detailed Implementation

[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0039] Combination Figure 1 The processing flow of the present invention includes the following steps:

[0040] The sensor array consists of M scalar sensors; the sensor array receives K far-field, narrowband target signals, and the angle of arrival (θ) of the k-th target is θ. k The direction finding range is [θ] min ,θ max );

[0041] The specific implementation steps of a method for simultaneously estimating the number of sources and DOA based on deep neural networks and clustering include:

[0042] (1) The array consists of M elements, and K far-field, narrowband target signals are incident on the array. The received signal of the m-th element is r. m (t); The observation data vector of M array elements is r(t)=[r1(t),r2(t),…,r M (t)] T Construct the covariance matrix in,(·) T This indicates the transpose operation, (·) H This represents the conjugate transpose operation, where N is the number of samples. The specific steps are as follows:

[0043] (1.1) Assume that the DOAs of the received K far-field signals are θ1, θ2, ..., θ K s(t) is a vector consisting of K information sources, s(t) = [s1(t), ... s2(t)]. K (t)] T , A=[a(θ1),a(θ2),…,a(θ K )],a(θ i ) is the steering vector at the i-th angle. (where d) m λ is the distance between the m-th array element and the first array element, λ is the wavelength of the signal, and j is the imaginary unit. The noise vector n(t) = [n1(t), ..., n M (t)] T The received signal array is: r(t) = As(t) + n(t);

[0044] (1.2) The estimated value of the covariance matrix is

[0045] (2) According to Obtain the training data-label set {(x1,y1),(x2,y2),…,(x...} of the autoencoder, consisting of F sets of data-label pairs. F ,y F )}, where x i Let y represent the i-th input vector. i This represents the corresponding output vector; and the training data-label set {(x1,Z1),(x2,Z2),…,(x...} of the parallel classifier consisting of S sets of data-label pairs. S Z S )}, where x i Let Z represent the i-th input vector. i This represents the corresponding output vector. The specific steps are as follows:

[0046] (2.1) Take The upper right triangle elements are arranged to form a A vector b cor =[R 1,2 ,R 1,3 ,…,R 1,M ,R 2,3 ,…,R 2,M ,…,R M-1,M ] T ; due to b cor It is a complex vector containing real and imaginary parts. We need to split and concatenate its real and imaginary parts to obtain an M(M-1) dimensional real vector b = [Real{b cor T},Imag{b cor T (where Real(·) is the real part operation and Imag(·) is the imaginary part operation), and then b is normalized to obtain the vector. Where mean(·) represents taking the mean, and ||·||2 represents the 2-norm;

[0047] (2.2) The data label corresponding to the autoencoder is an M×(M-1)×NUM dimensional vector y. i(where NUM is the number of parallel classifiers), thus obtaining a data-label pair as (x i ,y i ), where the subscript i is the sequence number; the data label corresponding to the parallel classifier is A vector of dimension Z i (where grid is the grid spacing of the deep neural network), further obtaining a data-label pair as (x i Z i ), where the subscript i is the sequence number. Z i It is a sparse vector in which most elements are equal to 0, and its non-zero positions are the target DOAs;

[0048] (2.3) Obtain observation data r(t) under different scenarios, and obtain the corresponding data according to step 1. Finally, the training data-label set of the autoencoder, consisting of F sets of data-label pairs, is obtained: {(x1,y1),(x2,y2),…,(x...} F ,y F )}, and the training data-label set {(x1,Z1),(x2,Z2),…,(x)} of a parallel classifier consisting of S sets of data-label pairs. S Z S )}.

[0049] (3) Using the training data-label set {(x1,y1),(x2,y2),…,(x...} F ,y F {(x1,Z1),(x2,Z2),…,(x)} and {(x1,Z1),(x2,Z2),…,(x)} S Z S Training a deep neural network. The specific steps are as follows:

[0050] (3.1) Using the training data-label set {(x1,y1),(x2,y2),…,(x F ,y F Training the autoencoder layer, assuming the autoencoder network has L layers (excluding the input layer), using net... l This represents the output of layer l, net. l =W l,l-1 h l-1 +b l l = 1, 2, ..., L, where W l,l-1 b represents the weight matrix between layer (l-1) and layer l. l Indicates the deviation of the l-th layer;

[0051] (3.2) Using the training data-label set {(x1,Z1),(x2,Z2),…,(x S Z STraining parallel classifier layers, assuming the total number of layers in the classifier's neural network is... (Excluding the input layer), use net l This represents the output of the l-th layer. in This represents the weight matrix between the (l-1)th layer and the lth layer. G represents the deviation of the l-th layer. l (·) is the activation function; for the output layer in Indicates the first Layer and First Weight matrix between layers Indicates the first Layer deviation.

[0052] (4) During the testing phase, obtain input data that is not in the training dataset. The input is fed into the deep neural network trained in step 3 to obtain the output spatial spectrum.

[0053] (5) Obtain the M-1 large spectral peaks of the spatial spectral vector p through spectral peak search, which constitute the spectral peak dataset {peak}. i {i = 1, 2, ..., M-1} and their corresponding angle values

[0054] (6) Expand the spectral peak dataset to obtain the expanded spectral peak dataset {peak}. i {i = 1, 2, ..., M} and their corresponding angle values The specific steps are as follows:

[0055] (6.1) At the angle At this point, virtual spectral peak data is added to the spectral peak dataset, and its amplitude is set to amp1, thereby obtaining the Mth spectral peak. M =amp1, and its corresponding angle is θ M =θ vs ;

[0056] (6.2) Obtain the extended spectral peak dataset {peak i {i = 1, 2, ..., M} and their corresponding angle values

[0057] (7) Use clustering algorithms to analyze the extended spectral peak dataset {peak} i Clustering is performed on the group i = 1, 2, ..., M to obtain an estimate of the number of sources. and the corresponding DOA estimate The specific steps are as follows:

[0058] (7.1) The spectral peak dataset {peak} expanded by clustering pairs i Perform binary clustering on the {i = 1, 2, ..., M};

[0059] (7.2) The clustering results are two types of data: signal and noise. The angle values ​​corresponding to the data in the signal category are obtained, and the virtual angle θ is filtered out. vs Finally, an estimate of the number of information sources was obtained. and the corresponding DOA estimate

[0060] Example:

[0061] The sensor array is a uniform linear array consisting of 10 elements; the target signal is a far-field, narrow-band signal. For the target signal power, The noise power is used to collect training data under uniform noise conditions; the Monte Carlo iteration is 500; the direction finding range is [θ]. min = -60°, θ max =60°).

[0062] The deep neural network in this embodiment of the invention consists of a multi-task autoencoder and six parallel classifiers. The neural network has a grid spacing of 1° and 120 neurons in the output layer. The autoencoder has M×(M-1) input neurons, M×(M-1) / 2 hidden layer neurons, and 6×M×(M-1) output layer neurons. Each classifier has M×(M-1) input neurons, 2×M×(M-1) / 3 hidden layer neurons, and 4×M×(M-1) / 9 hidden layer neurons. Each hidden layer is followed by a ReLU nonlinear activation function, and the output layer has 20 neurons.

[0063] Training dataset construction:

[0064] The autoencoder layer adds random noise for each target angle and collects data, with each group labeled xi, where the subscript i is the data sequence number, and the output label is y. i Because autoencoders are designed with linearity and additive properties, good performance in single-target scenarios guarantees good performance in spatial filtering. Therefore, training datasets are generated in two ways. Additionally, the network is trained simultaneously using data from scenarios with different signal-to-noise ratios (SNRs) and snapshot counts. The SNR ranges from -20dB to 0dB, with values ​​in 0.1 increments (200 different SNRs). The snapshot count ranges from 400 to 100 (13 different snapshot counts).

[0065] 1. Consider a target signal scene where the target angle search interval is [-60°, 60°) with a 1° interval. The dataset consists of 200 × 13 × 120 = 312000 covariance vectors and their labels, forming a data-label set {(x1, y1), (x2, y2), ..., (x...}. F1 ,y F1 )}, where F1=312000;

[0066] 2. Considering the scenario with zero target signals, i.e., the received data contains only noise, to ensure a similar data volume as the single-target scenario, we repeatedly generate noisy data, resulting in a final dataset consisting of 312,000 covariance vectors and their labels – a data label set {(x1,y1),(x2,y2),…,(x...}. F0 ,y F0 )}, where F0 = 312000;

[0067] The two cases combined result in a data set consisting of 624,000 covariance vectors and their labels – the label set {(x1,y1),(x2,y2),…,(x...}. F ,y F )}.

[0068] The classifier layer adds random noise to each set of target angles and collects data; each set of data is labeled as x. i Where the subscript i is the data sequence number, and the output label Z i Maximum number of sources K max =2, then the number of information sources K = {0,1,2}, generating the training dataset in three cases. Additionally, consider training the network simultaneously using data from scenarios with different signal-to-noise ratios (SNRs) and different snapshot counts. The SNR ranges from -20dB to 1dB, with values ​​in 1-degree intervals, resulting in 21 different SNRs. The snapshot count ranges from [400, 200, 150, 100, 90, 80, 70, 60, 50, 40, 30, 20, 10], resulting in 13 different snapshot counts.

[0069] 1. Consider two target signal scenarios, with a target angle search range of [-60°, 60°). The set of these two target angle intervals is Δ = {1°, 2°, ..., 119°}, covering scenarios where the target signals are close to each other and far apart. The first target angle range is θ1 ∈ [60°, 60° - Δ]. s ], where △ s This represents the angle between two targets selected from the Δ set, with the second angle set as θ2 = θ1 + Δ. sThe final dataset contains (119+118+…+1)×21×13=1949220 covariance vectors and their labels, forming a data-label set {(x1,Z1),(x2,Z2),…,(x...}}. s2 Z s2 )}, where s2=1949220;

[0070] 2. Considering a scenario with zero target signals, the values ​​are repeated 20 times. The final dataset contains 7140 × 21 × 13 = 1949220 covariance vectors and their labels, forming a data-label set {(x1, Z1), (x2, Z2), ..., (x...}. s0 Z s0 )}, where s0=1949220;

[0071] 3. Consider a target signal scene where the target angle search interval is [-60°, 60°) with an interval of 1°, repeated 59 times. The final dataset contains 59 × 120 × 21 × 13 = 1932840 covariance vectors and their labels, forming a data-label set {(x1, Z1), (x2, Z2), ..., (x...}. s1 Z s1 )}, where s1=1932840;

[0072] The three cases combined result in a data set consisting of 5,831,280 covariance vectors and their labels: {(x1,Z1),(x2,Z2),…,(x...}. s Z s )}.

[0073] During the testing phase, spectral peak search was performed on the output spatial spectral vector of the deep neural network to obtain a spectral peak dataset consisting of M-1 large spectral peaks. This dataset was then expanded by adding a virtual spectral peak with amplitude amp1 = 0.4 at 60° and virtual noise with amplitude amp0 = 0 at 61°. The K-means algorithm was used to cluster this expanded spectral peak dataset to obtain the number of sources. and the corresponding angle value This experiment examines the performance of the method of this invention (i.e., DNN With Clustering, abbreviated as DNN-WC) and compares it with traditional source number estimation methods (MDL and NUMDL). The performance metric is the probability of detection (PD). It is further compared with MDL-MUSIC, NU-MDL-MUSIC, and SSM-MUSIC algorithms, with the performance metric being the root mean square error (RMSE) of the DOA estimation. The performance changes of the algorithm under non-uniform noise conditions are also examined. The non-uniform noise power settings are as follows: the noise power of the first array element is set to the minimum and set to 1; the noise power of the second array element is set to 2; and the noise variance of the seventh array element is set to the maximum and marked as... The noise power of other array elements is in the range The noise is generated randomly according to a uniform distribution. In this case, the worst noise power ratio (WNPR) is...

[0074] The experiment examines the detection performance of the algorithm under K=0, 1, and 2 conditions. Unless otherwise specified, WNPR=2. When K=0, there is no target, and the received signal is only noise. In this case, the algorithm's PD, i.e., the accuracy of estimating the number of sources K=0, is examined. When K=1, there is a single target. Unless otherwise specified, the single target DOA is θ1=-15.2°, the number of snapshots N=1000, and the SNR=0dB. The algorithm's PD and RMSE are examined. When K=2, there are two targets. Unless otherwise specified, the two target DOAs are (θ1=-15.2°, θ2=32.1°), N=1000, and the SNR=0dB. The algorithm's PD and RMSE are examined. Note that when the estimated number of sources is not equal to the actual number of sources, we pad the estimated DOAs vector or the actual DOAs vector with the value 100 to account for the influence of the estimated number of sources when calculating RMSE.

[0075] The above experimental results show the effectiveness of the method of the present invention under different experimental conditions:

[0076] When K=0, there is no objective, such as Figure 2 As shown, the method of this invention, namely DNN-WC, can achieve correct detection with PD=1 in the range of small to large snapshots; while the traditional MDL method and NUMDL method cannot achieve correct detection.

[0077] When K=1, in the single-target case, such as Figure 3As shown, the method of the present invention can achieve correct detection with a PD of 1 in the range of SNR from -10dB to 18dB, and the RMSE of the DOA estimate is less than 1 degree. However, due to the influence of non-uniform noise, the traditional MDL method fails. The NUMDL method can only achieve correct detection with a PD of 1 when the SNR is greater than -4dB. The RMSE of the corresponding NUMDL-MUSIC algorithm is close to that of the method of the present invention only when the SNR is greater than -4dB.

[0078] When K=1, in the single-target case, such as Figure 4 As shown, both the method of this invention and the NUMDL method can correctly detect single targets from different angles; the DOA estimation accuracy of both the method of this invention and the NUMDL-MUSIC method is less than 1 degree; the MDL method and the MDL-MUSIC method fail due to non-uniform noise.

[0079] When K=1, in the single-target case, such as Figure 5 As shown, the method of the present invention is not affected by WNPR and has robust detection performance; while the NUMDL method cannot achieve 100% correct detection when WNPR is greater than 3, which causes the DOA estimation accuracy of NUMDL-MUSIC to deteriorate when WNPR is greater than 3.

[0080] When K=2, in the dual-objective case, such as Figure 6 As shown, both the method of this invention and the NUMDL method can correctly detect dual targets with different angular intervals; the DOA estimation accuracy of both the method of this invention and the NUMDL-MUSIC method is less than 1 degree.

[0081] When K=2, in the dual-objective case, such as Figure 7 The figure shows the change in algorithm performance with WNPR, and its conclusions are consistent with those in the single-objective case. Figure 5 Similarly, the method of this invention is not affected by WNPR and has robust detection performance; while the NUMDL method and NUMDL-MUSIC method degrade in performance when WNPR is greater than 3.

Claims

1. A method for simultaneously estimating the number of sources and DOA based on deep neural networks and clustering, characterized in that, The steps include the following: Step 1: The array consists of M elements, and K far-field, narrowband target signals are incident on the array; the received signal of the m-th element is r. m (t); The observation data vector of M array elements is r(t)=[r1(t),r2(t),…,r M (t)] T Construct the covariance matrix in,(·) T This indicates the transpose operation, (·) H This represents the conjugate transpose operation, where N is the number of samples; Step 2: According to Obtain the training data-label set {(x1,y1),(x2,y2),…(x...}} of the autoencoder, consisting of F sets of data-label pairs. i ,y i ),…,(x F ,y F )}, where x i Let y represent the i-th input vector. i This represents the corresponding label vector; and the training data-label set {(x1,Z1),(x2,Z2),…(x...} of the parallel classifier consisting of S sets of data-label pairs. i Z i ),…,(x S Z S )}, where x i Let Z represent the i-th input vector. i This represents the corresponding label vector; Step 3: Utilize the training data - label set {(x1,y1),(x2,y2),…,(x F ,y F {(x1,Z1),(x2,Z2),···,(x)} and {(x1,Z1),(x2,Z2),···,(x)} S Z S Training deep neural networks; Step 4: Testing phase, obtaining input data not in the training dataset. The input is fed into the deep neural network trained in step 3 to obtain the output spatial spectral vector p; Step 5: Obtain the M-1 large spectral peaks of the spatial spectral vector p through spectral peak search, which form the spectral peak dataset {peak}. i The values ​​of i = 1, 2, ..., M-1 and their corresponding angles. Step 6: Expand the spectral peak dataset to obtain the expanded spectral peak dataset {peak}. i {i = 1, 2, ..., M} and their corresponding angle values Step 7: Use a clustering algorithm to analyze the expanded spectral peak dataset {peak} i Clustering is performed on the group i = 1, 2, ..., M to obtain an estimate of the number of sources. and the corresponding DOA estimate 2. The method for simultaneous estimation of source number and DOA based on deep neural networks and clustering according to claim 1, characterized in that, Step 1 specifically involves: Step 1-1: Let the angles of the K received far-field signals be θ1, θ2, ..., θ K s(t) is a vector consisting of K information sources, s(t) = [s1(t), ... s2(t)]. K (t)] T , A=[a(θ1),a(θ2),…,a(θ K )],a(θ i ) is the steering vector at the i-th angle. Where d m λ is the distance between the m-th array element and the first array element, λ is the wavelength of the signal, j is the imaginary unit, and the noise vector n(t) = [n1(t), ..., n M (t)] T The array receives the signal vector as: r(t) = As(t) + n(t); Step 1-2: The estimated value of the covariance matrix is 3. The method for simultaneous estimation of source number and DOA based on deep neural networks and clustering according to claim 2, characterized in that, Step 2 specifically involves: Step 2-1: Take The upper right triangle elements are arranged to form a A vector b cor =[R 1,2 ,R 1,3 ,…,R 1,M ,R 2,3 ,…,R 2,M ,…,R M-1,M ] T b cor It is a complex vector containing real and imaginary parts. Decomposing and concatenating its real and imaginary parts yields an M(M-1) dimensional vector b = [Real{b cor T },Imag{b cor T }], where Real(·) is the real part operation and Imag(·) is the imaginary part operation; then b is normalized to obtain the vector. Where mean(·) represents taking the mean, and ||·||2 represents the 2-norm; Step 2-2: The data label corresponding to the autoencoder is an M×(M-1)×NUM dimensional vector y. i Where NUM is the number of parallel classifiers, thus obtaining a data-label pair as (x i ,y i ), where the subscript i is the sequence number; the data label corresponding to the parallel classifier is A vector of dimension Z i Where grid is the grid spacing of the deep neural network, and a data-label pair is further obtained as (x i Z i ), where the subscript i is the sequence number; Z i It is a sparse vector in which most elements are equal to 0, and its non-zero positions are the target DOA; Steps 2-3: Obtain observation data r(t) under different scenarios, and obtain the corresponding results based on Step 1. Finally, the training data-label set of the autoencoder, consisting of F sets of data-label pairs, is obtained: {(x1,y1),(x2,y2),…,(x...} F ,y F )}, and the training data-label set {(x1,Z1),(x2,Z2),…,(x)} of a parallel classifier consisting of S sets of data-label pairs. S Z S )}.

4. The method for simultaneous estimation of source number and DOA based on deep neural networks and clustering according to claim 3, characterized in that, Step 3 specifically involves: Step 3-1: Utilize the training data - label set {(x1,y1),(x2,y2),…,(x F ,y F Training the autoencoder layer: Let the number of layers in the autoencoder network be L, excluding the input layer, and use net... l This represents the output of layer l, net. l =W l,l-1 h l-1 +b l l = 1, 2, ..., L, where W l,l-1 b represents the weight matrix between layer (l-1) and layer l. l Indicates the deviation of the l-th layer; Step 3-2: Utilize the training data - label set {(x1,Z1),(x2,Z2),…,(x S Z S Train parallel classifier layers, assuming the total number of layers in the classifier's neural network is . Without an input layer, use net l This represents the output of the l-th layer. in This represents the weight matrix between the (l-1)th layer and the lth layer. G represents the deviation of the l-th layer. l (·) is the activation function; for the output layer in Indicates the first Layer and First Weight matrix between layers Indicates the first Layer deviation.

5. The method for simultaneous estimation of source number and DOA based on deep neural networks and clustering according to claim 4, characterized in that, Step 6 specifically involves: Step 6-1: At the angle At this point, virtual spectral peak data is added to the spectral peak dataset, and its amplitude is set to amp1, thereby obtaining the Mth spectral peak. M =amp1, and its corresponding angle is θ M =θ vs ; Step 6-2: Obtain the extended spectral peak dataset {peak} i {i = 1, 2, ..., M} and their corresponding angle values 6. The method for simultaneous estimation of source number and DOA based on deep neural networks and clustering according to claim 5, characterized in that, Step 7 specifically involves: Step 7-1: Cluster the expanded spectral peak dataset {peak} i Perform binary clustering on the {i = 1, 2, ..., M}; Step 7-2: The clustering results are divided into two categories: signal and noise. Obtain the angle values ​​corresponding to the data in the signal category and filter out the virtual angle θ. vs Finally, an estimate of the number of information sources was obtained. and the corresponding DOA estimate

Images