Horizontal array intelligent target orientation estimation method and system based on subarray division

Abstract

The invention discloses a horizontal array intelligent target orientation estimation method and system based on sub-array division, and relates to the technical field of target orientation estimation, a whole linear array is divided into a plurality of sub-arrays, CBF and MVDR beam domain features of the sub-arrays are input into a sub-array DOA estimation model based on multi-stage fusion multi-beam domain features, and a target orientation estimation model is obtained. Fusion of multiple beam forming methods is realized, and DOA estimation precision of a single subarray is ensured; then, the DOA estimation result of the single subarray, subarray information and array element parameters serve as input of a conditional convolution DOA estimation model based on multi-source feature fusion, conditional convolution in the model parameterizes a convolution kernel, convolution kernel parameters are dynamically generated according to input conditions, all convolution kernel parameters in a backbone network are adjusted in a self-adaptive mode, and the DOA estimation result of the single subarray, subarray information and array element parameters are obtained. And fusing the information to obtain a final DOA estimation result. According to the method, the whole linear array is divided into a plurality of sub-arrays, so that the multipath effect and noise interference can be effectively reduced, and the estimation precision of the target azimuth is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of target azimuth estimation, and more particularly to a method and system for horizontal array intelligent target azimuth estimation based on subarray division. Background Art

[0002] At present, using underwater acoustic information to accurately estimate the direction of arrival (DOA) of underwater acoustic targets is an important research direction in the field of array signal processing. The basic problem of DOA estimation is to determine the position of the signal of interest in space at the same time. Generally, the key information of the target can be obtained through the DOA estimation method, such as the direction angle of the target relative to the reference array element of the receiving array. The DOA process is generally achieved by processing array signals through beamforming technology. Array signals refer to spatial sampling signals received by a sensor array (such as a horizontal array) composed of multiple sensors placed at different spatial positions. Beamforming is a spatial processing technology for data received by multiple sensor arrays, which aims to enhance the signal reception capability in a specific direction while suppressing interference from other directions.

[0003] In the actual ocean environment, for a long horizontal array, when the noise source level radiated by the target is small, the array elements farther away from the target may not be able to receive the target noise. In addition, horizontal arrays placed on the seabed for a long time may have problems such as failure of some array elements or reduced sensitivity. If all array elements are used directly for azimuth estimation, the final accuracy will be affected. Therefore, large arrays usually adopt the method of sub-array division, for example, hundreds or thousands of frame elements are divided into more than a dozen sub-arrays, beamforming is used within the sub-arrays, and then the beamforming results between sub-arrays are combined to obtain the final result. By designing reasonable sub-array weights, DOA estimation results that are better than those of a single sub-array can usually be achieved.

[0004] Existing beamforming algorithms include conventional beamforming (CBF), minimum variance distortionless response (MVDR), etc. However, these algorithms have their own advantages and disadvantages. For example, MVDR is a beamforming technology with high resolution, but its robustness is low, and a small mismatch can lead to a large positioning error; CBF does not rely on the accurate estimation of the noise covariance matrix and is suitable for rapidly changing environments, but its accuracy is not as good as MVDR.

[0005] Therefore, how to provide a new technology to further optimize the existing DOA algorithm to improve the accuracy of target direction information estimation is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0006] In view of this, the present invention provides a method and system for horizontal array intelligent target azimuth estimation based on sub-array division to solve the problems existing in the background technology.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A method for estimating the azimuth of a horizontal array intelligent target based on subarray division, comprising:

[0009] Step 1: Divide the signal arrival direction into A sectors. Set the label as 1 when the target source exists in the sector, otherwise it is 0. The final label is represented as a one-dimensional matrix with a length of A dimensions and only one value is 1 and the other values ​​are 0.

[0010] Step 2: Randomly read a full line array data, and divide the full line array P into several sub-arrays [P1, P2, P3, .., P N ];

[0011] Step 3: Subarray data preprocessing: using two beamforming algorithms, CBF and MVDR, to process several subarrays into two beam domain characteristic single-frequency CBF output power P CBF (f, θ) and single-frequency diagonally loaded MVDR output power P Mv (f, θ), the two beam domain features are normalized and modulo-valued to obtain and As the input of the subarray DOA estimation model based on multi-stage fusion of multi-beam domain features;

[0012] Step 4: and Input it into the pre-trained subarray DOA estimation model based on multi-stage fusion of multi-beam domain features to obtain the DOA estimation results of each subarray of the full linear array.

[0013] Step 5: Perform steps 3 to 4 on all sub-arrays in step 2 to obtain the DOA estimation results of all sub-arrays [y1, y2, ..., y N ];

[0014] Step 6: All sub-array DOA estimation results [y1, y2, ..., y N ], several sub-array information [P1, P2, P3, .., P N ] and the array element parameters are input into the pre-trained conditional convolution DOA estimation model based on multi-source feature fusion to obtain the DOA estimation result of the full linear array.

[0015] Optionally, the step 3 specifically includes:

[0016] By scanning the existence range angles of several sub-array DOAs, the single-frequency CBF output power is obtained as follows:

[0017]

[0018] in is the weighted vector of CBF, obtained according to different weighting methods The values ​​are also different, for example, they can be uniformly weighted. R x is the cross-spectral density matrix of the received signal, R n With R x The unit noise matrix of the same size, SNR is the signal-to-noise ratio.

[0019] By scanning the existence range angles of several sub-array DOAs, the single-frequency diagonally loaded MVDR output power is obtained as follows:

[0020]

[0021] Where W MV (f, θ) is the optimal weight vector of MVDR, which can be obtained by using the Lagrange operator method. R x is the cross-spectral density matrix of the received signal, R n With R x The unit noise matrix of the same size, SNR is the signal-to-noise ratio, when no noise is added, R = R x +ρI, ρ is the diagonal loading, I is the size and R x Same identity matrix;

[0022] The beamforming output P CBF (f, θ) and P Mv (f, θ) is normalized and modulo value is obtained and Input into the subarray DOA estimation model based on multi-stage fusion of multi-beam domain features; the input dimension is N f ×N θ , as the input of the model, where N f is the number of frequency points, N θ is the number of scanning angles during beamforming, and the matrix form of the two inputs is

[0023]

[0024]

[0025] For each input matrix Corresponding to a label vector

[0026] Optionally, before processing in step 3, the following steps may be further included:

[0027] Modeling of a single full-line array receiving signal, K signals with center frequencies f S(t) = [s1(t), s2(t), ..., s K (t)] T Since the wave direction θ=[θ1,θ2,…,θ K ] T The signal is incident on a single uniform linear array. The number of sensors on the linear array is N, N>K, the array element spacing is d, the signal satisfies the far-field narrowband assumption, and the signal envelope does not change while passing through the array. X(t) = [x1(t), x2(t), ..., x K (t)] T For each array element receiving signal, the array receiving signal model is expressed as:

[0028] X(t)=B(f,θ)S(t)+N(t);

[0029] Among them, B(f, θ)=[a(f, θ1), a(f, θ2), .., a(f, θ K )] is the array manifold vector matrix, and the incoming wave direction is θ k The corresponding array manifold vector is a(f,θ k )=[exp(-jk T p1), exp(-jk T p2), ..., exp(-jk T p N )] T , the array element position is [p1, p2, ..., p N ], uniform linear array [p1, p2,,.., p N ]=[0,d,...,(N-1)d],k is the wave number vector,k=2πf / c·v(θ),b(θ) is the unit vector of the acoustic signal propagation direction,c is the signal propagation speed,N(t)=[n1(t),n2(t),...,n N (t)] T is the noise vector matrix received by each element of the linear array.

[0030] Optionally, step 4 specifically includes:

[0031] Step 4-1: Get The input is fed into a subarray DOA estimation model that fuses multi-beam domain features in multiple stages, and an encoder is used to encode the input features;

[0032] Step 4-2: Use two feature extraction paths to extract CBF beam domain features G respectively r and MVDR beam domain characteristics L r , the two feature extraction paths are composed of R cascade-designed downsampling modules, which are composed of Conv Block and Identity Block:

[0033] Step 4-3: For the CBF beam domain feature G obtained at each stage r and MVDR beam domain characteristics L r , the hierarchical feature fusion module is used to fuse the CBF beam domain features and MVDR beam domain features of different scales obtained at different stages, and the fused feature F at the R stage is obtained. R ;

[0034] Step 4-4: The feature F fused in the R stage R Input into the classifier to get the subarray DOA estimation result

[0035] Optionally, the encoder is specifically:

[0036]

[0037] Among them, e CBF , e MV Represents the low-level features obtained by encoding, ε represents the encoder, which consists of two-dimensional convolution Conv2d, one-dimensional convolution Conv1d, normalization layer LayerNorm, ReLU activation function, and maximum pooling MaxPool.

[0038] Optionally, the downsampling module is composed of a convolution block and an identity block, Conv Block is a convolution block, and IdentitvBlock is an identity block, and the specific structure is:

[0039] Conv Block(x)=ReLU(BatchNorm(Conv2d

[0040] (ReLU(BatchNorm(Conv2d(ReLU(BatchNorm(Conb2d(x))))))))+

[0041] BatchNorm(Conv2d(x)));

[0042] Identity Block=BatchNorm(Conv2d(ReLU(BatchNorm

[0043] (Conv2d(ReLU(BatchNorm(Conv2d(x))))))))+x;

[0044] in, represents the rth downsampling module for processing CBF features, r∈(1,R), represents the parameters of the rth downsampling module,

[0045] The specific extraction of CBF beam domain features and MVDR beam domain features is as follows:

[0046]

[0047]

[0048] Optionally, the specific calculation process of the hierarchical feature fusion module is as follows:

[0049] Type G r , L r and F r-1 , calculated by the following formula:

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056] in, represents element-wise multiplication, Generated by the channel attention mechanism, it represents the CBF pathway features with enhanced features in the channel dimension. Generated by the spatial attention mechanism, by paying attention to L r Different regions in the feature map are used to enhance the model's ability to extract MVDR pathway features. The F obtained from the previous stage of HFFM r-1 The intermediate features generated by downsampling, F r-1 Represents the fusion result of the hierarchical feature fusion module obtained in the previous stage; Conv1d represents a one-dimensional convolutional layer, Avgpo0l is an average pooling layer, Conv2d represents a two-dimensional convolutional layer, Concat represents a concatenation operation, GELU represents an activation function, and FC represents a linear layer; is the intermediate fusion result, composed of G r ,Lr , Calculated;

[0057] Will and Stitch together and pass Generate the fusion result F of the current stage r , It consists of the structure of an inverse residual multilayer perceptron.

[0058] Optionally, step 6 specifically includes:

[0059] Step 6-1: The calculated DOA estimation results of all sub-arrays [y1, y2, ..., y N ] and several sub-array information [P1, P2, P3, .., P N ] Input the backbone network and use two consecutive backbone modules to encode it to obtain the low-level encoding feature y in , where the backbone module consists of residual and convolution operations;

[0060] At the same time, the array element parameters, including the number of array elements S F and spacing S D Input to the first fully connected layer of the model to obtain the input feature f FC , where the fully connected layer contains n FC neurons, using ReLU activation function;

[0061] Step 6-2, f FC Input fully connected layer 1:

[0062] Fully connected layer 1 to f FC Processing to obtain the output feature f FC1 , where the fully connected layer 1 contains n FC1 neurons, using Sigmoid activation function;

[0063] f FC1 As the weight coefficient of the first convolution parameter generation block, for n FC1 The convolution kernels of the groups are respectively in n EXP1 The weighted summation is performed on the output channels to obtain a set of output channels with n EXP1 The convolution kernel of

[0064] In the conditional convolution layer 1, the convolution kernel is used to in Do convolution, the output is P CC1 ;

[0065] Step 6-3: Further, f FC Input fully connected layer 2:

[0066] Fully connected layer 2 pairs fFC Processing to obtain the output feature f FC2 , where the fully connected layer 2 contains n FC2 neurons, using Sigmoid activation function;

[0067] f FC2 As the weight coefficient of the second convolution parameter generation block, for n FC2 The convolution kernels of the groups are respectively in n EXP2 The weighted summation is performed on the output channels to obtain a set of output channels with n EXP2 The convolution kernel of

[0068] In the conditional convolution layer 2, the convolution kernel P is obtained. CC1 Do convolution, the output is P CC2 ;

[0069] Step 6-4, f FC Input fully connected layer 3:

[0070] Fully connected layer 3 pairs of f FC Processing to obtain the output feature f FC3 , where the fully connected layer 3 contains n FC3 neurons, using Sigmoid activation function;

[0071] f FC3 As the weight coefficient of the third convolution parameter generation block, for n FC3 The convolution kernels of the groups are respectively in n EXP3 The weighted summation is performed on the output channels to obtain a set of output channels with n EXP3 The convolution kernel of

[0072] In the conditional convolution layer 3, the convolution kernel P is obtained. CC2 Do convolution, the output is P CC3 ;

[0073] Step 6-5, f FC Input fully connected layer 4:

[0074] Fully connected layer 4 pairs of f FC Processing to obtain the output feature f FC4 , where the fully connected layer 4 contains n FC4 neurons, using Sigmoid activation function;

[0075] f FC4 As the weight coefficient of the fourth convolution parameter generation block, for n FC4 The convolution kernels of the groups are respectively in n EXP4 The weighted summation is performed on the output channels to obtain a set of output channels with n EXP4 The convolution kernel of

[0076] In the conditional convolution layer 4, the convolution kernel P is obtained. CC3 Do convolution, the output is P CC4 ;

[0077] Step 6-6, until f FC Input fully connected layer M:

[0078] The fully connected layer M is used for f FCM Processing to obtain the output feature f FCM , where FCM contains n FCM neurons, using Sigmoid activation function;

[0079] f FCM As the weight coefficient of the Mth convolution parameter generation block, for n FCM The convolution kernels of the groups are respectively in n EXPM The weighted summation is performed on the output channels to obtain a set of output channels with n EXPM The convolution kernel of

[0080] In the conditional convolution layer M, the convolution kernel P is used to obtain CC(M-1) Do convolution, the output is P CCM ;

[0081] Step 6-7, P CCM Input into the classifier to obtain the DOA estimation result of the entire linear array The classifier consists of a fully connected layer and a Softmax function, with P CCM As input, output

[0082] A horizontal array intelligent target azimuth estimation system based on subarray division, comprising:

[0083] The signal direction division module divides the signal arrival direction into A sectors. The label is set to 1 when the target source exists in the sector, otherwise it is set to 0. The final label is represented as a one-dimensional matrix with a length of A dimensions and only one value is 1 and the other values ​​are 0.

[0084] The subarray acquisition module randomly reads a full-line array data and divides the full-line array P into several subarrays [P1, P2, P3, .., P N ];

[0085] The subarray data preprocessing module uses two beamforming algorithms, CBF and MVDR, to process several subarrays into two beam domain characteristic single-frequency CBF output power P CBF (f, θ) and single-frequency diagonally loaded MVDR output power PMv(f, θ), the two beam domain characteristics are normalized and modulo-valued to obtain and As the input of the subarray DOA estimation model based on multi-stage fusion of multi-beam domain features;

[0086] Subarray DOA estimation module, and Input it into the pre-trained subarray DOA estimation model based on multi-stage fusion of multi-beam domain features to obtain the DOA estimation results of each subarray of the full linear array.

[0087] The subarray estimation result output module performs DOA estimation on all subarrays and obtains the DOA estimation results of all subarrays [y1, y2, ..., y N ];

[0088] The DOA estimation module of the full array takes all the DOA estimation results of the sub-arrays [y1, y2, ..., y N ], several sub-array information [P1, P2, P3, .., P N ] and the array element parameters are input into the pre-trained conditional convolution DOA estimation model based on multi-source feature fusion to obtain the DOA estimation result of the full linear array.

[0089] Through the above technical solutions, it can be known that compared with the prior art, the present invention discloses a method and system for intelligent target azimuth estimation of a horizontal array based on subarray division. First, a multi-stage fusion method is used to realize the fusion of multiple beamforming methods, thereby obtaining rich beam domain features and ensuring the accuracy of DOA estimation of a single subarray; then, the DOA estimation result, subarray information, and array element parameters of a single subarray are used as the input of a conditional convolution DOA estimation model based on multi-source feature fusion. The conditional convolution in the model can parameterize its convolution kernel, and dynamically generate convolution kernel parameters using a linear combination of several expert modules according to the input conditions, and adaptively adjust all convolution kernel parameters in the backbone network, and fuse this information to obtain the final DOA estimation result. The DOA estimation result obtained by the present invention not only integrates multiple beamforming methods, but also forms an adaptive subarray fusion based on subarray information and array element parameters, which can solve the problem of insufficient beam domain features generated by the current single beamforming method. In addition, by dividing the full linear array into multiple subarrays, the present method can effectively reduce multipath effects and noise interference, improve the estimation accuracy of target azimuth, and is suitable for target detection and positioning in complex environments. At the same time, the independence of the sub-arrays is utilized to achieve effective signal processing, thereby enhancing the robustness of the system in strong interference environments and ensuring the accuracy of target location information. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0091] Figure 1 A schematic diagram of a target position estimation method provided by the present invention;

[0092] Figure 2 A schematic diagram of the calculation of the data tag and target incident angle provided by the present invention;

[0093] Figure 3 A structural diagram of a subarray DOA estimation model based on multi-stage fusion of multi-beam domain features provided by the present invention;

[0094] Figure 4 Detailed diagrams of the encoder, convolution block and identity block modules in the subarray DOA estimation model provided by the present invention, wherein (1) is a detailed diagram of the encoder, (2) is a detailed diagram of the convolution block, and (3) is a detailed diagram of the identity block;

[0095] Figure 5 A detailed diagram of the HFFM module in the subarray DOA estimation model provided by the present invention;

[0096] Figure 6 The structural diagram of the conditional convolution DOA estimation model based on multi-source feature fusion provided by the present invention;

[0097] Figure 7 Detailed diagram of the backbone module in the DOA estimation model. DETAILED DESCRIPTION

[0098] 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.

[0099] The embodiment of the present invention discloses a method for estimating the azimuth of a horizontal array intelligent target based on subarray division, and its schematic diagram is shown in FIG. Figure 1 As shown, the method includes:

[0100] Step 1: The present invention defines the DOA estimation problem of the subarray and the full linear array as a classification problem. Figure 2As shown in the figure, the signal arrival direction is divided into A sectors. The label is 1 when the target source exists in the sector, otherwise it is 0. The final label is represented in the form of one-hot encoding, that is, a one-dimensional matrix with a length of A dimensions and only one value of 1 and other values ​​of 0.

[0101] Step 2: Randomly read a full line array data and divide the full line array P into several sub-arrays [P1, P2, P3, .., P N ], each sub-array can be overlapping or non-overlapping, which is set and adjusted according to the array division criteria and the actual situation of the array. For example, the division scheme can be: P1: [1, 2, 3, 4], P2: [5, 6, 7, 8], P3: [9, 10, 11, 12], P4: [13, 14, 15, 16];

[0102] Step 3: Subarray data preprocessing: using two beamforming algorithms, CBF and MVDR, to process the subarray data into two beam domain features P CBF (f,θ) and P MV (f,θ), the two beam domain features are normalized and modulo-valued to obtain and As the input of the subarray DOA estimation model based on multi-stage fusion of multi-beam domain features;

[0103] Step 4: and Input into the pre-trained subarray DOA estimation model based on multi-stage fusion multi-beam domain features. The model structure is as follows: Figure 3 As shown, the DOA estimation results of each sub-array of the full linear array are obtained.

[0104] Step 5: Perform steps 3 to 4 on all sub-arrays in step 2 to obtain the DOA estimation results of all sub-arrays [y1, y2, ..., y N ];

[0105] Step 6) replace [y1, y2, ..., y N ], subarray information [P1, P2, P3, .., P N ], array element parameters (including the number of array elements S F , spacing S D ) is input into the pre-trained conditional convolutional DOA estimation model based on multi-source feature fusion. The model structure is as follows Figure 6 As shown, the DOA estimation result of the full linear array is obtained;

[0106] Furthermore, step 3 specifically includes:

[0107] The present invention models a single full-line array receiving signal, assuming that K signals of center frequency f are S(t)=[s1(t), s2(t), ..., S K (t)] T Since the wave direction θ=[θ1,θ2,…,θ K ] T The signal is incident on a single uniform linear array. The number of sensors on the linear array is N (N>K), the array element spacing is d, the signal meets the far-field narrowband assumption, and the signal envelope does not change while passing through the array.

[0108] Let X(t)=[x1(t), x2(t),...,x K (t)] T The signal received by each array element generally includes the sound source signal and the noise signal after propagation through the channel. The array receiving signal model is expressed as

[0109] X(t)=B(f,θ)S(t)+N(t);

[0110] Among them, B(f, θ)=[a(f, θ1), a(f, θ2), .., a(f, θ K )] is the array manifold vector matrix, and the array manifold vector corresponding to the incoming wave direction θk is:

[0111] a(f,θ k )=[exp(-jk T p1), exp(-jk T p2), ..., exp(-jk T p N )] T , the array element position is [p1, p2, ..., p N ], uniform linear array [p1, p2, ..., p N ]=[0,d,...,(N-1)d],k is the wave number vector,k=2πffc·v(θ),v(θ) is the unit vector of the acoustic signal propagation direction,c is the signal propagation speed,N(t)=[n1(t),n2(t),...,n N (t)] T is the noise vector matrix received by each element of the linear array.

[0112] Furthermore, by scanning the possible existence range angle of DOA, the single-frequency CBF output power is obtained as

[0113]

[0114] in, is the weighted vector of CBF, obtained according to different weighting methods The values ​​are also different, for example, they can be uniformly weighted. R x is the cross-spectral density matrix of the received signal, R n With R x The same size unit matrix, SNR is the signal-to-noise ratio.

[0115] By scanning the possible existence range angle of DOA, the output power of the single-frequency diagonally loaded MVDR is obtained as follows:

[0116]

[0117] Among them, w MV (f,θ) is the optimal weight vector of MVDR, which can be obtained by Lagrange operator method. R x is the cross-spectral density matrix of the received signal, R n With R x The unit noise matrix of the same size, SNR is the signal-to-noise ratio, when no noise is added, R = R x +ρI, ρ is the diagonal loading, I is the size and R x Same identity matrix;

[0118] Furthermore, the beamforming output P CBF (f,θ) and P MV (f,θ) is normalized and modulo value is obtained and Input to the subarray DOA estimation model based on multi-stage fusion of multi-beam domain features. The input dimension is N f ×N θ , as the input of the model, where N f is the number of frequency points, N θ is the number of scanning angles during beamforming, and the matrix form of the two inputs is:

[0119]

[0120]

[0121] For each input matrix Corresponding to a label vector

[0122] Furthermore, step 4 specifically includes:

[0123] Step 4-1: The The input is fed into the subarray DOA estimation model based on multi-stage fusion of multi-beam domain features. The model first uses an encoder to encode the input features. The encoder structure is as follows: Figure 4 (1) As shown:

[0124]

[0125] Among them, e CBF , e MV Represents the low-level features obtained by encoding, ε represents the encoder, which consists of two-dimensional convolution, one-dimensional convolution, normalization layer, ReLU activation function, and maximum pooling;

[0126] Step 4-2: Further, two feature extraction paths are used to extract CBF beam domain features and MVDR beam domain features respectively. The two feature extraction paths are composed of R cascaded downsampling modules. The downsampling module is composed of a convolution block and an identity block. Its structure is as follows: Figure 4 (2) and Figure 4 (3) can be calculated by the following formula:

[0127]

[0128]

[0129] Conv Block(x)=ReLU(BatchNorm(Conv2d

[0130] (ReLU(BatchNorm(Conv2d(ReLU(BatchNorm(Conv2d(x))))))))+

[0131] BatchNorm(Conv2d(x)));

[0132] Identity Block=BatchNorm(Conv2d(ReLU(BatchNorm

[0133] (Conv2d(ReLU(BatchNorm(Conv2d(x))))))))+x;

[0134] in, represents the rth downsampling module for processing CBF features, r∈(1,R), Represents the parameters of the rth downsampling module;

[0135] Step 4-3: For each stage of G r and L r ,like Figure 5 As shown in Figure 2, the Hierarchical Feature Fusion Module (HFFM) is used to fuse the CBF features and MVDR features of different scales obtained at different stages.r-1 Indicates the fusion result of the HFFM module obtained in the previous stage, input G r , L r and F r-1 , can be calculated by the following formula:

[0136]

[0137]

[0138]

[0139]

[0140]

[0141]

[0142] in, represents element-wise multiplication, Generated by the channel attention mechanism, it represents the CBF pathway features with enhanced features in the channel dimension. Generated by the spatial attention mechanism, by paying attention to L r Different regions in the feature map are used to enhance the model's ability to extract MVDR pathway features. The F obtained from the previous stage of HFFM r-1 Intermediate features generated by downsampling;

[0143] Conv1d represents a one-dimensional convolutional layer, Avgp00l is an average pooling layer, Conv2d represents a two-dimensional convolutional layer, Concat represents a concatenation operation, GELU represents an activation function, and FC represents a linear layer;

[0144] is the intermediate fusion result, composed of G r , L r , Calculated;

[0145] Will and Stitch together and pass Generate the fusion result F of the current stage r ,

[0146] It consists of the structure of an inverse residual multilayer perceptron;

[0147] Step 4-4: Further, the feature F fused in the Rth stage R Input into the classifier to get the subarray DOA estimation result

[0148] Furthermore, step 6 specifically includes:

[0149] Step 6-1: Sub-array DOA estimation results calculated in step 5 are converted into [y1, y2, ..., y N ] and subarray information [P1, P2, P3, .., P N ] Input the backbone network and use two consecutive backbone modules to encode it to obtain the low-level encoding feature y in , where the backbone module consists of residual and convolution operations;

[0150] At the same time, the array element parameters (including the number of array elements S F , spacing S D ) is input to the first fully connected layer of the model to obtain the input feature f FC , where the fully connected layer contains n FC neurons, using ReLU activation function;

[0151] Step 6-2: Further, f FC Input fully connected layer 1:

[0152] Step 6-2-1, fully connected layer 1 to f FC Processing to obtain the output feature f FC1 , where the fully connected layer 1 contains n FC1 neurons, using Sigmoid activation function;

[0153] Step 6-2-2, f FC1 As the weight coefficient of the first convolution parameter generation block, for n FC1 The convolution kernels of the groups are respectively in n EXP1 The weighted summation is performed on the output channels to obtain a set of output channels with the number of nE XP1 The convolution kernel of

[0154] Step 6-2-3: In the conditional convolution layer 1, use the obtained convolution kernel to check P in (Input sub-matrix information) is convolved and the output is P CC1 ;

[0155] Step 6-3: Further, f FC Input fully connected layer 2:

[0156] Step 6-3-1, fully connected layer 2 to f FC Processing to obtain the output feature f FC2 , where the fully connected layer 2 contains n FC2 neurons, using Sigmoid activation function;

[0157] Step 6-3-2, f FC2As the weight coefficient of the second convolution parameter generation block, for n FC2 The convolution kernels of the groups are respectively in n EXP2 The weighted summation is performed on the output channels to obtain a set of output channels with n EXP2 The convolution kernel of

[0158] Step 6-3-3: Use the obtained convolution kernel to check P in the conditional convolution layer 2 CC1 Do convolution, the output is P CC2 ;

[0159] Step 6-4: Further, f FC Input fully connected layer 3:

[0160] Step 6-4-1, fully connected layer 3 pairs f FC Processing to obtain the output feature f FC3 , where the fully connected layer 3 contains n FC3 neurons, using Sigmoid activation function;

[0161] Step 6-4-2, f FC3 As the weight coefficient of the third convolution parameter generation block, for n FC3 The convolution kernels of the groups are respectively in n EXP3 The weighted summation is performed on the output channels to obtain a set of output channels with n EXP3 The convolution kernel of

[0162] Step 6-4-3: Use the obtained convolution kernel to check P in the conditional convolution layer 3 CC2 Do convolution, the output is P CC3 ;

[0163] Step 6-5: Further, f FC Input fully connected layer 4:

[0164] Step 6-5-1, fully connected layer 4 pairs f FC Processing to obtain the output feature f FC4 , where the fully connected layer 4 contains n FC4 neurons, using Sigmoid activation function;

[0165] Step 6-5-2, f FC4 As the weight coefficient of the fourth convolution parameter generation block, for n FC4 The convolution kernels of the groups are respectively in n EXP4 The weighted summation is performed on the output channels to obtain a set of output channels with n EXP4 The convolution kernel of

[0166] Step 6-5-3: Use the obtained convolution kernel to check P in the conditional convolution layer 4 CC3 Do convolution, the output is P CC4;

[0167] Step 6-6, until f FC Input fully connected layer M:

[0168] Step 6-6-1, fully connected layer M to f FCM Processing to obtain the output feature f FCM , where FCM contains n FCM neurons, using Sigmoid activation function;

[0169] Step 6-6-2, f FCM As the weight coefficient of the Mth convolution parameter generation block, for n FCM The convolution kernels of the groups are respectively in n EXPM The weighted summation is performed on the output channels to obtain a set of output channels with n EXPM The convolution kernel of

[0170] Step 6-6-3: Use the obtained convolution kernel to check P in the conditional convolution layer M CC(M-1) Do convolution, the output is P CCM ;

[0171] Step 6-7, P CCM Input into the classifier to obtain the DOA estimation result of the entire linear array The classifier consists of a fully connected layer and a Softmax function, with P CCM As input, output The details of the backbone module in the DOA estimation model are shown in the figure Figure 7 shown.

[0172] A horizontal array intelligent target azimuth estimation system based on subarray division, comprising:

[0173] The signal direction division module divides the signal arrival direction into A sectors. The label is set to 1 when the target source exists in the sector, otherwise it is set to 0. The final label is represented as a one-dimensional matrix with a length of A dimensions and only one value is 1 and the other values ​​are 0.

[0174] The subarray acquisition module randomly reads a full-line array data and divides the full-line array P into several subarrays [P1, P2, P3, .., P N ];

[0175] The subarray data preprocessing module uses two beamforming algorithms, CBF and MVDR, to process several subarrays into two beam domain characteristic single-frequency CBF output power P CBF (f, θ) and single-frequency diagonally loaded MVDR output power PMv(f, θ), the two beam domain characteristics are normalized and modulo-valued to obtain and As the input of the subarray DOA estimation model based on multi-stage fusion of multi-beam domain features;

[0176] Subarray DOA estimation module, and Input it into the pre-trained subarray DOA estimation model based on multi-stage fusion of multi-beam domain features to obtain the DOA estimation results of each subarray of the full linear array.

[0177] The subarray estimation result output module performs DOA estimation on all subarrays and obtains the DOA estimation results of all subarrays [y1, y2, ..., y N ];

[0178] The DOA estimation module of the full array takes all the DOA estimation results of the sub-arrays [y1,y2,...,y N ], several sub-array information [P1, P2, P3, .., P N ] and the array element parameters are input into the pre-trained conditional convolution DOA estimation model based on multi-source feature fusion to obtain the DOA estimation result of the full linear array.

[0179] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0180] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for estimating the azimuth of a horizontal array intelligent target based on subarray division, characterized in that: include: Step 1: Divide the signal arrival direction into A sectors. Set the label as 1 when the target source exists in the sector, otherwise it is 0. The final label is represented as a one-dimensional matrix with a length of A dimensions and only one value is 1 and the other values ​​are 0. Step 2: Randomly read a full line array data and divide the full line array P into several sub-arrays [P1, P2, P3, .., P N ]; Step 3: Subarray data preprocessing: using two beamforming algorithms, CBF and MVDR, to process several subarrays into two beam domain characteristic single-frequency CBF output power P CBF (f,θ) and single-frequency diagonally loaded MVDR output power P MV (f,θ), the two beam domain features are normalized and modulo-valued to obtain and As the input of the subarray DOA estimation model based on multi-stage fusion of multi-beam domain features; Step 4: and Input it into the pre-trained subarray DOA estimation model based on multi-stage fusion of multi-beam domain features to obtain the DOA estimation results of each subarray of the full linear array. Step 5: Perform steps 3 to 4 on all sub-arrays in step 2 to obtain the DOA estimation results of all sub-arrays [y1, y2, ..., y N ]; Step 6: All sub-array DOA estimation results [y1, y2, ..., y N ], several sub-array information [P1, P2, P3, .., P N ] and the array element parameters are input into the pre-trained conditional convolution DOA estimation model based on multi-source feature fusion to obtain the DOA estimation result of the full linear array.

2. According to the method for horizontal array intelligent target azimuth estimation based on subarray division according to claim 1, it is characterized in that: Before processing in step 3, it also includes: Modeling a single full-line array receiving signal, K signals with center frequencies f S(t) = [s1(t), s2(t), ..., s K (t)] T Since the wave direction θ=[θ1,θ2,...,θ K ] T The signal is incident on a single uniform linear array. The number of sensors on the linear array is N, N>K, the array element spacing is d, the signal satisfies the far-field narrowband assumption, and the signal envelope does not change while passing through the array. X(t)=[x1(t),x2(t),...,x K (t)] T For each array element receiving signal, the array receiving signal model is expressed as: X(t)=B(f,θ)S(t)+N(t); Among them, B(f,θ)=[a(f,θ1),a(f,θ2),..,a(f,θ K )] is the array manifold vector matrix, and the incoming wave direction is θ k The corresponding array manifold vector is a(f,θ k )=[exp(-jk T p1),exp(-jk T p2),...,exp(-jk T p N )] T , the array element position is [p1,p2,...,p N ], uniform linear array [p1,p2,...,p N ]=[0,d,...,(N-1)d], k is the wave number vector, k=2πf / c·v(θ), v(θ) is the unit vector of the acoustic signal propagation direction, c is the signal propagation speed, N(t)=[n1(t),n2(t),...,n N (t)] T is the noise vector matrix received by each element of the linear array.

3. The method for horizontal array intelligent target azimuth estimation based on subarray division according to claim 2 is characterized in that: The step 3 specifically includes: By scanning the existence range angles of several sub-array DOAs, the single-frequency CBF output power is obtained as follows: in, is the weight vector of CBF, R x is the cross-spectral density matrix of the received signal, R n With R x The unit noise matrix of the same size, SNR is the signal-to-noise ratio. By scanning the existence range angles of several sub-array DOAs, the single-frequency diagonally loaded MVDR output power is obtained as follows: Among them, w MV (f,θ) is the optimal weight vector of MVDR, R x is the cross-spectral density matrix of the received signal, R n With R x The unit noise matrix of the same size, SNR is the signal-to-noise ratio, when no noise is added, R = R x +ρI, ρ is the diagonal loading, I is the size and R x Same identity matrix; The beamforming output P CBF (f,θ) and P MV (f,θ) is normalized and modulo value is obtained and Input into the subarray DOA estimation model based on multi-stage fusion of multi-beam domain features; the input dimension is N f ×N θ , as the input of the model, where N f is the number of frequency points, N θ is the number of scanning angles during beamforming, and the matrix form of the two inputs is For each input matrix Corresponding to a label vector 4. The method for horizontal array intelligent target azimuth estimation based on subarray division according to claim 1 is characterized in that: The step 4 specifically includes: Step 4-1: Get The input is input into a subarray DOA estimation model based on multi-stage fusion of multi-beam domain features, and an encoder is used to encode the input features; Step 4-2: Use two feature extraction paths to extract CBF beam domain features G respectively r and MVDR beam domain characteristics L r , the two feature extraction paths consist of R cascaded downsampling modules, which consist of convolution blocks and identity blocks: Step 4-3: For the CBF beam domain feature G obtained in each stage r and MVDR beam domain characteristics L r , the hierarchical feature fusion module is used to fuse the CBF beam domain features and MVDR beam domain features of different scales obtained at different stages, and the fused feature F at the R stage is obtained. R ; Step 4-4: The feature F fused in the R stage R Input into the classifier to get the subarray DOA estimation result 5. The method for horizontal array intelligent target azimuth estimation based on subarray division according to claim 4 is characterized in that: The encoder is specifically: where e CBF ,e MV Represents the low-level features obtained by encoding, ε represents the encoder, which consists of two-dimensional convolution Conv2d, one-dimensional convolution Conv1d, normalization layer LayerNorm, ReLU activation function, and maximum pooling MaxPool.

6. The method for horizontal array intelligent target azimuth estimation based on subarray division according to claim 4 is characterized in that: The downsampling module is composed of a convolution block and an identity block, Conv Block is a convolution block, and Identity Block is an identity block. The specific structure is as follows: Conv Block(x)=ReLU(BatchNorm(Conv2d (ReLU(BatchNorm(Conv2d(ReLU(BatchNorm(Conv2d(x))))))))+BatchNorm(Conv2d(x))); Identity Block=BatchNorm(Conv2d(ReLU(BatchNorm(Conv2d(ReLU(BatchNorm(Conv2d(x))))))))+x; in, represents the rth downsampling module for processing CBF features, r∈(1,R), represents the parameters of the rth downsampling module, The specific extraction of CBF beam domain features and MVDR beam domain features is as follows:

7. The method for horizontal array intelligent target azimuth estimation based on subarray division according to claim 1 is characterized in that: The specific calculation process of the hierarchical feature fusion module is as follows: Type G r , L r and F r-1 , calculated by the following formula: in, represents element-wise multiplication, Generated by the channel attention mechanism, it represents the CBF pathway features with enhanced features in the channel dimension. Generated by the spatial attention mechanism, by paying attention to L r Different regions in the feature map are used to enhance the model's ability to extract MVDR pathway features. F obtained from the previous stage of the hierarchical feature fusion module r-1 The intermediate features generated by downsampling, F r-1 Indicates the fusion result of the hierarchical feature fusion module obtained in the previous stage; Conv1d represents a one-dimensional convolutional layer, Avgpool is an average pooling layer, Conv2d represents a two-dimensional convolutional layer, Concat represents a concatenation operation, GELU represents an activation function, and FC represents a linear layer; is the intermediate fusion result, composed of G r ,L r , Calculated; Will and Stitch together and pass Generate the fusion result F of the current stage r , It consists of the structure of an inverse residual multilayer perceptron.

8. The method for horizontal array intelligent target azimuth estimation based on subarray division according to claim 1 is characterized in that: The step 6 specifically includes: Step 6-1: Compute all the sub-array DOA estimation results [y1, y2, ..., y N ] and several sub-array information [P1, P2, P3, .., P N ] Input the backbone network and use two consecutive backbone modules to encode it to obtain the low-level encoding feature y in , where the backbone module consists of residual and convolution operations; At the same time, the array element parameters, including the number of array elements S F and spacing S D Input to the first fully connected layer of the model to obtain the input feature f FC , where the fully connected layer contains n FC neurons, using ReLU activation function; Step 6-2, f FC Input fully connected layer 1: Fully connected layer 1 to f FC Processing to obtain the output feature f FC1 , where the fully connected layer 1 contains n FC1 neurons, using Sigmoid activation function; f FC1 As the weight coefficient of the first convolution parameter generation block, for n FC1 The convolution kernels of the groups are respectively in n EXP1 The weighted summation is performed on the output channels to obtain a set of output channels with n EXP1 The convolution kernel of In the conditional convolution layer 1, the convolution kernel is used to in Do convolution, the output is P CC1 ; Step 6-3: Further, f FC Input fully connected layer 2: Fully connected layer 2 pairs f FC Processing to obtain the output feature f FC2 , where the fully connected layer 2 contains n FC2 neurons, using Sigmoid activation function; f FC2 As the weight coefficient of the second convolution parameter generation block, for n FC2 The convolution kernels of the groups are respectively in n EXP2 The weighted summation is performed on the output channels to obtain a set of output channels with n EXP2 The convolution kernel of In the conditional convolution layer 2, the convolution kernel P is obtained. CC1 Do convolution, the output is P CC2 ; Step 6-4, f FC Input fully connected layer 3: Fully connected layer 3 pairs of f FC Processing to obtain the output feature f FC3 , where the fully connected layer 3 contains n FC3 neurons, using Sigmoid activation function; f FC3 As the weight coefficient of the third convolution parameter generation block, for n FC3 The convolution kernels of the groups are respectively in n EXP3 The weighted summation is performed on the output channels to obtain a set of output channels with n EXP3 The convolution kernel of In the conditional convolution layer 3, the convolution kernel P is obtained. CC2 Do convolution, the output is P CC3 ; Step 6-5, f FC Input fully connected layer 4: Fully connected layer 4 pairs of f FC Processing to obtain the output feature f FC4 , where the fully connected layer 4 contains n FC4 neurons, using Sigmoid activation function; f FC4 As the weight coefficient of the fourth convolution parameter generation block, for n FC4 The convolution kernels of the groups are respectively in n EXP4 The weighted summation is performed on the output channels to obtain a set of output channels with n EXP4 The convolution kernel of In the conditional convolution layer 4, the convolution kernel P is obtained. CC3 Do convolution, the output is P CC4 ; Step 6-6, until f FC Input fully connected layer M: The fully connected layer M is used for f FCM Processing to obtain the output feature f FCM , where FCM contains n FCM neurons, using Sigmoid activation function; f FCM As the weight coefficient of the Mth convolution parameter generation block, for n FCM The convolution kernels of the groups are respectively in n EXPM The weighted summation is performed on the output channels to obtain a set of output channels with n EXPM The convolution kernel of In the conditional convolution layer M, the convolution kernel P is used to obtain CC(M-1) Do convolution, the output is P CCM ; Step 6-7, P CCM Input into the classifier to obtain the DOA estimation result of the entire linear array The classifier consists of a fully connected layer and a Softmax function, with P CCM As input, output 9. A horizontal array intelligent target azimuth estimation system based on subarray division, characterized in that: A horizontal array intelligent target azimuth estimation method based on subarray division is applied, comprising: The signal direction division module divides the signal arrival direction into A sectors. The label is set to 1 when the target source exists in the sector, otherwise it is set to 0. The final label is represented as a one-dimensional matrix with a length of A dimensions and only one value is 1 and the other values ​​are 0. The subarray acquisition module randomly reads a full-line array data and divides the full-line array P into several subarrays [P1, P2, P3, .., P N ]; The subarray data preprocessing module uses two beamforming algorithms, CBF and MVDR, to process several subarrays into two beam domain characteristic single-frequency CBF output power P CBF (f,θ) and single-frequency diagonally loaded MVDR output power P MV (f,θ), the two beam domain features are normalized and modulo-valued to obtain and As the input of the subarray DOA estimation model based on multi-stage fusion of multi-beam domain features; Subarray DOA estimation module, and Input it into the pre-trained subarray DOA estimation model based on multi-stage fusion of multi-beam domain features to obtain the DOA estimation results of each subarray of the full linear array. The subarray estimation result output module performs DOA estimation on all subarrays and obtains the DOA estimation results of all subarrays [y1, y2, ..., y N ]; The DOA estimation module of the full array takes all the DOA estimation results of the sub-arrays [y1,y2,...,y N ], several sub-array information [P1, P2, P3, .., P N ] and the array element parameters are input into the pre-trained conditional convolution DOA estimation model based on multi-source feature fusion to obtain the DOA estimation result of the full linear array.

Images