A sparse linear array direction of arrival estimation method based on a NaViT network model

Abstract

The application discloses a sparse linear array wave direction estimation method based on a NaViT network model, and comprises the following steps: defining a sparse linear sensor array, constructing a received signal vector of the sparse linear sensor array according to a received signal of each array element, and obtaining a sparse array covariance matrix; performing flat processing on the sparse array covariance matrix to generate a matrix patch, and mapping the matrix patch to a D latent space, presetting a special mark to form an entire covariance matrix; using position embedding to express a position description of the matrix patch according to position information of the matrix patch in the entire covariance matrix to obtain an input vector of position coding; constructing a NaViT network model, inputting the input vector of position coding and the sparse array covariance matrix into the NaViT network model, and using an attention layer and an encoder of the NaViT network model to identify a wave direction of a received signal of the sparse linear array, so that the wave direction of the received signal is obtained.

Description

Technical Field

[0001] This invention relates to the field of signal processing, and specifically to a method for estimating the direction of arrival (DOA) of a sparse linear array based on the NaViT network model. Background Technology

[0002] Direction of arrival (DOA) estimation is a crucial aspect of sensor array signal processing. Compared to uniform linear arrays, sparse linear arrays (SLAs) have become a popular array configuration for DOA estimation due to their increased degrees of freedom. To improve algorithm performance and reduce computational complexity, deep learning has been widely applied to DOA estimation in recent years. However, most existing methods employ convolutional neural networks (CNNs), which are unable to effectively learn the global features of the input data.

[0003] Most deep learning-based direction-of-arrival (DOA) estimation methods exhibit superior performance compared to traditional model-driven algorithms. However, most traditional deep learning algorithms are based on CNNs. CNNs capture features through local receptive fields, but DOA information often exists within the entire covariance matrix of the received signal. Therefore, CNNs still have certain limitations in modeling global information. Summary of the Invention

[0004] To address the aforementioned shortcomings of existing technologies, this invention provides a method for estimating the direction of arrival (DOA) of sparse linear arrays based on the NaViT network model.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0006] A method for estimating the direction of arrival (DOA) of a sparse linear array based on the NaViT network model is provided, which includes the following steps:

[0007] S1: Define a sparse linear sensor array, construct the received signal vector of the sparse linear sensor array based on the received signal of each array element, and obtain the sparse array covariance matrix.

[0008] S2: Flatten the covariance matrix of the sparse array to generate a matrix patch, and map the matrix patch to a... D A latent space is defined by a special label, which forms the entire covariance matrix.

[0009] S3: Based on the position information of the matrix patch in the entire covariance matrix, position embedding is used to express the position description of the matrix patch, and the position encoded input vector is obtained;

[0010] S4: Construct a NaViT network model, input the position-encoded input vector and the sparse array covariance matrix into the NaViT network model, and use the attention layer and encoder of the NaViT network model to identify and classify the direction of arrival of the received signal of the sparse linear array to obtain the direction of arrival of the received signal.

[0011] Further, step S1 includes:

[0012] S11: Define the number of elements in the sensor array. M Array element spacing d , K Let the signal source that conforms to the spatial far-field narrowband hypothesis be denoted as . , For the first K One signal source, K The number of signal sources;

[0013] S12: Based on the first array element, each received signal is represented as follows: , No. m Received signal of each array element Represented as:

[0014] ;

[0015] in, For the first i One received signal, i The number of the received signal. The wavelength of the received signal. m The array element's number. For the first i One signal source, For the first m The noise of each array element n Patch length;

[0016] S13: Construct the received signal vector of the sensor array;

[0017] ;

[0018] ;

[0019] ;

[0020] ;

[0021] in, For the received signal vector, For the received signal vector, For noise vectors, The direction matrix, For the firstK The steering vector of the nth signal source, for the nth signal source i The steering vector of the received signal Represented as:

[0022] ;

[0023] for M A sparse array of n elements, the nth k The steering vector of each signal source is represented as:

[0024] ;

[0025] in, For the elements of a sparse array, the set of elements of a sparse array. It is a subset of a uniform linear array;

[0026] The turning vector of a sparse array is represented as a part of a uniform linear array;

[0027] ;

[0028] in, The selection matrix for the sparse array;

[0029] S14: Write the received signal vector of the sensor array as a signal matrix. :

[0030] ;

[0031] ;

[0032] in, For the received signal matrix, For the noise matrix, For the first T The received signal corresponding to each snapshot;

[0033] S15: The direction of arrival (DOA) information is estimated using the covariance matrix of the received signal. The sparse array covariance matrix is ​​also known as the array autocorrelation matrix. Represented as:

[0034] ;

[0035] in, To receive the signal covariance matrix, It is the identity matrix. This represents the noise power of the sparse array.

[0036] Further, step S2 includes:

[0037] S21: The covariance matrix of the sparse array The covariance matrix is ​​split into three channels: real, imaginary, and phase, and is a sparse array. Shape changed , L The number of elements in a sparse array; n Different covariance matrices Segment and flatten into sequence matrices , p The patch size for each sequence matrix. The number of patches generated. P Patch length;

[0038] S22: Flatten the matrix patch into a one-dimensional vector, and then use a trainable linear projection to map these flattened vectors to a... D A dimensional potential space, where each matrix patch is represented as a... D dimensional vector;

[0039] ;

[0040] in, It is the first n A sequence of covariance matrices, To represent the learnable transformation matrix of a linear layer in matrix form, For the first N One matrix patch;

[0041] S23: Will n Different covariance matrices Concatenate them into a long input vector ;

[0042] ;

[0043] S24: In long input vectors A special mark is pre-set in the middle As input for subsequent extraction of direction-of-arrival information, special markers are used to capture the covariance matrix and global information features. Add to long input vector The middle part serves as the entire covariance matrix. ;

[0044] ;

[0045] Among them, matrix patch For the entire covariance matrix The matrix block in the middle.

[0046] Further, step S3 includes:

[0047] S31: Based on the entire covariance matrix The location information of the matrix patch is expressed using position embedding;

[0048] ;

[0049] in, pos The position of the matrix block. u For the entire covariance matrix latitude, For location embedding;

[0050] S32: Convert the entire covariance matrix With position embedding Adding them together yields the position-encoded input vector. ;

[0051] ;

[0052] in, The position vector of the matrix block, position embedding Position vector Element.

[0053] Further, step S4 includes:

[0054] S41: Construct the NaViT network model, and use the sparse array covariance matrix and position-encoded input vector The discriminator in the NaViT network model obtains the input vector. query matrix Key matrix Sum matrix ;

[0055] S42: The covariance matrix of the sparse array and position-encoded input vector Copy to query matrix Key matrix Sum matrix Previously, a linear transformation was performed using an attention layer, and the attention function of the NaViT network model was:

[0056] ;

[0057] ;

[0058] in, Query matrix Key matrix Sum matrix The linear transformation matrix, Let be the attention function. For the output of the attention layer;

[0059] S43: Connect multiple attention layers, multiply by a weight vector, and obtain the encoder output of the NaViT network model;

[0060] ;

[0061] ;

[0062] in, For the encoder output, A connection function for the output sequences of multiple attention layers. This is the weight vector;

[0063] S44: Output of encoder As a classification feature signal, the classification feature signal is classified to obtain the direction of arrival classification signal of the sparse linear array.

[0064] The beneficial effects of this invention are as follows: This invention utilizes the covariance matrix of the received signal as input, with the real part, imaginary part, and phase of the covariance matrix serving as three input channels. Under single signal source and low signal-to-noise ratio conditions, the recognition accuracy is superior to traditional methods. By dividing the matrix into smaller matrix blocks and packaging them into time series for classification, the computation speed is improved by more than four times. The network structure exhibits stable performance and good estimation results. Attached Figure Description

[0065] Figure 1 This is a flowchart of a method for estimating the direction of arrival (DOA) of a sparse linear array based on the NaViT network model. Detailed Implementation

[0066] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0067] like Figure 1 As shown, a method for estimating the direction of arrival (DOA) of a sparse linear array based on the NaViT network model includes the following steps:

[0068] S1: Define a sparse linear sensor array. Construct the received signal vector of the sparse linear sensor array based on the received signal of each array element, and obtain the sparse array covariance matrix. Step S1 specifically includes:

[0069] S11: Define the number of elements in the sensor array. M Array element spacing d , K Let the signal source that conforms to the spatial far-field narrowband hypothesis be denoted as . , For the first K One signal source, K The number of signal sources;

[0070] S12: Based on the first array element, each received signal is represented as follows: , No. m Received signal of each array element Represented as:

[0071] ;

[0072] in, For the first i One received signal, i The number of the received signal. The wavelength of the received signal. m The array element's number. For the first i One signal source, For the first m The noise of each array element n Patch length;

[0073] S13: Construct the received signal vector of the sensor array;

[0074] ;

[0075] ;

[0076] ;

[0077] ;

[0078] in, For the received signal vector, For the received signal vector, For noise vectors, The direction matrix, For the first K The steering vector of the nth signal source, for the nth signal source i The steering vector of the received signal Represented as:

[0079] ;

[0080] for M A sparse array of n elements, the nthk The steering vector of each signal source is represented as:

[0081] ;

[0082] in, For the elements of a sparse array, the set of elements of a sparse array. It is a subset of a uniform linear array;

[0083] The turning vector of a sparse array is represented as a part of a uniform linear array;

[0084] ;

[0085] in, The selection matrix for the sparse array;

[0086] S14: Write the received signal vector of the sensor array as a signal matrix. :

[0087] ;

[0088] ;

[0089] in, For the received signal matrix, For the noise matrix, For the first T The received signal corresponding to each snapshot;

[0090] S15: The direction of arrival (DOA) information is estimated using the covariance matrix of the received signal. The sparse array covariance matrix is ​​also known as the array autocorrelation matrix. Represented as:

[0091] ;

[0092] in, To receive the signal covariance matrix, It is the identity matrix. Let represent the noise power of the sparse array. Compared to the covariance matrix of a uniform array, the covariance matrix structure of a sparse array is altered, with the contribution of missing elements reflected as zero in the corresponding rows and columns.

[0093] S2: Flatten the covariance matrix of the sparse array to generate a matrix patch, and map the matrix patch to a... D A latent space is defined, pre-defined with a special label, forming the entire covariance matrix. Step S2 specifically includes:

[0094] S21: The covariance matrix of the sparse array The covariance matrix is ​​split into three channels: real, imaginary, and phase, and is a sparse array. Shape changed , L Let be the number of elements in the sparse array; to improve computational efficiency, n Different covariance matrices Segment and flatten into sequence matrices , p The patch size for each sequence matrix. The number of patches generated. P Patch length;

[0095] S22: Flatten the matrix patch into a one-dimensional vector, and then use a trainable linear projection to map these flattened vectors to a... D A dimensional potential space, where each matrix patch is represented as a... D dimensional vector;

[0096] ;

[0097] in, It is the first n A sequence of covariance matrices, To represent the learnable transformation matrix of a linear layer in matrix form, For the first N One matrix patch;

[0098] S23: Will n Different covariance matrices Concatenate them into a long input vector ;

[0099] ;

[0100] S24: During processing, long input vectors can capture global information and features of the covariance matrix. A special mark is pre-set in the middle As input for subsequent extraction of direction-of-arrival information, special markers are used to capture the covariance matrix and global information features. Add to long input vector The middle part serves as the entire covariance matrix. ;

[0101] ;

[0102] Among them, matrix patch For the entire covariance matrix The matrix block in the middle.

[0103] S3: To preserve the positional order of each matrix patch relative to the original matrix and ensure the model can understand the order relationships in the data, we introduce positional information, similar to the standard transformer model. Based on the positional information of the matrix patches in the entire covariance matrix, positional embedding is used to express the positional description of the matrix patches, resulting in a position-encoded input vector. Step S3 specifically includes:

[0104] S31: Based on the entire covariance matrix The location information of the matrix patch is expressed using position embedding;

[0105] ;

[0106] in, pos The position of the matrix block. u For the entire covariance matrix latitude, For position embedding, to improve the discriminability of position embeddings in odd and even dimensions, we used sine and cosine functions as encoders, respectively. The sine function was chosen as the encoder because it allows the model to easily learn to locate relative to its position, since any fixed offset can be represented by a linear function of the position information.

[0107] S32: Convert the entire covariance matrix With position embedding Adding them together yields the position-encoded input vector. ;

[0108] ;

[0109] in, The position vector of the matrix block, position embedding Position vector Element.

[0110] S4: Construct a NaViT network model. Input the position-encoded input vector and the sparse array covariance matrix into the NaViT network model. Use the attention layer and encoder of the NaViT network model to classify the direction of arrival (DOA) of the received signal from the sparse linear array, thus obtaining the DOA of the received signal. Step S4 specifically includes:

[0111] S41: Construct the NaViT network model, and use the sparse array covariance matrix and position-encoded input vector The discriminator in the NaViT network model obtains the input vector. query matrix Key matrix Sum matrix ;

[0112] S42: The covariance matrix of the sparse array and position-encoded input vector Copy to query matrix Key matrix Sum matrix Previously, a linear transformation was performed using an attention layer, and the attention function of the NaViT network model was:

[0113] ;

[0114] ;

[0115] in, Query matrix Key matrix Sum matrix The linear transformation matrix, Let be the attention function. For the output of the attention layer;

[0116] S43: Connect multiple attention layers, multiply by a weight vector, and obtain the encoder output of the NaViT network model;

[0117] ;

[0118] ;

[0119] in, For the encoder output, A connection function for the output sequences of multiple attention layers. This is the weight vector;

[0120] S44: Output of encoder As a classification feature signal, the classification feature signal is classified to obtain the direction of arrival classification signal of the sparse linear array received signal, and thus the direction of arrival of the received signal is obtained.

[0121] This invention utilizes the covariance matrix of the received signal as input, with the real part, imaginary part, and phase of the covariance matrix serving as three input channels. Under single-signal source and low signal-to-noise ratio conditions, it achieves superior recognition accuracy compared to traditional methods. By dividing the matrix into smaller matrix blocks and packaging them into time series for classification, the computation speed is increased by more than four times. The network structure exhibits stable performance and good estimation results.

Claims

1. A method for estimating the direction of arrival (DOA) of a sparse linear array based on a NaViT network model, characterized in that, Includes the following steps: S1: Define a sparse linear sensor array, construct the received signal vector of the sparse linear sensor array based on the received signal of each array element, and obtain the sparse array covariance matrix. S2: Flatten the covariance matrix of the sparse array to generate a matrix patch, and map the matrix patch to a... D A latent space is defined by a special label, which forms the entire covariance matrix. S3: Based on the position information of the matrix patch in the entire covariance matrix, position embedding is used to express the position description of the matrix patch, and the position encoded input vector is obtained; S4: Construct a NaViT network model, input the position-encoded input vector and the sparse array covariance matrix into the NaViT network model, and use the attention layer and encoder of the NaViT network model to identify and classify the direction of arrival of the received signal of the sparse linear array to obtain the direction of arrival of the received signal. Step S4 includes: S41: Construct the NaViT network model, and use the sparse array covariance matrix and position-encoded input vector The discriminator in the NaViT network model obtains the input vector. query matrix Key matrix Sum matrix ; S42: The covariance matrix of the sparse array and position-encoded input vector Copy to query matrix Key matrix Sum matrix Previously, a linear transformation was performed using an attention layer, and the attention function of the NaViT network model was: ； ； in, Query matrix Key matrix Sum matrix The linear transformation matrix, Let be the attention function. For the output of the attention layer; S43: Connect multiple attention layers, multiply by a weight vector, and obtain the encoder output of the NaViT network model; ； ； in, For the encoder output, A connection function for the output sequences of multiple attention layers. This is the weight vector; S44: Output of encoder As a classification feature signal, the classification feature signal is classified to obtain the direction of arrival classification signal of the sparse linear array.

2. The method for estimating the direction of arrival (DOA) of a sparse linear array based on the NaViT network model according to claim 1, characterized in that, Step S1 includes: S11: Define the number of elements in the sensor array. M Array element spacing d , K Let the signal source that conforms to the spatial far-field narrowband hypothesis be denoted as . , For the first K One signal source, K The number of signal sources; S12: Based on the first array element, each received signal is represented as follows: , No. m Received signal of each array element Represented as: ； in, For the first i One received signal, i The number of the received signal. The wavelength of the received signal. m The array element's number. For the first i One signal source, For the first m The noise of each array element n Patch length; S13: Construct the received signal vector of the sensor array; ； ； ； ； in, For the received signal vector, For the received signal vector, For noise vectors, The direction matrix, For the first K The steering vector of the nth signal source, for the nth signal source i The steering vector of the received signal Represented as: ； for M A sparse array of n elements, the nth k The steering vector of each signal source is represented as: ； in, For the elements of a sparse array, the set of elements of a sparse array. It is a subset of a uniform linear array; The turning vector of a sparse array is represented as a part of a uniform linear array; ； in, The selection matrix for the sparse array; S14: Write the received signal vector of the sensor array as a signal matrix. : ； ； in, For the received signal matrix, For the noise matrix, For the first T The received signal corresponding to each snapshot; S15: The direction of arrival (DOA) information is estimated using the covariance matrix of the received signal. The sparse array covariance matrix is ​​also known as the array autocorrelation matrix. Represented as: ； in, To receive the signal covariance matrix, It is the identity matrix. This represents the noise power of the sparse array.

3. The method for estimating the direction of arrival (DOA) of a sparse linear array based on the NaViT network model according to claim 2, characterized in that, Step S2 includes: S21: The covariance matrix of the sparse array The covariance matrix is ​​split into three channels: real, imaginary, and phase, and is a sparse array. Shape changed , L The number of elements in a sparse array; n Different covariance matrices Segment and flatten into sequence matrices , p The patch size for each sequence matrix. The number of patches generated. P Patch length; S22: Flatten the matrix patch into a one-dimensional vector, and then use a trainable linear projection to map these flattened vectors to a... D A dimensional potential space, where each matrix patch is represented as a... D dimensional vector; ； in, It is the first n A sequence of covariance matrices, To represent the learnable transformation matrix of a linear layer in matrix form, For the first N One matrix patch; S23: Will n Different covariance matrices Concatenate them into a long input vector ; ； S24: In long input vectors A special mark is pre-set in the middle As input for subsequent extraction of direction-of-arrival information, special markers are used to capture the covariance matrix and global information features. Add to long input vector The middle part serves as the entire covariance matrix. ; ； Among them, matrix patch For the entire covariance matrix The matrix block in the middle.

4. The method for estimating the direction of arrival (DOA) of a sparse linear array based on the NaViT network model according to claim 3, characterized in that, Step S3 includes: S31: Based on the entire covariance matrix The location information of the matrix patch is expressed using position embedding; ； in, pos The position of the matrix block. u For the entire covariance matrix latitude, For location embedding; S32: Convert the entire covariance matrix With position embedding Adding them together yields the position-encoded input vector. ; ； in, The position vector of the matrix block, position embedding Position vector Element.

Images