Sparse linear array direction of arrival estimation method based on NaViT network model

Abstract

The invention discloses a sparse linear array direction-of-arrival estimation method based on a NaViT network model, and the method comprises the steps: defining a sparse linear sensor array, constructing a receiving signal vector of the sparse linear sensor array according to a receiving signal of each array element, and obtaining a sparse array covariance matrix; flattening the sparse array covariance matrix to generate a matrix patch, mapping the matrix patch to a D-dimensional potential space, and presetting a special mark to form a whole covariance matrix; according to position information of a matrix patch in the whole covariance matrix, position description of the matrix patch is expressed by using position embedding, and an input vector of position coding is obtained; and constructing a NaViT network model, inputting the input vector of the position coding and the sparse array covariance matrix into the NaViT network model, carrying out identification by using an attention layer and an encoder of the NaViT network model, and classifying the direction of arrival of a sparse linear array receiving signal to obtain the direction of arrival of the receiving signal.

Description

Technical Field

[0001] The invention relates to the field of signal processing, and in particular to a sparse linear array direction of arrival estimation method based on a NaViT network model. Background Art

[0002] Direction of arrival estimation is one of the key issues in sensor array signal processing. Compared with uniform linear arrays, sparse linear array (SLA) structures have become a popular array configuration in the field of direction of arrival estimation due to the expansion of degrees of freedom. In order to improve algorithm performance and reduce computational complexity, deep learning has been widely used in the field of direction of arrival (DOA) estimation in recent years. However, most existing methods use convolutional neural networks (CNNs), which cannot learn the global features of input data well.

[0003] Most deep learning-based DOA estimation methods show superior performance compared to traditional model-driven algorithms. However, most traditional deep learning algorithms are implemented based on CNN. CNN captures features through local receptive fields, but DOA information often exists in the entire received signal covariance matrix. Therefore, CNN still has certain limitations in modeling global information. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a sparse linear array direction of arrival estimation method based on the NaViT network model.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: A sparse linear array direction of arrival estimation method based on a NaViT network model is provided, which comprises the following steps: S1: define a sparse linear sensor array, construct a received signal vector of the sparse linear sensor array according to the received signal of each array element, and obtain a sparse array covariance matrix; S2: Flatten the sparse array covariance matrix to generate matrix patches and map the matrix patches to a D dimensional latent space, presetting a special marker to form the entire covariance matrix; S3: According to 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 to obtain the input vector of the position encoding; 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 NaViT network model attention layer and encoder to identify and classify the direction of arrival of the sparse linear array received signal, and obtain the direction of arrival of the received signal.

[0006] Further, step S1 includes: S11: Define the number of elements in the sensor array M , Array element spacing d , K The signal source that meets the spatial far-field narrowband assumption is denoted as , For the K signal source, K is the number of signal sources; S12: Taking the first array element as the reference, each received signal is expressed as , No. m The received signal of the array element It is expressed as: ; in, For the i A received signal, i is the number of the received signal, j for, is the wavelength of the received signal, m is the array element number, For the i signal source, For the m The noise of the array element, n is the patch length; S13: constructing a received signal vector of the sensor array; ; ; ; ; in, is the received signal vector, is the received signal vector, is the noise vector, is the direction matrix, For the K The steering vector of the signal source is i The steering vector of the received signal It is expressed as: ; for M The sparse array of array elements, k The steering vector of a signal source is expressed as: ; in, For the elements of the sparse array, the collection of elements of the sparse array is a subset of a uniform linear array; Represent the steering vector of the sparse array as a portion of a uniform linear array; ; in, is the selection matrix for sparse arrays; S14: Write the received signal vector of the sensor array into a signal matrix : ; ; in, is the received signal matrix, is the noise matrix, For the T The received signal corresponding to each snapshot; S15: Estimate the direction of arrival information through the covariance matrix of the received signal. The sparse array covariance matrix is ​​also called the array autocorrelation matrix. It is expressed as: ; in, is the received signal covariance matrix, is the identity matrix, is the noise power of the sparse array.

[0007] Further, step S2 includes: S21: Convert sparse array to covariance matrix Split into three channels: real part, imaginary part and phase, sparse array covariance matrix The shape is changed to , L is the number of elements of the sparse array; n Different covariance matrices Split and flatten into a sequence matrix , p For each sequence matrix the patch size is, is the number of patches generated, P is the patch length; S22: Flatten the matrix patches into one-dimensional vectors, and then use trainable linear projection to map these flattened vectors into a D dimensional latent space, each matrix patch is represented as a D dimensional vector; ; in, It is nThe sequence of covariance matrices, is the learnable transformation matrix representing the linear layer in matrix form, For the N Matrix patches; S23: n Different covariance matrices Concatenate into one long input vector ; ; S24: On long input vectors Preset a special tag in , as the input for subsequent extraction of direction of arrival information, in order to capture the covariance matrix and global information features, the special mark Add to long input vector As the entire covariance matrix ; ; Among them, the matrix patch is the entire covariance matrix The matrix blocks in .

[0008] Further, step S3 includes: S31: Based on the entire covariance matrix The position information of the matrix patch in the matrix is ​​expressed by using position embedding; ; in, POS is the position of the matrix block, u is the entire covariance matrix The latitude, for position embedding; S32: The entire covariance matrix With positional embedding Add together to get the position-encoded input vector ; ; in, is the position vector of the matrix block, the position embedding is the position vector elements.

[0009] Further, step S4 includes: S41: Construct NaViT network model and convert sparse array covariance matrix and the position-encoded input vector Input into the NaViT network model, the discriminator NaViT network model obtains the input vector The query matrix , key matrix Sum Matrix ; S42: Convert sparse array covariance matrix and the position-encoded input vector Copy to query matrix , key matrix Sum Matrix Previously, the attention layer was used for linear transformation, and the attention function of the NaViT network model was: ; ; in, The query matrix , key matrix Sum Matrix The linear transformation matrix of is the attention function, is the output of the attention layer; S43: Connect multiple attention layers and multiply by a weight vector to obtain the encoder output of the NaViT network model; ; ; in, is the output of the encoder, The concatenation function for the output sequences of multiple attention layers, is the weight vector; S44: The output of the encoder The classification feature signals are used as classification feature signals, and the classification feature signals are classified to obtain the direction of arrival classification signals of the sparse linear array.

[0010] The beneficial effects of the present invention are as follows: the present invention uses the covariance matrix of the received signal as input, wherein the real part, imaginary part and phase of the covariance matrix are used as three channels of input, and under the conditions of a single signal source and a low signal-to-noise ratio, the recognition accuracy is better than that of the traditional method. By dividing the matrix into smaller matrix blocks and packaging them into time series for classification, the calculation speed is increased by more than four times. The network structure has stable performance and good estimation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a flowchart of the sparse linear array direction of arrival estimation method based on the NaViT network model. DETAILED DESCRIPTION

[0012] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0013] like Figure 1 As shown, a sparse linear array direction of arrival estimation method based on the NaViT network model includes the following steps: S1: define a sparse linear sensor array, construct a received signal vector of the sparse linear sensor array according to the received signal of each array element, and obtain a sparse array covariance matrix. Step S1 specifically includes: S11: Define the number of elements in the sensor array M , Array element spacing d , K The signal source that meets the spatial far-field narrowband assumption is denoted as , For the K signal source, K is the number of signal sources; S12: Taking the first array element as the reference, each received signal is expressed as , No. m The received signal of the array element It is expressed as: ; in, For the i A received signal, i is the number of the received signal, j for, is the wavelength of the received signal, m is the array element number, For the i signal source, For the m The noise of the array element, n is the patch length; S13: constructing a received signal vector of the sensor array; ; ; ; ; in, is the received signal vector, is the received signal vector, is the noise vector, is the direction matrix, For the K The steering vector of the signal source is i The steering vector of the received signal It is expressed as: ; for M The sparse array of array elements, k The steering vector of a signal source is expressed as: ; in, For the elements of the sparse array, the collection of elements of the sparse array is a subset of a uniform linear array; Represent the steering vector of the sparse array as a portion of a uniform linear array; ; in, is the selection matrix for sparse arrays; S14: Write the received signal vector of the sensor array into a signal matrix : ; ; in, is the received signal matrix, is the noise matrix, For the T The received signal corresponding to each snapshot; S15: Estimate the direction of arrival information through the covariance matrix of the received signal. The sparse array covariance matrix is ​​also called the array autocorrelation matrix. It is expressed as: ; in, is the received signal covariance matrix, is the identity matrix, is the noise power of the sparse array. Compared with the covariance matrix of a uniform array, the structure of the covariance matrix of a sparse array is changed, and the contribution of the missing elements is reflected as zero in the corresponding rows and columns.

[0014] S2: Flatten the sparse array covariance matrix to generate matrix patches and map the matrix patches to a D dimensional latent space, presetting a special marker to form the entire covariance matrix. Step S2 specifically includes: S21: Convert sparse array to covariance matrix Split into three channels: real part, imaginary part and phase, sparse array covariance matrix The shape is changed to , L is the number of elements in the sparse array; in order to improve the computational efficiency, n Different covariance matrices Split and flatten into a sequence matrix , p For each sequence matrix the patch size is, is the number of patches generated, P is the patch length; S22: Flatten the matrix patches into one-dimensional vectors, and then use trainable linear projection to map these flattened vectors into a D dimensional latent space, each matrix patch is represented as a D dimensional vector; ; in, It is n The sequence of covariance matrices, is the learnable transformation matrix representing the linear layer in matrix form, For the N Matrix patches; S23: n Different covariance matrices Concatenate into one long input vector ; ; S24: During the processing, the long input vector can capture the global information and characteristics of the covariance matrix. Preset a special tag in , as the input for subsequent extraction of direction of arrival information, in order to capture the covariance matrix and global information features, the special mark Add to long input vector As the entire covariance matrix ; ; Among them, the matrix patch is the entire covariance matrix The matrix blocks in .

[0015] S3: In order to preserve the position order of each matrix block relative to the original matrix and ensure that the model can understand the order relationship in the data, we introduce position information, similar to the standard transformer model. According to 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 to obtain the position encoded input vector. Step S3 specifically includes: S31: Based on the entire covariance matrix The position information of the matrix patch in the matrix is ​​expressed by using position embedding; ; in, POS is the position of the matrix block, u is the entire covariance matrix The latitude, is the position embedding; to improve the distinguishability of position embeddings in odd and even dimensions, we use the sin function and cos function as encoders respectively. The reason for choosing the sine function as the encoder is that it allows the model to easily learn to locate by relative position, because for any fixed offset, it can be represented by a linear function of the position information.

[0016] S32: The entire covariance matrix With positional embedding Add together to get the position-encoded input vector ; ; in, is the position vector of the matrix block, the position embedding is the position vector elements.

[0017] 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 NaViT network model attention layer and encoder to identify, classify the direction of arrival of the sparse linear array received signal, and obtain the direction of arrival of the received signal. Step S4 specifically includes: S41: Construct NaViT network model and convert sparse array covariance matrix and the position-encoded input vector Input into the NaViT network model, the discriminator NaViT network model obtains the input vector The query matrix , key matrix Sum Matrix ; S42: Convert sparse array covariance matrix and the position-encoded input vector Copy to query matrix , key matrix Sum Matrix Previously, the attention layer was used for linear transformation, and the attention function of the NaViT network model was: ; ; in, The query matrix , key matrix Sum Matrix The linear transformation matrix of is the attention function, is the output of the attention layer; S43: Connect multiple attention layers and multiply by a weight vector to obtain the encoder output of the NaViT network model; ; ; in, is the output of the encoder, The concatenation function for the output sequences of multiple attention layers, is the weight vector; S44: The encoder output As the 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 the direction of arrival of the received signal is obtained.

[0018] The present invention uses the covariance matrix of the received signal as input, where the real part, imaginary part and phase of the covariance matrix are used as three channels of input. Under the conditions of single signal source and low signal-to-noise ratio, the recognition accuracy is better than the traditional method. By dividing the matrix into smaller matrix blocks and packaging them into time series for classification, the calculation speed is increased by more than four times. The network structure has stable performance and good estimation effect.

Claims

1. A sparse linear array direction of arrival estimation method based on NaViT network model, characterized in that: The following steps are involved: S1: define a sparse linear sensor array, construct a received signal vector of the sparse linear sensor array according to the received signal of each array element, and obtain a sparse array covariance matrix; S2: Flatten the sparse array covariance matrix to generate matrix patches and map the matrix patches to a D dimensional latent space, presetting a special marker to form the entire covariance matrix; S3: According to 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 to obtain the input vector of the position encoding; 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 NaViT network model attention layer and encoder to identify and classify the direction of arrival of the sparse linear array received signal, and obtain the direction of arrival of the received signal.

2. The sparse linear array direction of arrival estimation method based on the NaViT network model according to claim 1, characterized in that: The step S1 comprises: S11: Define the number of elements in the sensor array M , Array element spacing d , K The signal source that meets the spatial far-field narrowband assumption is denoted as , For the K signal source, K is the number of signal sources; S12: Taking the first array element as the reference, each received signal is expressed as , No. m The received signal of the array element It is expressed as: ； in, For the i The received signal, i is the number of the received signal, j for, is the wavelength of the received signal, m is the array element number, For the i signal source, For the m The noise of the array element, n is the patch length; S13: constructing a received signal vector of the sensor array; ； ； ； ； in, is the received signal vector, is the received signal vector, is the noise vector, is the direction matrix, For the K The steering vector of the signal source is i The steering vector of the received signal It is expressed as: ； for M The sparse array of array elements, k The steering vector of a signal source is expressed as: ； in, For the elements of the sparse array, the collection of elements of the sparse array is a subset of a uniform linear array; Represent the steering vector of the sparse array as a portion of a uniform linear array; ； in, is the selection matrix for sparse arrays; S14: Write the received signal vector of the sensor array into a signal matrix : ； ； in, is the received signal matrix, is the noise matrix, For the T The received signal corresponding to each snapshot; S15: Estimate the direction of arrival information through the covariance matrix of the received signal. The sparse array covariance matrix is ​​also called the array autocorrelation matrix. It is expressed as: ； in, is the received signal covariance matrix, is the identity matrix, is the noise power of the sparse array.

3. The sparse linear array direction of arrival estimation method based on the NaViT network model according to claim 2, characterized in that: The step S2 comprises: S21: Convert sparse array to covariance matrix Split into three channels: real part, imaginary part and phase, sparse array covariance matrix The shape is changed to , L is the number of elements of the sparse array; n Different covariance matrices Split and flatten into a sequence matrix , p For each sequence matrix the patch size is, is the number of patches generated, P is the patch length; S22: Flatten the matrix patches into one-dimensional vectors, and then use trainable linear projection to map these flattened vectors into a D dimensional latent space, each matrix patch is represented as a D dimensional vector; ； in, It is n The sequence of covariance matrices, is the learnable transformation matrix representing the linear layer in matrix form, For the N Matrix patches; S23: n Different covariance matrices Concatenate into one long input vector ; ； S24: On long input vectors Preset a special tag in , as the input for subsequent extraction of direction of arrival information, in order to capture the covariance matrix and global information features, the special mark Add to long input vector As the entire covariance matrix ; ； Among them, the matrix patch is the entire covariance matrix The matrix blocks in .

4. The method for estimating the direction of arrival of a sparse linear array based on the NaViT network model according to claim 3, characterized in that: The step S3 comprises: S31: Based on the entire covariance matrix The position information of the matrix patch in the matrix is ​​expressed by using position embedding; ； in, POS is the position of the matrix block, u is the entire covariance matrix The latitude, for position embedding; S32: The entire covariance matrix With positional embedding Add together to get the position-encoded input vector ; ； in, is the position vector of the matrix block, the position embedding is the position vector elements.

5. The sparse linear array direction of arrival estimation method based on the NaViT network model according to claim 4, characterized in that: The step S4 comprises: S41: Construct NaViT network model and convert sparse array covariance matrix and the position-encoded input vector Input into the NaViT network model, the discriminator NaViT network model obtains the input vector The query matrix , key matrix Sum Matrix ; S42: Convert sparse array covariance matrix and the position-encoded input vector Copy to query matrix , key matrix Sum Matrix Previously, the attention layer was used for linear transformation, and the attention function of the NaViT network model was: ； ； in, The query matrix , key matrix Sum Matrix The linear transformation matrix of is the attention function, is the output of the attention layer; S43: Connect multiple attention layers and multiply by a weight vector to obtain the encoder output of the NaViT network model; ； ； in, is the output of the encoder, The concatenation function for the output sequences of multiple attention layers, is the weight vector; S44: The encoder output The classification feature signals are used as classification feature signals, and the classification feature signals are classified to obtain the direction of arrival classification signals of the sparse linear array.

Images