A pilot and data symbol fusion direction of arrival estimation method based on deep learning

Abstract

The application discloses a pilot and data symbol fusion direction of arrival (DOA) estimation method based on deep learning, and belongs to the technical field of wireless communication. The method constructs an end-to-end perception network architecture containing pilot feature extraction, data symbol feature extraction and cross-attention fusion modules. Specifically, channel estimation is carried out by using a pilot to extract spatial channel prior features; unknown data symbols are processed by a multi-layer self-attention network to extract high-order statistical features of non-Gaussian signals; a cross-attention mechanism is adopted to fuse pilot and data symbol features; and finally, a DOA estimation result is output through pooling. The application has good performance, low complexity and strong adaptive ability, and is suitable for a 6G communication and perception integrated system.

Description

Technical Field

[0001] This invention relates to a deep learning-based method for estimating direction of arrival (DOA) by fusing pilot and data symbols, belonging to the field of wireless communication technology. Background Technology

[0002] Integrated communication and sensing is a key technology for 6G. However, existing sensing solutions often require modification of the communication waveform, which not only incurs additional system overhead but also makes direct compatibility with wireless communication standards difficult. In the core sensing task of DOA estimation, traditional subspace-based DOA estimation methods typically follow the process of reconstructing the sample covariance matrix and subspace eigenvalue decomposition. These methods are mostly based on the Gaussian signal assumption, utilizing only the second-order statistics of the received signal. However, actual wireless communication systems transmit discretely modulated communication signals, exhibiting significant non-Gaussian characteristics. Traditional algorithms cannot effectively mine the high-order statistical features inherent in actual communication signals, resulting in severely limited sensing performance in scenarios with low signal-to-noise ratios or limited snapshots. Furthermore, the computational complexity of its matrix eigenvalue decomposition increases cubically with the number of antennas, making it difficult to apply to large-scale arrays.

[0003] In recent years, deep learning technology has been introduced into the field of DOA estimation. Most existing deep learning perception methods still focus on covariance matrix reconstruction, and their underlying input and optimization logic remains constrained by the Gaussian signal assumption, failing to truly achieve efficient perception based on actual non-Gaussian communication signals. Some deep learning DOA estimation methods have begun to use data symbols for perception, but they do not fully utilize the deterministic pilot information in the communication frame. Therefore, there is an urgent need for a deep learning DOA estimation method that can fuse deterministic pilots and random data symbols, ensuring low computational overhead while overcoming the bottleneck of insufficient utilization of non-Gaussian features and pilot priors in existing methods, thereby enhancing the system's perception robustness in complex multi-source and real-world deployment environments. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to overcome the problems of insufficient perception accuracy and excessive computational complexity in existing DOA estimation methods, which are limited by second-order statistics and do not fully utilize the prior information of pilots in communication frames. This results in problems such as low signal-to-noise ratio, few snapshots, and dense signal sources in complex real-world communication scenarios. The invention proposes a deep learning-based pilot and data symbol fusion DOA estimation method to achieve high-precision, low-complexity, and highly robust end-to-end perception of communication signals.

[0005] Technical solution: The deep learning-based pilot and data symbol fusion direction of arrival estimation method of the present invention includes: (1) constructing a simulation dataset based on the multi-source array signal receiving model and completing the training of the DOA estimation network; in the actual sensing stage, the receiver acquires the complete multi-source superimposed signal and, according to the frame structure preset by the communication protocol, splits the total received signal into pilot observation matrix and data observation matrix in the time domain;

[0006] (2) In the pilot feature extraction branch, the initial channel response is estimated for the pilot observation matrix by combining the known deterministic pilot sequences of both the transmitter and receiver, the channel prior information is extracted as the subsequent observation constraint, and it is mapped to the pilot query matrix; in the data symbol feature extraction branch, the data observation matrix is ​​sent to the multi-layer self-attention network module to extract the high-order statistical features contained in the unknown and random data symbols.

[0007] (3) Pilot features are used as query matrix and data features are used as keys and values. Pilot and data features are fused through cross-attention mechanism. The fused spatial features are used to eliminate the ambiguity of sequence arrangement. They are then processed by global mean pooling and regression mapping of fully connected network layer to finally output the DOA estimates of multiple sources in the space.

[0008] Furthermore, step (1) specifically includes:

[0009] (1.1) Considering a multi-source uplink transmission system, the receiver is equipped with a combination of... A linear array of antennas exists in space. An unknown narrowband far-field signal source; simulation data is generated using an array signal receiving model, and the total received signal matrix is... The model is as follows: in It is the set of DOAs to be estimated. For array manifold matrix, This is the path loss diagonal matrix. The matrix is ​​an additive white Gaussian noise matrix. Let be the total transmitted signal matrix; assuming the antenna spacing is half a wavelength, the steering vector of the k-th narrowband far-field source in the linear array is defined as: in This represents the position index of the m-th antenna.

[0010] (1.2) When generating the dataset, assume the total transmitted signal Composed of deterministic pilot signals and random data symbols, the model is defined as follows: in This represents a pilot sequence with known deterministic characteristics. A matrix representing unknown random data symbols. and These are the pilot index selection matrix and the data index selection matrix, respectively. For pilot snapshots, Let T be the number of data snapshots, and let T be the total number of snapshots, and let T satisfy the orthogonality condition. .

[0011] (1.3) When training the network, a loss function is constructed with the objective of minimizing the mean squared error based on angle ranking between the predicted angle and the true angle. The DOA estimation network is trained using the simulation dataset, and the loss function is defined as follows:

[0012]

[0013] in, For the network to predict the angle vector, The actual angle label vector. This indicates an operation to sort the elements of a vector in ascending order.

[0014] (1.4) The receiver uses a time index selection matrix to separate the signal and extract the pilot observation signal. With data observation signals in The channel matrix; The additive noise matrix of the pilot section. This is the additive noise matrix for the data portion.

[0015] Furthermore, step (2) specifically includes:

[0016] (2.1) Obtain the initial channel response estimation matrix using the linear minimum mean square error algorithm. The calculation formula is as follows: in, To observe the noise variance, Here is the regularization constant. It is an identity matrix.

[0017] (2.2) The initial channel response estimation matrix is ​​split into real and imaginary parts and concatenated to obtain the real-valued feature matrix. .

[0018] (2.3) Transform the real-valued characteristic matrix The input pilot projection module, which sequentially includes a fully connected layer, layer normalization, and a GELU activation function, maps these layers to the hidden layer dimensions defined by the network. Generate pilot query matrix .

[0019] (2.4) Observation of data signals The real and imaginary parts are concatenated, and a hidden layer of dimension is generated using a data projection module with the same structure. Initial data feature matrix .

[0020] (2.5) Input the initial data feature matrix sequentially Layer self-attention module, in the first layer In the layer, through a learnable weight matrix , and Input features from the previous layer Projection to generate query matrix Key matrix AND-value matrix Its generation formula is:

[0021] Subsequently, based on the generated query, key, and value matrix, the self-attention feature update is calculated:

[0022] through The cascaded processing outputs higher-order data features. .

[0023] Furthermore, step (3) specifically includes:

[0024] (3.1) Calculate fusion features based on cross-attention mechanism ,by As a query, data characteristics Using the key and value as keys, perform inner product attention operations to extract data features that match the pilot observation constraints:

[0025] (3.2) Calculate the adaptive gating coefficients and perform dynamic fusion, and then... and The input gating network is concatenated, wherein the gating network sequentially includes gating dimensions from... Dimensional reduction Fully connected layers, ReLU activation function, and increasing dimensionality to The fully connected layer and the sigmoid activation function. Gated network output. Interval weight matrix Calculate the weighted fusion spatial features: And The input is a fusion projection module, which sequentially includes a fully connected layer, layer normalization, and a GELU activation function.

[0026] (3.3) Features of fusion space Perform global mean pooling to obtain a fixed-length global feature vector. The final DOA estimate is obtained by mapping through a fully connected output layer. .

[0027] Furthermore, the network training of the method specifically involves:

[0028] The DOA estimation network is trained using a mini-batch gradient descent algorithm with end-to-end supervised learning. The mean squared error based on angle sorting is used as the loss function, defined as follows:

[0029] in, For the network to predict the angle vector, The actual angle label vector. This represents the operation of sorting vector elements in ascending order. Training uses the AdamW optimizer with dynamics, and the network hidden layer dimension... Set to 96, for the number of self-attention layers of 16QAM signals. Set it to 3.

[0030] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0031] This invention rationally combines deep learning with the DOA estimation task. It captures high-order statistical information through self-attention and fuses pilot and data symbol features through a cross-attention mechanism, achieving high-precision DOA estimation without additionally encroaching on communication service resources. The proposed method eliminates the need for explicit reconstruction of the covariance matrix, reducing computational complexity and making it suitable for large-scale arrays. In summary, this invention provides a communication-aware integrated end-to-end DOA estimation scheme with practical deployment prospects. Furthermore, the method is applicable to both uniform and sparse linear arrays, demonstrating good versatility. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the overall system architecture according to an embodiment of the present invention;

[0033] Figure 2 This is a detailed schematic diagram of the direction-of-arrival estimation network according to an embodiment of the present invention. Detailed Implementation

[0034] The present invention will now be described in detail with reference to the accompanying drawings and embodiments of the multi-source array signal receiving system.

[0035] I. System Model Applicable to This Embodiment

[0036] This example applies to a multi-source uplink transmission system, where the receiver is equipped with a... Linear arrays composed of antennas include two configurations: Uniform Linear Array (ULA) and Sparse Linear Array (SLA), which exist in space. An unknown narrowband far-field independent signal source. This embodiment uses... Uniform linear array and Sparse arrays are used as a typical application scenario. The set of DOAs to be estimated is defined as follows: The array in the total number of snapshots Internally received superimposed signal matrix It can be represented as: in, It is an array manifold matrix; It is a diagonal matrix containing the path loss and power of each source; The additive white Gaussian noise matrix satisfies zero mean and variance. ; This is the total transmitted signal matrix from multiple sources.

[0037] In practical wireless communication standards (such as 5G NR), the frame structure consists of a small number of deterministic pilot symbols (such as DMRS) and a large number of random data symbols (such as PUSCH) arranged according to a certain time index. An orthogonal time index selection matrix is ​​used. and The receiving end can receive the total signal. It is separated into two parts in the time domain: (1) corresponding to the pilot snapshot (length is Pilot observation signal: (2) Corresponding to data snapshots (length is Data observation signals: .in For the equivalent channel matrix, The pilot sequence is completely known to both the sender and receiver. It is an unknown data symbol sequence with non-Gaussian statistical properties (such as 16QAM or QPSK modulation).

[0038] II. Specific Steps of This Embodiment

[0039] like Figure 1 As shown, the DOA estimation network in this embodiment is an end-to-end deep learning network, and its complete execution flow includes three steps: feature extraction, dynamic fusion, and angle estimation output.

[0040] 1. Feature Extraction

[0041] Combination Figure 2 A detailed schematic diagram of the direction-of-arrival (DOA) estimation network is shown. First, pilot observation signals are used... and known pilot sequences The initial channel response estimation matrix is ​​calculated based on the linear minimum mean square error (LMMSE) criterion. : in, To prevent singularity during matrix inversion, a small regularization constant (which can be taken as...) is used. Compared to traditional least squares (LS) estimation, LMMSE can effectively suppress noise amplification when the pilot is extremely short.

[0042] Subsequently, the complex matrix Decompose it into real and imaginary parts and concatenate them to obtain the dimension. real-valued eigenmatrix .like Figure 2 As shown in the upper left part, The input pilot projection module consists of a fully connected layer, layer normalization, and a GELU activation function. This module maps the input to the hidden layer dimension defined by the network. In this embodiment Generate a pilot query matrix that includes reliable physical space constraints. .

[0043] For unknown data observation signals Similarly, it is split and concatenated into a real-valued matrix. ,like Figure 2 As shown in the lower left section, a data projection module with the same structure as the pilot projection module is fed in to generate a hidden layer with dimension [missing information]. Initial data feature matrix .

[0044] Then enter Layer self-attention module. Because the multi-source snapshot signals are independently and identically distributed in the time dimension, this network removes positional encoding to ensure permutation invariance. In the first... In the layer encoder, queries are generated using learnable weights. ,key AND value We can use a self-attention mechanism to mine higher-order statistical features between snapshots. The intermediate dimension of the feedforward neural network is expanded to: In this embodiment, for a 16QAM non-Gaussian modulated signal, the number of self-attention layers is... Ideally, three layers are used to ensure the capture of sufficiently deep fourth-order and higher non-zero statistical properties, while preventing model overfitting, ultimately outputting high-order data features. .

[0045] 2. Dynamic fusion

[0046] To accurately extract effective spatial components from massive random data features, this embodiment designs a cross-attention module. For example... Figure 2 As shown in the dashed box on the right, the pilot query matrix For query matrix (Query), based on data characteristics The fusion features are calculated using the key matrix and value matrix. .

[0047] To address drastic fluctuations in signal-to-noise ratio, an adaptive gating network is constructed. and After being spliced ​​together, the data is fed into the gating network, whose network structure consists of sequentially changing the dimensions from... Dimensional reduction Fully connected layers, ReLU activation function, increasing dimensionality to The fully connected layer and the sigmoid activation function. Gated network output. Interval weighting coefficients And thereby generate weighted spatial features: at last, Feature normalization is performed using a fusion projection module that includes Linear, LayerNorm, and GELU.

[0048] 3. Angle estimation output

[0049] After processing Perform global mean pooling to obtain a fixed-length global feature vector. This operation eliminates the ambiguity of the implicitly assigned source order within the network. The input is the output layer of the multilayer perceptron, and finally the regression is... DOA estimates for each source .

[0050] During offline network training, the mean squared error based on angle sorting is used as the loss function:

[0051] in, For the network to predict the angle vector, This is the actual angle label vector.

[0052] Training employs the AdamW optimizer, combined with a one-cycle learning rate scheduling strategy, to perform end-to-end parameter optimization on the DOA estimation network. (Network hidden layer dimensions...) Set to 96, for the number of self-attention layers of 16QAM signals. Set it to 3.

[0053] III. Core Collaborative Mechanism of the Invention

[0054] The pilot and data symbol fusion network proposed in this invention integrates traditional subspace DOA estimation with modern deep learning models. Deterministic pilot sequences provide reliable observation constraints for the model, while a large number of random data symbols generate rich higher-order statistics through the Transformer's self-attention mechanism. Working collaboratively through cross-attention and gating networks, these two methods overcome the limitations of traditional DOA estimation schemes, which are constrained by the Gaussian signal assumption and suffer from high computational complexity. This achieves high-accuracy and high-robustness DOA estimation while maintaining low overhead. The principle of this method is based on the DOA estimation characteristics of array signal processing, possessing clear physical significance and engineering feasibility.

[0055] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A deep learning-based method for estimating direction-of-arrival (DOA) by fusing pilot signals and data symbols, characterized in that... Includes the following steps: (1) Use the multi-source array signal receiving model to generate simulation data to train the deep learning-based DOA estimation network; the receiver obtains the total received signal and separates the pilot observation signal and the data observation signal from it according to the predetermined frame structure; (2) Based on the known deterministic pilot sequence and the pilot observation signal, perform initial channel response estimation and extract channel prior information to generate a pilot query matrix composed of pilot query vectors corresponding to each source; input the data observation signal into the multilayer self-attention module of the DOA estimation network to extract the high-order statistical features of the random data symbols; (3) The pilot query matrix and the high-order statistical features of data symbols are fused through the cross-attention mechanism. After global mean pooling and full connection layer processing, the network estimates and outputs the DOA angle information of multiple sources.

2. The deep learning direction-of-arrival estimation method based on deterministic pilot signals and random data symbols according to claim 1, characterized in that: Step (1) specifically includes: (1.1) Considering a multi-source uplink transmission system, the receiver is equipped with a combination of... A linear array of antennas exists in space. An unknown narrowband far-field signal source; simulation data is generated using an array signal receiving model, and the overall received signal model is as follows: ； in, It is the set of DOAs to be estimated. For array manifold matrix, This is the path loss diagonal matrix. The matrix is ​​an additive white Gaussian noise matrix. Let be the total transmitted signal matrix; assuming the antenna spacing is half a wavelength, the steering vector of the k-th narrowband far-field source in the linear array is defined as: ； in, This represents the position index of the m-th antenna; (1.2) When generating the dataset, assume the total transmitted signal Composed of deterministic pilot signals and random data symbols, the model is defined as follows: ； in, This represents a pilot sequence with known deterministic characteristics. A matrix representing unknown random data symbols. and These are the pilot time index selection matrix and the data time index selection matrix, respectively. For pilot snapshots, Let T be the number of data snapshots, and let T be the total number of snapshots, and let T satisfy the orthogonality condition. ; (1.3) When training the network, a loss function is constructed with the objective of minimizing the mean squared error based on angle ranking between the predicted angle and the true angle. The DOA estimation network is trained using the simulation dataset, and the loss function is defined as follows: ； in, For the network to predict the angle vector, The actual angle label vector. This indicates an operation to sort the elements of a vector in ascending order. (1.4) The receiver uses a time index selection matrix to separate the signal and extract the pilot observation signal. With data observation signals in The channel matrix; The additive noise matrix of the pilot section. This is the additive noise matrix for the data portion.

3. The deep learning-based pilot and data symbol fusion direction-of-arrival estimation method according to claim 2, characterized in that: Step (2) specifically includes: (2.1) Obtain the initial channel response estimation matrix using the linear minimum mean square error algorithm. The calculation formula is as follows: ； in, To observe the noise variance, Here is the regularization constant. It is the identity matrix; (2.2) The initial channel response estimation matrix is ​​split into real and imaginary parts and concatenated to obtain the real-valued feature matrix. ; (2.3) Transform the real-valued characteristic matrix Input the pilot projection module and map it to the hidden layer dimension defined by the network. Generate the pilot query matrix ; (2.4) Observe the data signal The real and imaginary parts are concatenated to generate the initial data feature matrix. ; (2.5) Input the initial data feature matrix sequentially Layer self-attention module, in the first layer In the layer, through a learnable weight matrix , and Input features from the previous layer Projection to generate query matrix Key matrix AND-value matrix Its generation formula is: Subsequently, based on the generated query, key, and value matrix, self-attention feature updates are calculated: ;through The cascaded processing outputs higher-order data features. .

4. The deep learning-based pilot and data symbol fusion direction-of-arrival estimation method according to claim 3, characterized in that: Step (3) specifically includes: (3.1) Calculate fusion features based on cross-attention mechanism ,by As a query, data characteristics As key and value: ； (3.2) Calculate the adaptive gating coefficients and perform dynamic fusion, and then... and By concatenating the input gating network, the weights are obtained. Generate fused spatial features ; (3.3) Features of fusion space Perform global mean pooling to obtain a fixed-length global feature vector. The final DOA estimate is obtained by mapping through a fully connected output layer. .

5. The method according to claim 4, characterized in that, The DOA estimation network is a Transformer network architecture with location encoding removed; the gating network sequentially includes elements that reduce the dimension from... Dimensional reduction Fully connected layers, ReLU activation function, and increasing dimensionality to The network consists of fully connected layers and a Sigmoid activation function; for 16QAM modulated signals, the hidden layer dimension of the network is... Set to 96, number of self-attention layers Set to 3 layers.

Images