Direction of arrival estimation method based on residual network

Abstract

The invention discloses a direction of arrival estimation method based on a residual network. A space encoder and an improved residual network cascade module are included. According to the system, complex signals received by an antenna array element are used as input, a covariance matrix is extracted, and a multi-channel characteristic pattern is constructed. Direction-of-arrival coarse estimation is carried out through space encoder partition, and further direction-of-arrival fine estimation is carried out through an improved residual network. Through reasonable combination of the deep neural network and the residual network, high estimation precision and operation speed can still be maintained in complex scenes of low signal-to-noise ratio, small signal source angle interval, multipath interference and the like. Compared with the traditional direction of arrival estimation method, the method provided by the invention has a remarkable advantage in the direction of arrival estimation precision under the complex test conditions of low signal-to-noise ratio and the like. Meanwhile, on the premise that high precision is kept, the method has the higher operation speed, and good real-time performance and engineering application prospects are embodied.

Description

Technical Field

[0001] This invention belongs to the field of communications and computers, and specifically relates to a method for estimating angle of arrival based on residual networks. Background Technology

[0002] In recent decades, the rapid development of mobile communication technology has profoundly transformed modern society's production and management models, as well as people's lifestyles. As the medium for information propagation in wireless communication systems, the characteristics of the wireless channel are crucial for wireless communication system design, communication quality analysis, and network performance simulation. In wireless communication systems, wireless transmission performance is directly determined by the channel propagation characteristics. Therefore, estimating wireless channel measurement parameters is of paramount importance for in-depth analysis of channel propagation characteristics and for conducting channel modeling in wireless communication systems. Among these parameters, the angle of arrival (DOA) is one of the most important in wireless communication.

[0003] With the rapid development of deep learning technology, Convolutional Neural Networks (CNNs), due to their powerful feature extraction capabilities in image recognition and speech processing, have gradually been introduced into the field of wireless communication, especially demonstrating superior performance in channel characteristic analysis and parameter estimation tasks. CNNs can automatically learn and extract local spatial features from input data through multi-layer convolutional structures, effectively capturing key feature information in signals. In the problem of DOA estimation, traditional algorithms such as MUSIC and ESPRIT typically rely on channel noise modeling and array structure assumptions, resulting in significant performance degradation in complex or non-ideal environments. In contrast, CNN-based DOA estimation methods can directly learn spatial spectral distribution or angular information from the feature representation of the received signal (such as the covariance matrix), improving robustness and estimation accuracy in complex environments such as low signal-to-noise ratio and multipath interference. Therefore, applying CNNs to DOA estimation not only opens up a new path for the integration of traditional signal processing and deep learning but also provides new technical support for building more intelligent wireless communication systems.

[0004] For complex application scenarios with multiple target signal sources coexisting, especially when the signal source angles are close or the signal-to-noise ratio is low, traditional algorithms such as MUSIC and ESPRIT suffer from reduced DOA estimation accuracy. Although deep learning methods introduce powerful nonlinear modeling capabilities, especially convolutional neural networks which show significant advantages in feature extraction, they still have many limitations in multi-target DOA estimation.

[0005] On the one hand, deep convolutional neural networks are prone to gradient vanishing or gradient explosion, which limits the training depth and stability of the model. On the other hand, when using CNNs directly for multi-source DOA estimation, it is necessary to model and classify all possible angle combinations, which leads to an exponential increase in the number of categories with the angle resolution and the number of targets, resulting in a huge computational burden and seriously affecting the training efficiency of the model and the real-time performance of the system. Summary of the Invention

[0006] Technical Problem: To address the aforementioned issues, this paper proposes an angle-of-arrival (DOA) estimation method based on residual networks. This improved DOA estimation architecture cascades a spatial encoder and an improved residual network. The architecture first extracts spatial features from the covariance matrix of the array-received signals using the spatial encoder to locate the angular region of the signal source. Then, it invokes the corresponding residual network for fine-grained angle estimation. This architecture achieves higher accuracy compared to traditional algorithms and DNN networks in multi-source signal estimation.

[0007] Technical Solution: The purpose of this invention is to provide a method for estimating angle of arrival (OA) based on residual networks. This method includes the following steps:

[0008] Step 1: Construct a spatial encoder based on a deep neural network (DNN). The spatial encoder takes the covariance matrix of the signal received by the antenna array as input to extract its spatial features and perform angle-of-arrival (AOA) block division.

[0009] Step 2: Construct an improved residual network based on ResNet residual network; used to achieve fine estimation of angle of arrival based on the angle of arrival blocks given by the spatial encoder;

[0010] Step 3: Train the spatial encoder and the improved residual network separately, and cascade them to form a cascaded residual network;

[0011] Step 4: Input the received signal of the antenna array element into the cascaded residual network of Step 3 to perform angle of arrival (DOA) estimation and obtain the DOA estimation result.

[0012] in,

[0013] The spatial encoder includes, in the order of input and output, the following:

[0014] The input layer is used to normalize the covariance matrix of the received signal of the antenna array elements to obtain normalized features;

[0015] Fully connected layers are used to perform linear transformations and dimension mappings on normalized features to enhance their expressive power.

[0016] The Softmax layer is used to normalize the activation values ​​output by the fully connected layer into a probability distribution.

[0017] The classification layer is used to classify the signal at the sector level based on the output of the Softmax layer structure. It divides the entire directional space into 6 fixed sectors and outputs the category label of the sector where the signal source is located for subsequent residual network calls.

[0018] The input to step 1 is the covariance matrix of the received signal of the antenna array element, calculated using the following formula:

[0019]

[0020] Where y(n) is the received signal of the antenna element in the nth frame, and L is the number of frames. (*) H This indicates the conjugate transpose.

[0021] The improved residual network comprises: four residual units, one convolutional layer, one input layer, and one fully connected layer.

[0022] The improved residual network includes:

[0023] Input layer: The two-dimensional feature map size of the signal received by the antenna array elements is uniformly M×M×C, where M is the number of array elements of the antenna array; C is 2, representing the real part and the imaginary part channels, and is normalized before input to improve training stability.

[0024] Convolutional layer: Used to receive two-dimensional feature map data from the input layer, and uses two-dimensional convolution operation with a kernel size of 3×3, padding size of 1, and stride of 1 to extract the basic features of the feature map data;

[0025] Four residual units are used to perform residual learning on the basic feature map output by the initial convolutional layer. Each residual unit includes multiple residual modules with identity mapping connections. Each residual module performs weighted fusion of the input features and the convolutionally transformed feature map through skip connections to obtain deep target mapping features and enhance the gradient propagation capability of the network during training.

[0026] The fully connected layer receives the target mapping features output by the residual block group, flattens them, and then inputs them into two fully connected sub-layers. The output dimension of the first fully connected sub-layer is 2048, and the output dimension of the second fully connected sub-layer is 210. These sub-layers are used to output the predicted probability distribution of the angle of arrival directions corresponding to the two signal sources in the target sector.

[0027] The residual module includes the following steps:

[0028] Define residual cells:

[0029] y l =h(x l )+F(x l W l )

[0030] x l+1 =f(y l )

[0031] Where x l y represents the input of the l-th residual module. l W represents the output of the l-th residual module. l Let f(*) represent the weight coefficients of the residual network, f(*) be the activation function, and h(x) be the weight coefficients of the residual network. l ) represents the target mapping function;

[0032] The deep feature x is obtained by accumulating L-1 consecutive residual units. L

[0033]

[0034] Where x L The output x of the l-th residual unit l With the summation of a series of residual functions, W i The weighting coefficients represent the residual function.

[0035] Step 3 includes: first, training a spatial encoder using the signal covariance matrix of the antenna array elements with partition labels; then, training an improved residual network using the original received signal feature map; and finally, cascading the two to form a cascaded residual network.

[0036] Step 4 includes: inputting the received signal of the antenna array element into the cascaded residual network of step 3 for angle of arrival (DOA) estimation. First, the sector label of the direction to which the angle of arrival belongs can be obtained through the spatial encoder. The label is then input into the corresponding improved residual network to obtain the angle of arrival (DOA) estimation result.

[0037] Beneficial Effects: This invention provides a DOA estimation method based on residual networks, which is a DOA estimation method based on a cascaded structure of a spatial encoder and a residual network. Under complex test conditions such as low signal-to-noise ratio (SNR), the proposed method shows significant advantages in DOA estimation accuracy. Simultaneously, while maintaining high accuracy, this method has a faster running speed, demonstrating good real-time performance and promising engineering application prospects. For increasingly complex wireless communication environments, traditional DOA estimation methods and deep neural networks often struggle to meet the practical requirements of high accuracy and low latency when facing low SNR, multipath interference, and multiple source angles being close together. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0039] Figure 2 To improve the residual network structure diagram;

[0040] Figure 3 A comparison of the DOA estimation error of four algorithms with varying signal-to-noise ratio under different angular intervals; Figure 3 (a) shows the comparison of DOA estimation errors when the angular interval is 3°. Figure 3 (b) shows the comparison of DOA estimation errors when the angular interval is 6°. Figure 3 (c) in the figure represents the comparison of DOA estimation errors when the angular interval is 9°;

[0041] Figure 4 A comparison of the running time of four algorithms as the signal-to-noise ratio changes; Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0043] To illustrate the technical solutions disclosed in this invention in detail, further explanation will be provided below with reference to the accompanying drawings and specific examples.

[0044] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0045] In one embodiment of the present invention, the receiving array structure adopts a uniform linear array.

[0046] The m-th receiving element can be represented as

[0047]

[0048] Among them, s i (t) represents the i-th incident source signal, n m y(t) represents the noise received by the m-th array element, λ represents the carrier wavelength, and the received vector at time t is defined as y(t) = [y1(t), y2(t), ... yt]. M (t)],θ k Let be the angle of arrival of the k-th incident source signal, and the corresponding covariance matrix is ​​expressed as:

[0049]

[0050] Where y(n) is the received signal of the antenna element in the nth frame, and L is the number of frames. (*) H This indicates the conjugate transpose.

[0051] like Figure 1As shown, a method for estimating the angle of arrival (OA) based on a convolutional neural network includes the following steps:

[0052] Step 1: Construct a spatial encoder based on a deep neural network (DNN) to divide the received signal into angle-of-arrival (AOA) blocks based on the covariance matrix features of the received signal;

[0053] Step 2: Construct an improved residual network based on the ResNet residual network; wherein the improved residual network includes: four residual units, one convolutional layer and one fully connected layer, used to achieve fine estimation of the angle of arrival based on the angle of arrival blocks given by the spatial encoder;

[0054] Step 3: Train the spatial encoder and the improved residual network separately, and cascade them to form a cascaded residual network;

[0055] Step 4: Input the received signal of the antenna array element into the cascaded residual network of Step 3 for DOA estimation and obtain the DOA estimation result.

[0056] like Figure 1 As shown, the spatial encoder includes the following components in the order of input and output:

[0057] The input layer (including the covariance matrix calculation module and the feature normalization module) is used to normalize the covariance matrix of the received signal of the antenna array element to obtain normalized features.

[0058] The fully connected layer (including the linear transformation module, ReLU activation module, and feature dimension mapping module) is used to perform linear transformation and dimension mapping on the normalized features to enhance the feature representation capability.

[0059] The Softmax layer (including a probability normalization module and a classification confidence output module) is used to normalize the activation values ​​output by the fully connected layer into a probability distribution form.

[0060] The classification layer (including the label encoding module) is used to classify the signal at the sector level according to the output of the Softmax layer structure. It divides the entire directional space into 6 fixed sectors and outputs the category label of the sector where the signal source is located for subsequent residual network to call.

[0061] like Figure 2 As shown, the residual module in the residual unit includes the following steps:

[0062] Define residual cells:

[0063] y l =h(x l )+F(x l W l )

[0064] xl+1 =f(y l )

[0065] Where x l y represents the input of the l-th residual module. l W represents the output of the l-th residual module. l Let f(*) represent the weight coefficients of the residual network, f(*) be the activation function, and h(x) be the weight coefficients of the residual network. l ) represents the target mapping function;

[0066] The deep feature x is obtained by accumulating L-1 consecutive residual units. L

[0067]

[0068] Where x L The output x of the l-th residual unit l With the summation of a series of residual functions, W i The weighting coefficients represent the residual function.

[0069] like Figure 2 As shown, the improved residual network includes four residual units, a convolutional layer, and a fully connected layer, which are used to achieve fine estimation of the angle of arrival based on the angle of arrival blocks given by the spatial encoder.

[0070] like Figure 2 As shown, the input layer of the improved residual network is a two-dimensional feature map of the signal received by the antenna array elements, with a uniform size of M×M×C, where M is the number of array elements of the antenna array, and C is 2, representing the real part and the imaginary part channels. Normalization processing is performed before input to improve training stability.

[0071] like Figure 2As shown, the convolutional layer has a kernel size of 3×3, a stride of 1, and a padding value of 1, used to achieve lossless feature map transfer. The improved residual network contains a total of 4 residual units. The ConvE is a downsampling module, the purpose of which is to ensure that the input and output modules have the same dimension, doubling the number of channels while halving the data dimension. In particular, when the input and output dimensions are the same, they will be directly connected by a solid line. In residual unit 1, the input, output, and stride of ConvA are 64, 64, and 1 respectively; the input, output, and stride of ConvB to D are 64, 64, and 1 respectively; and ConvE is not configured. In residual unit 2, the input, output, and stride of ConvA are 64, 128, and 2 respectively; the input, output, and stride of ConvB to D are 128, 128, and 1 respectively; and the input, output, and stride of ConvE are 64, 128, and 2 respectively. In residual unit 3, ConvA... The input channels, output channels, and step sizes are 128, 256, and 2, respectively; the input channels, output channels, and step sizes of ConvB to D are 256, 256, and 1, respectively; and the input channels, output channels, and step sizes of ConvE are 128, 256, and 2, respectively. In the residual unit 4, the input channels, output channels, and step sizes of ConvA are 256, 512, and 2, respectively; the input channels, output channels, and step sizes of ConvB to D are 512, 512, and 1, respectively; and the input channels, output channels, and step sizes of ConvE are 256, 512, and 2, respectively.

[0072] like Figure 2 As shown, the fully connected layer is used to receive the target mapping features output by the residual block group, flatten them, and then input them to two fully connected sub-layers. The output dimension of the first fully connected sub-layer is 2048, and the output dimension of the second fully connected sub-layer is 210. They are used to output the predicted probability distribution of the angle of arrival direction corresponding to the two signal sources in the target sector.

[0073] This invention sets up two experiments to evaluate the DOA estimation accuracy and estimation time of cascaded residual networks at different angular intervals.

[0074] In the first experiment, MUSIC, ESPRIT, and DNN were selected as control groups. The experiment used an 8-element uniform linear array with a fixed number of snapshots of 512 and a signal-to-noise ratio range of -10dB to 20dB (with a step size of 5dB). The RMSE performance of this network was compared with that of MUSIC, ESPRIT, and DNN networks for three angle interval scenarios of 3°, 6°, and 9°.

[0075] like Figure 3The figure shows the RMSE performance of this network compared to MUSIC, ESPRIT, and DNN networks under three source angle spacing scenarios: 3°, 6°, and 9°. This invention achieves excellent estimation results when the signal-to-noise ratio is low and the source angle spacing is small. Under low signal-to-noise ratio conditions, this network exhibits a significant advantage in angle-of-arrival estimation accuracy.

[0076] In the second experiment, MUSIC, ESPRIT, and DNN were again selected as control groups. The experiment used an 8-element uniform linear array with a fixed number of snapshots of 512 and a signal-to-noise ratio range of -10dB to 20dB (with a step size of 5dB). The estimation time of the cascaded residual network was compared and tested with that of the MUSIC, ESPRIT, and DNN networks.

[0077] like Figure 4 The figure shows a comparison of the running time of the cascaded residual network with that of MUSIC, ESPRIT, and DNN networks as the signal-to-noise ratio changes. We found that the cascaded residual network outperforms MUSIC, ESPRIT, and DNN networks in overall running time. This invention reduces the average running time by more than half compared to traditional algorithms and DNN networks for the same data scale, demonstrating good real-time performance and promising engineering application prospects.

[0078] All aspects of this invention not described in detail are well-known to those skilled in the art. The preferred embodiments of this invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of this invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of this invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for estimating angle of arrival (OA) based on residual networks, characterized in that, The method includes the following steps: Step 1: Construct a spatial encoder based on a deep neural network (DNN). The spatial encoder takes the covariance matrix of the signal received by the antenna array as input to extract its spatial features and perform angle-of-arrival (AOA) block division. Step 2: Construct an improved residual network based on ResNet residual network; used to achieve fine estimation of angle of arrival based on the angle of arrival blocks given by the spatial encoder; Step 3: Train the spatial encoder and the improved residual network separately, and cascade them to form a cascaded residual network; Step 4: Input the received signal of the antenna array element into the cascaded residual network of Step 3 to perform angle of arrival (DOA) estimation and obtain the DOA estimation result.

2. The method for estimating angle of arrival based on residual networks according to claim 1, characterized in that, The spatial encoder includes, in the order of input and output, the following: The input layer is used to normalize the covariance matrix of the received signal of the antenna array elements to obtain normalized features; Fully connected layers are used to perform linear transformations and dimension mappings on normalized features to enhance their expressive power. The Softmax layer is used to normalize the activation values ​​output by the fully connected layer into a probability distribution. The classification layer is used to classify the signal at the sector level based on the output of the Softmax layer structure. It divides the entire directional space into 6 fixed sectors and outputs the category label of the sector where the signal source is located for subsequent residual network calls.

3. The method for estimating angle of arrival based on residual networks according to claim 1, characterized in that, The input to step 1 is the covariance matrix of the received signal of the antenna array element, calculated using the following formula: Where y(n) is the received signal of the antenna element in the nth frame, and L is the number of frames. (*) H This indicates the conjugate transpose.

4. The method for estimating angle of arrival based on residual networks according to claim 1, characterized in that, The improved residual network includes: four residual units, one convolutional layer, one input layer, and one fully connected layer.

5. The angle of arrival estimation method based on residual networks according to claim 4, characterized in that, The improved residual network includes: Input layer: The two-dimensional feature map size of the signal received by the antenna array elements is uniformly M×M×C, where M is the number of array elements of the antenna array; C is 2, representing the real part and the imaginary part channels, and is normalized before input to improve training stability. Convolutional layer: Used to receive two-dimensional feature map data from the input layer, and uses two-dimensional convolution operation with a kernel size of 3×3, padding size of 1, and stride of 1 to extract the basic features of the feature map data; Four residual units are used to perform residual learning on the basic feature map output by the initial convolutional layer. Each residual unit includes multiple residual modules with identity mapping connections. Each residual module performs weighted fusion of the input features and the convolutionally transformed feature map through skip connections to obtain deep target mapping features and enhance the gradient propagation capability of the network during training. The fully connected layer receives the target mapping features output by the residual block group, flattens them, and then inputs them into two fully connected sub-layers. The output dimension of the first fully connected sub-layer is 2048, and the output dimension of the second fully connected sub-layer is 210. These sub-layers are used to output the predicted probability distribution of the angle of arrival directions corresponding to the two signal sources in the target sector.

6. The method for estimating angle of arrival based on residual networks according to claim 5, characterized in that, The residual module includes the following steps: Define residual cells: y l =h(x l )+F(x l ,W l ) x l+1 =f(y l ) Where x l y represents the input of the l-th residual module. l W represents the output of the l-th residual module. l Let f(*) represent the weight coefficients of the residual network, f(*) be the activation function, and h(x) be the weight coefficients of the residual network. l ) represents the target mapping function; The deep feature x is obtained by accumulating L-1 consecutive residual units. L Where x L The output x of the l-th residual unit l With the summation of a series of residual functions, W i The weighting coefficients represent the residual function.

7. The method for estimating angle of arrival based on residual networks according to claim 1, characterized in that, Step 3 includes: first, training a spatial encoder using the signal covariance matrix of the antenna array elements with partition labels; then, training an improved residual network using the original received signal feature map; and finally, cascading the two to form a cascaded residual network.

8. The angle of arrival estimation method based on residual networks according to claim 1, characterized in that, Step 4 includes: inputting the received signal of the antenna array element into the cascaded residual network of step 3 for angle of arrival (DOA) estimation. First, the sector label of the direction to which the angle of arrival belongs can be obtained through the spatial encoder. The label is then input into the corresponding improved residual network to obtain the angle of arrival (DOA) estimation result.

Images