Angle of Arrival Estimation System Based on Deep Learning

By designing a dual-branch neural network, combining classification and regression networks, and optimizing C-Branch, the accuracy reduction problem of traditional DoA estimation methods in model defects and source number changes is solved, and high resolution and robust DoA is achieved It is estimated that the computational complexity is reduced.

CN114487988BActive Publication Date: 2024-11-08FUDAN UNIVERSITY

Patent Information

Application Number
CN202111653969.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-30
Publication Date
2024-11-08
Estimated Expiration
2041-12-30

AI Technical Summary

Technical Problem

The traditional approach to reach an angle of arrival (DoA) estimation method reduces accuracy when model defects are defective and expands to scenes where the number of sources changes, making it difficult to achieve high resolution.

Method used

Design a dual-branch neural network (TB-Net), combining classification and regression networks, and using them in parallel, optimize C-Branch, considering gain and phase inconsistency, mutual coupling between sensors, and antenna position deviation, reducing computational complexity through feature extraction network sharing The C-Branch and R-Branch parallel structure eliminate data dependencies.

Benefits of technology

It improves the robustness and accuracy of DoA estimation, reduces the amount of calculation, can maintain high accuracy when model defects exist, and significantly improves estimation accuracy at low signal-to-noise ratio.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114487988B_ABST
    Figure CN114487988B_ABST
Patent Text Reader

Abstract

The present invention belongs to the field of wireless communication technology, and is specifically a system for estimating the angle of arrival based on deep learning. The system of the present invention includes a signal model, a data preprocessing module, and a dual-branch neural network model; the signal model is used to model the input signal; the data preprocessing module is used to preliminarily extract signal features; the dual-branch neural network model includes a feature extraction network and a parallel prediction network, and the parallel prediction network includes a classification branch, a regression branch, and an output layer; the feature extraction network is used to extract deep features of the input data and output them to the classification branch and the regression branch, the classification branch gives the approximate range of DoA, i.e., the mask, the regression branch gives the error of DoA relative to the grid point, and the output layer combines the results of the two branches to give the final DoA estimation result. The simulation results show that compared with the model-based deep learning method and the existing deep learning method, the present invention can achieve higher DoA estimation accuracy in the presence of model defects, and the size is only 1.8MB.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to an angle of arrival (DoA) estimation system based on deep learning. Background Art

[0002] In recent decades, Direction of Arrival (DoA) estimation has been widely studied in the fields of acoustics, radar, sonar, and wireless communications. Traditional DoA estimation methods, such as Multiple Signal Classification (MUSIC) and Signal Parameter Estimation Based on Rotation Invariant Technique (ESPIRIT) rely on accurate signal models, and their DoA estimation accuracy will be significantly reduced when the model is imperfect.

[0003] In recent years, with the rapid development of deep learning, DoA estimation algorithms based on neural networks have been proposed. Due to the data-driven characteristics, these methods are robust to model defects. These methods are generally divided into two types: regression networks and classification networks.

[0004] Under the regression network, different structures are proposed for estimating DoA values. For example, D.Hu et al. proposed an end-to-end algorithm, and L.Wu et al. used a deep convolutional network to recover the spatial spectrum. However, the network structure is heavily dependent on the number of sources, which makes it difficult to expand to scenarios where the number of sources changes. GKPapageorgiou et al. used neural networks for signal denoising, and A.Barthelme et al. used a deep neural network (DNN) to reconstruct the covariance matrix. AM Ahmed et al. proposed a DNN to map the received signal to a larger dimension, which can be equivalent to using a larger antenna array, thereby improving the DoA resolution.

[0005] For classification networks, the most common architecture is a grid-based model, which divides the angular domain into several sectors and then, for each sector, determines whether there is an incoming signal. Z. Liu et al. proposed an autoencoder with a multi-layer classifier to construct a spatial spectrum. H. Xiang et al. proposed two frameworks to separate coherent signals. GK Papageorgiou et al. proposed a deep convolutional neural network (CNN) with a two-dimensional convolutional layer to improve the DoA estimation accuracy under low signal-to-noise ratio (SNR). The grid-based model can improve the stability of training, and the structure is universal for scenarios with varying number of sources, but it is difficult to achieve high resolution due to the limited number of grids. Summary of the invention

[0006] The purpose of the present invention is to provide an arrival angle estimation system based on deep learning, which can reduce the amount of calculation by lightweighting the network while ensuring that the system is more robust to model defects.

[0007] The present invention combines the classification network and regression network structures that are widely used in DoA estimation by current deep learning, uses the two structures in parallel, designs a two-branch neural network that combines classification and regression in parallel, denoted as TB-Net, and optimizes the C-Branch, thereby improving the accuracy of C-Branch rough estimation of DoA.

[0008] The arrival angle estimation system based on deep learning provided by the present invention has a structure as follows: Figure 1 As shown, it mainly includes three parts: signal model, data preprocessing module, and dual-branch neural network model;

[0009] The signal model models the input signal to facilitate subsequent processing. Compared with the general signal model, the present invention takes into account the three model defects of gain and phase inconsistency, mutual coupling between sensors and antenna position deviation, thereby optimizing the model, that is, adding four parameters on the basis of the original model, representing gain deviation, phase deviation, mutual coupling deviation between sensors and antenna position deviation respectively;

[0010] The data preprocessing module is used to initially extract signal features for subsequent network processing of data. It mainly takes the upper triangular area of ​​the covariance matrix of the input data, reconstructs the matrices for its real and imaginary parts respectively, and then splices them together to obtain the processed data;

[0011] The dual-branch neural network model includes a feature extraction network and a parallel prediction network. The parallel prediction network includes a classification branch, a regression branch and an output layer. The feature extraction network uses 5 convolutional layers (consisting of a convolution part, a normalization part and a ReLU activation function part) to extract the deep features of the input data and output them to the classification branch and the regression branch. The classification branch gives the approximate range of DoA, namely the mask (Mask), and the regression branch gives the error (Deviation) of DoA relative to the grid point. The output layer combines the results of the two branches to give the final DoA estimation result.

[0012] In the network structure of the present invention, the computational complexity is reduced by sharing the feature extraction network, and the parallel structure of C-Branch and R-Branch eliminates the data dependency between networks, so TB-Net can give DoA estimation in one evaluation. The specific network structure is as follows Figure 2As shown, it includes a feature extraction network and a parallel prediction network. The feature extraction network consists of five convolutional layers, each of which includes a convolution part, a normalization part, and a ReLu activation function part. The number of input data channels is 2, and the output obtained after passing through the feature extraction network is the characteristic value of 128 channels. The classification branch and the regression branch are composed of convolutional layers and activation layers. The activation functions are Sigmoid function and Tanh function respectively. The output is a vector containing 121 elements. The result given by the classification branch represents the mask of DoA. The elements in the vector are 1 or 0, indicating whether there is a target DoA value in the sector of the angle corresponding to the element index. The regression branch gives the error value between the grid point and DoA corresponding to each element index of C-Branch that is 1. Finally, the output layer multiplies the element index of 1 in the result of the classification branch by the grid spacing and adds the error value given by the regression branch to obtain the final DoA result. During training, the classification branch uses binary cross entropy as the loss function, and the regression branch uses l2 norm as the loss function. The loss function of the classification branch is multiplied by 0.1 and added to the loss function of the regression branch to obtain the total loss function, which is used to train and optimize the entire network.

[0013] The specific operation steps of the system are as follows:

[0014] Firstly, the input signal is modeled and three model defects are considered, including gain and phase inconsistency, mutual coupling between sensors and antenna position deviation. Then the model is modified by adding four parameters to the original model, representing gain deviation, phase deviation, mutual coupling deviation between sensors and antenna position deviation.

[0015] Preprocess the input signal by first taking the covariance matrix of the input data, then taking the upper triangular area of ​​the covariance matrix, reconstructing the matrices for its real and imaginary parts respectively, and then splicing them together to obtain the processed data as the input of the network.

[0016] The modeled input signal is preprocessed and output to the feature extraction network. The feature extraction network gives the characteristics of the covariance matrix of the input signal and outputs it to C-Branch and R-Branch.

[0017] C-Branch outputs a mask vector m = [m1, m2, ..., m G ], the value of each element belongs to {0, 1}, indicating a rough DoA estimate, that is, the DoA may be at the grid position corresponding to the mask index of 1; R-Branch outputs a deviation vector d = [d1, d2, ..., d G ], giving the DoA deviation relative to the grid corresponding to the element index, which is used to refine the DoA estimate.

[0018] The output layer combines the output results of C-Branch and R-Branch, and first finds the index p of the K peaks in m = [p1, ..., p k ], multiplied by the grid size to get a rough DoA estimate, and then added to the corresponding deviation value in d to get a precise DoA estimate.

[0019] During training, binary cross entropy (BCE) is used as the loss function of C-Branch, l2 norm is used as the loss function of R-Branch, and the total loss is set to L = 0.1 × l BCE +l2.

[0020] The system generates data sets using a 16-element uniform linear array with half-wavelength element spacing, and randomly generates two sources in the range of [-60°, 60°]. The data sets used for training, validation, and testing are 100,000, 20,000, and 20,000, respectively.

[0021] The present invention implements TB-Net in Pytorch. During the training process, the initial learning rate is set to 0.001, and the learning rate is adjusted to 0.9 times the previous learning rate every 30 epochs. During the training process, the Adam optimizer is used to update the network parameters. The total training epoch is set to 300, and the candidate model with the highest DoA estimation accuracy is selected as the final model.

[0022] Features of the present invention

[0023] Compared with the traditional arrival angle estimation system, the arrival angle estimation system constructed by the present invention has higher robustness against model defects. When the input signal model has defects such as gain and phase inconsistency, mutual coupling between sensors and antenna position deviation, the DoA estimation result can still be guaranteed to have high accuracy.

[0024] Secondly, the network model of the present invention is relatively lightweight, which greatly reduces the amount of calculation. Specifically, the dual-branch neural network model of this system only uses convolutional layers, in which the feature extraction network is composed of only 5 convolutional layers, and the classification branch and regression branch share a feature extraction network. The size of the final network model is only 1.8MB. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 The structure of the proposed deep learning-based arrival angle estimation system.

[0026] Figure 2 This is the network structure of the dual-branch neural network proposed in the present invention.

[0027] Figure 3 To use l BCEComparison of the mean absolute error (MAE) of the arrival angle estimation results of the l2 classification branch.

[0028] Figure 4 To show the impact of introducing regression branches and not introducing them on the results. DETAILED DESCRIPTION

[0029] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0030] The first step is to prepare the input data.

[0031] The input signal modeling process is as follows: Let s k (t) is the kth input signal, θ k is DoA, and the received signal is expressed as:

[0032]

[0033] Among them, n(t)~CN(0,σ 2 ), is Gaussian noise, K is the number of sources, a(θ k ) is the array response vector. Assuming a uniform linear array (ULA) is used, we have:

[0034]

[0035] Where M is the number of antennas in the ULA, d is the spacing between adjacent antennas, and λ is the wavelength.

[0036] Consider three model defects, namely gain and phase inconsistency (e g , e pha ), mutual coupling between sensors (e m ) and the deviation of the antenna position (e pos ), a(θ k ) is rewritten as:

[0037]

[0038] The input signal preprocessing process is as follows: The covariance matrix of x(t) can be approximated as:

[0039]

[0040] Where N represents the number of snapshots. Then, the input value of the proposed neural network is expressed as follows:

[0041] u=[Vec(Real(Triu(R))),Vec(Imag(Triu(R))))], (5)

[0042] Among them, Triu(·) represents the upper triangular part of the matrix, and Vec(·) reconstructs the matrix into a vector.

[0043] Convolutional layers are widely used in neural networks. The output of the i-th layer can be expressed as:

[0044]

[0045] Among them, u i is the input feature, W i is the convolution kernel, b i is the bias, and nonlinear function f(·) (such as ReLU, Sigmoid, Tanh, etc.) is used for spatial mapping. To train the neural network, W i Updated by back propagation under a certain loss function.

[0046] In order to reduce the displacement of internal covariates and accelerate convergence, batch normalization (BN) can be performed before activation, where y i can be rewritten as:

[0047]

[0048] where v i is W in formula (6) i ×u i +b i , E[] means taking the average value, Var[] means taking the root mean square, ε means a smaller attempt, ε=10 -6 .

[0049] The TB-Net designed by the present invention includes a feature extraction network and a parallel prediction network. The parallel prediction network includes C-Branch, R-Branch and an output layer. The parameters of TB-Net include convolution kernel parameters, step size, normalization and activation function. The convolution kernel is represented by parameters (C_IN, C_OUT, H, W), where C_IN represents the number of input channels, C_OUT represents the number of output channels, H represents the kernel height, and W represents the kernel width.

[0050] The feature extraction network consists of 5 convolutional layers with convolution kernel parameters of (2,8,1,5), (8,32,1,5), (32,64,1,5), (64,128,1,5), and (128,128,1,3). The step size of the 5 convolutional layers is 2, and they are activated by the ReLU function after batch normalization (BN). The C-Branch convolution kernel parameters are (128,121,1,1), the step size is 1, and the activation function is the Sigmoid function. The R-Branch convolution kernel parameters are (128,121,1,1), the step size is 1, and the activation function is the Tanh function.

[0051] The implementation process of each layer of the network is as follows:

[0052] The feature extraction network extracts features from the covariance matrix and outputs them to C-Branch and R-Branch to achieve feature reuse and reduce computational complexity. The parameters of the network are determined through experiments. The results show that the network composed of 5 convolutional layers has the best performance. In addition, BN is used to accelerate convergence. Experiments show that using BN in the first five layers can achieve the best stability and highest accuracy.

[0053] In the parallel prediction network, C-Branch and R-Branch work in parallel. G represents the number of grids, and the output of C-Branch is a mask vector m = [m1, m2, ..., m G ], whose ith element indicates that the DoA may be near the ith grid. The output of R-Branch is a deviation vector d = [d1, d2, ..., d G ], where d i Represents the DoA deviation relative to the i grid.

[0054] For model optimization, the total loss function is set to:

[0055] L=0.1×l c +l r , (8)

[0056] Among them, l c is the C-Branch loss function, l r is the R-Branch loss function, and the coefficient 0.1 is the optimal coefficient obtained after multiple experimental tests.

[0057] C-Branch is implemented as follows: a source in the sector [θ-0.5°, θ+0.5°] (θ = -60°, -59°, ..., 60°) is approximated by a grid θ, and the result of C-Branch is a vector with 121 elements, whose values ​​belong to {0,1}. The subscripts of non-zero elements are used to roughly estimate DoA.

[0058] In C-Branch, Sigmoid is used as the activation function of the output layer, which maps the result to [0,1]. Binary cross entropy (BCE) is used as the loss function to optimize the neural network, that is:

[0059]

[0060] Among them, m i is the tag value, is the output of the network.

[0061] The implementation of R-Branch is as follows: R-Branch consists of a convolutional layer with 121 output channels, which is consistent with C-Branch. For the i-th channel, the output is the deviation on the i-th grid in C-Branch. This deviation is only valid in m i It is only effective when =1.

[0062] Since the grid size in C-Branch is Δθ, the deviation is limited to [-0.5Δθ, 0.5Δθ]. Therefore, weighted Tanh is used as the activation function, that is:

[0063] d i =0.5Δθ×Tanh(v i ). (10)

[0064] For training, we use l2 as the loss function to optimize the neural network, namely:

[0065]

[0066] Where, d i Indicates the actual deviation, Represents the output of R-Branch.

[0067] For the output layer, it is designed to combine m and d to obtain the DoA estimate. First, Find the index p of K (number of sources) peaks in [p1, ..., p k ] and obtain a rough DoA estimate by multiplying it with the grid size. Then, the final DoA estimate is obtained by adding the bias selected in d according to p, that is:

[0068]

[0069] in, represents the DoA estimate, and Δθ represents the grid size.

[0070] The above is the entire implementation process of the present invention. For the present invention, an experiment was conducted during the testing process to illustrate the advantages of the present invention.

[0071] The experiment used a 16-element uniform linear array with half-wavelength element spacing to generate the data set. Two sources with the same power were randomly generated in [-60°, 60°]. The data set sizes for training, validation, and testing were 100,000, 20,000, and 20,000, respectively.

[0072] We use the mean absolute error (MAE) to measure the performance of the algorithm, namely:

[0073]

[0074] Among them, N T Indicates the number of test samples.

[0075] In the experiment, l2 and l BCE The optimized C-Branches are compared and the results are as follows Figure 3 As shown. It can be seen that l BCE The optimized network achieves better accuracy than the l2-optimized network. At a signal-to-noise ratio (SNR) of 10dB, the accuracy is improved by about 44.7%.

[0076] Figure 4 The impact of introducing R-Branch is shown. It can be seen that the accuracy of DoA estimation improves with the increase of SNR. When the SNR is low, the rough estimate of C-Branch will be far away from the true DoA value, which significantly reduces the MAE. As the SNR increases, since the rough DoA estimate obtained by C-Branch has almost no error, the deviation given by R-Branch gradually dominates the estimation accuracy. This phenomenon is obvious when the SNR is >2dB. Compared with C-Branch, the accuracy of TB-Net is improved by about 36.4% at a SNR of 10dB.

Claims

1. An arrival angle estimation system based on deep learning, characterized in that: It includes three parts: signal model, data preprocessing module and dual-branch neural network model; among which: The signal model is used to model the input signal; wherein, three model defects, namely, gain and phase inconsistency, mutual coupling between sensors, and antenna position deviation, are considered, and the model is optimized, and four parameters are added to the original model, representing gain deviation, phase deviation, mutual coupling deviation between sensors, and antenna position deviation, respectively; The data preprocessing module is used to preliminarily extract signal features, including taking the upper triangular area of ​​the covariance matrix of the input data, reconstructing the matrices for the real part and the imaginary part respectively, and then splicing them together to obtain processed data; The dual-branch neural network model includes a feature extraction network and a parallel prediction network, and the parallel prediction network includes a classification branch, a regression branch and an output layer; the feature extraction network is used to extract deep features of the input data and output them to the classification branch and the regression branch. The classification branch gives the approximate range of DoA, i.e., the mask, and the regression branch gives the error of DoA relative to the grid point. The output layer combines the results of the two branches to give the final DoA estimation result.

2. The arrival angle estimation system based on deep learning according to claim 1, characterized in that: In the dual-branch neural network model: The feature extraction network consists of five convolutional layers, each of which includes a convolution part, a normalization part and a ReLu activation function part. The number of input data channels is 2, and the output obtained after passing through the feature extraction network is the characteristic value of 128 channels; The classification branch and regression branch are both composed of a convolution layer and an activation layer. The activation functions are Sigmoid function and Tanh function respectively. The output is a vector containing 121 elements. The result given by the classification branch represents the DoA mask. The elements in the vector are 1 or 0, indicating whether there is a target DoA value in the sector of the angle corresponding to the element index. The regression branch gives the error value between the grid point corresponding to each element index of C-Branch that is 1 and the DoA. The output layer multiplies the index of the element with a value of 1 in the result of the classification branch by the grid spacing and adds the error value given by the regression branch to obtain the final DoA result; During training, the classification branch uses binary cross entropy as the loss function, and the regression branch uses l2 norm as the loss function. The loss function of the classification branch is multiplied by 0.1 and added to the loss function of the regression branch to obtain the total loss function, which is used to train and optimize the entire network.

3. The arrival angle estimation system based on deep learning according to claim 2, characterized in that: The system works as follows: (1) First, the input signal is modeled and three model defects are considered, including gain and phase inconsistency, mutual coupling between sensors, and antenna position deviation. The model is modified by adding four parameters to the original model, representing gain deviation, phase deviation, mutual coupling deviation between sensors, and antenna position deviation. (2) Preprocessing the input signal by first obtaining the covariance matrix of the input data, then obtaining the upper triangular area of ​​the covariance matrix, reconstructing the matrices for the real and imaginary parts respectively, and then splicing them together to obtain the processed data; (3) Output the preprocessed data to the feature extraction network, which gives the features of the covariance matrix of the input signal and outputs them to C-Branch and R-Branch; C-Branch outputs a mask vector m = [m1, m2, ..., m G ], the value of each element belongs to {0, 1}, indicating a rough DoA estimate, that is, the DoA may be at the grid position corresponding to the mask index of 1; R-Branch outputs a deviation vector d = [d1, d2, ..., d G ], giving the DoA deviation relative to the grid corresponding to the element index, which is used to refine the DoA estimate; The output layer combines the output results of C-Branch and R-Branch, and first finds the index p of the K peaks in m = [p1, ..., p k ], multiplied by the grid size to get a rough DoA estimate, and then added to the corresponding deviation value in d to get a precise DoA estimate; During training, binary cross entropy is used as the loss function of C-Branch, l2 norm is used as the loss function of R-Branch, and the total loss is set to L = 0.1 × l BCE +l2.

4. The arrival angle estimation system based on deep learning according to claim 1, characterized in that: The process of input signal modeling is as follows: Assume k (t) is the kth input signal, θ k is DoA, and the received signal is expressed as: Among them, n(t)~CN(0,σ 2 ), is Gaussian noise, K is the number of sources, a(θ k ) is the array response vector; assuming a uniform linear array (ULA), we have: Where M is the number of antennas in the ULA, d is the spacing between adjacent antennas, and λ is the wavelength; Consider three model defects, namely gain e g and phase e pha Inconsistency, mutual coupling between sensors m The deviation from the antenna position e pos , a(θ k ) is rewritten as: Where i=1,…,M.

5. The arrival angle estimation system based on deep learning according to claim 4, characterized in that: The input signal preprocessing process is as follows: The covariance matrix of x(t) is approximately: Where N represents the number of snapshots; then, the input value of the proposed neural network is expressed as follows: u=[Vec(Real(Triu(R))),Vec(Imag(Triu(R))))] (5) Among them, Triu(·) represents the upper triangular part of the matrix, and Vec(·) reconstructs the matrix into a vector.

Citation Information

Patent Citations

  • Arrival angle estimation method based on adversarial regularization deep neural network

    CN111767791A

  • Lightweight multi-head age estimation method based on face feature learning

    CN113298004A

  • Single-stage small-sample-object detection method based on decoupled metric

    US11205098B1

  • KR20210110081A

Cited By

  • Uplink and downlink combined arrival angle estimation method and system based on reinforcement learning

    CN121356639A