A DOA estimation method based on multi-time joint perception signal enhancement

Abstract

The application discloses a DOA estimation method based on multi-time joint sensing signal enhancement, which comprises the following steps: based on a water surface scene, a sensing channel model is constructed for a moving target; in order to improve the DOA estimation accuracy of the water surface object, multi-time joint denoising is used to optimize the time-varying channel estimation result; for time-varying channel estimation, a deep learning algorithm based on Double U-Net is used for denoising in two dimensions of space and angle; using the denoised channel matrix, the MUSIC algorithm is used for DOA estimation. The application uses the Double U-Net algorithm based on the deep convolutional neural network to extract the complex features of the channel from the multi-time joint sensing observation data, realizes channel denoising and enhancement, and further improves the DOA estimation accuracy, so that the accurate identification and tracking of the water surface object are effectively realized. Compared with the traditional method, the DOA estimation method provided by the application has higher estimation accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of DOA estimation in the 5G-A / 6G synaesthesia and intelligent computing integrated scenario, and specifically relates to a DOA estimation method based on multi-time joint perception signal enhancement. Background Art

[0002] The fifth generation of mobile communications (5G-A) and the sixth generation of mobile communications (6G) are envisioned as multi-purpose systems that can provide users with integrated communication, perception, and intelligent computing services. One of the key technologies for providing this service is direction of arrival (DOA) estimation. DOA estimation can help the system determine the direction of the signal source, thereby optimizing signal reception and processing, and improving communication quality and sensing performance. In this context, the importance of direction of arrival estimation as one of the key technologies is becoming increasingly prominent. This is because DOA estimation can effectively extract the target's location information, enabling accurate identification and tracking of surface objects. However, the complexity of the surface environment poses significant challenges to signal processing. The unique electromagnetic properties of the water surface, the influence of fluctuations, and the reflection, scattering, and multipath effects encountered during signal propagation all significantly affect the accuracy of DOA estimation.

[0003] Existing DOA estimation methods, such as the conventional beamforming (CBF) method, are essentially a simple extension of the time-domain Fourier spectrum estimation method to the spatial domain. While the principle is simple, its resolution is limited and it is constrained by the array's Rayleigh limit. The multiple signal classification (MUSIC) method and the estimation of signal parameters via rotational invariance technique (ESPRIT) method both utilize the properties of signal subspaces to achieve super-resolution direction finding. However, these two methods are covariance-based and require a sufficient number of data samples (snapshots) to accurately estimate the true covariance matrix. With the recent development of artificial intelligence (AI), many deep learning-based DOA estimation methods have emerged, such as using a denoising autoencoder (DAE) to learn the mapping from sample covariance to the true covariance matrix before performing DOA estimation.

[0004] To improve the accuracy of DOA estimation and identify and locate surface objects, especially within the framework of integrated synaesthesia technology, it is necessary to construct a channel model suitable for the surface environment. This model fully considers the unique electromagnetic characteristics of the surface, including the impact of fluctuations on signals and the impact of multipath on signal quality. This serves as a basis for optimizing DOA estimation performance. Accurate channel modeling can provide reliable channel state information for DOA estimation, thereby improving target recognition accuracy and signal-to-noise ratio, and further optimizing the collaborative operation of communication and perception. Summary of the Invention

[0005] From the perspective of improving DOA estimation accuracy, this paper proposes a channel estimation method based on multi-time joint sensing signal enhancement for the mobility characteristics of surface sensing detection targets. This method decomposes DOA estimation in the Integrated Sensing and Communications (ISAC) system into two processes: multi-time joint channel denoising and MUSIC spectrum estimation. Its performance is superior to traditional algorithms.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a DOA estimation method based on multi-time joint perception signal enhancement, characterized by comprising:

[0007] Step 1: Determine the path loss formula for different scenarios, such as overwater and in suburban areas, and construct a time-varying perception channel model for mobile targets.

[0008] Step 2: Based on the time-varying perceptual channel model in step 1, obtain the noise-free channel H gt (t), preprocess it using least squares to obtain the LS estimation result And sample the values ​​at T moments;

[0009] Step 3: Based on the noiseless channel in step 2 And LS estimation results Offline training of the Double U-Net network;

[0010] Step 4: Obtain the new real channel to be estimated online, perform least squares processing on it and sample the results at T moments and input them into the offline trained Double U-Net network for joint denoising to obtain the denoised channel matrix

[0011] Step 5: Channel matrix after denoising The joint MUSIC algorithm is used for DOA estimation.

[0012] Furthermore, the specific steps of step 1 include:

[0013] Step 11: Path loss is divided into two categories: line-of-sight and non-line-of-sight. For water surface scenarios, the roughness factor C is introduced, and the calculation formula for the path loss on the water surface is:

[0014]

[0015] Among them, n OS represents the path loss factor on the water surface, I0 represents the indicator function of the reflection path, Γ0 is the reflection coefficient, is the phase difference between the direct path and the reflected path, I1 represents the indicator function of the scattered path, Γ1 is the scattering coefficient, is the phase difference between the direct path and the scattered path, λ represents the wavelength, and j is the imaginary unit;

[0016] Step 12: Based on the sparsity of the perception channel, the classic SV channel model is used, and combined with the mobility characteristics of the moving target, a time-varying perception channel model is constructed as

[0017]

[0018] Where t is the time variable, γ is the scalar coefficient, They represent the arrival angle and departure angle of the lth path at time t, α l (t) is the attenuation of the lth multipath, The steering vector representing the arrival angle of the lth path at time t, The steering vector representing the departure angle of the lth path at time t.

[0019] Furthermore, the expression of the roughness factor C in step 11 is:

[0020]

[0021] Among them, σ r represents the roughness, ψ is the ground contact angle, which is the complementary angle of the arrival angle θ, and λ represents the wavelength.

[0022] Furthermore, step 2 specifically includes the following steps:

[0023] Step 21: Generate the original uncontaminated data set using the time-varying perceptual channel model, denoted as the true channel matrix H gt (t);

[0024] Step 22: For the real channel matrix H gt (t) Use LS method for preprocessing to obtain the initial rough estimate

[0025] Step 23: For the real channel matrix H gt (t) and an initial rough estimate Sample T moments separately to obtain a noise-free channel And LS estimation results

[0026] Furthermore, the specific steps of step 3 include:

[0027] Step 31: Initialize the training parameters, where the total number of training rounds is E, the learning rate is η, the side length of the convolution kernel is k, the training batch size is B, the weight coefficient α, the number of training layers is L, and the random initial vector T at time T is the parameter Θ in the Double U-Net network;

[0028] Step 32: Noiseless Channel As training labels, LS estimation results As the data set to be trained;

[0029] Step 33: LS estimation results Perform DFT transformation to obtain the angle domain result, input it into the Double U-Net network to obtain the angle domain noise estimation value and the spatial domain noise estimation value;

[0030] Step 34: Combine the LS estimation results, the angular domain noise estimation value, and the spatial domain noise estimation value After multiple iterations until the loss function value is minimized, the trained Double U-Net network is obtained.

[0031] Furthermore, the channel matrix after Double U-Net denoising is The expression is

[0032]

[0033] in, represents the spatial noise estimate, Θ = {W, b} is used as a parameter, including the weight W and the offset b.

[0034] Furthermore, the loss function is

[0035]

[0036] Among them, ||D t || is the number of samples included.

[0037] Furthermore, the step 4 specifically includes:

[0038] Step 41: Obtain a new estimated true noiseless channel online, perform LS estimation on it and sample T time instants;

[0039] Step 42: Input the above results and the results after DFT transformation into the trained Double U-Net network to obtain the denoised channel matrix

[0040] Furthermore, the step 5 specifically includes:

[0041] Step 51: Obtain the denoised channel matrix at T moments Divide into T groups according to time, denoted as H t (t=1,L,T), calculate the covariance matrix respectively:

[0042] R t =H t (H t ) H ;

[0043] Step 52: Perform eigenvalue decomposition on the covariance matrix to decompose it into signal subspace and noise subspace:

[0044]

[0045] in, is the signal subspace, is the noise subspace, and are their corresponding eigenvalue matrices respectively;

[0046] Step 53: Based on the noise subspace For the receiving direction θ r (t), the MUSIC spectrum function is defined as

[0047]

[0048] And by searching the spectral function P MUSIC (θ r (t)) and estimate the arrival angle θ r (t);

[0049] Step 54: Use the number of missed detections as a performance indicator to detect and estimate the arrival angle θ r (t).

[0050] Furthermore, the specific steps of step 54 are:

[0051] If there are L detection targets in the group at the current t-th moment, the angle of each target is estimated to obtain the estimated angle With real angle By comparison, the average angle error is

[0052]

[0053] like It is considered that there is a missed detection in this group of data.

[0054] Beneficial effects:

[0055] 1. Compared with the DOA estimation scenario, the present invention uses the steps of step 1 and considers the perception channel wireless model based on the surface mobile detection target, which is different from the traditional communication perception scenario;

[0056] 2. Compared with the traditional DOA estimation method, the present invention uses steps 2 to 5, considers the impact of low signal-to-noise ratio (SNR) on DOA estimation in water scenes, and addresses the DOA estimation problem under time-varying perceptual channels. It decomposes the problem into two stages: multi-time joint channel denoising based on the Double U-Net algorithm and multi-time joint DOA estimation.

[0057] 3. Compared with the traditional method, the present invention uses steps 3 to 5. The channel denoising ratio based on the Double U-Net algorithm has higher accuracy than the traditional LS, denoising convolutional neural network (DnCNN) algorithm and single-time Double U-Net. At the same time, the DOA estimation accuracy based on this method is also improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 1 is a flow chart of a DOA estimation method based on multi-time joint perception signal enhancement provided in an embodiment of the present invention;

[0059] Figure 2 is a schematic diagram of a water surface sensing system provided in an embodiment of the present invention;

[0060] Figure 3 Schematic diagram of the Double U-Net network structure provided in an embodiment of the present invention;

[0061] Figure 4 1 is a schematic diagram of NMSE results of channel estimation using different methods provided in an embodiment of the present invention as SNR changes;

[0062] Figure 5 1 is a schematic diagram of the number of missed detections when the number of paths is 5 and different methods are used for DOA estimation as the SNR changes. DETAILED DESCRIPTION

[0063] The present invention will be further explained below with reference to the accompanying drawings.

[0064] like Figure 1 As shown, the present invention provides a DOA estimation method based on multi-time joint perception signal enhancement, comprising:

[0065] Step 1: Determine the path loss formula for the water surface scenario and construct a time-varying perception channel model for the mobile target. Step 2: Based on the time-varying perception channel model in step 1, obtain the noise-free channel H gt (t), preprocess it using least squares to obtain the LS estimation result And sample the values ​​at T moments.

[0066] Step 3: Based on the noiseless channel in step 2 And LS estimation results The Double U-Net network is trained offline.

[0067] Step 4: Obtain the new real channel to be estimated online, perform least squares processing on it and sample the results at T moments and input them into the offline trained Double U-Net network for joint denoising to obtain the denoised channel matrix

[0068] Step 5: Channel matrix after denoising The joint MUSIC algorithm is used for DOA estimation.

[0069] Step 1 specifically includes the following steps:

[0070] Step 11: Path loss is divided into two categories: line-of-sight and non-line-of-sight. For water surface scenarios, the roughness factor C is introduced to obtain the water surface path loss calculation formula:

[0071]

[0072] Among them, n OS represents the path loss factor on the water surface, I0 represents the indicator function of the reflection path, Γ0 is the reflection coefficient, is the phase difference between the direct path and the reflected path, I1 represents the indicator function of the scattered path, Γ1 is the scattering coefficient, is the phase difference between the direct path and the scattered path, λ is the wavelength, and j is the imaginary unit.

[0073] Step 12: Based on the sparsity of the perception channel, the classic SV channel model is used, and combined with the mobility characteristics of the moving target, a time-varying perception channel model is constructed as

[0074]

[0075] Where t is the time variable, γ is the scalar coefficient, They represent the arrival angle and departure angle of the lth path at time t, αl (t) is the attenuation of the lth multipath, The steering vector representing the arrival angle of the lth path at time t, The steering vector representing the departure angle of the lth path at time t.

[0076] The expression of the roughness factor C in step 11 is

[0077]

[0078] Among them, σ r represents the roughness, ψ is the ground contact angle, which is the complementary angle of the arrival angle θ, and λ represents the wavelength.

[0079] In step 12, based on the ULA model, the steering vector a Ω (θ(t)) can be expressed as

[0080]

[0081] Among them, N Ω is the number of antenna arrays, d and λ represent the spacing between adjacent antennas and the signal wavelength, respectively. Therefore, by steering vector a Ω The formula (θ(t)) can be calculated and

[0082] like Figure 2 As shown, this embodiment is located in a water surface scene. There are 5 detection targets in the scene. The base station is set at the position (0, 0, 20). There are 5 data points of the detection targets in the rectangular area formed by the four points (-200, 50, 0), (-200, 250, 0), (200, 50, 0), and (200, 250, 0). For multi-target perception of objects on the water surface, the perception method of synaesthesia is divided into active perception and passive perception according to whether the party receiving the signal knows the signal sent by the sender in advance. When active perception is used and the receiver and sender are the same antenna, the parameters are set to the base station height h = 20m and the number of antennas N. t 、N r = 64, the carrier frequency is f = 2.7 GHz, the distance between the water surface and the base station is R = 10 m, the length and width of the water surface are both W = 500 m, and the departure and arrival angles of the antenna are θ. Objects are detected through echo signals reflected from them. For moving targets, the target has an initial velocity v (m / s). Accordingly, the channel transfer matrix H also changes in real time, becoming a function of time H(t).

[0083] Step 2: The specific steps include:

[0084] Step 21: Considering the target movement scenario of water surface detection, the time-varying perception channel model is used to generate the original uncontaminated data set, which is recorded as the real channel matrix H gt (t).

[0085] Step 22: Considering the low complexity of LS estimation, the real channel matrix H gt (t) Perform LS estimation to obtain an initial rough estimate

[0086] Step 23: Due to the time-varying nature of the perceptual channel, joint denoising is performed using temporal correlation. Specifically, for the real channel matrix H gt (t) and an initial rough estimate Sample T moments separately to obtain a noise-free channel And LS estimation results

[0087] Step 3: The specific steps include:

[0088] Step 31: Initialize the training parameters, where the total number of training rounds is E, the learning rate is η, the side length of the convolution kernel is k, the training batch size is B, the weight coefficient α, the number of training layers is L, and the random initial vector Θ in the Double U-Net network at T moments.

[0089] Step 32: Noiseless Channel As training labels, LS estimation results As the data set to be trained; Step 33: define the loss function as:

[0090]

[0091] Among them, ||D t || is the number of samples included.

[0092] Step 34: LS estimation results Perform DFT transformation to obtain the angle domain result, input it into the Double U-Net network, and obtain the angle domain noise estimation value and the spatial domain noise estimation value.

[0093] Step 35: Combine the LS estimation results, the angular domain noise estimation value, and the spatial domain noise estimation value After multiple iterations until the loss function value is minimized, the trained Double U-Net network is obtained.

[0094] In this embodiment, the total number of training rounds is E=100, and the learning rate is η=10 -2, the convolution kernel has a side length of k = 3, the training batch size is B = 128, the weight coefficient α = 0.5, the number of training layers is L = 9, and the goal of the Double U-Net algorithm is to minimize the estimated denoised channel matrix With noiseless channel The NMSE value between , therefore the above loss function is defined.

[0095] like Figure 3 As shown, the Double U-Net model of this embodiment adopts a dual-path encoder-decoder architecture. The LS channel estimates of T time sampling points are simultaneously input as multiple channels, and channel denoising is achieved through joint processing. One path processes the angle domain data after DFT, while the other path directly processes the original spatial domain data. The two paths are processed in parallel, extracting features from the angle domain and spatial domain respectively. Finally, the results are combined to achieve comprehensive noise suppression. The specific output of the Double U-Net algorithm can be expressed as follows:

[0096]

[0097] in, and Respectively represent the LS estimation values ​​in the spatial domain and the angular domain The noise estimate obtained by the U-Net network is represents the spatial noise estimate, Θ = {W, b} is used as a parameter, including the weight W and the offset b.

[0098] Step 4: The specific steps include:

[0099] Step 41: Obtain a new estimated true noiseless channel online, perform LS estimation on it and sample T time instants;

[0100] Step 42: Input the above results and the results after DFT transformation into the trained Double U-Net network to obtain the denoised channel matrix

[0101] Select SNR (dB) values ​​of -10, -5, 0, 5, and 10, and calculate the NMSE values ​​for the LS estimation results, DnCNN, Double U-Net, and multi-time joint estimation results. The results are as follows Figure 4 shown.

[0102] Step 5: The specific steps include:

[0103] Step 51: Obtain the denoised channel matrix at T moments Divide into T groups according to time, denoted as H t(t=1,L,T), calculate the covariance matrix respectively:

[0104] R t =H t (H t ) H .

[0105] Step 52: Perform eigenvalue decomposition on the covariance matrix to decompose it into signal subspace and noise subspace:

[0106]

[0107] in, is the signal subspace, is the noise subspace, and are their corresponding eigenvalue matrices respectively.

[0108] Step 53: Based on the noise subspace For the receiving direction θ r (t), the MUSIC spectrum function is defined as

[0109]

[0110] And by searching the spectral function P MUSIC (θ r (t)) and estimate the arrival angle θ r (t).

[0111] Step 54: Use the number of missed detections as a performance indicator to detect and estimate the arrival angle θ r (t).

[0112] In step 54, the specific steps of using the number of missed detections as the performance indicator include:

[0113] If there are L detection targets in the group at the current t-th moment, the angle of each target is estimated to obtain the estimated angle With real angle By comparison, the average angle error is

[0114]

[0115] like It is considered that there is a missed detection in this set of data. Therefore, the DOA estimation result can be tested.

[0116] Select SNR (dB) values ​​of -10, -5, 0, 5, and 10. DOA estimation is performed on the LS estimation results, DnCNN and DoubleU-Net results at a single moment, and the LS estimation results and DoubleU-Net estimation results at multiple moments. A total of 2500 tests are performed, and the missed detection results are as follows: Figure 5 shown.

[0117] This paper compares the estimation performance of multi-time joint estimation based on the DoubleU-Net algorithm with other traditional algorithms, such as the DnCNN and LS algorithms, and methods without multi-time joint estimation at different SNRs. The horizontal axis represents the SNR, and the vertical axis represents the mean square error between the obtained channel estimate and the standard value, as well as the number of missed detections. As the SNR increases, the mean square error of the channel estimation decreases, and the number of missed detections in the DOA estimation also decreases. Comparing our proposed Double U-Net algorithm with other algorithms, our algorithm has better performance under the same SNR conditions. This shows that the DOA estimation method proposed in this embodiment effectively solves the perception, positioning, and tracking problems in the scenario of mobile targets on the surface of water, and improves the accuracy of DOA estimation.

[0118] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A DOA estimation method based on multi-time joint perception signal enhancement, characterized in that: include: Step 1: Determine the path loss formula for different water surface scenarios and construct a time-varying perception channel model for the mobile target; Step 2: Based on the time-varying perceptual channel model in step 1, obtain the noise-free channel H gt (t), preprocess it using least squares to obtain the LS estimation result And sample the values ​​at T moments; Step 3: Based on the noiseless channel in step 2 And LS estimation results Offline training of the Double U-Net network; Step 4: Obtain the new real channel to be estimated online, perform least squares processing on it and sample the results at T moments and input them into the offline trained Double U-Net network for joint denoising to obtain the denoised channel matrix Step 5: Channel matrix after denoising The joint MUSIC algorithm is used for DOA estimation; The specific steps of step 1 include: Step 11: Path loss is divided into two categories: line-of-sight and non-line-of-sight. For water surface scenarios, the roughness factor C is introduced, and the calculation formula for the path loss on the water surface is: Among them, n OS represents the path loss factor on the water surface, I0 represents the indicator function of the reflection path, Γ0 is the reflection coefficient, is the phase difference between the direct path and the reflected path, I1 represents the indicator function of the scattered path, Γ1 is the scattering coefficient, is the phase difference between the direct path and the scattered path, λ represents the wavelength, and j is the imaginary unit; Step 12: Based on the sparsity of the perception channel, the classic SV channel model is used, and combined with the mobility characteristics of the moving target, a time-varying perception channel model is constructed as Where t is the time variable, γ is the scalar coefficient, They represent the arrival angle and departure angle of the lth path at time t, α l (t) is the attenuation of the lth multipath, The steering vector representing the arrival angle of the lth path at time t, The steering vector representing the departure angle of the lth path at time t; The expression of the roughness factor C in step 11 is: Among them, σ r represents the roughness, ψ is the ground contact angle, which is the complementary angle of the arrival angle θ, and λ represents the wavelength; The step 5 specifically includes: Step 51: Obtain the denoised channel matrix at T moments Divide into T groups according to time, denoted as H t (t=1,…,T), calculate the covariance matrix respectively: R t =H t (H t ) H ; Step 52: Perform eigenvalue decomposition on the covariance matrix to decompose it into signal subspace and noise subspace: in, is the signal subspace, is the noise subspace, and are their corresponding eigenvalue matrices respectively; Step 53: Based on the noise subspace For the receiving direction θ r (t), the MUSIC spectrum function is defined as And by searching the spectral function P MUSIC (θ r (t)) and estimate the arrival angle θ r (t); Step 54: Use the number of missed detections as a performance indicator to detect and estimate the arrival angle θ r (t); The specific steps of step 54 are: If there are L detection targets in the group at the current t-th moment, the angle of each target is estimated to obtain the estimated angle With real angle By comparison, the average angle error is like It is considered that there is a missed detection in this group of data.

2. The DOA estimation method based on multi-time joint perception signal enhancement according to claim 1, characterized in that: Step 2: The specific steps include: Step 21: Generate the original uncontaminated data set using the time-varying perceptual channel model, denoted as the true channel matrix H gt (t); Step 22: For the real channel matrix H gt (t) Use LS method for preprocessing to obtain the initial rough estimate Step 23: For the real channel matrix H gt (t) and an initial rough estimate Sample T moments separately to obtain a noise-free channel And LS estimation results 3. The DOA estimation method based on multi-time joint perception signal enhancement according to claim 1, characterized in that: The specific steps of step 3 include: Step 31: Initialize the training parameters, where the total number of training rounds is E, the learning rate is η, the side length of the convolution kernel is k, the training batch size is B, the weight coefficient α, the number of training layers is L, and the random initial vector T at time T is the parameter Θ in the Double U-Net network; Step 32: Noiseless Channel As training labels, LS estimation results As the data set to be trained; Step 33: LS estimation results Perform DFT transformation to obtain the angle domain result, input it into the Double U-Net network to obtain the angle domain noise estimation value and the spatial domain noise estimation value; Step 34: Combine the LS estimation results, the angular domain noise estimation value, and the spatial domain noise estimation value After multiple iterations until the loss function value is minimized, the trained Double U-Net network is obtained.

4. The DOA estimation method based on multi-time joint perception signal enhancement according to claim 3, characterized in that: The channel matrix after Double U-Net denoising The expression is in, represents the spatial noise estimate, Θ = {W, b} is used as a parameter, including the weight W and the offset b.

5. The DOA estimation method based on multi-time joint perception signal enhancement according to claim 3, characterized in that: The loss function is Among them, ||D t || is the number of samples included.

6. The DOA estimation method based on multi-time joint perception signal enhancement according to claim 1, characterized in that: The step 4 specifically includes: Step 41: Obtain a new estimated true noiseless channel online, perform LS estimation on it and sample T time instants; Step 42: Input the above results and the results after DFT transformation into the trained Double U-Net network to obtain the denoised channel matrix

Images