Sparse Channel Estimation Method, Apparatus, Device and Storage Medium

Abstract

The present invention belongs to the field of wireless communication technologies, and discloses a sparse channel estimation method, apparatus, device, and storage medium. The method includes: inputting a signal to be processed into an angular resampling network to obtain a plurality of sampling angles; constructing a sparse channel representation dictionary of the signal to be processed according to the sampling angles; inputting the sparse channel representation dictionary and the signal to be processed into a deep unfolding neural network, and obtaining a sparse channel estimation value of the signal to be processed through the signal to be processed, the sparse channel representation dictionary, and the pre-trained learnable parameters in the deep unfolding neural network. The present invention adaptively determines the sampling angles close to the true arrival angles through resampling, constructs a sparse channel representation dictionary in the channel angular domain based on this, realizes a more effective channel sparse representation, and the deep unfolding learning network quickly converges based on the sparse channel representation dictionary in the channel angular domain to obtain a more accurate channel estimation value, solving the technical problem that it is difficult to efficiently perform sparse channel estimation on the basis of using a large bandwidth and a high-dimensional antenna array.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a sparse channel estimation method, device, equipment and storage medium. Background Art

[0002] Shannon's information theory is an important foundation for the design of 6G communication systems, revealing two main ways to increase system capacity: increasing system bandwidth and improving spectrum efficiency. As one of the key physical layer technologies that expands in frequency, bandwidth, and spatial dimensions, broadband massive multiple input multiple output (MIMO) systems are a hot topic in the evolution of 6G. The key to fully leveraging its technical advantages lies in effective channel estimation, which is the basis for the implementation of technologies such as channel equalization, precoding, beamforming, and resource allocation. It is crucial to ensure the effective transmission of broadband massive MIMO signals and obtain broadband massive MIMO system gain.

[0003] Due to the use of large bandwidth and high-dimensional antenna arrays, channel estimation faces challenges such as complex channel characteristics, large pilot overhead, and high algorithm complexity. The contradiction between the limited system pilot resources and the extremely large dimensions of the base station antenna array and the high channel estimation accuracy requirements is particularly prominent. By fully exploring the sparse characteristics of wireless multipath channels in the time domain, frequency domain, and spatial domain, channel estimation is modeled as a sparse signal reconstruction problem under the theoretical framework of compressed sensing (CS). The advantage of compressed sensing technology in reconstructing high-dimensional signals with a high probability based on a small number of observations is fully utilized to achieve the purpose of reducing the required pilot length and improving the accuracy of channel estimation. This is one of the main solutions to the channel estimation problem in large-scale MIMO systems. However, with the increase in the scale of the wireless channel matrix, the shortcomings of this method, such as intensive calculation, slow convergence, and failure to guarantee the global optimal solution, have become bottlenecks in the application of compressed sensing in channel estimation.

[0004] The above contents are only used to assist in understanding the technical solution of the present invention and do not constitute an admission that the above contents are prior art. Summary of the invention

[0005] The main purpose of the present invention is to provide a sparse channel estimation method, aiming to solve the technical problem that it is difficult to efficiently perform sparse channel estimation in the prior art.

[0006] To achieve the above object, the present invention provides a sparse channel estimation method, the method comprising the following steps:

[0007] Inputting the signal to be processed into an angle resampling network to obtain multiple sampling angles;

[0008] Construct a sparse channel representation dictionary of the signal to be processed according to the sampling angle;

[0009] Input the sparse channel representation dictionary and the signal to be processed into a deep unfolded neural network, and obtain the sparse channel estimation value of the signal to be processed through the pre-trained learnable parameters, the signal to be processed, and the sparse channel representation dictionary in the deep unfolded neural network.

[0010] Optionally, the angle resampling network includes a conversion layer, a feature extraction layer, and a resampling prediction layer;

[0011] The step of inputting the signal to be processed into the angle resampling network to obtain multiple sampling angles includes:

[0012] Convert the signal to be processed into an amplitude feature signal through the conversion layer;

[0013] Extract the direction-of-arrival (DOA) feature of the amplitude feature signal through the feature extraction layer;

[0014] Based on the DOA feature, perform prediction through the resampling prediction layer to obtain the predicted true DOA, and perform sampling according to the true DOA to obtain multiple sampling angles.

[0015] Optionally, the step of performing prediction through the resampling prediction layer based on the DOA feature to obtain the predicted true DOA and performing sampling according to the true DOA to obtain multiple sampling angles includes:

[0016] Compare the weights of the DOA features to obtain the predicted true DOA, and obtain a preset sampling grid and a preset sampling fraction;

[0017] Determine the sampling range according to the predicted true DOA and the sampling grid, and obtain the sampling interval according to the sampling range and the preset sampling fraction;

[0018] Perform resampling within the sampling range according to the sampling interval to obtain multiple sampling angles.

[0019] Optionally, before inputting the signal to be processed into the angle resampling network to obtain multiple sampling angles, it further includes:

[0020] Obtain an angle resampling data set, where the angle resampling data set includes multiple corresponding training signals and true DOAs;

[0021] Input the training signal into the initial angle resampling network to obtain the initial resampling angle;

[0022] Construct a Gaussian mixture loss function based on the Gaussian distribution of the true angle of arrival, the probability that each sampling point is equal to the true angle of arrival, the angle of each sampling point, and a preset standard deviation;

[0023] Optimize the initial angle resampling network according to the Gaussian mixture loss function to obtain an angle resampling network.

[0024] Optionally, the step of inputting the sparse channel representation dictionary and the signal to be processed into a deep unfolded neural network to obtain the sparse channel estimation value of the signal to be processed through the pre-trained learnable parameters, the signal to be processed, and the sparse channel representation dictionary in the deep unfolded neural network includes:

[0025] Input the sparse channel representation dictionary and the signal to be processed into a deep unfolded neural network, and the deep unfolded neural network includes multiple cascaded layers;

[0026] Calculate the sparse channel estimation value of the next cascaded layer according to the sparse channel representation dictionary, the signal to be processed, the sparse channel estimation value, learnable parameters, and parameters to be estimated of the current cascaded layer;

[0027] When the next cascaded layer is a preset cascaded layer, use the sparse channel estimation value of the next cascaded layer as the sparse channel estimation value of the signal to be processed.

[0028] Optionally, the parameters to be estimated include Gaussian white noise parameters, sparse channel estimation values, and Gaussian prior distribution parameters;

[0029] The step of calculating the sparse channel estimation value of the next cascaded layer according to the sparse channel representation dictionary, the signal to be processed, the sparse channel estimation value, learnable parameters, and parameters to be estimated of the current cascaded layer includes:

[0030] Calculate the sparse channel estimation value of the current cascaded layer according to the sparse channel representation dictionary, the signal to be processed, the learnable parameters, Gaussian white noise parameters, Gaussian prior distribution parameters, and sparse channel estimation value of the previous cascaded layer;

[0031] Calculate the Gaussian prior distribution parameter of the current cascaded layer according to the Gaussian posterior distribution parameter and the sparse channel estimation value of the current cascaded layer;

[0032] Calculate the Gaussian white noise parameter of the current cascaded layer according to the sparse channel representation dictionary, the signal to be processed, the sparse channel estimation value of the current cascaded layer, the learning parameters, the sparse channel estimation value, and the Gaussian posterior distribution parameter of the previous cascaded layer.

[0033] Calculate the sparse channel estimate of the next cascaded layer based on the sparse channel representation dictionary, the signal to be processed, and the learnable parameters, Gaussian white noise parameters, Gaussian prior distribution parameters, and sparse channel estimate of the current cascaded layer.

[0034] Optionally, before inputting the sparse channel representation dictionary and the signal to be processed into the deep unfolding neural network to obtain the sparse channel estimate of the signal to be processed through the pre-trained learnable parameters, the signal to be processed, and the sparse channel representation dictionary in the deep unfolding neural network, it further includes:

[0035] Input multiple training signals and the sparse channel representation dictionary into the initial deep unfolding neural network to obtain multiple sparse channel training values;

[0036] Calculate the estimated channel matrix of the training signal according to the sparse channel training values;

[0037] Construct a feature loss function according to the number of training signals and the estimated channel matrix;

[0038] Optimize the learnable parameters of the initial deep unfolding neural network according to the feature loss function to obtain the deep unfolding neural network.

[0039] In addition, to achieve the above object, the present invention also proposes a sparse channel estimation device, and the sparse channel estimation device includes:

[0040] A resampling module, configured to input a signal to be processed into an angular resampling network to obtain multiple sampling angles;

[0041] A sparse channel representation dictionary construction module, configured to construct a sparse channel representation dictionary of the signal to be processed according to the sampling angles;

[0042] A sparse channel estimation module, configured to input the sparse channel representation dictionary and the signal to be processed into the deep unfolding neural network, and obtain the sparse channel estimate of the signal to be processed through the pre-trained learnable parameters, the signal to be processed, and the sparse channel representation dictionary in the deep unfolding neural network.

[0043] In addition, to achieve the above object, the present invention also proposes a sparse channel estimation device, and the sparse channel estimation device includes: a memory, a processor, and a sparse channel estimation program stored on the memory and executable on the processor, and the sparse channel estimation program is configured to implement the steps of the sparse channel estimation method as described above.

[0044] In addition, to achieve the above object, the present invention also provides a storage medium, on which a sparse channel estimation program is stored. When the sparse channel estimation program is executed by a processor, the steps of the sparse channel estimation method as described above are implemented.

[0045] The present invention adaptively determines sampling angles related to the true angle of arrival through resampling, constructs a channel angular domain sparse channel representation dictionary based on the resampled sampling angles, realizes a more effective channel sparse representation, inputs the effective channel sparse representation into an unfolded deep learning network, can quickly converge to calculate a more accurate channel estimation value, and then constructs a channel angular domain sparse channel representation dictionary based on this to achieve a more effective channel sparse representation. The deep unfolded learning network quickly converges based on the channel angular domain sparse channel representation dictionary to obtain a more accurate channel estimation value, solving the technical problem of difficult efficient sparse channel estimation on the basis of using a large bandwidth and a high-dimensional antenna array. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a schematic structural diagram of a sparse channel estimation device in a hardware operating environment related to the solution of an embodiment of the present invention;

[0047] Figure 2 is a schematic flowchart of the first embodiment of the sparse channel estimation method of the present invention;

[0048] Figure 3 is a schematic diagram of a cascade layer in an embodiment of the sparse channel estimation method of the present invention;

[0049] Figure 4 is a schematic flowchart of the second embodiment of the sparse channel estimation method of the present invention;

[0050] Figure 5 is a schematic diagram of an angle resampling network in an embodiment of the sparse channel estimation method of the present invention;

[0051] Figure 6 is a schematic flowchart of the third embodiment of the sparse channel estimation method of the present invention;

[0052] Figure 7 is a schematic diagram of a deep unfolded neural network in an embodiment of the sparse channel estimation method of the present invention;

[0053] Figure 8 is a schematic block diagram of the first embodiment of the sparse channel estimation device of the present invention.

[0054] The realization, functional features and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0055] It should be understood that the specific embodiments described herein are for explaining the present invention and not for limiting the present invention.

[0056] Referring to Figure 1 , Figure 1 is a schematic structural diagram of a sparse channel estimation device for the hardware operating environment involved in the solution of the embodiment of the present invention.

[0057] As Figure 1 shown, the sparse channel estimation device may include: a processor 1001, such as a Central Processing Unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display) and an input unit such as a keyboard (Keyboard). Optionally, the user interface 1003 may further include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be a high-speed Random Access Memory (RAM) memory or a stable non-volatile memory (Non-Volatile Memory, NVM), such as a disk memory. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0058] Those skilled in the art can understand that Figure 1 the structure shown in

[0059] does not constitute a limitation on the sparse channel estimation device, and may include more or fewer components than shown, or combine some components, or have different component arrangements. Figure 1 shown, in the memory 1005 as a storage medium, there may be included an operating system, a network communication module, a user interface module, and a sparse channel estimation program.

[0060] In Figure 1 the sparse channel estimation device shown, the network interface 1004 is mainly used for data communication with a network server; the user interface 1003 is mainly used for data interaction with a user; the processor 1001 and the memory 1005 in the sparse channel estimation device of the present invention may be provided in the sparse channel estimation device. The sparse channel estimation device calls the sparse channel estimation program stored in the memory 1005 through the processor 1001 and executes the sparse channel estimation method provided by the embodiment of the present invention.

[0061] An embodiment of the present invention provides a sparse channel estimation method. Refer to Figure 2 , Figure 2 which is a schematic flowchart of the first embodiment of a sparse channel estimation method of the present invention.

[0062] In this embodiment, the sparse channel estimation method includes the following steps:

[0063] Step S10: Input the signal to be processed into the angle resampling network to obtain multiple sampling angles.

[0064] It can be understood that the signal to be processed can be a received signal containing multiple pilot symbols received by the antenna array at the base station end.

[0065] It should be noted that consider an uplink broadband massive MIMO communication system. The base station is equipped with a uniform linear antenna array (ULA) with dimension M (M >> 1) to serve a single-antenna user, and orthogonal frequency division multiplexing (OFDM) is used as the signal transmission mechanism. Assume that λ c is the wavelength, f c is the carrier frequency, and f s is the transmission bandwidth. To estimate the uplink channel, a pilot-assisted channel estimation method is adopted, that is, among the N subcarriers included in each OFDM symbol, K (K << N) subcarriers are used to transmit pilot symbols, and the pilot position identifier is defined as At this time, the received signal containing K pilot symbols received by the antenna array at the base station end, and this received signal Y ψ can be expressed as:

[0066] Y ψ = H ψ X ψ + N ψ where, represents the pilot matrix, where ξ k satisfies the uniform distribution U(0, 2π), and diag{·} represents the diagonalization operation symbol, represents the Gaussian white noise, which satisfies the complex Gaussian distribution CN(0, σ -1 ), where, represents the uplink channel matrix, where, represents the broadband antenna array response vector, and where, β l represents the l-th complex channel gain, L represents the number of multipaths between the base station and the user, and τ lDenote the time delay of the \(l\)-th path between the user and the first antenna element of the base station antenna, \(\theta\) l Denote the true angle of arrival of the \(l\)-th path Denote the frequency interval Denote the normalized angle of arrival, \(d\) α Denote the distance between antenna elements

[0067] It should be further noted that, based on the above received signal \(Y\) Ψ , due to the limited number of scatterers, the wideband millimeter-wave massive MIMO channel is sparse in the angular domain. Assume where Since \(S\) is large, the true angle of arrival is a subset of. Let be the sparse channel representation dictionary of the channel and the sparse representation matrix of the sparse channel representation dictionary of the channel. Then the above received signal \(Y\) Ψ (the signal to be processed) channel \(H\) Ψ can be expressed as:

[0068]

[0069] where

[0070]

[0071]

[0072]

[0073] where \(VEC(\cdot)\) represents the vectorization operation represents the Kronecker product, and \(\odot\) represents element-wise multiplication. From the above formula, it can be clearly obtained to require the channel of the signal to be processed, then it is required to have the sparse channel representation dictionary of the channel and the sparse representation matrix of the sparse channel representation dictionary of the channel. Because and \(X\) ψ is known, so the sparse representation matrix ψ can be calculated by solving \(w\) where \(w\) ψ represents the estimated value of the sparse channel

[0074] Furthermore, the above formula can be expanded to obtain:

[0075]

[0076] where Based on the above formula, the sparse channel representation dictionary of the signal to be processed \(y\) ψ and the signal to be processed and Gaussian white noise N ψ Calculate the estimated value of the sparse channel.

[0077] It should be noted that the sparse channel representation dictionary is the basis for calculating the estimated value of the sparse channel. A more effective sparse channel representation dictionary can calculate the estimated value of the sparse channel more accurately. The sparse channel representation dictionary is constructed through multiple sampling angles. In order to obtain a more accurate sparse channel representation dictionary for more effective sparse representation, the sampling angles for constructing the sparse channel representation dictionary need to be closer to the true angles.

[0078] It should be understood that when receiving the signal to be processed, since interference noise will be generated during the signal transmission process and the interference noise also has an arrival direction, it is impossible to determine the true arrival angle of the signal to be processed based on the interference noise.

[0079] It should be noted that the angle resampling network can be a neural network that predicts the true arrival angle of the signal to be processed and samples multiple sampling angles according to the predicted true arrival angle.

[0080] It can be understood that the angle resampling network needs to be trained before actually using the angle resampling. The trained angle resampling network can predict the true arrival angle of the signal to be processed more accurately.

[0081] It should be understood that the sampling angles are multiple angles obtained by angle sampling around the predicted true arrival angle obtained by the angle resampling network.

[0082] Step S20: Construct the sparse channel representation dictionary of the signal to be processed according to the sampling angles.

[0083] It should be noted that based on the formula Formula It is easy to know that the sparse channel representation dictionary is obtained by performing the Kronecker product and element-wise multiplication on the angles of each sampling angle. The process of calculating through each sampling angle can be the construction process of the sparse channel representation dictionary.

[0084] Among them, the sampling angles can be those in constructing the sparse channel representation dictionary where there are a total of S sampling angles.

[0085] Step S30: Input the sparse channel representation dictionary and the signal to be processed into the deep unfolded neural network, and obtain the estimated value of the sparse channel of the signal to be processed through the pre-trained learnable parameters, the signal to be processed, and the sparse channel representation dictionary in the deep unfolded neural network.

[0086] It should be noted that the deep unfolded neural network can be a cascaded neural network, and the learnable parameters can be the fixed parameters obtained after training in the deep unfolded neural network. The optimal value of the learnable parameters is determined through training, and this optimal value is used as the learnable parameters for finally applying the deep unfolded neural network.

[0087] It should be further noted that the entire embodiment includes two stages. One stage is to train the angle resampling network through the training signal and the true angle of arrival of the training signal. After the angle resampling network is trained, the output result is input into the deep unfolded neural network, and the deep unfolded neural network is trained to obtain the trained learnable parameters. The other stage is that after the angle resampling network and the deep unfolding network are trained, the received signal is input into the trained angle resampling network and the deep unfolding network to obtain the sparse channel estimation value of the signal.

[0088] In a specific implementation, when the angle resampling network has been trained and the cascaded neural network needs to be trained at this time, the training signal is input into the angle resampling signal to obtain the training sampling angle, a training sparse channel representation dictionary is constructed, and the cascaded neural network is input according to the training sparse channel representation dictionary. The learnable parameters of the cascaded neural network are optimized according to the recognition result of the cascaded neural network.

[0089] It should be further noted that the construction of the deep unfolded neural network: In the formula the sparse channel estimation value w ψ after expansion, the following formula can be obtained:

[0090]

[0091] where max{·} represents taking the maximum value, eig{·} represents finding the eigenvalue, ||·|| 2 represents finding the 2-norm, and Re{·} represents taking the real part of the complex number. For the sake of simplicity of representation, the subscript Ψ is omitted in the above formula, and the lower bound of the distribution P(y|ω, σ) in the previous formula is expressed as:

[0092]

[0093] To promote the sparsity of ω, ω is set to have a two-level prior model. In the first layer, ω satisfies a Gaussian distribution with parameter α, that is where is a non-negative parameter to promote the sparsity of ω. In the second layer, the prior distributions of α and σ satisfy a gamma distribution, that is and P(σ) = Gamma(σ|C, d), where a = b = c = d = 10 -4 .

[0094] While performing sparse channel estimation, all parameter sets to be estimated are defined as Since it is difficult to obtain the exact posterior distribution of the parameters to be estimated in Θ, the posterior approximation distribution of Θ is factorized as q(Θ) = q w (w)q α (α)q σ (σ), where q x (x) represents the posterior distribution of factor x. The lower bound of the likelihood of maximizing q(Θ) is obtained by alternately updating the approximate posterior distributions of the parameters to be estimated, where the optimization expression of q(Θ) is:

[0095]

[0096] where Θ n represents the nth parameter to be estimated in Θ, represents the expectation with respect to q(Θ) / q(Θ n ), and q(Θ) / q(Θ n ) represents the q(Θ) distribution that does not include q(Θ n ). For each parameter to be estimated, other parameters are kept unchanged during the iteration process step by step. After the iteration converges, the estimated value of each parameter is equal to the expectation of its posterior distribution.

[0097] For the update of q w (w) among the parameters to be estimated, the following formula can be referred to:

[0098]

[0099] where ∝ represents the proportional operation; q w (w) follows a Gaussian distribution, that is, q w (w) ∼ CN(w|μ, ∑), and the calculation of its mean and variance can refer to the following formula:

[0100]

[0101]

[0102] For the update of q α (α) among the parameters to be estimated, the following formula can be referred to:

[0103]

[0104] The following formula can be further derived:

[0105]

[0106] where

[0107] q in the parameters to be estimated σ (σ) can be updated with reference to the following formula:

[0108]

[0109] ∝(MK + C - l)lnσ - σ(χ + d)

[0110] where tr{·} represents the trace of a matrix. Further, the following formula can be obtained:

[0111]

[0112] Based on the iterative update of the above parameters to be estimated, each parameter to be estimated is used as the input parameter to be estimated for each cascade layer of the cascaded neural network to form a deep unfolded neural network. For details, reference can be made to Figure 3 , which is the neural network structure of the t-th cascade layer in the figure, where α (t) , σ (t) , μ (t) are the iterative parameters of the t-th layer, where μ (t) represents the sparse channel estimation value, that is, the value to be finally obtained; the α (t) , σ (t) , μ (t) input to the neural network of the t-th cascade layer are used as the input of the t-th cascade layer after calculation of σ (t-1) , α (t+1) and μ (t+1) until the μ' calculated in the last layer is the final required sparse channel estimation value.

[0113] It should be emphasized that the iterative calculation method of each parameter to be estimated in the deep unfolded neural network will change. The iterative calculation formula of each parameter to be estimated in the deep unfolded neural network can be found in the third embodiment in detail.

[0114] In this embodiment, the angle resampling network is used to perform resampling to determine the sampling angle, and a channel angular domain sparse channel representation dictionary is constructed based on the resampled sampling angle to achieve a more effective channel sparse representation. The effective channel sparse representation is input into the unfolded deep learning network, and the more accurate channel estimation value can be quickly converged and calculated.

[0115] Reference Figure 4 , Figure 4 is the flowchart of the second embodiment of a sparse channel estimation method of the present invention.

[0116] Based on the above first embodiment, the sparse channel estimation method in this embodiment in step S10 includes:

[0117] Step S11: Convert the signal to be processed into an amplitude feature signal through the conversion layer.

[0118] It can be understood that the angle resampling network includes a conversion layer, a feature extraction layer, and a resampling prediction layer. The structure of the angle resampling network can refer to Figure 5 , which shows the data input and output during the training of the angle resampling network. The signal to be processed Y and the true arrival angle θ of the signal to be processed are input into the DFT conversion layer, then through the feature extraction layer, and finally through the resampling prediction layer to obtain the resampled angle, so as to obtain the sparse channel representation dictionary. The resampled angle can be expressed as Through The constructed sparse channel representation dictionary is

[0119] It should be understood that the conversion layer can be a discrete Fourier transform (DFT) conversion layer (DFT Transform Layer).

[0120] It is worth noting that converting the signal to be processed into an amplitude feature signal can enhance the direction feature of the true arrival angle, so that the arrival angle direction of the signal to be processed can be obtained quickly.

[0121] In a specific implementation, converting the signal to be processed Y into an amplitude feature signal can refer to the following formula:

[0122]

[0123] Where F s represents the S-dimensional DFT matrix, represents the (s, k) - th value of and for the k - th column of , a relatively large value can be obtained only when

[0124] Step S12: Extract the arrival angle direction feature of the amplitude feature signal through the feature extraction layer.

[0125] It can be understood that the feature extraction layer is used to extract the arrival angle direction feature in the amplitude feature signal. The arrival angle direction feature can be understood as the feature that can represent the arrival angle in the signal to be processed and the carried interference noise.

[0126] It should be understood that at this time, the arrival angle direction feature obtained through feature extraction may also include the arrival angle of interference noise. The feature extraction layer needs to assign weights to each arrival angle direction feature. The weight of each arrival angle direction feature can be calculated by the magnitude of each arrival angle direction feature, and the larger the arrival angle direction feature, the greater its weight.

[0127] It should be noted that the feature extraction layer may include 4 residual neural networks, and each residual neural network may include a one-dimensional convolutional layer, a BatchNorm layer (uniform distribution layer), and a ReLU activation function (linear rectification function, also known as the rectified linear unit, which is a commonly used activation function in artificial neural networks and usually refers to non-linear functions represented by the ramp function and its variants).

[0128] Step S13: Based on the angle-of-arrival direction feature, perform prediction through the resampling prediction layer to obtain a predicted true angle of arrival, and sample according to the true angle of arrival to obtain multiple sampling angles.

[0129] It can be understood that the predicted true angle of arrival may be that through the resampling prediction layer, it is predicted that this angle-of-arrival direction feature is the angle of arrival of the signal to be processed, rather than the angle-of-arrival direction feature of interference noise.

[0130] It should be noted that the sampling prediction layer may obtain the predicted true angle of arrival by comparing the weights of the angle-of-arrival direction features. For example, all the weights are compared, and the angle-of-arrival direction feature with the largest weight is used as the true angle of arrival; it may also be a preset weight threshold, and the angle-of-arrival features greater than the weight threshold are used as the true angle of arrival. At this time, the true angle of arrival may be one or more, and resampling is performed for each true angle of arrival respectively.

[0131] It should be further noted that a preset sampling grid point and a preset sampling fraction are obtained; a sampling range is determined according to the predicted true angle of arrival and the sampling grid point, and a sampling interval is obtained according to the sampling range and the preset sampling fraction; resampling is performed within the sampling range according to the sampling interval to obtain multiple sampling angles.

[0132] It should be understood that there will also be a certain angular difference between the predicted true angle of arrival found through the weight and the true angle of arrival, but at this time, the angular difference between the predicted true angle of arrival and the true angle of arrival is relatively small.

[0133] It should be understood that the preset sampling grid point may be a manually set angular difference between the predicted true angle of arrival and the true angle of arrival. For example, if the predicted true angle of arrival is 55°, and the manually set angular difference is 5 degrees, then the preset sampling grid point is 5°, and the sampling range is 50° - 60°.

[0134] It should be further understood that the preset sampling fraction may be the number of angles to be sampled. For example, if the sampling range is 50° - 60° and the preset sampling fraction is 20, then the sampling interval is 0.5°, and the sampling angles are 50.5°, 51°, 51.5°... 60°.

[0135] It should be noted that the angle resampling neural network predicts the true angle of arrival according to the weights of the angle of arrival direction features, determines a sampling range around the predicted true angle of arrival direction, and performs resampling within this range, which can include the true angle of arrival with a higher probability, thereby obtaining a more effective sparse channel representation dictionary.

[0136] It should be further noted that before inputting the signal to be processed into the angle resampling network to obtain multiple sampling angles, an angle resampling data set is acquired. The angle resampling data set includes a plurality of training signals and true angles of arrival that correspond one by one; the training signals are input into the initial angle resampling network to obtain initial resampled angles; a Gaussian mixture loss function is constructed according to the Gaussian distribution of the true angle of arrival, the probability that each sampling point is equal to the true angle of arrival, the angle of each sampling point, and a preset standard deviation; the initial angle resampling network is optimized according to the Gaussian mixture loss function to obtain the angle resampling network.

[0137] Among them, the calculation formula of the Gaussian mixture loss function is as follows:

[0138]

[0139] Among them, p s represents the probability that the s-th sampling point of each training signal in the angle resampling data set is close to the true angle of arrival, where represents the Gaussian distribution of the l-th true angle of arrival, θ l represents the l-th true angle of arrival, where ψ s represents the angle of the s-th sampling point, and ζ represents the preset standard deviation, which can be used to balance robustness and accuracy.

[0140] It should be noted that the angle resampling network is trained with multiple training signals and the true angles of arrival of the signals, a loss function is constructed according to the training results, and the initial angle resampling network is optimized through the loss function, so that the angle resampling network can more accurately estimate the true angle of arrival, further obtain sampling angles closer to the true angle of arrival, and thus obtain a more effective sparse channel representation dictionary.

[0141] In this embodiment, the predicted true angle of arrival is obtained through the angle resampling network, and multiple sampling angles are obtained by sampling around the true angle of arrival, so that the sampling range can preferably include the true angle or be closer to the true angle, thereby constructing a more accurate sparse channel representation dictionary.

[0142] Reference Figure 6 , Figure 6 is the schematic flowchart of the third embodiment of a sparse channel estimation method of the present invention.

[0143] Based on the above first embodiment, the sparse channel estimation method in this embodiment in step S30 includes:

[0144] Step S31: Input the sparse channel representation dictionary and the signal to be processed into a deep unfolded neural network, where the deep unfolded neural network includes multiple cascaded layers.

[0145] It can be understood that the cascaded layer structure of the deep unfolded neural network can refer to Figure 7 , where each Layer in the figure represents a cascaded layer in the deep unfolded neural network. The input of each cascaded layer includes the signal to be processed, the sparse channel representation dictionary, and the learnable parameters of the current layer the parameter to be estimated of the current layer (the Gaussian prior distribution parameter α (t) , the Gaussian white noise parameter σ (t) , and the sparse channel estimation value μ (t) ). Iteratively update the parameter to be estimated of the current layer, and use the updated parameter to be estimated as the input of the next layer.

[0146] It should be noted that the calculation of the parameter to be estimated by each cascaded layer can be an iterative update of the parameter to be estimated. The calculation of each cascaded layer can be regarded as an iteration of the parameter to be estimated. The total number of cascaded layers in the deep unfolded neural network is preset. The total number of layers can be 18 layers, 20 layers, or 25 layers, etc., which can be adjusted according to the actual situation.

[0147] It should be noted that before actually using the deep unfolded neural network, it is necessary to train the deep unfolded neural network to obtain the optimal learnable parameters of the deep unfolded neural network. The training process can be to input multiple training signals and the sparse channel representation dictionary into the initial deep unfolded neural network to obtain multiple sparse channel training values. The sparse channel training value can be understood as the sparse channel estimation value of the training signal during the training process;

[0148] Calculate the estimated channel matrix of the training signal according to the sparse channel training value. The specific calculation can refer to the following formula:

[0149]

[0150]

[0151] where H ψ represents the estimated channel matrix, X ψ is known, and w ψ represents the sparse channel training value;

[0152] Construct a feature loss function based on the number of the training signals and the estimated channel matrix. The loss function can adopt the normalized mean squared error (NMSE) as the error function, and the loss function can be expressed by the following formula;

[0153]

[0154] where M t represents the number of the training signals, represents the m t -th estimated channel matrix;

[0155] Optimize the learnable parameters of the initial deep unfolded neural network according to the feature loss function to obtain the deep unfolded neural network. Refer to Figure 7 , in the figure represents the learnable parameters, and

[0156] It should be noted that the initial deep unfolded neural network is trained by the already trained sparse channel representation dictionary and the training signals, so as to optimize the learning parameters of each cascade layer in the initial deep unfolded neural network to obtain the optimized learning parameters, and determine the learning parameter values in the deep unfolded neural network according to the optimized learning parameters, so that the deep unfolded neural network can perform sparse channel estimation more quickly and accurately in the sparse channel estimation process.

[0157] Step S32: Calculate the sparse channel estimation value of the next cascade layer according to the sparse channel representation dictionary, the signal to be processed, and the sparse channel estimation value, learnable parameters and parameters to be estimated of the current cascade layer.

[0158] It should be noted that the parameters to be estimated in each cascade layer of the deep unfolded neural network include Gaussian white noise parameters, sparse channel estimation values, and Gaussian prior distribution parameters.

[0159] It should be further noted that when calculating the sparse channel estimation value of the next cascade layer according to the sparse channel representation dictionary, the signal to be processed, the sparse channel estimation value, learnable parameters and parameters to be estimated of the current cascade layer, in the specific implementation, μ (t) is the sparse channel estimation value of the current cascade layer, is the learnable parameter of the current cascade layer, σ (t) is the Gaussian white noise parameter of the current cascade layer, ∑ (t) is the Gaussian posterior distribution parameter of the current cascade layer, α (t+1 () is the Gaussian prior distribution parameter of the next cascade layer, Λ (t) is the diagonal matrix of the Gaussian prior distribution parameter of the previous cascade layer.

[0160] Then, calculate the sparse channel estimation value of the current cascade layer based on the sparse channel representation dictionary, the signal to be processed, the learnable parameters of the previous cascade layer, the Gaussian white noise parameters, the Gaussian prior distribution parameters, and the sparse channel estimation value. The calculation of the sparse channel estimation value can refer to the following formula:

[0161]

[0162] Calculate the Gaussian prior distribution parameters of the current cascade layer based on the Gaussian posterior distribution parameters and the sparse channel estimation value of the current cascade layer. The calculation of the Gaussian prior distribution parameters can refer to the following formula:

[0163]

[0164] Calculate the Gaussian white noise parameters of the current cascade layer based on the sparse channel representation dictionary, the signal to be processed, the sparse channel estimation value of the current cascade layer, the learning parameters of the previous cascade layer, the sparse channel estimation value, and the Gaussian posterior distribution parameters. The calculation of the Gaussian white noise parameters can refer to the following formula:

[0165]

[0166] Wherein,

[0167]

[0168] Calculate the sparse channel estimation value of the next cascade layer based on the sparse channel representation dictionary, the signal to be processed, the learnable parameters of the current cascade layer, the Gaussian white noise parameters, the Gaussian prior distribution parameters, and the sparse channel estimation value (the calculation formula of the sparse channel estimation value of the next cascade layer can be obtained according to the calculation formula of the sparse channel estimation value of the current cascade layer).

[0169] It can be understood that

[0170] It can be understood that when the current layer is the first layer, the parameter to be estimated can be the corresponding value set artificially or the initial parameter to be estimated (the initial parameters to be estimated are all 0).

[0171] It should be noted that by further describing the specific method for calculating the sparse channel estimation value of each cascade layer in the deep unfolding neural network, the sparse channel estimation value calculated by each cascade layer is used for the calculation of the sparse channel estimation value of the next cascade layer, and the calculation is iterated layer by layer, so as to obtain the sparse channel value more quickly and effectively.

[0172] Step S33: When the next cascade layer is the preset cascade layer, use the sparse channel estimation value of the next cascade layer as the sparse channel estimation value of the signal to be processed.

[0173] It is understandable that the preset cascade layer can be the last cascade layer of the artificially set deep unfolding neural network.

[0174] It should be understood that when the next cascade layer is the preset cascade layer, the sparse channel estimation value of the next layer can be calculated, and other related parameters to be estimated are no longer updated correspondingly.

[0175] In this embodiment, by further describing the iterative calculation of the parameters to be estimated in each cascade layer of the deep unfolding neural network, the sparse channel estimation value calculated by each cascade layer is used to calculate the sparse channel estimation value of the next cascade layer, and the iterative calculation is performed layer by layer. After the update of the last cascade layer, the calculation steps of the sparse channel estimation value are obtained, and the sparse channel value can be obtained more quickly and effectively.

[0176] In addition, an embodiment of the present invention also proposes a storage medium, on which a sparse channel estimation program is stored. When the sparse channel estimation program is executed by a processor, the steps of the sparse channel estimation method described above are implemented.

[0177] Refer to Figure 8 , Figure 8 which is the structural block diagram of the first embodiment of the sparse channel estimation device of the present invention.

[0178] As Figure 8 shown, the sparse channel estimation device proposed in an embodiment of the present invention includes:

[0179] A resampling module 10, configured to input a signal to be processed into an angle resampling network to obtain a plurality of sampling angles;

[0180] A sparse channel representation dictionary construction module 20, configured to construct a sparse channel representation dictionary of the signal to be processed according to the sampling angles;

[0181] A sparse channel estimation module 30, configured to input the sparse channel representation dictionary and the signal to be processed into a deep unfolding neural network, and obtain a sparse channel estimation value of the signal to be processed through the pre-trained learnable parameters, the signal to be processed, and the sparse channel representation dictionary in the deep unfolding neural network.

[0182] In this embodiment, by adaptively determining the sampling interval related to the true arrival angle, and then constructing a channel angle domain sparse channel representation dictionary based on this, more effective channel sparse representation is realized. Based on the effective channel sparse representation, a more accurate channel estimation value is obtained through the fast convergence calculation of the unfolded deep learning network.

[0183] In one embodiment, the resampling module 10 is further configured to convert the signal to be processed into an amplitude feature signal through the conversion layer;

[0184] Extract the direction-of-arrival (DOA) feature of the amplitude feature signal through the feature extraction layer;

[0185] Based on the DOA feature, perform prediction through the resampling prediction layer to obtain the predicted true DOA, and sample according to the true DOA to obtain multiple sampling angles.

[0186] In one embodiment, the resampling module 10 is further configured to compare the weights of the DOA features to obtain the predicted true DOA, and obtain a preset sampling grid and a preset sampling fraction;

[0187] Determine the sampling range according to the predicted true DOA and the sampling grid, and obtain the sampling interval according to the sampling range and the preset sampling fraction;

[0188] Perform resampling within the sampling range according to the sampling interval to obtain multiple sampling angles.

[0189] In one embodiment, the resampling module 10 is further configured to obtain an angular resampling data set, where the angular resampling data set includes multiple corresponding training signals and true DOAs;

[0190] Input the training signal into the initial angular resampling network to obtain the initial resampling angle;

[0191] Construct a Gaussian mixture loss function according to the Gaussian distribution of the true DOA, the probability that each sampling point is equal to the true DOA, the angle of each sampling point, and a preset standard deviation;

[0192] Optimize the initial angular resampling network according to the Gaussian mixture loss function to obtain the angular resampling network.

[0193] In one embodiment, the sparse channel estimation module 30 is further configured to input the sparse channel representation dictionary and the signal to be processed into a deep unfolding neural network, and the deep unfolding neural network includes multiple cascaded layers;

[0194] Calculate the sparse channel estimation value of the next cascaded layer according to the sparse channel representation dictionary, the signal to be processed, the sparse channel estimation value, learnable parameters, and parameters to be estimated of the current cascaded layer;

[0195] When the next cascaded layer is a preset cascaded layer, use the sparse channel estimation value of the next cascaded layer as the sparse channel estimation value of the signal to be processed.

[0196] In one embodiment, the sparse channel estimation module 30 is further configured to calculate the sparse channel estimation value of the current cascaded layer according to the sparse channel representation dictionary, the signal to be processed, the learnable parameters of the previous cascaded layer, the Gaussian white noise parameters, the Gaussian prior distribution parameters, and the sparse channel estimation value;

[0197] Calculate the Gaussian prior distribution parameters of the current cascaded layer according to the Gaussian posterior distribution parameters and the sparse channel estimation value of the current cascaded layer;

[0198] Calculate the Gaussian white noise parameters of the current cascaded layer according to the sparse channel representation dictionary, the signal to be processed, the sparse channel estimation value of the current cascaded layer, the learning parameters of the previous cascaded layer, the sparse channel estimation value, and the Gaussian posterior distribution parameters;

[0199] Calculate the sparse channel estimation value of the next cascaded layer according to the sparse channel representation dictionary, the signal to be processed, and the learnable parameters, Gaussian white noise parameters, Gaussian prior distribution parameters, and sparse channel estimation value of the current cascaded layer.

[0200] In one embodiment, the sparse channel estimation module 30 is further configured to input a plurality of training signals and the sparse channel representation dictionary into an initial deep unfolded neural network to obtain a plurality of sparse channel training values;

[0201] Calculate the estimated channel matrix of the training signal according to the sparse channel training value;

[0202] Construct a feature loss function according to the number of the training signals and the estimated channel matrix;

[0203] Optimize the learnable parameters of the initial deep unfolded neural network according to the feature loss function to obtain a deep unfolded neural network.

[0204] It should be understood that the above is only an example for illustration and does not constitute any limitation to the technical solution of the present invention. In specific applications, those skilled in the art can set according to needs, and the present invention does not limit this.

[0205] It should be noted that the above-described work process is only illustrative and does not limit the protection scope of the present invention. In practical applications, those skilled in the art can select some or all of them according to actual needs to achieve the purpose of the solution of this embodiment, and no limitation is made here.

[0206] In addition, it should be noted that in this text, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or system. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or system including such element.

[0207] The serial numbers of the above embodiments of the present invention are only for description and do not represent the superiority or inferiority of the embodiments.

[0208] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as a read-only memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions for causing a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0209] The above are only the preferred embodiments of the present invention and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A sparse channel estimation method, characterized in that, the sparse channel estimation method includes: Inputting the signal to be processed into an angular resampling network to obtain multiple sampling angles; Constructing a sparse channel representation dictionary of the signal to be processed according to the sampling angles; Inputting multiple training signals and the sparse channel representation dictionary into an initial deep unfolding neural network to obtain multiple sparse channel training values; Calculating the estimated channel matrix of the training signal according to the sparse channel training values; Constructing a feature loss function according to the number of the training signals and the estimated channel matrix; Optimizing the learnable parameters of the initial deep unfolding neural network according to the feature loss function to obtain a deep unfolding neural network; Inputting the sparse channel representation dictionary and the signal to be processed into the deep unfolding neural network, and the deep unfolding neural network includes multiple cascaded layers; Calculating the sparse channel estimation value of the next cascaded layer according to the sparse channel representation dictionary, the signal to be processed, and the sparse channel estimation value, learnable parameters, and parameters to be estimated of the current cascaded layer, where the parameters to be estimated include Gaussian white noise parameters, sparse channel estimation values, and Gaussian prior distribution parameters; When the next cascaded layer is a preset cascaded layer, using the sparse channel estimation value of the next cascaded layer as the sparse channel estimation value of the signal to be processed; The calculating the sparse channel estimation value of the next cascaded layer according to the sparse channel representation dictionary, the signal to be processed, and the sparse channel estimation value, learnable parameters, and parameters to be estimated of the current cascaded layer includes: Calculating the sparse channel estimation value of the current cascaded layer according to the sparse channel representation dictionary, the signal to be processed, and the learnable parameters, Gaussian white noise parameters, Gaussian prior distribution parameters, and sparse channel estimation value of the previous cascaded layer; Calculating the Gaussian prior distribution parameter of the current cascaded layer according to the Gaussian posterior distribution parameter and the sparse channel estimation value of the current cascaded layer; Calculating the Gaussian white noise parameter of the current cascaded layer according to the sparse channel representation dictionary, the signal to be processed, the sparse channel estimation value of the current cascaded layer, the learning parameters of the previous cascaded layer, the sparse channel estimation value, and the Gaussian posterior distribution parameter; Calculating the sparse channel estimation value of the next cascaded layer according to the sparse channel representation dictionary, the signal to be processed, and the learnable parameters, Gaussian white noise parameters, Gaussian prior distribution parameters, and sparse channel estimation value of the current cascaded layer.

2. The sparse channel estimation method according to claim 1, characterized in that, the angular resampling network includes a conversion layer, a feature extraction layer, and a resampling prediction layer; The inputting the signal to be processed into an angular resampling network to obtain multiple sampling angles includes: Converting the signal to be processed into an amplitude feature signal through the conversion layer; Extracting the arrival angle direction feature of the amplitude feature signal through the feature extraction layer; Based on the arrival angle direction feature, predicting through the resampling prediction layer to obtain a predicted true arrival angle, and sampling according to the true arrival angle to obtain multiple sampling angles.

3. The sparse channel estimation method according to claim 2, characterized in that, Performing prediction through the resampling prediction layer based on the angle-of-arrival direction feature to obtain a predicted true angle of arrival, and sampling according to the true angle of arrival to obtain a plurality of sampling angles, including: Comparing the weights of the angle-of-arrival direction features to obtain a predicted true angle of arrival, and obtaining a preset sampling grid and a preset sampling fraction; Determining a sampling range according to the predicted true angle of arrival and the sampling grid, and obtaining a sampling interval according to the sampling range and the preset sampling fraction; Performing resampling within the sampling range according to the sampling interval to obtain a plurality of sampling angles.

4. The sparse channel estimation method according to claim 1, wherein, Before inputting the signal to be processed into the angle resampling network to obtain a plurality of sampling angles, it further includes: Obtaining an angle resampling data set, where the angle resampling data set includes a plurality of training signals and true angles of arrival that correspond one by one; Inputting the training signal into the initial angle resampling network to obtain an initial resampling angle; Constructing a Gaussian mixture loss function according to the Gaussian distribution of the true angle of arrival, the probability that each sampling point is equal to the true angle of arrival, the angle of each sampling point, and a preset standard deviation; Optimizing the initial angle resampling network according to the Gaussian mixture loss function to obtain an angle resampling network.

5. A sparse channel estimation device, wherein, The sparse channel estimation device includes: A resampling module, configured to input a signal to be processed into an angle resampling network to obtain a plurality of sampling angles; A sparse channel representation dictionary construction module, configured to construct a sparse channel representation dictionary of the signal to be processed according to the sampling angles; A sparse channel estimation module, configured to input a plurality of training signals and the sparse channel representation dictionary into an initial deep unfolding neural network to obtain a plurality of sparse channel training values; Calculating an estimated channel matrix of the training signal according to the sparse channel training values; Constructing a feature loss function according to the number of the training signals and the estimated channel matrix; Optimizing the learnable parameters of the initial deep unfolding neural network according to the feature loss function to obtain a deep unfolding neural network; Inputting the sparse channel representation dictionary and the signal to be processed into the deep unfolding neural network, where the deep unfolding neural network includes a plurality of cascaded layers; Calculating the sparse channel estimation value of the next cascaded layer according to the sparse channel representation dictionary, the signal to be processed, the sparse channel estimation value, the learnable parameters, and the parameters to be estimated of the current cascaded layer, where the parameters to be estimated include Gaussian white noise parameters, sparse channel estimation values, and Gaussian prior distribution parameters; When the next cascaded layer is a preset cascaded layer, using the sparse channel estimation value of the next cascaded layer as the sparse channel estimation value of the signal to be processed; The calculating the sparse channel estimation value of the next cascaded layer according to the sparse channel representation dictionary, the signal to be processed, the sparse channel estimation value, the learnable parameters, and the parameters to be estimated of the current cascaded layer includes: Calculate the sparse channel estimation value of the current cascaded layer according to the sparse channel representation dictionary, the signal to be processed, the learnable parameters of the previous cascaded layer, the Gaussian white noise parameters, the Gaussian prior distribution parameters, and the sparse channel estimation value; Calculate the Gaussian prior distribution parameters of the current cascaded layer according to the Gaussian posterior distribution parameters and the sparse channel estimation value of the current cascaded layer; Calculate the Gaussian white noise parameters of the current cascaded layer according to the sparse channel representation dictionary, the signal to be processed, the sparse channel estimation value of the current cascaded layer, the learning parameters of the previous cascaded layer, the sparse channel estimation value, and the Gaussian posterior distribution parameters; Calculate the sparse channel estimation value of the next cascaded layer according to the sparse channel representation dictionary, the signal to be processed, and the learnable parameters, Gaussian white noise parameters, Gaussian prior distribution parameters, and sparse channel estimation value of the current cascaded layer.

6. A sparse channel estimation device, characterized in that, the device includes: a memory, a processor, and a sparse channel estimation program stored on the memory and executable on the processor, the sparse channel estimation program being configured to implement the sparse channel estimation method according to any one of claims 1 to 4.

7. A storage medium, characterized in that, a sparse channel estimation program is stored on the storage medium, and when the sparse channel estimation program is executed by a processor, it implements the sparse channel estimation method according to any one of claims 1 to 4.

Images