A Deep Learning-Based Channel Estimation Method in Asymmetric Architectures

Abstract

This invention discloses a deep learning-based channel estimation method for asymmetric all-digital communication systems, comprising: designing an uplink receiving array based on a nested sparse array topology to recover the received signal of the sparse array into a virtual array signal of full dimension; decomposing the obtained virtual signal into a grid part and an off-grid part based on Taylor formula, and designing a two-step-angle estimation net (Ts-AEnet) to estimate the signal angle of arrival; solving a least-squares problem to estimate the signal path gain and reconstructing the downlink channel. This invention proposes an uplink receiving array based on a nested sparse array topology in asymmetric all-digital communication systems, minimizing information loss caused by missing antennas, and proposes a deep learning-based Ts-AEnet angle estimation network, which reduces the complexity of channel estimation while ensuring the system's estimation accuracy.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a deep learning-based channel estimation method in an asymmetric architecture. Background Technology

[0002] In the currently widely deployed fifth-generation mobile communication systems and the upcoming sixth-generation mobile communication systems, massive MIMO (Multi-Input Multi-Output) technology has become one of the key technologies due to its excellent performance in terms of capacity, spectral efficiency, and reliability. To meet the ever-increasing demands for data rate and reliability, there is a desire to further expand the scale of MIMO arrays. However, the resulting hardware costs, data processing burden, and power consumption requirements limit the further development of massive MIMO technology. To reduce the deployment cost and hardware complexity of large-scale all-digital architectures, and also to reduce signal loss due to low resolution in hybrid beamforming architectures, the paper "Design and Implementation of a Full-digital Beamforming Array with Nonreciprocal Tx / Rx Beam Patterns" (IEEE ANTENNAS WIRELESS PROPAGATION LETTERS, Vol. 19, No. 11, November 2020) proposes an asymmetric structure for decoupling the transceiver radio frequency chain. This special architecture not only reduces the data processing burden of the communication system but also lowers the overall system power consumption. Meanwhile, due to the inconsistency in the number of uplink and downlink radio links, channel reciprocity in time-division duplex mode cannot be directly applied, which poses a challenge to downlink channel estimation in asymmetric systems. Currently, researchers have proposed downlink channel estimation methods for asymmetric systems, but these methods are mostly based on iterative algorithms to estimate the signal's angle of arrival, leading to significant computational complexity. For example, the Newton-orthogonal matched pursuit algorithm, which achieves good estimation accuracy, requires designing a complete dictionary matrix within the signal angle range, and then reconstructing the downlink channel by iteratively calculating and traversing the dictionary matrix. Although this method offers good estimation performance, its high complexity makes it impractical for real-world communication applications. Therefore, to reduce the computational complexity in channel estimation, many researchers have begun to explore using deep learning from artificial intelligence for angle estimation. The paper "Deep Networks for Direction-of-Arrival Estimation in Low SNR" (IEEE TRANSACTIONS ON SIGNAL PROCESSING, Vol. 69, 2021) treats angle estimation as a multi-classification task and designs an angle estimation method based on Convolutional Neural Network (CNN). This estimation method can effectively reduce the estimation complexity by training the neural network online and deploying it offline.Therefore, applying deep learning to channel estimation in asymmetric systems has great research potential. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, this invention provides a deep learning-based channel estimation method in an asymmetric architecture, which overcomes the shortcomings of existing technologies and leverages the advantages of artificial intelligence. This method can achieve comparable estimation performance while avoiding the excessive computational load of traditional iterative algorithms, and effectively reduces the hardware and energy consumption costs of base stations.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a deep learning-based channel estimation method in an asymmetric architecture, comprising:

[0006] An uplink receiving array based on a nested sparse array topology is designed. The signal received by the sparse array is restored to a full-dimensional virtual array signal using a virtual array interpolation method. The obtained virtual signal is decomposed into a grid part and an off-grid part using Taylor formula. A two-step angle estimation network Ts-AEnet is designed to estimate the signal angle of arrival. The received signal dataset is input into a preprocessing layer for processing. The neural network is trained offline on the training set using supervised learning. Then, the trained Ts-AEnet is deployed online to estimate the signal angle, and the least squares problem is solved to estimate the signal path gain and reconstruct the downlink channel.

[0007] As a preferred embodiment of the deep learning-based channel estimation method in the asymmetric architecture described in this invention, the uplink receiving array based on a nested sparse array topology includes a uniform linear array with M antennas and a single-antenna user; in the uniform linear array, each antenna is connected to a transmit radio frequency chain for downlink signal transmission, but only N (N < M) antennas are connected to a receive radio frequency chain for uplink signal reception, denoted by the index S = {p1, p2, ..., p...}. N}, where p i ∈{0,1,…,M-1}, and when i=j, p i ≠p j p i and p j Let each represent any element in S; consider a multipath scenario where a base station antenna array receives signals from L independent, identically distributed paths; model the uplink channel as a millimeter-wave channel, expressed as:

[0008]

[0009] Where, α l Let θ represent the gain of the l-th path.l Let j represent the arrival angle of the l-th path, and a UL (θ l ) represents the uplink steering vector of the l-th path at the base station, with dimensions N×1, and is defined as:

[0010]

[0011] Where d represents the distance between two adjacent antenna elements, and λ represents the wavelength of the electromagnetic wave; if the user sends an all-1 pilot signal, the signal received at the base station is:

[0012]

[0013] Where ω is obeyed Additive white Gaussian noise, I represents noise power. N It is an identity matrix of dimension N. This represents a Gaussian distribution with mean μ and covariance Σ.

[0014] As a preferred embodiment of the deep learning-based channel estimation method in the asymmetric architecture described in this invention, the uplink receiving array based on the nested sparse array topology further includes a nested array structure as the uplink receiving array of the asymmetric system; the nested array consists of two subarrays with different antenna element spacings, wherein subarray 1 consists of n antenna elements and subarray 2 consists of m antenna elements; subarray 1 is a uniform linear array with an antenna spacing of d; a subarray 2 is nested between two adjacent antenna elements of subarray 2, so the antenna spacing is (n+1)d;

[0015] Set the antenna index for the nested array structure. Where subarray 1 has indices {0, 1, ..., n-1} and subarray 2 has indices {n, 2n+1, ..., m(n+1)-1}, the current index set is... The corresponding antenna is connected to the uplink receiving RF chain; the total number of base station antennas is M = m(n+1), and the number of uplink receiving array antennas is N = m+n.

[0016] As a preferred embodiment of the deep learning-based channel estimation method in the asymmetric architecture described in this invention, the virtual array interpolation method includes calculating the covariance matrix of the received signal as follows:

[0017]

[0018] Wherein, the superscript H indicates the conjugate transpose; in practice, R y Obtained through sampling covariance, and Where K is the number of samples;

[0019] Extracting elements from the covariance matrix to recover the complete signal; due to the existence of The calculation process, thus the covariance matrix R y Located at the pth i line p j The elements of the column contain the phase difference portion in Define the nonnegative element p in it. i -p j The set is:

[0020]

[0021] Next, extract the set. The covariance element values ​​corresponding to the elements in the middle are used to fill the signal gaps in the sparse array received signal caused by missing antenna elements; when the antenna with index "3" in the uplink receiving array is not connected to the receiving RF chain, it cannot receive a signal. This is achieved by finding the set... The R corresponding to element "3" y The covariance element values ​​fill in the missing received signal at the current position, constructing a virtual uniform linear array of dimension M, where the reconstructed virtual array signal is represented as:

[0022]

[0023] in, Let A be the guiding matrix of the reconstruction array of dimension M×L. DL (θ)=[a DL (θ1),a DL (θ2),…,a DL (θ L [)] represents the downlink array steering matrix for all L path signals. This represents the downlink array steering vector of the l-th path signal; This represents a power vector with dimension L×1; The noise term represents the reconstruction, and e0 represents the matrix I when reconstructing the virtual signal according to the above method. N The deformation.

[0024] As a preferred embodiment of the deep learning-based channel estimation method in the asymmetric architecture described in this invention, the decomposition of the obtained virtual signal includes constructing a Q-dimensional overcomplete angle set {θ1, θ2, ..., θ...} Q}, set {θ1,θ2,…,θ Q Divide an angular range into Q equal grid values, and use a first-order Taylor formula to array the guiding matrix. The l-th column is decomposed into the grid portion and the off-grid portion, as follows:

[0025] a(θ l )≈a(θ ql )+a'(θ ql )Δ ql

[0026] Where, θ ql ∈{θ1,θ2,…,θ Q} represents distance θ l The most recent grid angle value, a(θ) ql ) represents the array steering vector at the grid angle, a'(θ) ql ) represents a(θ ql ) at θ ql The derivative at Δ ql =θ l -θ ql Indicates the angle value from the grid;

[0027] Based on the observed virtual signals Rewritten as:

[0028]

[0029] Where, Φ=[a DL (θ1),a DL (θ2),…,a DL (θ Q ] represents the down-row array steering matrix of dimension M×Q in the grid value, a DL (θ ql ) represents the downlink array steering vector at the signal angle at the grid angle; F = [a' DL (θ1),a' DL (θ2),…,a' DL (θ Q )] represents the derivative matrix corresponding to Φ, a' DL (θ ql ) represents a DL (θ ql ) at θ ql The derivative at point; z = [z(θ1), z(θ2), ..., z(θ)] Q )] T Let z(θ) represent the Q-dimensional spectrum in the grid angular space. ql ) indicates that the grid angle is θ ql The signal power value at that time, since there are only L paths of signal, z is only at L positions {θ}. q1 ,θ q2 ,…,θ qL There exists a corresponding value at} Right now Other values ​​are 0; Θ represents a diagonal matrix of dimension Q, with diagonal elements containing L angle values ​​from the grid.

[0030] Given a grid angle distribution on a finite number of angle grids, a grid angle estimation network is designed to estimate the grid angle values ​​contained in z. The network output value does not represent the true power value, but only contains angular position information presented in binary vector form. Therefore, the angle is estimated by finding L maximum values ​​from the output angular spatial spectrum.

[0031] As a preferred embodiment of the deep learning-based channel estimation method in the asymmetric architecture described in this invention, the two-step angle estimation network includes two sub-networks: a grid angle estimation neural network and an off-grid angle estimation neural network, which respectively estimate the grid angle and the off-grid angle.

[0032] The design of the two-step angle estimation network Ts-AEnet for estimating the signal angle of arrival involves designing a convolutional neural network. The network comprises a preprocessing layer, an input layer, four convolutional layers, a fully connected layer, and an output layer. The network expression is as follows:

[0033]

[0034] Preprocessing layer The process of generating virtual signals involves the covariance of the received signals and filling in the missing holes in the physical array to obtain virtual signals.

[0035] The input layer receives the preprocessed signal data. Since neural networks can only process real numbers, they convert the virtual signal... Combination of real and imaginary parts As input data, among which, Indicate extraction The real part, Indicate extraction The imaginary part is then converted into a tensor of dimension M×1×2 and passed to the network input layer.

[0036] Convolutional layer function Where Conv(·) represents the convolution operation, x represents the input data, and {Ξ k ,ξ k} represents the weights and biases of the k-th convolutional layer, with a kernel size of c. l ×c l c l This represents the size of the l-th convolutional kernel; a batch normalization layer is connected after the convolutional layer to speed up training; ReLU is set as the activation function.

[0037] Fully connected layer Where Fc(·) represents a fully connected operation, and x represents the input data. This represents the weights and biases of the k-th fully connected layer; Flatten(·) represents the flattening operation, setting ReLU as the activation function;

[0038] The output layer is set as a dense layer with Q neurons, and outputs a Q-dimensional spatial spectral vector at the grid angle;

[0039] Since the angle from the grid is a continuous value, we treat the grid angle estimation as a regression problem and design a deep neural network to estimate the non-zero diagonal elements of Θ, which correspond to the angle values ​​from the grid.

[0040] The distance-from-grid angle estimation network consists of one input layer, seven fully connected layers, and one output layer, denoted as follows:

[0041]

[0042] in, Similar to the fully connected layer described above; the input layer receives the preprocessed signal data and converts the virtual signal... Cascade of real and imaginary parts As input data, it is passed to the network input layer; concatenate[·] indicates a cascade operation; the output layer consists of L neurons, representing L angle values ​​from the grid.

[0043] The designed on-grid angle estimation network and off-grid angle estimation network are combined to form the complete Ts-AEnet.

[0044] As a preferred embodiment of the deep learning-based channel estimation method in the asymmetric architecture described in this invention, the offline training of two neural networks using supervised learning includes: generating a training dataset for the grid angle estimation network with a sample size of D1; and constructing a Q-dimensional overcomplete angle set. in, It is the interval of the angle set, θ max It is the maximum angle; from Select each set of angles and calculate the corresponding signal y based on the random signal-to-noise ratio level to obtain the virtual array signal according to the preprocessed signal; define the input data. For each Y i The corresponding L angle values ​​are converted into a Q-dimensional binary vector b, where only L elements are set to 1, and all other elements are 0, which serve as the labels for the input data; the training dataset is defined.

[0045] In the training dataset The network for estimating grid angles is trained offline using supervised learning; network parameters are updated via backpropagation to minimize the loss function; and a binary cross-entropy function is used as the loss function.

[0046]

[0047] Generate a training dataset for the off-grid angle estimation network with D2 samples; from The continuous angles are selected in a concentrated manner, and the corresponding signal y is calculated based on the random signal-to-noise ratio level. The virtual array signal is obtained according to the preprocessed signal, and the input data is defined as follows: The label is represented as c = {Δ ql}, l=1,2,…,L, representing the difference between the true value and the nearest grid value; then the training dataset for the current stage is denoted as

[0048] In the training dataset The network is trained offline on a grid angle estimation network. The goal of training the network is to minimize the mean squared error between the predicted and ground truth values, and the loss function is expressed as:

[0049]

[0050] The two trained sub-networks are merged into a complete Ts-AEnet, which is then deployed online. The received signals are processed and input into the network to obtain the estimated signal angle.

[0051] As a preferred embodiment of the deep learning-based channel estimation method in the asymmetric architecture described in this invention, the step of solving the least squares problem to estimate the signal path gain includes:

[0052] After obtaining the estimated signal angle, solve the least squares problem. The estimated signal path gain is calculated as follows:

[0053]

[0054] in, These represent the estimated path and angle of arrival, respectively. This represents the updirection matrix, where ρ is a positive number used to ensure the matrix is ​​full rank and prevent it from being impossible to invert.

[0055] Finally, the downlink channel is reconstructed, specifically as follows:

[0056] A computer device includes a memory and a processor, the memory storing a computer program, characterized in that the processor executes the computer program to implement the steps of a deep learning-based channel estimation method in an asymmetric architecture.

[0057] A computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of a deep learning-based channel estimation method in an asymmetric architecture.

[0058] The beneficial effects of this invention are as follows: First, in the uplink reception process of an asymmetric communication system, this invention adopts a nested array topology as the uplink receiving array. Compared with the random antenna selection scheme, the sparse antenna structure proposed in this invention can minimize the information loss caused by missing antennas, and improve the accuracy of channel estimation while reducing system hardware complexity and energy consumption.

[0059] Secondly, this invention proposes a signal angle estimation scheme based on deep learning, which uses a neural network to replace the angle estimation part in the traditional channel estimation method. When the scheme is deployed online to estimate the signal angle, it only needs to calculate the multiplication and addition operations of the weights and biases with the data in the forward propagation, as well as some nonlinear operations, avoiding matrix inversion operations. The complexity is lower than that of existing iterative estimation algorithms and dictionary matrix-based estimation algorithms, and the final estimation performance is comparable to them.

[0060] Finally, the angle estimation network Ts-AEnet in this invention considers dividing the signal angle estimation into two parts, and designs two corresponding neural networks to estimate the two parts of the angle value based on the characteristics of the angle values ​​in the grid angle and the angle away from the grid, i.e. the characteristics of the classification task and the regression task, respectively, so as to estimate the complete angle. Compared with the general deep learning estimation scheme that only considers the angle value in the grid, the Ts-AEnet scheme in this invention reduces the limitation of the angle grid size on the angle estimation accuracy as much as possible. Attached Figure Description

[0061] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a schematic diagram of an uplink receiving antenna structure based on a nested array, designed using a deep learning-based channel estimation method in an asymmetric architecture, as provided in one embodiment of the present invention.

[0063] Figure 2 This is a flowchart of a specific embodiment of the Ts-AEnet angle estimation network designed using a deep learning-based channel estimation method in an asymmetric architecture, as provided in one embodiment of the present invention.

[0064] Figure 3Simulation results comparing the root mean square error performance of the present invention with existing asymmetric system channel estimation methods for estimating signal angles under different signal-to-noise ratio settings.

[0065] Figure 4 The simulation diagram shows a comparison of the normalized mean square error performance of the present invention and existing asymmetric system channel estimation methods for estimating downlink channels under different signal-to-noise ratio settings. Detailed Implementation

[0066] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0067] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0068] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0069] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0070] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0071] Unless otherwise clearly defined and limited in the present invention, the terms "installation, connection, and coupling" shall be understood in a broad sense. For example, it may be a fixed connection, a detachable connection, or an integral connection; it may also be a mechanical connection, an electrical connection, or a direct connection, and may also be indirectly connected through an intermediate medium, or may be the communication inside two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0072] Embodiment 1

[0073] Referring to Figures 1-4 , this is the first embodiment of the present invention. This embodiment provides a channel estimation method based on deep learning in an asymmetric architecture, including:

[0074] The present invention proposes a channel estimation method based on deep learning in an asymmetric all-digital communication system. First, an uplink receiving array based on a nested sparse array topology is designed to restore the received signals of the sparse array into virtual array signals with a complete dimension; then, the obtained virtual signals are decomposed into on-grid and off-grid parts using the Taylor formula, and a two-step angle estimation network Ts-AEnet is designed to estimate the angle of arrival of the signals; finally, the least squares problem is solved to estimate the signal path gain and reconstruct the downlink channel. The present invention proposes an uplink receiving array based on a nested sparse array topology in an asymmetric all-digital communication system, which can greatly reduce the information loss caused by antenna loss, and proposes a Ts-AEnet angle estimation network based on deep learning, which not only reduces the complexity of channel estimation but also improves the system estimation accuracy.

[0075] The following further elaborates on the present invention with reference to the accompanying drawings:

[0076] The embodiment of the present invention proposes a channel estimation method based on deep learning in an asymmetric all-digital communication system. The asymmetric all-digital system includes a uniform linear array equipped with M antennas and a single-antenna user; in this uniform linear array, each antenna is connected to a transmitting radio frequency chain for downlink signal transmission, but only N (N < M) antennas are connected to receiving radio frequency chains for uplink signal reception, and the index is expressed as where, p i ∈{0, 1, …, M - 1};

[0077] Considering a multi-path scenario, the base station antenna array receives signals from L independently and identically distributed paths; the uplink channel is modeled as a millimeter-wave channel, expressed as:

[0078]

[0079] where, α l represents the gain of the l-th path, θl Let a represent the arrival angle of the l-th path. UL (θ l ) represents the uplink steering vector of the l-th path at the base station, with dimensions N×1, and is defined as

[0080]

[0081] in,[·] T The matrix transpose, where d represents the spacing between two adjacent antenna elements and λ represents the electromagnetic wave wavelength; if the user transmits an all-1 pilot signal, the signal received at the base station is:

[0082]

[0083] Where ω is obeyed Additive white Gaussian noise, I represents noise power. N It is an identity matrix of dimension N. This represents a Gaussian distribution with mean μ and covariance Σ. The specific process includes the following steps:

[0084] Step 1: Design an uplink receiving array based on a nested sparse array topology, and then use a virtual array interpolation method to recover the signal received by the sparse array into a virtual array signal with full dimensions. The specific implementation steps are as follows:

[0085] Step 1.1: Employ a nested array structure as the uplink receiver array for the asymmetric system, such as... Figure 1 As shown; the nested array consists of two subarrays with different antenna element spacings, where subarray 1 consists of n antenna elements and subarray 2 consists of m antenna elements; subarray 1 is a uniform linear array with an antenna spacing of d; two adjacent antenna elements of subarray 2 can be nested together, so the antenna spacing is (n+1)d.

[0086] Set the antenna index for the nested array structure. Where subarray 1 has indices {0, 1, ..., n-1} and subarray 2 has indices {n, 2n+1, ..., m(n+1)-1}, this index set... The corresponding antennas are connected to the uplink receiving RF chain; therefore, the total number of base station antennas is M = m(n+1), and the number of uplink receiving array antennas is N = m+n.

[0087] Step 1.2: The covariance matrix of the received signal is calculated using equation (3):

[0088]

[0089] in,(·) HR represents the conjugate transpose; in practice, R is usually obtained by sampling covariance. y ,Right now Where K is the number of samples.

[0090] Step 1.3: Extract the elements from the covariance matrix to recover the complete signal; due to the existence of in equation (4) The calculation process, thus the covariance matrix R y Located at the pth i line p j The elements of the column contain the phase difference portion in Define the nonnegative element p in it. i -p j The set is:

[0091]

[0092] Next, the covariance element values ​​corresponding to the elements in set V are extracted to fill the signal gaps in the sparse array received signal caused by missing antenna elements; for example, when the antenna with index "3" in the uplink receiving array is not connected to the receiving RF chain, it cannot receive a signal, and thus the set is found. The R corresponding to element "3" y The covariance element values ​​are used to fill in the missing received signal at that position; thus, a virtual uniform linear array of dimension M can be constructed, and the reconstructed virtual array signal is represented as:

[0093]

[0094] in, This represents the guiding matrix of a reconstruction array with dimension M×L. This represents the downlink array steering matrix for all L path signals. This represents the downlink array steering vector of the l-th path signal; This represents a power vector with dimension L×1; The noise term represents the reconstruction, and e0 represents the matrix I when reconstructing the virtual signal according to the above method. N The deformation.

[0095] Step 2: Decompose the obtained virtual signal into in-grid and off-grid components using Taylor's formula, and design a two-step angle estimation network (Ts-AEnet) to estimate the signal's angle of arrival. This network includes two sub-networks: an in-grid angle estimation neural network and an off-grid angle estimation neural network, which estimate the in-grid angle and the off-grid angle, respectively, including the following steps:

[0096] Step 2.1: Construct a Q-dimensional overcomplete set of angles {θ1, θ2, ..., θ} Q The set divides an angular range into Q grid values ​​on average, and then uses the first-order Taylor formula to guide the array in equation (6) into a matrix. The l-th column is decomposed into the grid portion and the off-grid portion;

[0097]

[0098] Where, θ ql ∈{θ1,θ2,…,θ Q} represents distance θ l The most recent grid angle value, a(θ) ql ) represents the array steering vector at the grid angle, a'(θ) ql ) represents a(θ ql ) at θ ql The derivative at Δ ql =θl-θ ql Indicates the angle value from the grid;

[0099] Based on equation (7), the virtual signal observed in equation (6) is... Rewritten as:

[0100]

[0101] Where, Φ=[a DL (θ1),a DL (θ2),…,a DL (θ Q ] represents the down-row array steering matrix of dimension M×Q in the grid value, a DL (θ ql ) represents the downlink array steering vector at the signal angle at the grid angle; F = [a' DL (θ1),a' DL (θ2),…,a' DL (θ Q )] represents the derivative matrix corresponding to Φ, a' DL (θ ql ) represents a DL (θ ql ) at θ ql The derivative at point; z = [z(θ1), z(θ2), ..., z(θ)] Q )] T Let z(θ) represent the Q-dimensional spectrum in the grid angular space. ql ) indicates that the grid angle is θ ql The signal power value at that time, since there are only L paths of signal, z is only at L positions {θ}. q1 ,θ q2,…,θ qL There exists a corresponding value at} Right now Other values ​​are 0; Θ represents a diagonal matrix of dimension Q, with diagonal elements containing L angle values ​​from the grid.

[0102] Step 2.2: As described in Step 2.1, since the grid angles are distributed on a finite number of angle grids, grid angle estimation will be treated as a classification problem; design a grid angle estimation network to estimate the grid angle values ​​contained in z in equation (8); it should be noted that the output value of the network does not represent the true power value, it only contains the angle position information presented in the form of a binary vector; therefore, the angle is estimated by finding L maximum values ​​from the output angle space spectrum.

[0103] Design a convolutional neural network for grid angle estimation, consisting of a preprocessing layer, an input layer, four convolutional layers, one fully connected layer, and an output layer. The network expression is as follows:

[0104]

[0105] Preprocessing layer The process includes the generation of the virtual signal in step 1.3 above; specifically, it involves calculating the covariance of the received signal in equation (4) and filling the missing holes in the physical array to obtain the virtual signal in (6).

[0106] The input layer receives the preprocessed signal data. Since neural networks can only process real numbers, they convert the virtual signal... Combination of real and imaginary parts As input data, among which, Indicate extraction The real part, Indicate extraction The imaginary part is then converted into a tensor of dimension M×1×2 and passed to the network input layer.

[0107] Convolutional layer function Where Conv(·) represents the convolution operation, x represents the input data, and {Ξ k ,ξ k} represents the weights and biases of the k-th convolutional layer, with a kernel size of c. l ×c l c l This represents the size of the l-th convolutional kernel; a batch normalization layer is connected after the convolutional layer to speed up training; ReLU is set as the activation function.

[0108] Fully connected layer Where Fc(·) represents a fully connected operation, and x represents the input data. This represents the weights and biases of the k-th fully connected layer; Flatten(·) represents the flattening operation, setting ReLU as the activation function;

[0109] The output layer is set as a dense layer with Q neurons, and outputs a Q-dimensional spatial spectral vector at the grid angle;

[0110] Step 2.3: As described in Step 2.1, the angle away from the grid is a continuous value. Treating the grid angle estimation as a regression problem, a deep neural network is designed to estimate the non-zero diagonal elements of Θ in equation (8), which correspond to the angle away from the grid. Since the range of the angle away from the grid is... Within the range, it varies within a small range, unlike the grid value which takes values ​​over a large angular range; therefore, considering the feature that the grid angle value is easy to extract, a deep neural network with a simple structure and easy training is used to construct the grid angle estimation network.

[0111] The distance-from-grid angle estimation network consists of one input layer, seven fully connected layers, and one output layer, and is represented as follows:

[0112]

[0113] in, Similar to the fully connected layer shown in step 2.3; the input layer receives the preprocessed signal data and converts the virtual signal... Cascade of real and imaginary parts As input data, it is passed to the network input layer; concatenate[·] indicates a cascade operation; the output layer consists of L neurons, representing L angle values ​​from the grid.

[0114] Step 2.4: Combine the in-mesh angle estimation network and off-mesh angle estimation network designed in Steps 2.2 and 2.3 to form the complete Ts-AEnet, as shown below. Figure 2 As shown;

[0115] Step 3: After processing the received signal dataset in the preprocessing layer, two neural networks are trained offline using supervised learning on the training set. Then, the trained Ts-AEnet is deployed online to estimate the signal angle. Next, the least squares problem is solved to estimate the signal path gain. Finally, the downlink channel is reconstructed, including the following steps:

[0116] Step 3.1: Generate a training dataset for the grid angle estimation network with a sample size of D1; construct a Q-dimensional overcomplete angle set. in, It is the interval of the angle set, θ max It is the maximum angle; from Select each set of angles and calculate the corresponding signal y according to (3) at the random signal-to-noise ratio level. Then, preprocess the signal according to steps 1.2 and 1.3 to obtain the virtual array signal; define the input data. For each Y i Convert the corresponding L angle values ​​into a Q-dimensional binary vector b, where only L elements are set to 1 and the rest to 0, serving as the labels for the input data; define the training dataset.

[0117] In the training dataset The network for estimating grid angles is trained offline using supervised learning; network parameters are updated via backpropagation to minimize the loss function; and a binary cross-entropy function is used as the loss function.

[0118]

[0119] Step 3.2: Generate the training dataset for the off-grid angle estimation network, with a sample size of D2; compared to Step 3.1, the difference in generating the training dataset for the off-grid angle estimation network lies in the format of the dataset. The continuous angles are selected, and the corresponding signal y is calculated according to (3) at the random signal-to-noise ratio level. Then, the virtual array signal is obtained by preprocessing the signal according to steps 1.2 and 1.3. The input data is defined as The label is represented as c = {Δ ql}, l=1,2,…,L, representing the difference between the true value and the nearest grid value; then the training dataset for this stage is denoted as:

[0120] In the training dataset The network is trained offline on a grid angle estimation network. The goal of training the network is to minimize the mean squared error between the predicted and ground truth values, and the loss function is expressed as:

[0121]

[0122] Step 3.3: Merge the two trained sub-networks into a complete Ts-AEnet, deploy it online, process the received signal and input it into the network to obtain the estimated signal angle;

[0123] Step 3.4: After obtaining the estimated signal angle, solve the least squares problem. The estimated signal path gain is calculated as follows:

[0124]

[0125] in, These represent the estimated path and angle of arrival, respectively. This represents the updirection matrix, where ρ is a positive number used to ensure the matrix is ​​full rank and prevent it from being impossible to invert.

[0126] Finally, the downlink channel is rebuilt.

[0127] Figure 3 and Figure 4 The figure shows the simulation results of this invention, with the following parameter settings: the base station has M=110 antennas, of which only N=20 are used for uplink signal reception; the number of paths L=2.

[0128] Figure 3 The horizontal axis represents the signal-to-noise ratio (SNR) in decibels, and the vertical axis represents the root mean square error of the estimated angle. The five curves in the figure represent: the proposed Ts-AEnet scheme, which shows the effect obtained according to the scheme designed in this invention; the random scheme, which shows the channel estimation performance obtained when randomly selecting uplink receiving antennas; and the Newtonized orthogonal matching pursuit algorithm, the multiple signal classification algorithm, and the convolutional neural network algorithm, all based on existing channel estimation methods. As can be seen from the figure, the estimation performance of the scheme proposed in this invention is significantly better than other schemes.

[0129] Figure 4 The horizontal axis represents the signal-to-noise ratio (SNR) in decibels, and the vertical axis represents the normalized mean square error of channel estimation. The five curves in the figure are the same as those described above. Similarly, it can be seen from this figure that the channel estimation performance of the proposed scheme is superior to other schemes.

[0130] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0131] Example 2

[0132] The second embodiment of the present invention differs from the first embodiment in that:

[0133] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0134] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0135] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0136] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0137] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0138] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0139] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A deep learning-based channel estimation method in an asymmetric architecture, characterized in that: include, Design an uplink receiving array based on a nested sparse array topology, and use a virtual array interpolation method to recover the signal received by the sparse array into a virtual array signal with full dimensions. The obtained virtual signal is decomposed into a grid part and an off-grid part using the Taylor formula. A two-step angle estimation network Ts-AEnet is designed to estimate the signal angle of arrival. The two-step angle estimation network design includes two sub-networks: a grid angle estimation neural network and an off-grid angle estimation neural network, which estimate the grid angle and the off-grid angle, respectively. Design a convolutional neural network for estimating grid angles, consisting of a preprocessing layer, an input layer, four convolutional layers, a fully connected layer, and an output layer. The off-grid angle estimation network consists of one input layer, seven fully connected layers, and one output layer. The designed on-grid angle estimation network and off-grid angle estimation network are combined to form the complete Ts-AEnet. After the received signal dataset is processed by the preprocessing layer, a neural network is trained offline using supervised learning on the training set. Then, the trained Ts-AEnet is deployed online to estimate the signal angle and solve the least squares problem to estimate the signal path gain, thus reconstructing the downlink channel.

2. The channel estimation method based on deep learning in an asymmetric architecture as described in claim 1, characterized in that: The uplink receiving array based on a nested sparse array topology includes a uniform linear array with M antennas and a single-antenna user. In the uniform linear array, each antenna is connected to a transmit RF chain for downlink signal transmission, but only N (N < M) antennas are connected to a receive RF chain for uplink signal reception, denoted by the index. Where, p i ∈{0,1,…,M-1}, and when i=j, p i ≠p j p i and p j They all said Any element in the equation; consider a multipath scenario where a base station antenna array receives signals from L independent, identically distributed paths; model the uplink channel as a millimeter-wave channel, expressed as: Where, α l Let θ represent the gain of the l-th path. l Let a represent the arrival angle of the l-th path. UL (θ l ) represents the uplink steering vector of the l-th path at the base station, with dimensions N×1, and is defined as: Where d represents the distance between two adjacent antenna elements, and λ represents the wavelength of the electromagnetic wave; if the user sends an all-1 pilot signal, the signal received at the base station is: Where ω is obeyed Additive white Gaussian noise, I represents noise power. N It is an identity matrix of dimension N. This represents a Gaussian distribution with mean μ and covariance Σ.

3. The channel estimation method based on deep learning in an asymmetric architecture as described in claim 2, characterized in that: The uplink receiving array based on the nested sparse array topology also includes an uplink receiving array that uses a nested array structure as an asymmetric system. The nested array consists of two subarrays with different antenna element spacings. Subarray 1 consists of n antenna elements and subarray 2 consists of m antenna elements. Subarray 1 is a uniform linear array with an antenna spacing of d. Subarray 2 is nested between two adjacent antenna elements, so the antenna spacing is (n+1)d. Set the antenna index for the nested array structure. Where subarray 1 has indices {0, 1, ..., n-1} and subarray 2 has indices {n, 2n+1, ..., m(n+1)-1}, the current index set is... The corresponding antenna is connected to the uplink receiving RF chain; The total number of base station antennas is M = m(n+1), and the number of uplink receiving array antennas is N = m+n.

4. The channel estimation method based on deep learning in an asymmetric architecture as described in claim 3, characterized in that: The virtual array interpolation method includes calculating the covariance matrix of the received signal as follows: Wherein, the superscript H indicates the conjugate transpose; in practice, R y Obtained through sampling covariance, and Where K is the number of samples; Extracting elements from the covariance matrix to recover the complete signal; due to the existence of The calculation process, thus the covariance matrix R y Located at the pth i line p j The elements of the column contain the phase difference portion in Define the nonnegative element p in it. i -p j The set is: Next, extract the set. The covariance element values ​​corresponding to the elements in the middle are used to fill the signal gaps in the sparse array received signal caused by missing antenna elements; when the antenna with index "3" in the uplink receiving array is not connected to the receiving RF chain, no signal can be received. This is achieved by finding the set... The R corresponding to element "3" in the middle y The covariance element values ​​fill in the missing received signal at the current position, constructing a virtual uniform linear array of dimension M, where the reconstructed virtual array signal is represented as: in, Let A be the guiding matrix of the reconstruction array of dimension M×L. DL (θ)=[a DL (θ1),a DL (θ2),…,a DL (θ L [)] represents the downlink array steering matrix for all L path signals. This represents the downlink array steering vector of the l-th path signal; This represents a power vector with dimension L×1; The noise term represents the reconstruction, and e0 represents the matrix I when reconstructing the virtual signal according to the above method. N The deformation.

5. The channel estimation method based on deep learning in an asymmetric architecture as described in claim 4, characterized in that: The decomposition of the obtained virtual signal includes constructing a Q-dimensional overcomplete angle set {θ1, θ2, ..., θ...} Q }, set {θ1,θ2,…,θ Q Divide an angular range into Q equal grid values, and use a first-order Taylor formula to array the guiding matrix. The l-th column is decomposed into the grid portion and the off-grid portion, as follows: a(θ l )≈a(θ ql )+a'(θ ql )D ql Where, θ ql ∈{θ1,θ2,…,θ Q } represents distance θ l The most recent grid angle value, a(θ) ql ) represents the array steering vector at the grid angle, a'(θ) ql ) represents a(θ ql ) at θ ql The derivative at Δ ql =θ l -θ ql Indicates the angle value from the grid; Based on the observed virtual signals Rewritten as: Where, Φ=[a DL (θ1),a DL (θ2),…,a DL (θ Q ] represents the down-row array steering matrix of dimension M×Q in the grid value, a DL (θ ql ) represents the downlink array steering vector at the signal angle at the grid angle; F = [a' DL (θ1),a' DL (θ2),…,a' DL (θ Q )] represents the derivative matrix corresponding to Φ, a' DL (θ ql ) represents a DL (θ ql ) at θ ql The derivative at point; z = [z(θ1), z(θ2), ..., z(θ)] Q )] T Let z(θ) represent the Q-dimensional spectrum in the grid angular space. ql ) indicates that the grid angle is θ ql The signal power value at that time, since there are only L paths of signal, z is only at L positions {θ}. q1 ,θ q2 ,…,θ qL There exists a corresponding value at} Right now Other values ​​are 0; Θ represents a diagonal matrix of dimension Q, with diagonal elements containing L angle values ​​from the grid. Given a grid angle distribution on a finite number of angle grids, a grid angle estimation network is designed to estimate the grid angle values ​​contained in z. The network output value does not represent the true power value, but only contains angular position information presented in binary vector form. Therefore, the angle is estimated by finding L maximum values ​​from the output angular spatial spectrum.

6. The channel estimation method based on deep learning in an asymmetric architecture as described in claim 5, characterized in that: Design a convolutional neural network for grid angle estimation. The network consists of a preprocessing layer, an input layer, four convolutional layers, one fully connected layer, and an output layer. The network expression is as follows: Preprocessing layer The process of generating virtual signals involves the covariance of the received signals and filling in the missing holes in the physical array to obtain virtual signals. The input layer receives the preprocessed signal data. Since neural networks can only process real numbers, they convert the virtual signal... Combination of real and imaginary parts As input data, among which, Indicate extraction The real part, Indicate extraction The imaginary part is then converted into a tensor of dimension M×1×2 and passed to the network input layer. Convolutional layer function Where Conv(·) represents the convolution operation, and x represents the input data. This represents the weights and biases of the k-th convolutional layer, with a kernel size of c. l ×c l c l This represents the size of the l-th convolutional kernel; a batch normalization layer is connected after the convolutional layer to speed up training; ReLU is set as the activation function. Fully connected layer Where Fc(·) represents a fully connected operation, and x represents the input data. This represents the weights and biases of the k-th fully connected layer; Flatten(·) represents the flattening operation, setting ReLU as the activation function; The output layer is set as a dense layer with Q neurons, and outputs a Q-dimensional spatial spectral vector at the grid angle; Since the angle from the grid is a continuous value, we treat the grid angle estimation as a regression problem and design a deep neural network to estimate the non-zero diagonal elements of Θ, which correspond to the angle values ​​from the grid. The distance-from-grid angle estimation network consists of one input layer, seven fully connected layers, and one output layer, denoted as follows: in, Similar to the fully connected layer described above; the input layer receives the preprocessed signal data and converts the virtual signal... Cascade of real and imaginary parts As input data, it is passed to the network input layer; concatenate[·] indicates a cascade operation; the output layer consists of L neurons, representing L angle values ​​from the grid. The designed on-grid angle estimation network and off-grid angle estimation network are combined to form the complete Ts-AEnet.

7. The channel estimation method based on deep learning in an asymmetric architecture as described in claim 6, characterized in that: The offline training of two neural networks using supervised learning includes: Generate a training dataset for the grid angle estimation network with a sample size of D1; Construct a Q-dimensional overcomplete angle set in, It is the interval of the angle set, θ max It is the maximum angle; from Select each set of angles and calculate the corresponding signal y based on the random signal-to-noise ratio level to obtain the virtual array signal according to the preprocessed signal; define the input data. For each Y i The corresponding L angle values ​​are converted into a Q-dimensional binary vector b, where only L elements are set to 1, and all other elements are 0, which serve as the labels for the input data; the training dataset is defined. In the training dataset The network for estimating grid angles is trained offline using supervised learning; network parameters are updated via backpropagation to minimize the loss function; and a binary cross-entropy function is used as the loss function. Generate a training dataset for the off-grid angle estimation network with D2 samples; from The continuous angles are selected in a concentrated manner, and the corresponding signal y is calculated based on the random signal-to-noise ratio level. The virtual array signal is obtained according to the preprocessed signal, and the input data is defined as follows: The label is represented as c = {Δ ql }, l=1,2,…,L, representing the difference between the true value and the nearest grid value; then the training dataset for the current stage is denoted as In the training dataset The network is trained offline on a grid angle estimation network. The goal of training the network is to minimize the mean squared error between the predicted and ground truth values, and the loss function is expressed as: The two trained sub-networks are merged into a complete Ts-AEnet, which is then deployed online. The received signals are processed and input into the network to obtain the estimated signal angle.

8. The channel estimation method based on deep learning in an asymmetric architecture as described in claim 7, characterized in that: The method of solving the least squares problem to estimate the signal path gain includes... After obtaining the estimated signal angle, solve the least squares problem. The estimated signal path gain is calculated as follows: in, These represent the estimated path and angle of arrival, respectively. This represents the updirection matrix, where ρ is a positive number used to ensure the matrix is ​​full rank and prevent it from being impossible to invert. Finally, the downlink channel is reconstructed, specifically as follows:

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that... Similar to 5, when the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images