Beam focusing method and system based on convolution expansion self-coding element learning

Abstract

The invention discloses a beam focusing method based on convolution expansion self-coding element learning and a system thereof. The beam focusing method based on convolution expansion self-coding element learning comprises the following steps: establishing a system model; after the establishment of the system model is completed, determining a system optimization problem; and optimizing the system according to the determined system optimization problem, and outputting the maximum spectrum efficiency of the system. By elaborately designing the expansion coefficient and filling of the expansion convolution, the receptive field of the network is expanded, so that the global understanding of the whole input is realized as much as possible, and the spectrum efficiency of the system is improved. Besides, the pre-coding matrix is input into the network for optimization after dimensionality reduction is carried out on the pre-coding matrix by designing the auto-encoder, so that the search space and the convergence time are reduced, the algorithm complexity is reduced, and the network depth is increased by the design of the auto-encoder, so that the learning ability of the algorithm to a near-field channel is enhanced, and the spectrum efficiency of the system is improved.

Description

Technical Field

[0001] The present application relates to the field of mobile communications, and in particular, to a beam focusing method and system based on convolutional dilation autoencoder meta-learning. Background Art

[0002] In order to achieve comprehensive signal coverage of the Sixth Generation Mobile Networks (6G) communication, future 6G communication requires higher spectral efficiency (spectral efficiency, SE), lower power consumption, and solve the path loss and blocking effect of the communication system during signal propagation. Driven by this vision, the intelligent reflecting surface (IRS) technology is considered to be a key technology to achieve 6G performance indicators. However, due to the double path loss attenuation when the IRS reflects the signal, the physical / electrical size of the IRS usually needs to be designed to be large enough to obtain a good performance gain of the IRS-assisted communication. Therefore, the communication assisted by the extremely large-scale IRS (XL-IRS) aims to offset the path loss that occurs during signal transmission, thereby improving the reliability and transmission rate of the communication link. However, with the increase in the number of IRS reflection units, the Rayleigh distance in the communication system gradually expands, making the near-field scenario more common in wireless networks, and it is necessary to consider using near-field spherical waves to model the electromagnetic radiation field. Therefore, the information contained in the near-field channel is more complex than that in the far-field channel. This complexity increases the computational overhead of the system, affects the overall communication performance and efficiency, and thus places higher requirements on signal processing. Extremely Large-scale Array (ELAA) is a common feature of several key candidate technologies for 6G. Due to the huge number of antennas in ELAA, the near-field area is expanded, so near-field spherical waves are needed to accurately model the electromagnetic radiation field. For the near-field communication scenario where ELAA is deployed at the base station (BS), ELAA deployed at the BS can significantly improve the spectrum efficiency SE of the wireless network. Therefore, people have conducted extensive research on near-field communication where ELAA is deployed at the BS. However, there will be path loss and blocking effect problems during signal propagation. In order to expand the signal coverage, the introduction of IRS technology has become an effective solution to the above problems in view of the advantages of IRS-assisted communication. Most existing methods of IRS-assisted communication focus on the assumption of traditional far-field uniform plane waves, that is, for the same channel path, all IRS reflection units have the same arrival / departure angles (Angles of Arrival, AoA / Angles of Departure, AoD). However, in practical applications, when the size of the IRS is large enough, the near-field area of ​​the IRS-assisted communication link will be further expanded, and the BS and / or user equipment (UE) may no longer be located in the far-field area of ​​the XL-IRS, so it is necessary to consider using near-field spherical waves to accurately model the electromagnetic radiation field.For the near-field communication scenario assisted by XL-IRS, the prior art has studied the near-field modeling and performance analysis of the XL-IRS-assisted wireless communication system, proving the necessity of near-field modeling of wireless communication under the assistance of XL-IRS. The existing theory derives the two key parameters of the width and depth of the reflected beam in the communication of the IRS-assisted system, proving that the near-field beamforming design must consider the situation of the spherical wavefront. At present, there are studies discussing the near-field single-beam codebook design of XL-IRS, and proposing a hierarchical training scheme, and designing a variable single-beam hierarchical codebook suitable for near and far fields. At the same time, a multi-beam design scheme based on block coordinate descent and a dominant minimization algorithm are proposed to solve the equal modulus constraints of XL-IRS, and point out that if both the transmitter and the receiver are in the radiation near field, the IRS can establish a high-rank channel between the transmitter and the receiver to achieve maximum freedom and improve the communication capacity. At the same time, the prior art considers the near-field channel model of the BS to IRS link, and designs beamforming at the BS to improve the XL-IRS beam training performance. A low-complexity near-field beam focusing method is proposed. A deep reinforcement learning algorithm is first used to determine the active beamforming target projection and the corresponding passive beamforming matrix. Then, the codebook designed based on the geometric channel state information (CSI) is used to calculate the active beamforming matrix, reducing the computational complexity. However, the above-mentioned method is mainly for near-field channel modeling and performance analysis, which provides an effective theoretical guidance for the research of near-field communications. However, in order to improve the service quality of users, the communication system needs to have a higher spectral efficiency SE, which requires carefully designed beam focusing to generate high-quality beams. At the same time, due to the large number of reflection units, the information contained in the near-field channel is more complex, and the computational complexity of the precoding matrix at the BS and the phase shift matrix at the XL-IRS is increased.

[0003] Therefore, how to provide a beam focusing optimization method that strikes a balance between system spectrum efficiency SE performance and computational complexity to meet the needs of future 6G communication applications has become an urgent problem to be solved in this field. Summary of the invention

[0004] The present application proposes a beam focusing method based on convolutional dilation autoencoder meta-learning, comprising the following steps: establishing a system model; after completing the establishment of the system model, determining the system optimization problem; optimizing the system according to the determined system optimization problem, and outputting the maximum spectral efficiency of the system.

[0005] As described above, the beam focusing method based on convolutional dilation autoencoder meta-learning, wherein the establishment of the system model includes the following sub-steps: determining a multi-user communication system; based on the multi-user communication system, performing far-field channel modeling and near-field channel modeling.

[0006] As described above, the beam focusing method based on convolutional expansion autoencoder meta-learning, wherein the multi-user communication system includes a base station equipped with M antennas, an ultra-large-scale IRS with N reflection units, and K single-antenna user devices.

[0007] The beam focusing method based on convolution expansion autoencoder meta-learning as described above, wherein determining the multi-user communication system further includes determining the signal y received by the k-th user equipment k , expressed as:

[0008]

[0009] G∈C N×M represents the channel matrix between the base station and the ultra-large-scale IRS, C represents a set of complex numbers, and w k Denotes the precoding matrix W∈C M×K In the kth column at BS, w i Denotes the precoding matrix W∈C M×K In the i-th column at BS, s k represents the data flow of the kth UE, s i represents the data flow of the i-th UE, H∈C K×N The channel matrix between the ultra-large-scale IRS and the user equipment is represented as, is a cascade channel, Θ = diag(θ 1 ,θ 2 ,…,θ N ) is represented as the phase shift matrix of XL-IRS, represents the reflection coefficient of the nth reflection unit, where p n ∈[0,2π] represents the phase shift caused by the nth XL-IRS unit, h k represents the transpose of the kth row of H, represents the complex Gaussian white noise received by the kth UE.

[0010] As described above, the beam focusing method based on convolutional dilation autoencoder meta-learning, in which the system is optimized according to a determined system optimization problem, outputs the system's maximum spectral efficiency and includes the following sub-steps: determining the network configuration and parameter settings; determining the input parameters after completing the network configuration and parameter settings; initializing the information after determining the input parameters; and entering a three-layer nested loop after completing the information initialization.

[0011] A beam focusing system based on convolutional dilation autoencoder meta-learning specifically includes: a system establishment unit, an optimization problem determination unit and an optimization unit; the system establishment unit is used to establish a system model; the optimization problem determination unit is used to determine the system optimization problem; the optimization unit is used to optimize the system according to the determined system optimization problem and output the maximum spectrum efficiency of the system.

[0012] As described above, the beam focusing system based on convolutional dilation autoencoder meta-learning, wherein the system establishment unit establishes the system model including the following sub-steps: determining a multi-user communication system; and based on the multi-user communication system, performing far-field channel modeling and near-field channel modeling.

[0013] As described above, the beam focusing system based on convolutional expansion autoencoder meta-learning, wherein the multi-user communication system determined in the system establishment unit includes a base station equipped with M antennas, an ultra-large-scale IRS with N reflection units, and K single-antenna user devices.

[0014] The beam focusing system based on convolution expansion autoencoder meta-learning as described above, wherein determining the multi-user communication system further includes determining the signal y received by the k-th user equipment k , expressed as:

[0015]

[0016] G∈C N×M represents the channel matrix between the base station and the ultra-large-scale IRS, C represents a set of complex numbers, and w k Denotes the precoding matrix W∈C M×K In the kth column at BS, w i Denotes the precoding matrix W∈C M×K In the i-th column at BS, s k represents the data flow of the kth UE, s i represents the data flow of the i-th UE, H∈C K×N The channel matrix between the ultra-large-scale IRS and the user equipment is represented as, is a cascade channel, Θ = diag(θ 1 ,θ 2 ,…,θ N ) is represented as the phase shift matrix of XL-IRS, represents the reflection coefficient of the nth reflection unit, where p n ∈[0,2π] represents the phase shift caused by the nth XL-IRS unit, h k represents the transpose of the kth row of H, represents the complex Gaussian white noise received by the kth UE.

[0017] As described above, the beam focusing system based on convolutional dilation autoencoder meta-learning, wherein the optimization unit optimizes the system according to the determined system optimization problem, and the output system maximum spectral efficiency includes the following sub-steps: determining the network configuration and parameter settings; determining the input parameters after completing the network configuration and parameter settings; initializing the information after determining the input parameters; and entering a three-layer nested loop after completing the information initialization.

[0018] This application has the following beneficial effects:

[0019] (1) The Convolutional Dilated Autoencoder Meta Learning (CDAML) algorithm used in this application utilizes the characteristics of local connection and weight sharing to reduce the weight parameters that need to be learned for XL-IRS-assisted near-field communication, thereby achieving a balance between the system SE performance and computational complexity.

[0020] (2) This application expands the network's receptive field by carefully designing the expansion coefficient and padding of the dilated convolution, thereby achieving a global understanding of the entire input as much as possible and improving the system's spectral efficiency. In addition, by designing an autoencoder to reduce the dimension of the precoding matrix before inputting it into the network for optimization, the search space and convergence time are reduced, the algorithm complexity is reduced, and the design of the autoencoder increases the network depth, thereby enhancing the algorithm's ability to learn near-field channels and improving the system's spectral efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0022] Figure 1 It is a flowchart of a beam focusing method based on convolution dilation autoencoder meta-learning provided in an embodiment of the present application;

[0023] Figure 2 is a schematic diagram of an XL-IRS-assisted multi-user communication system provided according to this embodiment;

[0024] Figure 3 is a schematic diagram of a near-field channel model from an XL-IRS to a k-th user according to this embodiment;

[0025] Figure 4 is a schematic diagram of a specific structure of a CDAML algorithm network provided according to this embodiment;

[0026] Figure 5 It is a schematic diagram of the internal structure of a beam focusing system based on convolutional dilation autoencoder meta-learning provided in an embodiment of the present application. DETAILED DESCRIPTION

[0027] The following is a clear and complete description of the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0028] This application is based on the problem of maximizing the spectrum efficiency SE of a near-field communication system assisted by XL-IRS. Under the constraints of the equal modulus of the transmit power and the XL-IRS reflection unit, the precoding matrix at the BS and the phase shift matrix at the XL-IRS are jointly optimized. At the same time, an unsupervised learning and a convolution-expanded autoencoder meta-learning algorithm without pre-training are proposed. First of all, this application is an early attempt at a joint optimization algorithm for beam focusing in an XL-IRS-assisted near-field multi-user communication system, and studies are conducted under the assumptions of perfect and imperfect CSI, respectively.

[0029] Secondly, the optimization algorithm proposed in this application directly uses the matrix form of the precoding matrix and the phase shift matrix as the input of the network, and unlike the traditional method of horizontally splicing the imaginary part to the real part according to the specified dimension, this application vertically splices the real part and the imaginary part, and through the careful design of the convolution kernel and padding of the convolutional neural network, the nonlinear phase information between the real and imaginary parts of the near-field channel matrix is ​​fully retained. Then, by carefully designing the expansion coefficient and padding of the dilated convolution, the receptive field of the network is expanded, so as to achieve a global understanding of the entire input as much as possible and improve the system spectral efficiency SE. In addition, by designing an autoencoder to reduce the dimension of the precoding matrix and then input it into the network for optimization, the search space and convergence time are reduced, the algorithm complexity is reduced, and the design of the autoencoder increases the network depth, thereby enhancing the algorithm's learning ability for the near-field channel and improving the system spectral efficiency.

[0030] Note: Unless otherwise noted, matrices are denoted by bold uppercase letters (e.g., H), vectors are denoted by bold lowercase letters (e.g., h), and scalars are denoted by plain letters (e.g., d). C denotes the set of complex numbers. h k Represents the kth element of a vector. [·] -1 , [·] T ,[·] H , Tr(·), diag(·), |·| and ||·|| 2 They represent inverse, transpose, conjugate transpose, trace, diagonal operation, modulus operation and L respectively. 2 Norm. I N is an N×N identity matrix. Indicates that the mean is u and the covariance is is a complex Gaussian random distribution. E[·] represents the statistical expectation. O represents the asymptotic time complexity.

[0031] Embodiment 1

[0032] like Figure 1 As shown, the beam focusing method based on convolution expansion autoencoder meta-learning provided in the embodiment of the present application specifically includes the following steps:

[0033] Step S1: Establishing a system model.

[0034] Wherein step S1 comprises the following sub-steps:

[0035] Step S11: Determine a multi-user communication system.

[0036] Before determining the system model, it is necessary to first determine the XL-IRS-assisted multi-user communication system, such as Figure 2 As shown in FIG. 4 , an XL-IRS-assisted multi-user communication system consists of a BS equipped with M antennas, an XL-IRS with N reflection units, and K single-antenna UEs.

[0037] The BS uses a uniform linear array (ULA), and the XL-IRS uses a uniform planar array (UPA). A three-dimensional coordinate system is established with the center of the XL-IRS plane as the origin. 1 and N 2 Respectively represent the number of reflection units in the z-axis and y-axis directions, that is, N = N 1 ×N 2 , d is the distance between adjacent reflection units. The data stream of the kth UE is marked as s k , satisfying E{|s k | 2}=1.

[0038] Considering that the direct connection between BS and UE is blocked by buildings, UE can only communicate with BS through XL-IRS. The channel matrix between BS and XL-IRS is G∈C N×M , the channel matrix between XL-IRS and UE is H∈C K×N , then the signal y received by the kth UE k It can be expressed as:

[0039]

[0040] And define is a cascade channel, θ = diag(θ 1 ,θ 2 ,…,θ N ) is represented as the phase shift matrix of XL-IRS, represents the reflection coefficient of the nth reflection unit, where pn ∈[0,2π] represents the phase shift caused by the nth XL-IRS unit, w i Denotes the precoding matrix W∈C M×K In the i-th column at BS, s k represents the data flow of the kth UE, s i represents the data flow of the i-th UE, h k represents the transpose of the kth row of H, represents the complex Gaussian white noise received by the kth UE.

[0041] Step S12: Based on the multi-user communication system, far-field channel modeling and near-field channel modeling are performed.

[0042] The Rayleigh distance is proportional to the square of the maximum aperture of the array. For an array with a fixed carrier wavelength and spacing between reflective units, the Rayleigh distance increases accordingly as the number of reflective units increases. Therefore, the near-field area will be expanded, and on this basis, the BS and / or UE are more likely to be located in the near-field area. For XL-IRS-assisted communication systems, as long as either the BS or the UE is located in the near-field area of ​​the XL-IRS, the channel should be accurately modeled using near-field spherical waves. To comply with the typical deployment strategy of XL-IRS, this embodiment considers the scenario where the UE is located in the near-field area of ​​the XL-IRS and the BS is located in the far-field area of ​​the XL-IRS.

[0043] Based on the above, step S12 includes the following sub-steps:

[0044] Step S121: Perform far-field channel modeling.

[0045] Therefore, the far-field channel from BS to XL-IRS can be modeled as the Saleh-Valenzuela channel model, and considering a line-of-sight (LoS) path and L non-line-of-sight (NLoS) paths, the far-field channel from BS to XL-IRS is expressed as:

[0046]

[0047] Where κ is the Rice fading factor, α l is the complex gain of the lth path, φ in,l and Respectively represent the azimuth and elevation angles of the signal arriving at XL-IRS. l represents the AoD of the lth path at the BS, where l = 0 represents a LoS path and l > 0 represents a NLoS path. a(γ) and are the far-field array steering vectors of BS and XL-IRS, respectively. Under the far-field plane wavefront assumption, a(γ) and They can be expressed as:

[0048]

[0049]

[0050] Where λ is the signal wavelength.

[0051] Step S122: Perform near-field channel modeling.

[0052] Since this embodiment considers that the UE is located in the near field area of ​​the XL-IRS, such as Figure 3 As shown, it is the near-field channel model from XL-IRS to the kth user.

[0053] In near-field channel modeling, the array steering vector must consider both angle information and distance information, and the angle and distance from each reflector in the XL-IRS to the UE antenna are different, that is, the phase of each reflector in the XL-IRS needs to be calculated independently. Therefore, the channel from the XL-IRS to the kth UE can be modeled as:

[0054]

[0055] Similarly, l = 0 indicates a LoS path, and l > 0 indicates a NLoS path. l is the complex gain of the lth path, r k,l represents the spatial coordinate vector of the lth path from XL-IRS to UE, φ re,k,l and They represent the azimuth and elevation angles of the signal leaving the XL-IRS, respectively. Represents the XL-IRS near-field array steering vector, which focuses the signal energy on The indicated position is expressed based on the near-field spherical wave model as:

[0056]

[0057] Where r is the distance between the scatterer (or UE) and the center of the XL-IRS array, represents the reflection unit index of each XL-IRS, where In the lth path, the scatterer (or UE) and the reflector (n 1 ,n 2 ) is expressed as:

[0058]

[0059] Where r l represents the distance between the scatterer (or UE) and the center of XL-IRS in the lth path.

[0060] Step S2: After the system model is established, the system optimization problem is determined.

[0061] In formula (1), is the signal that the kth UE expects to receive, and the interference between the kth UE and other UEs is considered as noise. k With unit power, the signal-to-noise ratio of the kth UE can be expressed as:

[0062]

[0063] Using spectrum efficiency SE to evaluate system performance can be quantified as follows:

[0064]

[0065] Where ω k represents the weight of the kth user. By optimizing W and Θ, the system spectrum efficiency SE is maximized. Therefore, the optimization problem of this embodiment can be described as:

[0066]

[0067] Where P t is the maximum total transmit power of BS, C 1 The constraint is that the total transmission power of the BS is limited, C 2 The constraint is the isomodus constraint of the XL-IRS reflection unit, that is, the reflection coefficient of the XL-IRS reflection unit is on the unit circle, which simplifies the hardware implementation of XL-IRS.

[0068] Step S3: Optimize the system according to the determined system optimization problem and output the maximum spectrum efficiency of the system.

[0069] This embodiment uses the CDAML algorithm network to extract and learn nonlinear phase features in the near-field channel, thereby improving the spectrum efficiency SE of the XL-IRS-assisted near-field communication system.

[0070] Specifically, from the far-field channel modeling from BS to XL-IRS and the near-field channel modeling from XL-IRS to UE, it can be known that the main difference between far-field communication and near-field communication modeling lies in the array steering vector, which is not only related to the angle but also to the distance. In order to enable the CDAML algorithm network to better learn the phase information related to the angle and distance in the near-field channel matrix, this embodiment uses the matrix form of the phase shift matrix and the precoding matrix as the input of the CDAML algorithm network. Compared with expanding the matrix into one dimension as the network input, using the matrix form as the input can better extract the relevant features between the various parameters of the near-field channel.

[0071] At the same time, considering that the precoding matrix is ​​a complex matrix, deep learning frameworks such as tensorflow, keras and pytorch do not support complex input. A common solution is to divide the complex matrix into two parts, the real part and the imaginary part, and then horizontally splice the imaginary part to the real part according to the specified dimension. However, this method of processing complex matrices cannot well preserve the relevant information between the real part and the imaginary part, especially the nonlinear phase information in near-field communication. Therefore, in this embodiment, the real part and the imaginary part of the precoding matrix are vertically spliced ​​as the input of the CDAML network, fully retaining the nonlinear phase information about angles and distances in near-field communication. Specifically, the dimension of the input precoding complex matrix is ​​M×K, and the precoding matrix is ​​converted to B×c×M×K by adding an extra dimension. After the real part and the imaginary part are vertically spliced, the dimension is B×c×2×(M×K), where B represents the batch size and c represents the number of channels.

[0072] like Figure 4 As shown in the figure, the specific structure of the CDAML algorithm network is shown in the figure. The CDAML network architecture is mainly composed of CNN. 1 and CNN 2 It consists of two sub-networks.

[0073] CNN 1 The structure consists of two layers of dilated convolutional layers (DC) with a convolution kernel of 3×3, a maximum pooling layer (pool), a fully connected layer (FC) and a Lambda layer. A ReLU activation function is added after each dilated convolutional layer (DC) to enable the network to learn and represent nonlinear relationships. Dilated convolution can expand the receptive field without increasing the number of parameters of the convolution kernel, thereby avoiding excessive parameters and calculations, and helping to improve the efficiency and generalization ability of the model. Applying it to the beam focusing optimization of XL-IRS-assisted near-field communication can capture larger-sized input data information and improve the spectral efficiency SE of the XL-IRS-assisted near-field communication system. The maximum pooling layer can reduce the amount of data while retaining the most important information, further reducing the computational complexity of the network. The function of the fully connected layer is to flatten the output of the previous layer into a one-dimensional vector. To ensure that the output reflection unit angle is in the range of 0 to 2π and satisfies the equal modulus constraint C of the XL-IRS reflection unit 2 , the output of the fully connected layer is passed through the sigmoid function to control the output within the range of 0 to 1. Then, it passes through the Lambda layer.

[0074] CNN 1 Taking the phase shift matrix Θ as the network input, the final output of the Lambda layer is expressed as ε=2π is the amplification factor, is the output of the fully connected layer, δ(·) represents the sigmoid function. The reflection coefficient of the final updated reflection unit is expressed as:

[0075]

[0076] In the formula Represents the updated reflection coefficient of the nth reflection unit. Finally, through diagonal operation Get the updated phase shift matrix

[0077] CNN 2 The network structure consists of an autoencoder, two convolutional layers with 2×2 convolution kernels, and a fully connected layer. The autoencoder consists of two convolutional layers with 2×2 convolution kernels and two average pooling layers, and each convolutional layer is connected to an average pooling layer. The autoencoder is responsible for reducing the matrix dimension from B×c×2×(M×K) to B×c×2×K 2 dimension, focusing only on the valid signals received by each UE. In this way, the information of the BS antenna domain can be mapped to the UE signal space, retaining the relationship between the signals of each UE, while ignoring the redundant BS antenna space redundancy. Therefore, the computational complexity can be reduced and the processing efficiency can be improved. At the same time, the application of the autoencoder deepens the algorithm network depth, so that more complex near-field channel features can be extracted. Two layers of convolutional layers with a convolution kernel of 2×2 are connected after the autoencoder.

[0078] Specifically, the specific convolution process of the convolution layer is that for the matrix after the dimensionality reduction processing of the autoencoder, it is first filled with zeros, and the matrix dimension changes to B×c×4×(K 2 +2), and then pass through the first layer of convolution with a 2×2 convolution kernel. During the convolution process, the convolution kernel can extract the real and imaginary features separately, as well as the combined features of the real and imaginary parts. Therefore, the CDAML network can better extract the nonlinear phase characteristics of the real and imaginary parts of the complex matrix, and improve the system spectrum efficiency SE of XL-IRS assisted near-field communication. Finally, the second layer of convolution with a 2×2 convolution kernel restores the matrix to B×c×2×K 2 The function of the fully connected layer is to flatten the output into a one-dimensional vector. At the same time, considering the power constraint C 1 , CNN 2 The output of the fully connected layer is a one-dimensional vector of size 2×M×K, and then the real and imaginary parts are concatenated to obtain at last The power-constrained precoding matrix is ​​obtained by calculation using formula (12):

[0079]

[0080] Based on the above, step S3 specifically includes the following sub-steps:

[0081] Step S31: Perform CDAML algorithm network configuration and parameter setting.

[0082] As shown in Table 1, it is the network configuration and parameter settings of the CDAML algorithm.

[0083] Table 1

[0084]

[0085] Step S32: After completing the network configuration and parameter setting, the input parameters are determined.

[0086] The input parameters include sample channel matrices G and H, transmit power P t , sub-cycle number I i , Main loop times I o , iterative cycle number I e ,CNN 1 Learning rate l Θ ,CNN 2 Learning rate l W .

[0087] Step S33: Initialize the information after determining the input parameters.

[0088] Initializing the information includes:

[0089] Step S331: Randomly initialize θ Θ ,θ W , W (0,1) , Θ (0,1) .

[0090] θ Θ and θ W Respectively represent CNN 1 and CNN 2 The network parameter set, W (0,1) represents the initial precoding matrix of the input, that is, the precoding matrix of the first sub-cycle in the first main cycle, Θ (0,1) Represents the initial phase shift matrix of the input, that is, the phase shift matrix of the 1st sub-cycle in the 1st main cycle.

[0091] Step S332: (0,1) Perform normalization.

[0092] Formula 12 is used for normalization.

[0093] Step S333: Initialize the maximum spectrum efficiency SE.

[0094] Initialize the maximum spectrum efficiency SE, denoted as MAX=0.

[0095] Step S334: Initialize the loss function to 0.

[0096] Step S335: Let θ (0,j) =Θ (0,1) , W (0,j) =W (0,1) .

[0097] W (0,j) and θ (0,j) They respectively represent the precoding matrix and phase shift matrix of the 1st sub-cycle in the jth main cycle.

[0098] Step S34: After completing information initialization, enter a three-layer nested loop.

[0099] The three levels of nested loops include sub-loops, main loops, and iteration loops.

[0100] The function of the sub-loop is to be 1 and CNN 2 Loop optimization Θ and W.

[0101] Specifically, Θ (i-1,j) ,i=1,2,3,…I i ; j = 1, 2, 3, ... I o Enter CNN 1 subnetwork, and through CNN 1 Subnetwork output Θ (i,j) . (i-1,j) ,i=1,2,3,…I i ; j = 1, 2, 3, ... I o Enter CNN 2 subnetwork, and through CNN 2 Subnetwork output W (i,j) . Use formula 12 to normalize W (i,j) .

[0102] The main loop is used to accumulate the loss function. i Sub-cycle, repeat I i To achieve unsupervised learning, the loss function of the jth main loop is expressed as the negative value of SE:

[0103] L j =-R(W * ,Θ * ; H, G)(13)

[0104] Where W * and θ *They respectively represent the precoding matrix and phase shift matrix after each main cycle update.

[0105] The repetition I i sub-cycles, Θ * =Θ (Ii,j) , W * =W (Ii,j) .

[0106] Calculate L using formula (13) j , then accumulate the loss function in Represents the accumulated loss function after i main cycles.

[0107] If the accumulated loss function is greater than the maximum spectral efficiency SE(-L j >MAX), then MAX=-L j , output Θ = Θ * , W=W * .

[0108] The iteration loop is mainly responsible for updating CNN 1 and CNN 2 The parameters in the sub-network have I in each iteration cycle o A main loop, for the accumulated loss function Take the average and record it as

[0109]

[0110] Backpropagation using the Adam optimizer to update the CNN 1 and CNN 2 The network parameter set θ Θ and θ W The whole optimization process has I e Iteration loop, the update interval is set to a constant n o , in each n o Update the CNN once in an iterative cycle 1 and CNN 2 The network parameters in .

[0111] Finally, W and Θ are returned, and the spectrum efficiency SE and CPU running time are output.

[0112] The CDAML network not only has the advantages of the CNN network, but can also effectively capture the spatial structure and local correlation of the input data, thereby realizing the extraction and learning of nonlinear phase features in the near-field channel and improving the spectrum efficiency SE of the XL-IRS-assisted near-field communication system. At the same time, it has the advantages of meta-learning and three-layer nested loops, can quickly adapt to new XL-IRS-assisted near-field channels, and has the characteristics of no need for pre-training and unsupervised learning. It iterates the optimal precoding matrix W and phase shift matrix Θ, and finally outputs the maximum spectrum efficiency SE value of the XL-IRS-assisted near-field communication system.

[0113] This embodiment analyzes and compares the time complexity of the AO algorithm, the GMML algorithm, and the proposed CDAML algorithm, as well as the contrast algorithm.

[0114] First, we analyze the CNN in the CDAML algorithm 1 Complexity. CNN 1 The complexity of is mainly composed of calculating SE and embedding CNN network. SE is calculated according to formulas (8) and (9). Specifically, The required complexity is O(MN), and because it is necessary to calculate the signal-to-noise ratio γ of K UEs k ,so The calculation of needs to be repeated K times, and finally the SE calculated according to formula (9) is repeated K times. Therefore, the overall complexity of calculating SE is O(K 2 MN). The complexity of the embedded CNN network mainly comes from the fully connected layer, that is, O(N 2 ). Therefore, CNN 1 The complexity of a single iteration can be expressed as O(K 2 MN+N 2 ). Then analyze CNN 2 The complexity of CNN 2 The main components are the calculation formula (12), calculation of SE, autoencoder and embedded CNN network. Therefore, CNN 2 The complexity of a single iteration can be expressed as O(MK 2 +K 2 MN+MK+MK 3 )=O(K 2 MN+MK 3 Finally, considering three layers of nested loop iterations, the overall complexity of the CDAML algorithm is O(I e I o I i (K 2 MN+N 2 +MK 3 )).

[0115] The time complexity of AO algorithm and GMML algorithm is O(L 1 (L 2 KM 3 +L 3 K 2 N 2 )) and O(I e I o I i (K 2 MN+HN+HK 2 )). Therefore, the AO algorithm has the highest time complexity among the three algorithms. At the same time, when the number of reflection units is small, the time complexity of the CDAML algorithm proposed in this embodiment is equivalent to that of the GMML algorithm, and when the number of reflection units is large, the time complexity of the CDAML algorithm proposed in this embodiment is slightly higher than that of the GMML algorithm. This is because when the number of neurons H in the hidden layer of the GMML algorithm is not negligible compared with the number of reflection units, the complexity of the CDAML algorithm is equivalent to that of the GMML algorithm, and when the number of neurons in the hidden layer of the GMML algorithm is much smaller than the number of reflection units, the complexity of the CDAML algorithm is greater than that of the GMML algorithm. Although the time complexity of the CDAML algorithm is slightly higher than that of the GMML algorithm, the GMML algorithm has a simple network structure and an input method of expanding the matrix into a one-dimensional vector, and the imaginary part of the complex matrix is ​​horizontally spliced ​​after the real part, so it is difficult to fully extract the nonlinear phase characteristics of the near-field channel, resulting in a system SE lower than that of the CDAML algorithm.

[0116] Embodiment 2

[0117] like Figure 5 As shown, it is a schematic diagram of the internal structure of the beam focusing system based on convolutional dilation autoencoder meta-learning provided in an embodiment of the present application, which specifically includes: a system establishment unit 510, an optimization problem determination unit 520, and an optimization unit 530.

[0118] The system establishing unit 510 is used to establish a system model.

[0119] Before determining the system model, it is necessary to first determine the XL-IRS-assisted multi-user communication system, such as Figure 2 As shown in FIG. 4 , an XL-IRS-assisted multi-user communication system consists of a BS equipped with M antennas, an XL-IRS with N reflection units, and K single-antenna UEs.

[0120] Among them, the uniform linear array (ULA) is used at the BS, and the uniform planar array (UPA) is used at the XL-IRS. A three-dimensional coordinate system is established with the center of the XL-IRS plane as the coordinate origin. 1 and N2 Respectively represent the number of reflection units in the z-axis and y-axis directions, that is, N = N 1 ×N 2 , d is the distance between adjacent reflection units. The data stream of the kth UE is marked as s k , satisfying E{|s k | 2}=1.

[0121] Considering that the direct connection between BS and UE is blocked by buildings, UE can only communicate with BS through XL-IRS. The channel matrix between BS and XL-IRS is G∈C N×M , the channel matrix between XL-IRS and UE is H∈C K×N , then the signal y received by the kth UE k It can be expressed as:

[0122]

[0123] And define is a cascade channel, θ = diag(θ 1 ,θ 2 ,…,θ N ) is represented as the phase shift matrix of XL-IRS, represents the reflection coefficient of the nth reflection unit, where p n ∈[0,2π] represents the phase shift caused by the nth XL-IRS unit, C represents a set of complex numbers, and w k Denotes the precoding matrix W∈C M×K In the kth column at BS, w i Denotes the precoding matrix W∈C M×K In the i-th column at BS, s k represents the data flow of the kth UE, s i represents the data flow of the i-th UE, h k represents the transpose of the kth row of H, represents the complex Gaussian white noise received by the kth UE.

[0124] The Rayleigh distance is proportional to the square of the maximum aperture of the array. For an array with a fixed carrier wavelength and spacing between reflective units, the Rayleigh distance increases accordingly as the number of reflective units increases. Therefore, the near-field area will be expanded, and on this basis, the BS and / or UE are more likely to be located in the near-field area. For XL-IRS-assisted communication systems, as long as either the BS or the UE is located in the near-field area of ​​the XL-IRS, the channel should be accurately modeled using near-field spherical waves. To comply with the typical deployment strategy of XL-IRS, this embodiment considers the scenario where the UE is located in the near-field area of ​​the XL-IRS and the BS is located in the far-field area of ​​the XL-IRS.

[0125] Therefore, the far-field channel from BS to XL-IRS can be modeled as the Saleh-Valenzuela channel model, and considering a line-of-sight (LoS) path and L non-line-of-sight (NLoS) paths, the far-field channel from BS to XL-IRS is expressed as:

[0126]

[0127] Where κ is the Rice fading factor, α l is the complex gain of the lth path, φ in,l and Respectively represent the azimuth and elevation angles of the signal arriving at XL-IRS. l represents the AoD of the lth path at the BS, where l = 0 represents a LoS path and l > 0 represents a NLoS path. a(γ) and are the far-field array steering vectors of BS and XL-IRS, respectively. Under the far-field plane wavefront assumption, a(γ) and They can be expressed as:

[0128]

[0129]

[0130] Where λ is the signal wavelength.

[0131] Since this embodiment considers that the UE is located in the near field area of ​​the XL-IRS, such as Figure 3 As shown, it is the near-field channel model from XL-IRS to the kth user.

[0132] In near-field channel modeling, the array steering vector must consider both angle information and distance information, and the angle and distance from each reflector in the XL-IRS to the UE antenna are different, that is, the phase of each reflector in the XL-IRS needs to be calculated independently. Therefore, the channel from the XL-IRS to the kth UE can be modeled as:

[0133]

[0134] Similarly, l = 0 indicates a LoS path, and l > 0 indicates a NLoS path. l is the complex gain of the lth path, r k,l represents the spatial coordinate vector of the lth path from XL-IRS to UE, φ re,k,l and They represent the azimuth and elevation angles of the signal leaving the XL-IRS, respectively. Represents the XL-IRS near-field array steering vector, which focuses the signal energy on The indicated position is expressed based on the near-field spherical wave model as:

[0135]

[0136] Where r is the distance between the scatterer (or UE) and the center of the XL-IRS array, represents the reflection unit index of each XL-IRS, where In the lth path, the scatterer (or UE) and the reflector (n 1 ,n 2 ) is expressed as:

[0137]

[0138] Where r l represents the distance between the scatterer (or UE) and the center of XL-IRS in the lth path.

[0139] The optimization problem determination unit 520 is used to determine the system optimization problem.

[0140] In formula (1), is the signal that the kth UE expects to receive, and the interference between the kth UE and other UEs is considered as noise. k With unit power, the signal-to-noise ratio of the kth UE can be expressed as:

[0141]

[0142] Using spectrum efficiency SE to evaluate system performance can be quantified as follows:

[0143]

[0144] Where ω k represents the weight of the kth user. By optimizing W and Θ, the system spectrum efficiency SE is maximized. Therefore, the optimization problem of this embodiment can be described as:

[0145]

[0146] Where P t is the maximum total transmit power of BS, C 1 The constraint is that the total transmission power of the BS is limited, C 2 The constraint is the isomodus constraint of the XL-IRS reflection unit, that is, the reflection coefficient of the XL-IRS reflection unit is on the unit circle, which simplifies the hardware implementation of XL-IRS.

[0147] The optimization unit 530 is used to optimize the system according to the determined system optimization problem and output the maximum spectrum efficiency of the system.

[0148] This embodiment uses the CDAML algorithm network to extract and learn nonlinear phase features in the near-field channel, thereby improving the spectrum efficiency SE of the XL-IRS-assisted near-field communication system.

[0149] Specifically, from the far-field channel modeling from BS to XL-IRS and the near-field channel modeling from XL-IRS to UE, it can be known that the main difference between far-field communication and near-field communication modeling lies in the array steering vector, which is not only related to the angle but also to the distance. In order to enable the CDAML algorithm network to better learn the phase information related to the angle and distance in the near-field channel matrix, this embodiment uses the matrix form of the phase shift matrix and the precoding matrix as the input of the CDAML algorithm network. Compared with expanding the matrix into one dimension as the network input, using the matrix form as the input can better extract the relevant features between the various parameters of the near-field channel.

[0150] At the same time, considering that the precoding matrix is ​​a complex matrix, deep learning frameworks such as tensorflow, keras and pytorch do not support complex input. A common solution is to divide the complex matrix into two parts, the real part and the imaginary part, and then horizontally splice the imaginary part to the real part according to the specified dimension. However, this method of processing complex matrices cannot well preserve the relevant information between the real part and the imaginary part, especially the nonlinear phase information in near-field communication. Therefore, in this embodiment, the real part and the imaginary part of the precoding matrix are vertically spliced ​​as the input of the CDAML network, fully retaining the nonlinear phase information about angles and distances in near-field communication. Specifically, the dimension of the input precoding complex matrix is ​​M×K, and the precoding matrix is ​​converted to B×c×M×K by adding an extra dimension. After the real part and the imaginary part are vertically spliced, the dimension is B×c×2×(M×K), where B represents the batch size and c represents the number of channels.

[0151] like Figure 4 As shown in the figure, the specific structure of the CDAML algorithm network is shown in the figure. The CDAML network architecture is mainly composed of CNN. 1 and CNN 2 It consists of two sub-networks.

[0152] CNN 1The structure consists of two layers of dilated convolutional layers (DC) with a convolution kernel of 3×3, a maximum pooling layer (pool), a fully connected layer (FC) and a Lambda layer. A ReLU activation function is added after each dilated convolutional layer (DC) to enable the network to learn and represent nonlinear relationships. Dilated convolution can expand the receptive field without increasing the number of parameters of the convolution kernel, thereby avoiding excessive parameters and calculations, and helping to improve the efficiency and generalization ability of the model. Applying it to the beam focusing optimization of XL-IRS-assisted near-field communication can capture larger-sized input data information and improve the spectral efficiency SE of the XL-IRS-assisted near-field communication system. The maximum pooling layer can reduce the amount of data while retaining the most important information, further reducing the computational complexity of the network. The function of the fully connected layer is to flatten the output of the previous layer into a one-dimensional vector. To ensure that the output reflection unit angle is in the range of 0 to 2π and satisfies the equal modulus constraint C of the XL-IRS reflection unit 2 , the output of the fully connected layer is passed through the sigmoid function to control the output within the range of 0 to 1. Then, it passes through the Lambda layer.

[0153] CNN 1 Taking the phase shift matrix Θ as the network input, the final output of the Lambda layer is expressed as ε=2π is the amplification factor, is the output of the fully connected layer, δ(·) represents the sigmoid function. The reflection coefficient of the final updated reflection unit is expressed as:

[0154]

[0155] In the formula Represents the updated reflection coefficient of the nth reflection unit. Finally, through diagonal operation Get the updated phase shift matrix

[0156] CNN 2 The network structure consists of an autoencoder, two convolutional layers with 2×2 convolution kernels, and a fully connected layer. The autoencoder consists of two convolutional layers with 2×2 convolution kernels and two average pooling layers, and each convolutional layer is connected to an average pooling layer. The autoencoder is responsible for reducing the matrix dimension from B×c×2×(M×K) to B×c×2×K 2dimension, focusing only on the valid signals received by each UE. In this way, the information of the BS antenna domain can be mapped to the UE signal space, retaining the relationship between the signals of each UE, while ignoring the redundant BS antenna space redundancy. Therefore, the computational complexity can be reduced and the processing efficiency can be improved. At the same time, the application of the autoencoder deepens the algorithm network depth, so that more complex near-field channel features can be extracted. Two layers of convolutional layers with a convolution kernel of 2×2 are connected after the autoencoder.

[0157] Specifically, the specific convolution process of the convolution layer is that for the matrix after the dimensionality reduction processing of the autoencoder, it is first filled with zeros, and the matrix dimension changes to B×c×4×(K 2 +2), and then pass through the first layer of convolution with a 2×2 convolution kernel. During the convolution process, the convolution kernel can extract the real and imaginary features separately, as well as the combined features of the real and imaginary parts. Therefore, the CDAML network can better extract the nonlinear phase characteristics of the real and imaginary parts of the complex matrix, and improve the system spectrum efficiency SE of XL-IRS assisted near-field communication. Finally, the second layer of convolution with a 2×2 convolution kernel restores the matrix to B×c×2×K 2 The function of the fully connected layer is to flatten the output into a one-dimensional vector. At the same time, considering the power constraint C 1 , CNN 2 The output of the fully connected layer is a one-dimensional vector of size 2×M×K, and then the real and imaginary parts are concatenated to obtain at last The power-constrained precoding matrix is ​​obtained by calculation using formula (12):

[0158]

[0159] Based on the above, the optimization unit 403 performs the following sub-steps:

[0160] Step T1: Perform CDAML algorithm network configuration and parameter setting.

[0161] As shown in Table 1, it is the network configuration and parameter settings of the CDAML algorithm.

[0162] Table 1

[0163]

[0164] Step T2: Determine input parameters after completing network configuration and parameter setting.

[0165] The input parameters include sample channel matrices G and H, transmit power P t , sub-cycle number I i , Main loop times I o , iterative cycle number I e,CNN 1 Learning rate l Θ ,CNN 2 Learning rate l W .

[0166] Step T3: Initialize the information after determining the input parameters.

[0167] Initializing the information includes:

[0168] Step T31: Randomly initialize θ Θ ,θ W , W (0,1) , Θ (0,1) .

[0169] θ Θ and θ W Respectively represent CNN 1 and CNN 2 The network parameter set, W (0,1) represents the initial precoding matrix of the input, that is, the precoding matrix of the first sub-cycle in the first main cycle, Θ (0,1) Represents the initial phase shift matrix of the input, that is, the phase shift matrix of the 1st sub-cycle in the 1st main cycle.

[0170] Step T32: For W (0,1) Perform normalization.

[0171] Formula 12 is used for normalization.

[0172] Step T33: Initialize the maximum spectrum efficiency SE.

[0173] Initialize the maximum spectrum efficiency SE, denoted as MAX=0.

[0174] Step T4: Initialize the loss function to 0.

[0175] Step T35: Let θ (0,j) =Θ (0,1) , W (0,j) =W (0,1) .

[0176] W (0,j) and θ (0,j) They respectively represent the precoding matrix and phase shift matrix of the 1st sub-cycle in the jth main cycle.

[0177] Step T4: After completing information initialization, enter a three-layer nested loop.

[0178] The three levels of nested loops include sub-loops, main loops, and iteration loops.

[0179] The function of the sub-loop is to be1 and CNN 2 Loop optimization Θ and W.

[0180] Specifically, Θ (i-1,j) ,i=1,2,3,…I i ; j = 1, 2, 3, ... I o Enter CNN 1 subnetwork, and through CNN 1 Subnetwork output Θ (i,j) . (i-1,j) ,i=1,2,3,…I i ; j = 1, 2, 3, ... I o Enter CNN 2 subnetwork, and through CNN 2 Subnetwork output W (i,j) . Use formula 12 to normalize W (i,j) .

[0181] The main loop is used to accumulate the loss function. i Sub-cycle, repeat I i To achieve unsupervised learning, the loss function of the jth main loop is expressed as the negative value of SE:

[0182] L j =-R(W * ,Θ * ; H, G)(13)

[0183] Where W * and θ * They respectively represent the precoding matrix and phase shift matrix after each main cycle update.

[0184] The repetition I i Sub-cycles, Calculate L using formula (13) j , then accumulate the loss function in Represents the accumulated loss function after i main cycles.

[0185] If the accumulated loss function is greater than the maximum spectral efficiency SE(-L j >MAX), then MAX=-L j , output Θ = Θ * , W=W * .

[0186] The iteration loop is mainly responsible for updating CNN 1 and CNN 2 The parameters in the sub-network have I in each iteration cycle o A main loop, for the accumulated loss function Take the average and record it as

[0187]

[0188] Backpropagation using the Adam optimizer to update the CNN 1 and CNN 2 The network parameter set θ Θ and θ W The whole optimization process has I e Iteration loop, the update interval is set to a constant n o , in each n o Update the CNN once in an iterative cycle 1 and CNN 2 The network parameters in .

[0189] Finally, W and Θ are returned, and the spectrum efficiency SE and CPU running time are output.

[0190] The CDAML network not only has the advantages of the CNN network, but can also effectively capture the spatial structure and local correlation of the input data, thereby realizing the extraction and learning of nonlinear phase features in the near-field channel and improving the spectrum efficiency SE of the XL-IRS-assisted near-field communication system. At the same time, it has the advantages of meta-learning and three-layer nested loops, can quickly adapt to new XL-IRS-assisted near-field channels, and has the characteristics of no need for pre-training and unsupervised learning. It iterates the optimal precoding matrix W and phase shift matrix Θ, and finally outputs the maximum spectrum efficiency SE value of the XL-IRS-assisted near-field communication system.

[0191] This embodiment analyzes and compares the time complexity of the AO algorithm, the GMML algorithm, and the proposed CDAML algorithm, as well as the contrast algorithm.

[0192] First, we analyze the CNN in the CDAML algorithm 1 Complexity. CNN 1 The complexity of is mainly composed of calculating SE and embedding CNN network. SE is calculated according to formulas (8) and (9). Specifically, The required complexity is O(MN), and because it is necessary to calculate the signal-to-noise ratio γ of K UEs k ,so The calculation of needs to be repeated K times, and finally the SE calculated according to formula (9) is repeated K times. Therefore, the overall complexity of calculating SE is O(K 2 MN). The complexity of the embedded CNN network mainly comes from the fully connected layer, that is, O(N 2 ). Therefore, CNN 1 The complexity of a single iteration can be expressed as O(K 2MN+N 2 ). Then analyze CNN 2 The complexity of CNN 2 The main components are the calculation formula (12), the calculation of SE, the autoencoder and the embedded CNN network. Therefore, CNN 2 The complexity of a single iteration can be expressed as O(MK 2 +K 2 MN+MK+MK 3 )=O(K 2 MN+MK 3 Finally, considering three layers of nested loop iterations, the overall complexity of the CDAML algorithm is O(I e I o I i (K 2 MN+N 2 +MK 3 )).

[0193] The time complexity of AO algorithm and GMML algorithm is O(L 1 (L 2 KM 3 +L 3 K 2 N 2 )) and O(I e I o I i (K 2 MN+HN+HK 2 )). Therefore, the AO algorithm has the highest time complexity among the three algorithms. At the same time, when the number of reflection units is small, the time complexity of the CDAML algorithm proposed in this embodiment is equivalent to that of the GMML algorithm, and when the number of reflection units is large, the time complexity of the CDAML algorithm proposed in this embodiment is slightly higher than that of the GMML algorithm. This is because when the number of neurons H in the hidden layer of the GMML algorithm is not negligible compared with the number of reflection units, the complexity of the CDAML algorithm is equivalent to that of the GMML algorithm, and when the number of neurons in the hidden layer of the GMML algorithm is much smaller than the number of reflection units, the complexity of the CDAML algorithm is greater than that of the GMML algorithm. Although the time complexity of the CDAML algorithm is slightly higher than that of the GMML algorithm, the GMML algorithm has a simple network structure and an input method of expanding the matrix into a one-dimensional vector, and the imaginary part of the complex matrix is ​​horizontally spliced ​​after the real part, so it is difficult to fully extract the nonlinear phase characteristics of the near-field channel, resulting in a system SE lower than that of the CDAML algorithm.

[0194] This application has the following beneficial effects:

[0195] (1) The CDAML algorithm used in this application utilizes the characteristics of local connection and weight sharing to reduce the weight parameters that need to be learned for XL-IRS-assisted near-field communication, thereby achieving a balance between system SE performance and computational complexity.

[0196] (2) This application expands the network's receptive field by carefully designing the expansion coefficient and padding of the dilated convolution, thereby achieving a global understanding of the entire input as much as possible and improving the system's spectral efficiency. In addition, by designing an autoencoder to reduce the dimension of the precoding matrix before inputting it into the network for optimization, the search space and convergence time are reduced, and the algorithm complexity is reduced.

[0197] Although the present application is described with reference to examples, this is for illustrative purposes only and is not intended to limit the present application, and changes, additions and / or deletions to the embodiments may be made without departing from the scope of the present application.

[0198] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A beam focusing method based on convolutional dilation autoencoder meta-learning, characterized in that: The following steps are involved: Establish system model; After completing the establishment of the system model, determine the system optimization problem; The system is optimized according to the determined system optimization problem and the maximum spectrum efficiency of the system is output.

2. The beam focusing method based on convolutional dilation autoencoder meta-learning as claimed in claim 1, characterized in that: The establishment of the system model includes the following sub-steps: Identify multi-user communication systems; Based on the multi-user communication system, far-field channel modeling and near-field channel modeling are performed.

3. The beam focusing method based on convolutional dilation autoencoder meta-learning as claimed in claim 2, characterized in that: The multi-user communication system includes a base station equipped with M antennas, a very large-scale IRS with N reflection units, and K single-antenna user devices.

4. The beam focusing method based on convolutional dilation autoencoder meta-learning as claimed in claim 3, characterized in that: Determining the multi-user communication system further includes determining the signal y received by the kth user equipment k , expressed as: G∈C N×M represents the channel matrix between the base station and the ultra-large-scale IRS, C represents a set of complex numbers, and w k Denotes the precoding matrix W∈C M×K In the kth column at BS, w i Denotes the precoding matrix W∈C M×K In the i-th column at BS, s k represents the data flow of the kth UE, s i represents the data flow of the i-th UE, H∈C K×N The channel matrix between the ultra-large-scale IRS and the user equipment is represented as, is a cascade channel, θ=diag(θ1,θ2,…,θ N ) is represented as the phase shift matrix of XL-IRS, represents the reflection coefficient of the nth reflection unit, where p n ∈[0,2π] represents the phase shift caused by the nth XL-IRS unit, h k represents the transpose of the kth row of H, represents the complex Gaussian white noise received by the kth UE.

5. The beam focusing method based on convolutional dilation autoencoder meta-learning as claimed in claim 1, characterized in that: Optimizing the system according to the determined system optimization problem and outputting the system maximum spectrum efficiency includes the following sub-steps: Determine network configuration and parameter settings; After completing the network configuration and parameter setting, determine the input parameters; Initialize the information after determining the input parameters; After completing the information initialization, enter the three-layer nested loop.

6. A beam focusing system based on convolutional dilation autoencoder meta-learning, characterized in that: Specifically include: System establishment unit, optimization problem determination unit and optimization unit; A system building unit, used for building a system model; An optimization problem determination unit, used for determining a system optimization problem; The optimization unit is used to optimize the system according to the determined system optimization problem and output the maximum spectrum efficiency of the system.

7. The beam focusing system based on convolutional dilation autoencoder meta-learning as claimed in claim 6, characterized in that: The system building unit establishes the system model including the following sub-steps: Identify multi-user communication systems; Based on the multi-user communication system, far-field channel modeling and near-field channel modeling are performed.

8. The beam focusing system based on convolutional dilation autoencoder meta-learning as claimed in claim 7, characterized in that: The multi-user communication system determined in the system establishment unit includes a base station equipped with M antennas, an ultra-large-scale IRS with N reflection units, and K single-antenna user equipments.

9. The beam focusing system based on convolutional dilation autoencoder meta-learning as claimed in claim 8, characterized in that: Determining the multi-user communication system further includes determining the signal y received by the kth user equipment k , expressed as: G∈C N×M The channel matrix between the base station and the ultra-large-scale IRS is represented by, C represents a complex number set, and w k Denotes the precoding matrix W∈C M×K In the kth column at BS, w i Denotes the precoding matrix W∈C M×K In the i-th column at BS, s k represents the data flow of the kth UE, s i represents the data flow of the i-th UE, H∈C K×N The channel matrix between the ultra-large-scale IRS and the user equipment is represented as, is a cascade channel, θ=diag(θ1,θ2,…,θ N ) is represented as the phase shift matrix of XL-IRS, represents the reflection coefficient of the nth reflection unit, where p n ∈[0,2π] represents the phase shift caused by the nth XL-IRS unit, h k represents the transpose of the kth row of H, represents the complex Gaussian white noise received by the kth UE.

10. The beam focusing system based on convolution dilation autoencoder meta-learning as claimed in claim 6, characterized in that: The optimization unit optimizes the system according to the determined system optimization problem, and outputs the maximum spectrum efficiency of the system, including the following sub-steps: Determine network configuration and parameter settings; After completing the network configuration and parameter setting, determine the input parameters; Initialize the information after determining the input parameters; After completing the information initialization, enter the three-layer nested loop.

Images