Near-field channel estimation method and device for IRS-assisted multi-user MIMO system based on DRN

Abstract

The invention discloses a near-field channel estimation method and device for an IRS-assisted multi-user MIMO system based on a DRN. The method specifically comprises the following steps: constructing an IRS-assisted multi-user MIMO near-field communication system model; establishing near-field channel models from the base station to the IRS, from the IRS to the user and from the base station to the user; based on the near-field channel model, deriving mean square error expressions of a least square (LS) algorithm and a minimum mean square error (MMSE) algorithm, and calculating a Cramer-Rao bound (CRLB) of the received signal; and on the basis of the deep residual network and federated learning, constructing an FL-DRN-NFCE network, namely, a federated learning deep residual network near-field channel estimation network, and performing IRS-assisted channel estimation of the near-field multi-user MIMO system. According to the invention, complex modes and subtle differences in the signals can be captured, the precision of near-field channel estimation is improved, the pilot frequency overhead is reduced, the calculation complexity is reduced, and the efficiency of channel estimation is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and channel estimation, and in particular to a near-field (NF) channel estimation (CE) method for an IRS (Intelligent reflecting surface) assisted multi-user MIMO (Multiple-input multiple-output) system based on a DRN (Deep residual network). Background Art

[0002] Channel estimation plays a vital role in the optimization process of communication systems. Its accuracy is directly related to the effectiveness of subsequent optimization strategies, and accurate channel state information is the cornerstone of this process. However, in communication networks assisted by intelligent reflecting surfaces (IRS), the lack of signal processing capabilities of the IRS itself greatly increases the difficulty of channel estimation, making it a key issue that needs to be solved in this field. Although traditional channel estimation methods can meet the needs to a certain extent, they generally have limitations such as high pilot overhead and insufficient estimation accuracy. In order to overcome these challenges, channel estimation methods based on deep learning have gradually emerged in recent years and have become the focus of attention in academia and industry.

[0003] The Deep Residual Network (DRN) is a deep learning neural network architecture that solves the gradient vanishing and gradient exploding problems in the deep network training process by introducing residual connections, making it easier to train and optimize deeper networks. Federated learning (FL) is a distributed machine learning method that allows multiple participants to jointly train a shared machine learning model while maintaining data privacy and localization, reducing the risk of data leakage. Its core advantage is that it does not share original data, uses decentralized data to train models, protects data privacy, and improves data utilization efficiency. It is suitable for a variety of scenarios.

[0004] In this context, DRN is introduced into the channel estimation of IRS-assisted multi-user multiple-input multiple-output (MIMO) near-field communication system. By utilizing the powerful capabilities of DRN in areas such as image denoising, accurate estimation of the near-field channel state is achieved. It can show a channel estimation performance that is superior to traditional channel estimation methods while significantly reducing pilot overhead. Summary of the invention

[0005] The object of the present invention is to provide a near-field channel estimation method and device for an IRS-assisted multi-user MIMO system based on a DRN with low pilot overhead and high estimation accuracy.

[0006] The technical solution to achieve the purpose of the present invention is: a near-field channel estimation method for an IRS-assisted multi-user MIMO system based on DRN, comprising the following steps:

[0007] Step 1: Construct a system model of IRS-assisted multi-user MIMO near-field communication;

[0008] Step 2: Establish near-field channel models from base station to IRS, IRS to user, and base station to user;

[0009] Step 3: Based on the near-field channel models from the base station to the IRS, from the IRS to the user, and from the base station to the user, the mean square error expressions of the least squares LS algorithm and the minimum mean square error MMSE algorithm are derived, and the Cramer-Rao lower bound CRLB of the received signal is calculated;

[0010] Step 4: Based on the deep residual network and federated learning, a FL-DRN-NFCE network, i.e., a federated learning deep residual network near-field channel estimation network, is constructed to perform channel estimation of the IRS-assisted near-field multi-user MIMO system.

[0011] A near-field channel estimation device for an IRS-assisted multi-user MIMO system based on a DRN, the device is used to implement the near-field channel estimation method for an IRS-assisted multi-user MIMO system based on a DRN, the device comprises a system model building module, a near-field channel model building module, a CRLB calculation module and a channel estimation module, wherein:

[0012] System model building module, used to build the system model of IRS-assisted multi-user MIMO near-field communication;

[0013] A near-field channel model building module is used to build near-field channel models from base station to IRS, from IRS to user, and from base station to user;

[0014] The CRLB calculation module derives the mean square error expressions of the least squares LS algorithm and the minimum mean square error MMSE algorithm based on the near-field channel models from the base station to the IRS, the IRS to the user, and the base station to the user, and calculates the Cramer-Rao lower bound CRLB of the received signal;

[0015] The channel estimation module, based on deep residual network and federated learning, constructs the FL-DRN-NFCE network, that is, the federated learning deep residual network near-field channel estimation network, to perform channel estimation of IRS-assisted near-field multi-user MIMO system.

[0016] A mobile terminal comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the near-field channel estimation method of the IRS-assisted multi-user MIMO system based on DRN is implemented.

[0017] Compared with the prior art, the present invention has the following significant advantages:

[0018] (1) It not only focuses on the far-field channel model, but also considers the channel estimation problem of the near-field channel model, and simultaneously considers the direct channel and the cascade channel, the line-of-sight component and the non-line-of-sight component, thus improving the network's discrimination ability and enabling it to better distinguish similar signals;

[0019] (2) It can capture complex patterns and subtle differences in signals in areas where signals may overlap or exhibit similar characteristics, improving channel estimation accuracy;

[0020] (3) It reduces pilot overhead, reduces computational complexity, lightens computational load, and improves channel estimation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The present invention is a flowchart of a near-field channel estimation method for an IRS-assisted multi-user MIMO system based on DRN.

[0022] Figure 2 Schematic diagram of a system model for IRS-assisted multi-user MIMO near-field communication in an embodiment of the present invention.

[0023] Figure 3 Schematic diagram of the structure of the federated learning deep residual network near-field channel estimation network designed in an embodiment of the present invention.

[0024] Figure 4 The figure is a schematic diagram of the accuracy of the region classifier designed in the embodiment of the present invention under different SNRs.

[0025] Figure 5 This is a schematic diagram comparing the standardized mean square error NMSE of the federated learning deep residual network near-field channel estimation network designed in an embodiment of the present invention with the least squares LS algorithm and the minimum mean square error MMSE algorithm under different SNRs. DETAILED DESCRIPTION

[0026] The present invention discloses a near-field (NF) channel estimation (CE) method for a multi-user multiple-input multiple-output (MIMO) system assisted by an intelligent reflecting surface (IRS) based on a deep residual network (DRN). Firstly, a system model of multi-user MIMO near-field communication assisted by IRS is constructed; then, different from the far-field channel model, a NF channel model from base station (BS) to IRS, IRS to user, and BS to user is established; further, based on the NF channel model, the mean square error expression of traditional CE algorithms, such as the least square (LS) and minimum mean square error (MMSE) algorithms, is derived, and a Cramer-Rao lower bound is provided; finally, based on DRN and federated learning (FL), a CE framework suitable for IRS-assisted NF multi-user MIMO system, namely, FL-DRN-NFCE network, is proposed. The simulation analysis results show that the FL-DRN-NFCE network designed by the present invention can achieve better channel estimation accuracy than the traditional LS and MMSE algorithms while saving five-sixths of the pilot overhead. In addition, the designed regional classifier can achieve a classification accuracy of about 95% in the area with 5000 samples and high signal-to-noise ratio.

[0027] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0028] like Figure 1 As shown, the present invention provides a near-field channel estimation method for an IRS-assisted multi-user MIMO system based on DRN, comprising the following steps:

[0029] Step 1: Construct a system model of IRS-assisted multi-user MIMO near-field communication;

[0030] Step 2: Establish near-field channel models from base station to IRS, IRS to user, and base station to user;

[0031] Step 3: Based on the near-field channel models from the base station to the IRS, from the IRS to the user, and from the base station to the user, the mean square error expressions of the least squares LS algorithm and the minimum mean square error MMSE algorithm are derived, and the Cramer-Rao lower bound CRLB of the received signal is calculated;

[0032] Step 4: Based on the deep residual network and federated learning, a FL-DRN-NFCE network, i.e., a federated learning deep residual network near-field channel estimation network, is constructed to perform channel estimation of the IRS-assisted near-field multi-user MIMO system.

[0033] As a specific example, in step 1, a system model of IRS-assisted multi-user MIMO near-field communication is constructed, such as Figure 2 As shown, the details are as follows:

[0034] Step 1.1, deploy IRS between the base station and multiple single-antenna users to enhance communication, where the base station is equipped with an antenna array of M elements, and the IRS consists of N reflective elements, both of which are uniform planar arrays (UPA); since a single user is usually limited to information in a specific channel scenario, the neural network model trained based on a data set collected by a single user may fail when the user spans different areas. In order to overcome this limitation and ensure that the neural network model performs well under a wide range of channel conditions, the entire system is divided into T independent areas, where t = 1, 2, ..., T, and the number of users in the tth area is defined as K t ; Users distributed in the same area show highly similar channel characteristics, while users distributed in different areas show different channel characteristics;

[0035] Step 1.2: Set G and f t,k 、h t,k They represent the channels from the base station to the IRS, from the IRS to the kth user in the tth area, and from the base station to the kth user in the tth area, respectively, where k = 1, 2, ..., K t , then the received signal at the kth user in the tth area is:

[0036] y t,k =h t,k vx+f t,k ΘGvx+w t,k (1)

[0037] Where x is the transmitted signal at the base station, v is the beamforming vector at the base station, is the reflection coefficient matrix of IRS, diag{·} represents the diagonalization operation, is a 1×N row vector, w t,k is the additive Gaussian white noise at the kth user in the tth region, which has a mean of 0 and a variance of The complex Gaussian distribution of

[0038] because so Therefore, the core task of channel estimation is to accurately estimate the following expression:

[0039]

[0040] Formula (1) can be re-expressed as:

[0041] y t,k =θH t,k vx+w t,k (3)

[0042] in

[0043] As a specific example, in step 2, near-field channel models from the base station to the IRS, from the IRS to the user, and from the base station to the user are established as follows:

[0044] Step 2.1: Establish a near-field channel model from the base station to the kth user in the tth area:

[0045] Assume that the M antennas in the UPA of the base station are arranged in M xb ×M zb array, that is, M = M xb ×M zb ,in and And the antenna spacing of the UPA of the base station is set to d in two dimensions. xb and d zb ; Set θ b and φ b They represent the azimuth and elevation of the user relative to the base station in the xz plane, and the distance between the user and the UPA center of the base station is r b , then the coordinates of the kth user in the tth area are expressed as:

[0046] u b =(r b cosθ b sinφ b ,r b sinθ b sinφ b ,r b cosφ b ) (4)

[0047] The coordinates of the (m,n)th antenna in the UPA of the base station are:

[0048] b m,n =(nd xb ,0,md zb ) (5)

[0049] in

[0050] Thus, the distance between the (m,n)th antenna in the UPA of the base station and the kth user in the tth area is obtained as:

[0051]

[0052] The following second-order Taylor expansion is applied to equation (6):

[0053]

[0054] Finally, formula (6) can be approximated as:

[0055]

[0056] where h1 and h2 depend only on n and m, respectively;

[0057] Therefore, the near-field line-of-sight multiple-input single-output MISO channel can be modeled as:

[0058]

[0059] Among them, λ represents the signal wavelength, β b Indicates h t,k The channel gain of

[0060] By removing the constant phase The final phase of the array response vector, i.e., h1 and h2, can be obtained, so the array response vector of the channel between the base station and the user is:

[0061]

[0062] in represents the Kronecker product, [a x (θ b ,φ b ,r b )] n and [a z (φ b ,r b )] m They are:

[0063]

[0064] Since scatterers in the near-field environment can cause signal refraction or reflection, the signal reaches the user through a non-line-of-sight path. In this case, the channel from the base station to the scatterer can be analogized to a multiple-input single-output MISO channel model. Therefore, the channel h from the base station to the kth user in the tth area is t,k It is expressed as:

[0065]

[0066] Among them, L b represents the number of scatterers between the base station and the user, l = 1, 2, ..., L b ; They represent the azimuth, elevation, distance and channel gain of the lth scatterer between the base station and the user respectively;

[0067] Step 2.2: Establish the near-field channel model from IRS to the kth user in the tth region:

[0068] The UPA of the IRS is placed in the xz plane, parallel to the UPA of the base station. For the kth user in the tth area, its coordinate relative to the UPA plane of the IRS is:

[0069] u a =(r a cosθ a sinφ a ,r a sinθ a sinφ a ,r a cosφ a ) (14)

[0070] where θ a , r a They represent the azimuth, elevation and distance of the user relative to the xz plane of the IRS, respectively;

[0071] The coordinates of the (p,q)th element in the UPA of IRS are:

[0072]

[0073] Where N = N xa ×N za , d xa and d za are the element spacings along the x and z axes in the UPA of IRS, respectively;

[0074] The distance between the kth user in the tth region and the (p,q)th element of the IRS is obtained as:

[0075]

[0076] Therefore, the array response vector of the channel between the IRS and the user is:

[0077]

[0078] in

[0079]

[0080] Therefore, the channel between the kth user in the tth region and the IRS is expressed as:

[0081]

[0082] where β a represents f t,k The channel gain, L a represents the number of scatterers between the IRS and the user, denote the azimuth, elevation, distance and channel gain of the lth scatterer between the IRS and the user respectively;

[0083] Step 2.3: Establish the near-field channel model from the base station to the IRS:

[0084] The near-field channel between the base station and the IRS is expressed as:

[0085]

[0086] Where L ba represents the number of scatterers between the base station and the IRS, denote the azimuth, elevation, distance and channel gain of the lth scatterer between the base station and the IRS relative to the base station plane, respectively. represent the azimuth, elevation, distance and channel gain of the lth scatterer between the base station and the IRS relative to the IRS plane, respectively, and for the near-field line-of-sight component G LoS ,have:

[0087]

[0088] in Represents G LoS The channel gain, θ ba , r ba Respectively represent the azimuth, elevation and distance of the center of the base station relative to the center of the IRS, θ ab , r ab They represent the azimuth, elevation and distance of the center of the IRS relative to the center of the base station, respectively.

[0089] As a specific example, in step 3, based on the near-field channel models from the base station to the IRS, from the IRS to the user, and from the base station to the user, the mean square error expressions of the least squares LS algorithm and the minimum mean square error MMSE algorithm are derived, and the Cramer-Rao lower bound CRLB of the received signal is calculated, as follows:

[0090] Step 3.1, derive the mean square error expression of the least squares LS algorithm and the minimum mean square error MMSE algorithm:

[0091] In order to estimate the channel, the base station sends a preset pilot signal p to the user with the assistance of IRS. q , where q = 1, 2, ..., Q, the received pilot signal at the user is:

[0092]

[0093] where θ q is the phase shift of the IRS in the current time slot, w t,k,q is the noise at the user in the current time slot;

[0094] After Q consecutive time slots, the continuous received pilot signal is:

[0095]

[0096] in w t,k =[w t,k,1 ,w t,k,2 ,…,w t,k,Q ] T , (·) T Represents a transpose operation;

[0097] For subsequent channel estimation, equation (26) can be rewritten as:

[0098]

[0099] Through vectorization operation, equation (27) can be re-expressed as:

[0100]

[0101] Where vec(·) represents a vectorized operation;

[0102] Therefore, the least squares LS estimator is:

[0103]

[0104] in represents a pseudo-inverse operation;

[0105] The minimum mean square error MMSE estimator of the received signal is:

[0106]

[0107] in

[0108]

[0109] in represents the cross-correlation matrix between the real channel and the least squares LS estimated channel, represents the autocorrelation matrix of the least squares LS estimated channel, represents the noise variance, represents the signal variance;

[0110] Step 3.2, calculate the Cramer-Rao lower bound CRLB:

[0111] CRLB is used to calculate the best estimation accuracy that can be achieved theoretically, which can evaluate the effectiveness of the proposed channel estimation algorithm;

[0112] For the channel estimation problem (28), it can be divided into two parts, where the real part can be expressed as:

[0113]

[0114] The imaginary part can be expressed as:

[0115]

[0116] in

[0117]

[0118] in represents the real part operation, Indicates the operation of taking the imaginary part;

[0119] Therefore, CRLB is also divided into two parts, as follows:

[0120]

[0121] in Indicates the expected operation;

[0122] Because w t,k,u The mean is 0 and the variance is The complex Gaussian distribution of The conditional probability density function is:

[0123]

[0124] The Fisher information matrix of formula (32) is:

[0125]

[0126] in and Respectively The i-th and j-th elements of ;

[0127] Therefore, the CRLB representation of the real part is:

[0128]

[0129] because so It is expressed as:

[0130]

[0131] because where tr(·) represents the trace operation, so, can be calculated as:

[0132]

[0133] set up is the matrix vv H The M eigenvalues ​​of is the matrix Φ H The N+1 eigenvalues ​​of Φ give:

[0134]

[0135] Further we can get:

[0136]

[0137] therefore, The CRLB representation of the real part of is:

[0138]

[0139] Since the real and imaginary parts are of the same form, the CRLB representation of the imaginary part is:

[0140]

[0141] In summary, the CRLB of the received signal formula (28) is:

[0142]

[0143] As a specific example, in step 4, based on the deep residual network and federated learning, a FL-DRN-NFCE network, i.e., a federated learning deep residual network near-field channel estimation network, is constructed to perform channel estimation of an IRS-assisted near-field multi-user MIMO system, such as Figure 3 As shown, the details are as follows:

[0144] Step 4.1. Design a region classifier, including four convolutional layers and one linear layer. The number of channels in the first three layers gradually increases from the first layer to the third layer, as follows:

[0145] The input of the regional classifier network is the received pilot signal, and the initial number of channels is 2, which effectively reduces the parameters and computational complexity of the subsequent linear layer;

[0146] The first convolutional layer increases the number of channels to 32, enabling the network to capture a wider range of features;

[0147] The second convolutional layer increases the number of channels to 64, enhancing the network's ability to represent complex signal patterns;

[0148] The third convolutional layer increases the number of channels to 128, improving the network’s ability to discriminate between similar signals, especially in areas where signals may overlap or exhibit similar characteristics, making it easier to capture complex patterns and subtle differences in signals.

[0149] The fourth convolutional layer reduces the dimension of the feature map received from the previous layer and reduces the number of channels to 2, which effectively reduces the parameters and computational complexity of the subsequent linear layers;

[0150] The four convolutional layer networks can effectively extract signal features while reducing the computational load by reducing dimensionality;

[0151] The linear layer maps the signal to the category label, achieving accurate and efficient region classification of the signal;

[0152] Step 4.2: Design a federated learning deep residual network near-field channel estimation network, including two convolutional layers, one average pooling layer, one linear layer, and three residual blocks, as follows:

[0153] The first layer is a convolutional layer, which expands the height and width dimensions of the input signal from Q×1 to (Q+2)×(1+2) to meet the feature extraction requirements of the subsequent convolutional layer and the dimensionality reduction requirements of the average pooling layer;

[0154] The second to fourth layers consist of three consecutive residual blocks. The first residual block uses a 1×1 convolution kernel to reduce the dimension of the feature map, the second residual block uses a standard convolution layer for feature extraction, and the third residual block uses another 1×1 convolution kernel to expand the number of channels in the feature map. This design strategy has significant advantages in reducing computational load and improving model efficiency. In order to maintain dimensional consistency in the residual connection, an additional 1×1 convolution layer is used for dimensional adjustment to maintain dimensional consistency in the residual connection. In the second residual block, the dimension of the input feature map is (Q+2)×(1+2)×64, and the size of the output feature map is (Q+2)×(1+2)×128. The number of channels is first reduced to 32, and the number of parameters used for feature extraction in the intermediate convolution layer is halved, thereby reducing the computational complexity. The final 1×1 convolution layer only changes the number of channels, reducing the computational complexity.

[0155] The fifth layer is a 1×1 convolutional layer, which is used to further reduce the number of channels in the input feature map to 2;

[0156] The sixth layer is the average pooling layer, which is used to reduce the height and width of the input feature map from (Q+2)×(1+2) to

[0157] The fifth and sixth layers are used to reduce the input dimension of the seventh linear layer, thereby reducing the parameters and computational complexity of the linear layer;

[0158] The seventh layer is a linear layer, which is used to output the prediction signal, and the output dimension is (N+1)×M×2.

[0159] The present invention also provides a near-field channel estimation device for an IRS-assisted multi-user MIMO system based on DRN, which is used to implement the near-field channel estimation method for an IRS-assisted multi-user MIMO system based on DRN. The device includes a system model construction module, a near-field channel model establishment module, a CRLB calculation module and a channel estimation module, wherein:

[0160] System model building module, used to build the system model of IRS-assisted multi-user MIMO near-field communication;

[0161] A near-field channel model building module is used to build near-field channel models from base station to IRS, from IRS to user, and from base station to user;

[0162] The CRLB calculation module derives the mean square error expressions of the least squares LS algorithm and the minimum mean square error MMSE algorithm based on the near-field channel models from the base station to the IRS, the IRS to the user, and the base station to the user, and calculates the Cramer-Rao lower bound CRLB of the received signal;

[0163] The channel estimation module, based on deep residual network and federated learning, constructs the FL-DRN-NFCE network, that is, the federated learning deep residual network near-field channel estimation network, to perform channel estimation of IRS-assisted near-field multi-user MIMO system.

[0164] The present invention also provides a mobile terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the near-field channel estimation method of the DRN-based IRS-assisted multi-user MIMO system is implemented.

[0165] The present invention is further described in detail below with reference to specific embodiments.

[0166] Example

[0167] This embodiment verifies the accuracy of the regional classifier and the estimation performance of the federated learning deep residual network near-field channel estimation network through simulation embodiments.

[0168] Figure 4 The simulation curves of the accuracy of the designed regional classifier under different SNRs are shown. As can be seen from the figure, as the number of samples increases, the classification accuracy also increases. However, when the number of samples is greater than or equal to 4000, further increasing the number of samples will not significantly improve the accuracy. In the high signal-to-noise ratio area and when the number of samples is 5000, the accuracy of the designed regional classifier network can reach about 95%.

[0169] Figure 5 The comparison results of the standardized mean square error NMSE of the designed federated learning deep residual network near-field channel estimation network with the least squares LS algorithm and the minimum mean square error MMSE algorithm at different SNRs are shown. As can be seen from the figure, the designed federated learning deep residual network near-field channel estimation network can achieve better estimation accuracy than the least squares LS algorithm and the minimum mean square error MMSE algorithm. In addition, with the advantages of federated learning, the designed global neural network is better than the single-region neural network. Finally, the effectiveness of the designed federated learning deep residual network near-field channel estimation network is verified by CRLB.

[0170] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A near-field channel estimation method for an IRS-assisted multi-user MIMO system based on DRN, characterized in that: The following steps are involved: Step 1: Construct a system model of IRS-assisted multi-user MIMO near-field communication; Step 2: Establish near-field channel models from base station to IRS, IRS to user, and base station to user; Step 3: Based on the near-field channel models from the base station to the IRS, from the IRS to the user, and from the base station to the user, the mean square error expressions of the least squares LS algorithm and the minimum mean square error MMSE algorithm are derived, and the Cramer-Rao lower bound CRLB of the received signal is calculated; Step 4: Based on the deep residual network and federated learning, a FL-DRN-NFCE network, i.e., a federated learning deep residual network near-field channel estimation network, is constructed to perform channel estimation of the IRS-assisted near-field multi-user MIMO system.

2. The near-field channel estimation method for an IRS-assisted multi-user MIMO system based on DRN according to claim 1, characterized in that: The system model for building IRS-assisted multi-user MIMO near-field communication described in step 1 is as follows: Step 1.1, deploy IRS between the base station and multiple single-antenna users to enhance communication, where the base station is equipped with an antenna array of M elements, and the IRS consists of N reflective elements, both of which are uniform planar arrays (UPA); divide the entire system into T independent regions, where t = 1, 2, …, T, and define the number of users in the tth region as K t ; Step 1.2: Set G and f t,k 、h t,k They represent the channels from the base station to the IRS, from the IRS to the kth user in the tth area, and from the base station to the kth user in the tth area, respectively, where k = 1, 2, ..., K t , then the received signal at the kth user in the tth area is: y t,k =h t,k vx+f t,k ΘGvx+w t,k (1) Where x is the transmitted signal at the base station, v is the beamforming vector at the base station, is the reflection coefficient matrix of IRS, diag{·} represents the diagonalization operation, is a 1×N row vector; w t,k is the additive Gaussian white noise at the kth user in the tth region, with a mean of 0 and a variance of The complex Gaussian distribution of because so Therefore, the core task of channel estimation is to estimate the following expression: Formula (1) is reformulated as: y t,k =θH t,k vx+w t,k (3) in 3. The near-field channel estimation method for the IRS-assisted multi-user MIMO system based on DRN according to claim 2, characterized in that: In step 2, a near-field channel model from the base station to the user is established, as follows: Step 2.1: Establish a near-field channel model from the base station to the kth user in the tth area: Assume that the M antennas in the UPA of the base station are arranged in M xb ×M zb array, that is, M = M xb ×M zb ,in and And the antenna spacing of the UPA of the base station is set to d in two dimensions. xb and d zb ; Set θ b and φ b They represent the azimuth and elevation of the user relative to the base station in the xz plane, and the distance between the user and the UPA center of the base station is r b , then the coordinates of the kth user in the tth area are expressed as: u b =(r b cosθ b sinφ b ,r b sinθ b sinφ b ,r b cosφ b ) (4) The coordinates of the (m,n)th antenna in the UPA of the base station are: b m,n =(nd xb ,0,md zb ) (5) in Thus, the distance between the (m,n)th antenna in the UPA of the base station and the kth user in the tth area is obtained as: The following second-order Taylor expansion is applied to equation (6): Finally, formula (6) is approximated as: where h1 and h2 depend only on n and m, respectively; Therefore, the near-field line-of-sight multiple-input single-output MISO channel is modeled as: Among them, λ represents the signal wavelength, β b Indicates h t,k The channel gain of By removing the constant phase The final phase of the array response vector is obtained, namely h1 and h2, so the array response vector of the channel between the base station and the user is: in represents the Kronecker product, [a x (θ b ,φ b ,r b )] n and [a z (φ b ,r b )] m They are: Since scatterers in the near-field environment can cause signal refraction or reflection, the signal reaches the user through a non-line-of-sight path. In this case, the channel from the base station to the scatterer is analogous to the multiple-input single-output MISO channel model. Therefore, the channel h from the base station to the kth user in the tth area is t,k It is expressed as: Among them, L b represents the number of scatterers between the base station and the user, l = 1, 2, ..., L b ; They represent the azimuth, elevation, distance and channel gain of the lth scatterer between the base station and the user respectively.

4. The near-field channel estimation method for the IRS-assisted multi-user MIMO system based on DRN according to claim 3, characterized in that: In step 2, a near-field channel model from IRS to user is established as follows: Step 2.2: Establish the near-field channel model from IRS to the kth user in the tth region: The UPA of the IRS is placed in the xz plane, parallel to the UPA of the base station. For the kth user in the tth area, its coordinate relative to the UPA plane of the IRS is: u a =(r a cosθ a sinφ a ,r a sinθ a sinφ a ,r a cosφ a ) (14) where θ a , r a They represent the azimuth, elevation and distance of the user relative to the xz plane of the IRS, respectively; The coordinates of the (p,q)th element in the UPA of IRS are: Where N = N xa ×N za , d xa and d za are the element spacings along the x and z axes in the UPA of IRS, respectively; The distance between the kth user in the tth region and the (p,q)th element of the IRS is obtained as: Therefore, the array response vector of the channel between the IRS and the user is: in Therefore, the channel between the kth user in the tth region and the IRS is expressed as: where β a represents f t,k The channel gain, L a represents the number of scatterers between the IRS and the user, They represent the azimuth, elevation, distance and channel gain of the lth scatterer between the IRS and the user respectively.

5. The near-field channel estimation method for the IRS-assisted multi-user MIMO system based on DRN according to claim 4, characterized in that: In step 2, a near-field channel model from the base station to the IRS is established as follows: Step 2.3: Establish the near-field channel model from the base station to the IRS: The near-field channel between the base station and the IRS is expressed as: Where L ba represents the number of scatterers between the base station and the IRS, denote the azimuth, elevation, distance and channel gain of the lth scatterer between the base station and the IRS relative to the base station plane, respectively. represent the azimuth, elevation, distance and channel gain of the lth scatterer between the base station and the IRS relative to the IRS plane, respectively, and for the near-field line-of-sight component G LoS ,have: in Represents G LoS The channel gain, θ ba , r ba Respectively represent the azimuth, elevation and distance of the center of the base station relative to the center of the IRS, θ ab , r ab They represent the azimuth, elevation and distance of the center of the IRS relative to the center of the base station, respectively.

6. The near-field channel estimation method for the IRS-assisted multi-user MIMO system based on DRN according to claim 5, characterized in that: In step 3, based on the near-field channel models from the base station to the IRS, from the IRS to the user, and from the base station to the user, the mean square error expressions of the least squares LS algorithm and the minimum mean square error MMSE algorithm are derived as follows: Step 3.1, derive the mean square error expression of the least squares LS algorithm and the minimum mean square error MMSE algorithm: In order to estimate the channel, the base station sends a preset pilot signal p to the user with the assistance of IRS. q , where q = 1, 2, ..., Q, the received pilot signal at the user is: where θ q is the phase shift of the IRS in the current time slot, w t,k,q is the noise at the user in the current time slot; After Q consecutive time slots, the continuous received pilot signal is: in w t,k =[w t,k,1 ,w t,k,2 ,…,w t,k,Q ] T , (·) T Represents a transpose operation; Formula (26) is rewritten as: Through vectorization operation, equation (27) is re-expressed as: Where vec(·) represents a vectorized operation; Therefore, the least squares LS estimator is: in represents a pseudo-inverse operation; The minimum mean square error MMSE estimator of the received signal is: in in represents the cross-correlation matrix between the real channel and the least squares LS estimated channel, represents the autocorrelation matrix of the least squares LS estimated channel, represents the noise variance, represents the signal variance.

7. The near-field channel estimation method for the IRS-assisted multi-user MIMO system based on DRN according to claim 6, characterized in that: In step 3, the Cramer-Rao lower bound CRLB of the received signal is calculated as follows: Step 3.2, calculate the Cramer-Rao lower bound CRLB: For the channel estimation problem (28), it is divided into two parts: the real part and the imaginary part, where the real part is expressed as: The imaginary part is expressed as: in in represents the real part operation, Indicates the operation of taking the imaginary part; Therefore, CRLB is also divided into two parts, as follows: in Indicates the expected operation; Because w t,k,u The mean is 0 and the variance is The complex Gaussian distribution of The conditional probability density function is: The Fisher information matrix of formula (32) is: where h′ t,k,u,i and h′ t,k,u,j Respectively represent h′ t,k,u The i-th and j-th elements of ; Therefore, the CRLB representation of the real part is: because so It is expressed as: because and tr(A T )=tr(A), where tr(·) represents the trace operation, so is calculated as: set up is the matrix vv H The M eigenvalues ​​of is the matrix Φ H The N+1 eigenvalues ​​of Φ give: Further: therefore, The CRLB representation of the real part of is: Since the real and imaginary parts are of the same form, the CRLB representation of the imaginary part is: In summary, the CRLB of the received signal formula (28) is:

8. The near-field channel estimation method for the IRS-assisted multi-user MIMO system based on DRN according to claim 7, characterized in that: In step 4, based on the deep residual network and federated learning, the FL-DRN-NFCE network, i.e. the federated learning deep residual network near-field channel estimation network, is constructed to perform channel estimation of the IRS-assisted near-field multi-user MIMO system, as follows: Step 4.

1. Design a region classifier, including four convolutional layers and one linear layer. The number of channels in the first three layers gradually increases from the first layer to the third layer, as follows: The input of the regional classifier network is the received pilot signal, and the initial number of channels is 2; The first convolutional layer increases the number of channels to 32; The second convolutional layer increases the number of channels to 64. The third convolutional layer increases the number of channels to 128; The fourth convolutional layer reduces the dimension of the feature map received from the previous layer and reduces the number of channels to 2; The linear layer maps the signal to the category label to achieve regional classification of the signal; Step 4.2: Design a federated learning deep residual network near-field channel estimation network, including two convolutional layers, one average pooling layer, one linear layer, and three residual blocks, as follows: The first layer is a convolutional layer, which expands the height and width dimensions of the input signal from Q×1 to (Q+2)×(1+2) to meet the feature extraction requirements of the subsequent convolutional layer and the dimensionality reduction requirements of the average pooling layer; The second to fourth layers consist of three consecutive residual blocks. The first residual block uses a 1×1 convolution kernel to reduce the dimension of the feature map. The second residual block uses a standard convolution layer for feature extraction. The third residual block uses another 1×1 convolution kernel to expand the number of channels in the feature map and uses an additional 1×1 convolution layer for dimensionality adjustment to maintain dimensional consistency in the residual connection. In the second residual block, the dimension of the input feature map is (Q+2)×(1+2)×64, and the size of the output feature map is (Q+2)×(1+2)×128. The number of channels is first reduced to 32, halving the number of parameters used for feature extraction in the intermediate convolution layer. The final 1×1 convolution layer only changes the number of channels. The fifth layer is a 1×1 convolutional layer, which is used to further reduce the number of channels in the input feature map to 2; The sixth layer is the average pooling layer, which is used to reduce the height and width of the input feature map from (Q+2)×(1+2) to The fifth and sixth layers are used to reduce the input dimension of the seventh linear layer; The seventh layer is a linear layer, which is used to output the prediction signal, and the output dimension is (N+1)×M×2.

9. A near-field channel estimation device for an IRS-assisted multi-user MIMO system based on DRN, characterized in that: The device is used to implement the near-field channel estimation method of the IRS-assisted multi-user MIMO system based on DRN according to any one of claims 1 to 8, and the device includes a system model construction module, a near-field channel model establishment module, a CRLB calculation module and a channel estimation module, wherein: System model building module, used to build the system model of IRS-assisted multi-user MIMO near-field communication; A near-field channel model building module is used to build near-field channel models from base station to IRS, from IRS to user, and from base station to user; The CRLB calculation module derives the mean square error expressions of the least squares LS algorithm and the minimum mean square error MMSE algorithm based on the near-field channel models from the base station to the IRS, the IRS to the user, and the base station to the user, and calculates the Cramer-Rao lower bound CRLB of the received signal; The channel estimation module, based on deep residual network and federated learning, constructs the FL-DRN-NFCE network, that is, the federated learning deep residual network near-field channel estimation network, to perform channel estimation of IRS-assisted near-field multi-user MIMO system.

10. A mobile terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the near-field channel estimation method for the DRN-based IRS-assisted multi-user MIMO system is implemented as described in any one of claims 1 to 8.

Images