A deep learning-based channel estimation method for RIS-aided communication systems

Abstract

This invention relates to the field of wireless communication technology, and particularly to a channel estimation method for a RIS-assisted communication system based on deep learning. The method includes: constructing a channel estimation model for the RIS-assisted communication system, the model employing an AMRN network; activating some RIS elements, where L is less than N; the user equipment transmitting channel information and forwarding it to the base station through the activated RIS elements; and performing preliminary channel estimation on the channel information received by the base station to obtain a preliminary channel estimation result H. LS ; H LS The input is fed into a pre-trained AMRN network to obtain the channel estimation result. This invention enhances the capture and enhancement of local high-frequency features of the channel by using depthwise separable convolution to extract local features and combining them with a context-aware nonlinear weight generation mechanism. Furthermore, by fusing high-frequency and low-frequency information, it effectively improves the accuracy and applicability of RIS channel estimation and reduces computational complexity.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and more specifically to a channel estimation method for a RIS-assisted communication system based on deep learning. Background Technology

[0002] Millimeter wave communication (mmWave), as an important wireless communication technology, utilizes the 30GHz to 300GHz frequency band, providing greater bandwidth and higher data transmission rates. With the rapid increase in data traffic, millimeter wave communication plays a crucial role in future high-capacity, high-speed communication systems. However, millimeter wave signals are easily affected by obstacles and environmental factors during propagation, leading to signal attenuation and quality degradation, which poses a significant challenge to channel estimation.

[0003] To address this issue, Intelligent Reflector (RIS) technology has been proposed as an effective solution. By adjusting the tunable elements on the reflector, RIS can improve the signal propagation path, enhance signal strength, and optimize coverage, showing significant advantages, especially in complex environments. However, channel estimation for RIS systems still faces challenges, and traditional estimation methods often fail to meet the requirements for high accuracy.

[0004] In recent years, deep learning has shown many advantages in RIS channel estimation. However, traditional deep learning methods generally have high training costs and require a lot of computing resources. Therefore, there is an urgent need for a low-cost and high-precision RIS channel estimation scheme. Summary of the Invention

[0005] In view of this, the present invention proposes a channel estimation method for RIS-assisted communication systems based on deep learning to solve the above problems; the method includes: constructing a channel estimation model for the RIS-assisted communication system, wherein the model adopts an AMRN network; activating L RIS elements, where L is less than N; the user equipment transmits channel information and forwards it to the base station through the activated RIS elements; performing preliminary channel estimation on the channel information received by the base station to obtain a preliminary channel estimation result H. LS ; H LS The input is fed into a pre-trained AMRN network to obtain the channel estimation result; where RIS is a reconfigurable smart surface technology and AMRN is an attention-enhanced multi-scale residual network.

[0006] The AMRN network processes the preliminary channel estimation results through the following steps:

[0007] Step 1. Use a dual-branch data processing method to process H LS Multi-scale feature extraction is performed to obtain the channel matrix;

[0008] Step 2. Integrate, refine, and enhance the channel matrix after multi-scale feature extraction to obtain the enhanced channel matrix;

[0009] Step 3. Perform an upsampling operation on the enhanced channel matrix to obtain a high-resolution matrix H;

[0010] The dual-branch structure includes branch 1 and branch 2. The preliminary channel estimation results are input into branch 1 and branch 2 respectively. Branch 1 performs global feature extraction on the preliminary channel estimation results, and branch 2 performs deep learning on the preliminary channel estimation results to obtain a channel matrix with high-frequency information. The global features of branch 1 and the channel matrix with high-frequency information of branch 2 are fused to obtain the final channel matrix.

[0011] The beneficial effects of this invention include:

[0012] This invention recovers a high-resolution channel matrix while maintaining low pilot overhead by performing multi-scale feature extraction on the channel with only some RIS elements activated. It enhances the capture and enhancement of local high-frequency features of the channel by introducing a nonlinear weight generation mechanism, thereby extracting high-frequency information. By using depthwise separable convolution to adjust the importance of feature channels, it extracts low-frequency information and reduces computational complexity. Furthermore, by employing a method that fuses high-frequency and low-frequency information for channel estimation, it improves the overall accuracy and applicability of RIS channel estimation, achieving approximately 12% higher accuracy than traditional models. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the communication system model described in Example 1;

[0014] Figure 2 This is a schematic diagram of the channel estimation frame structure used in this invention;

[0015] Figure 3 This is a network architecture diagram of the AMRN network described in this invention;

[0016] Figure 4 This is a structural diagram of the LEAM described in this invention;

[0017] Figure 5 A structural comparison diagram of standard convolution and depthwise separable convolution;

[0018] Figure 6 This is a schematic diagram illustrating the iteratively amplified content described in this invention;

[0019] Figure 7 This is a flowchart of the upsampling operation described in this invention;

[0020] Figure 8This is a comparison chart of the training of the channel estimation model of the RIS-assisted communication system described in this invention and the model using the traditional SRCNN network framework;

[0021] Figure 9 This is a comparison chart showing the training of the channel estimation model of the RIS-assisted communication system with the GFE branch added and the channel estimation model of the RIS-assisted communication system without the GFE branch, as described in this invention. Detailed Implementation

[0022] To make the objectives, technical solutions, features, and advantages of the present invention clearer, the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0023] Example 1:

[0024] The communication system considered in this embodiment is as follows: Figure 1 As shown, a channel estimation model for a RIS-assisted communication system is presented. The RIS-assisted communication system includes user equipment, a base station, and N RIS components.

[0025] S1. Construct a channel estimation model for a RIS-assisted communication system, using an AMRN network; activate L RIS elements, where L is less than N; the user equipment transmits channel information and forwards it to the base station through the activated RIS elements; perform preliminary channel estimation on the channel information received by the base station to obtain the preliminary channel estimation result H. LS .

[0026] Specifically, the AMRN network model is as follows: Figure 3 As shown, in the uplink of RIS-assisted multi-user millimeter-wave (mmWave) communication, a coordinate system is established, in which K user equipment (UE) are located in the xy plane, each equipped with one antenna; the base station (BS) is located in the xz plane, equipped with M antennas; the RIS element is located in the yz plane, consisting of N reflector elements; L RIS elements are activated, and the number of activated RIS elements L is less than the total number of RIS elements N; the RIS-assisted multi-user millimeter-wave communication system can be divided into three channels: the RIS-BS channel G∈C M×N The channel h from the k-th UE to the RIS k ∈C N×1 UE-BS channel D∈C M×1 Assuming the direct link is blocked by a building, only the cascaded channels of UE-RIS and RIS-BS are considered; the Saleh-Valenzuela channel model is used to model the acquired channel information. Assuming both the base station and RIS adopt a UPA structure, the formula for the channel matrix G is:

[0027]

[0028] Among them, LG This represents the number of paths between RIS and BS. For path l i Gain on, This indicates the azimuth angle at RIS. Indicates the elevation angle at RIS. This indicates the azimuth angle at BS. Indicates the elevation angle at BS; equipped with M (M = M x ×M y ) directional vectors of BS antenna array The formula is:

[0029]

[0030] Where λ represents the carrier wavelength, and d represents the physical distance between adjacent antennas (m x ,m y () represents the location coordinates of the base station.

[0031] Furthermore, equipped with N (N = N) x ×N y ) element RIS antenna array direction vector Represented as:

[0032]

[0033] Among them, (n x n y ) represents the position coordinates of the RIS element.

[0034] Furthermore, the channel h between the k-th UE and the RIS k Represented as:

[0035]

[0036] Among them, L H This represents the number of paths between the k-th UE and the RIS; Indicates the lth i Gain on each path; Indicates the lth i The azimuth angle of the path from RIS Indicates the lth i The elevation angle of RIS along the path.

[0037] Furthermore, Figure 2 This is a schematic diagram of the channel estimation frame structure used in this invention; wherein, a frame is divided into D data blocks, the first data block is used for channel estimation, and the other data blocks are used for data transmission; the data block used for channel estimation is divided into Q time slots; in the q-th time slot, the signal received by the base station for:

[0038]

[0039] Among them, w q It has a mean of zero and a variance of δ. 2 Gaussian noise, The pilot signal transmitted by the k-th UE in the q-th time slot. The reflection coefficient matrix of the RIS element, θ i (i = 1, 2, ..., N) represents the phase of the i-th RIS element, β i Indicates the switching state of the i-th RIS:

[0040]

[0041] Furthermore, by disabling some RIS components, pilot overhead can be effectively reduced; after Q time slots, the signal received by the base station is for:

[0042]

[0043] Y = [y1, y2, ..., y Q ]

[0044]

[0045]

[0046] In order to distinguish the pilots of different UEs, an orthogonal pilot sequence is assigned to each UE. When i≠j, When i = j P k This refers to the signal transmission power.

[0047] Furthermore, the signal Y received by the base station from the Kth UE k for:

[0048]

[0049] in,

[0050] Furthermore, combining Gdiag(θ)h k =Gdiag(h k Let H = θ. k =Gdiag(h k ), Y k Rewritten as:

[0051] Y k =H k θ+w k

[0052] Furthermore, using the LS algorithm, the preliminary estimated channel matrix H is obtained. LS :

[0053] H LS =Y k θ +

[0054] Where, θ + =θ H (θθ H ) -1 θ + It is the pseudoinverse of θ, θ H It is the Hermite matrix of θ.

[0055] S2. A dual-branch data processing method is used for H. LS Multi-scale feature extraction is performed to obtain the channel matrix.

[0056] Specifically, the dual-branch structure includes branch 1 and branch 2; after obtaining the input information, it is input to branch 1 and branch 2 respectively; branch 1 performs global feature extraction, and branch 2 learns and extracts high-frequency information; the output results of branch 1 and branch 2 are fused as the dual-branch output result; branch 1 includes a 5*5 convolutional layer to support branch 1 in capturing global features of the channel matrix; branch 2 includes the following sequentially connected layers: a first convolutional layer, B LEAM (Local Enhancement Attention Modules), a second convolutional layer, and a third convolutional layer;

[0057] Furthermore, there are two branches: Branch 1, GFE (Global Feature Extraction); and Branch 2, LFE (Local Feature Extraction). GFE is responsible for learning and preserving the low-frequency information of the channel matrix, representing the global structure of the channel. LFE is responsible for learning and extracting high-frequency information, capturing the details of the channel, such as edges and textures.

[0058] Furthermore, channel information is processed in GFE using a large 5x5 convolution kernel. GFE focuses on extracting low-frequency information, adjusting the importance of feature channels, and helping to capture global structural information in the channel.

[0059] Specifically, DSConv (Depthwise separable convolution) consists of DWConv (Depthwise convolution) and PWConv (Pointwise convolution); depthwise convolution is used to extract spatial features, and pointwise convolution is used to extract channel features; depthwise separable convolution groups convolutions along the feature dimension, performs independent depthwise convolutions on each channel, and aggregates all channels using a 1×1 convolution before the output.

[0060] The differences between DWConv and standard convolution are as follows: Figure 5 As shown: DWConv's convolution kernel is in single-channel mode, requiring convolution on each channel of the input to obtain an output feature map with the same number of channels as the input feature map. That is, the number of input feature map channels = the number of convolution kernels = the number of output feature maps.

[0061] PWConv uses a 1×1 convolution kernel for dimensionality increase; however, due to the characteristic of channel-wise convolution, the number of output feature maps may be too small, thus affecting the effectiveness of information. Therefore, DSConv further performs pointwise convolution to solve this problem.

[0062] Furthermore, in LFE, for the input H LS Convolutional operations are performed to initially extract features from the channel matrix; subsequent multi-layer LEAM (Local Enhancement Attention Module) layers are used to further extract high-frequency information and capture local detail features; each LEAM enhances feature extraction through convolution and attention mechanisms.

[0063] Step 1. Transfer H LS The input is fed into the first convolutional layer for convolution, resulting in a shallow feature matrix H1.

[0064] Step 2. Pass H1 through B LEAMs to obtain the deep feature matrix H2.

[0065] Step 3. Input H2 into the second convolutional layer for convolution to obtain the channel matrix H3.

[0066] Step 4. Mix H3 with H LS The residuals are connected to obtain the fusion matrix H4.

[0067] Step 5. Input H4 into the third convolutional layer for convolution to obtain the channel matrix H5 with high-frequency information.

[0068] Specifically, the LEAM structure is as follows: Figure 4As shown, LEAM is used to enhance the ability to capture high-frequency local information. Data processing includes:

[0069] Step 1. Process the feature matrix X input to the module. in A linear transformation is performed to obtain feature information Q, K, and V; where Q is the query vector of the current feature, used to capture the importance of this position to other positions; K is the key vector of the feature, used to calculate the similarity with Q; and V is the feature value.

[0070] Q,K,V=FC(X in )

[0071] Among them, X in It is the input of LEAM, and FC indicates a fully connected layer.

[0072] Step 2. Perform local information fusion on Q, K, and V respectively to obtain Q. l K l V s .

[0073] Specifically, DSConv is used to aggregate local information for Q, K, and V:

[0074] Q l =DSConv(Q)

[0075] K l =DSConv(K)

[0076] V s =DSConv(V)

[0077] Step 3. Calculate Q l With K l The Hadamard Product is calculated, and the result is transformed to obtain the nonlinear weight vector Cafe; the transformations include Swish, linear transformation, and Tanh operation; further, the transformation formula is:

[0078] Cafe t =FC(Swish(FC(Q) l ⊙K l )))

[0079]

[0080] Step 4. Calculate Cafe and V s The Hadamard product is calculated and a linear transformation is performed on the result to obtain the matrix FC(Cafe⊙V). s This matrix incorporates enhancements in local spatial features and context weights;

[0081] Step 5. Fusion Xin With FC (Cafe⊙V) s Output deep feature matrix X out ;include:

[0082] X out =FC(Cafe⊙V s )+X in

[0083] S3. Integrate, refine, and enhance the channel matrix after multi-scale feature extraction to obtain the enhanced channel matrix.

[0084] Specifically, the dual-branch output is obtained and passed through three convolutional layers in sequence to obtain the enhanced channel matrix.

[0085] S4. Perform an upsampling operation on the enhanced channel matrix to obtain a high-resolution matrix H.

[0086] The flowchart for the upsampling operation is as follows: Figure 7 As shown; the upsampling operation includes the following steps:

[0087] Step 1. Determine the magnification factor N / L.

[0088] Specifically, determine the amplification factor: if N / l = 2, then a 2x upsampling is required.

[0089] Step 2. Iteratively amplify the dimensions of the integrated channel matrix according to the amplification factor to obtain the amplified channel matrix.

[0090] Step 3. Perform a 3*3 convolution operation on the amplified channel data to obtain the upsampling result.

[0091] Specifically, the iterative amplification is as follows: Figure 6 As shown, it includes:

[0092] Step 1. Convolve the enhanced channel matrix to expand the channel matrix channels.

[0093] Step 2. Perform a pixel shuffle operation on the expanded channel matrix to restore the number of channels and increase the column dimension.

[0094] Step 3. Iterate through steps 1 and 2. When the amplification factor is reached, output the amplified channel matrix.

[0095] The magnification factor for the Pixel Shuffle operation is 2x or 3x.

[0096] Specifically, for example, using 1×1 convolution expands the number of channels from 2 to 4; using the Pixel Shuffle operation, the magnification factor is set to 2; Pixel Shuffle restores the number of channels from 4 to 2 and increases the column dimension by 2 times, thereby achieving column upsampling by rearranging the channels; using iterative methods to achieve larger magnification factors, such as a 4x magnification factor can be obtained through two iterations with a magnification factor of 2; a 6x magnification factor can be obtained through two iterations with magnification factors of 2 and 3 respectively.

[0097] Example 2:

[0098] The simulation parameters are set as follows:

[0099] Operating frequency band: 28GHz;

[0100] Number of antennas for the base station: M = 16;

[0101] Number of RIS units: N = 256;

[0102] Number of UE antennas: 1;

[0103] Number of LEAM modules: B = 15;

[0104] Channel data for a millimeter-wave RIS-assisted communication system was generated using the SimRIS channel simulator, and training, validation, and test sets were constructed. The dataset contained 1000 samples, with the training set, test set, and validation set accounting for 70%, 15%, and 15%, respectively.

[0105] 196 of the 256 units in the RIS are selected as activation units because the pilot signal sequence is orthogonal and the received signal can be effectively separated. The data unit of each user obtains a preliminary low-dimensional channel matrix estimate by least squares method. The low-dimensional channel matrix is ​​then further processed by the AMRN network framework and recovered to obtain a high-precision channel matrix.

[0106] The normalized mean square error (NMSE) between the channel estimation matrix and the true channel matrix is ​​used as a performance metric, defined as follows:

[0107]

[0108] Where H represents the actual channel matrix. This represents the estimated channel matrix. It represents the mathematical expectation.

[0109] The proposed RIS-assisted communication system channel estimation model is compared with a model using the traditional SRCNN network framework. The comparison results are as follows: Figure 8 As shown:

[0110] Specifically, during training, the NMSE value of the RIS-assisted communication system channel estimation model is lower than that of the SRCNN model, indicating that its estimation accuracy is higher. After 40 training iterations, the NMSE of both models tends to stabilize, with the NMSE of the model using the traditional SRCNN network being approximately -11.5 dB. The NMSE of the RIS-assisted communication system channel estimation model using the AMRN network proposed in this invention is approximately -12.9 dB. The accuracy of the RIS-assisted communication system channel estimation model is about 12% higher than that of the model using the SRCNN network.

[0111] Furthermore, the GFE branch introduced in the RIS-assisted communication system channel estimation model is used to support the RIS-assisted communication system channel estimation model in extracting global information. The RIS-assisted communication system channel estimation model with and without the GFE branch are trained separately, and the NMSE comparison results are as follows: Figure 9 As shown:

[0112] Specifically, the model with the GFE branch exhibits a lower NMSE during training. The GFE branch effectively captures the global features of the channel matrix, reducing the mismatch between high-frequency and low-frequency information, thus converging faster during training and achieving a lower NMSE.

[0113] Finally, it should be noted that the above description only depicts some embodiments of the present invention. For those skilled in the art, various changes, modifications, substitutions, and variations can be conceived of these embodiments without departing from the principles and spirit of the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents, and all the above-mentioned behaviors should be covered within the scope of protection of the present invention.

Claims

1. A channel estimation method for a RIS-assisted communication system based on deep learning, wherein the RIS-assisted communication system includes user equipment, a base station, and N RIS components, characterized in that, include: A channel estimation model for a RIS-assisted communication system is constructed, using an AMRN network. Activate L RIS elements, where L is less than N; The user equipment transmits channel information and forwards it to the base station through the activated RIS element; the base station performs preliminary channel estimation on the received channel information to obtain the preliminary channel estimation result. ;Will The input is fed into a pre-trained AMRN network to obtain the channel estimation results; Among them, RIS is a reconfigurable smart surface technology, and AMRN is an attention-enhanced multi-scale residual network. The AMRN network processes the preliminary channel estimation results as follows: S1. A dual-branch data processing method is used to... Multi-scale feature extraction is performed to obtain the channel matrix; S2. The channel matrix is ​​integrated, refined, and enhanced to obtain the enhanced channel matrix; S3. Perform an upsampling operation on the enhanced channel matrix to obtain a high-resolution matrix. ; The dual-branch structure includes branch 1 and branch 2. The preliminary channel estimation results are input into branch 1 and branch 2 respectively. Branch 1 performs global feature extraction on the preliminary channel estimation results, while branch 2 performs deep learning on the preliminary channel estimation results to obtain a channel matrix with high-frequency information. The global features from branch 1 and the high-frequency information channel matrix from branch 2 are fused to obtain the final channel matrix. Branch 2 consists of a first convolutional layer, B LEAMs, a second convolutional layer, and a third convolutional layer. The LEAMs are local enhancement attention modules. Branch 2 processes the preliminary channel estimation results, including: S321. Will The input is fed into the first convolutional layer for convolution to obtain the shallow feature matrix H1; S322. Pass H1 through B LEAMs to obtain the deep feature matrix H2; the LEAM data processing includes: Step 1. Process the shallow feature matrix input to the module. A linear transformation is performed to obtain feature information Q, K, and V; where Q is the query vector of the current feature, K is the key vector of the feature, and V is the feature value. Step 2. Perform local information fusion on Q, K, and V respectively to obtain... , , ; Step 3. Calculation and The Hadamard product is calculated, and the result is transformed to obtain the nonlinear weight vector. The transformations include: Swish, linear transformations, and Tanh operations; Step 4. Calculation and The Hadamard product is calculated and a linear transformation is performed on the result to obtain the matrix. This matrix incorporates enhancements in local spatial features and context weights; Step 5. Fusion and Output deep feature matrix ; S323. Input H2 into the second convolutional layer for convolution to obtain the channel matrix H3; S324. Connect H3 with Residual connections yield the fusion matrix H4; S325. Input H4 into the third convolutional layer for convolution to obtain the channel matrix H5 with high-frequency information.

2. The channel estimation method for a RIS-assisted communication system based on deep learning according to claim 1, characterized in that, Preliminary channel estimation of the channel information includes: modeling the acquired channel information using the Saleh-Valenzuela channel model, and performing preliminary channel estimation using the LS algorithm to obtain a low-dimensional channel matrix. .

3. The channel estimation method for a RIS-assisted communication system based on deep learning according to claim 1, characterized in that, Branch 1 includes a 5x5 convolutional layer used to capture global features of the channel matrix.

4. The channel estimation method for a RIS-assisted communication system based on deep learning according to claim 1, characterized in that, The process of integrating, refining, and enhancing the channel matrix includes: passing the channel matrix through three convolutional layers in sequence to obtain the enhanced channel matrix.

5. The channel estimation method for a RIS-assisted communication system based on deep learning according to claim 1, characterized in that, Upsampling operations include: Step 1. Determine the magnification factor N / L; Step 2. Iteratively amplify the dimensions of the integrated channel matrix according to the amplification factor to obtain the amplified channel matrix; Step 3. Perform a 3*3 convolution operation on the amplified channel matrix to obtain the upsampling result.

6. The channel estimation method for a RIS-assisted communication system based on deep learning according to claim 5, characterized in that, Iterative amplification of the dimensions of the integrated channel matrix based on the amplification factor includes: Step 1. Convolve the enhanced channel matrix to expand the channel matrix channels; Step 2. Perform a pixel shuffle operation on the expanded channel matrix to restore the number of channels and increase the column dimension; Step 3. Iterate through steps 1 and 2, and when the amplification factor is reached, output the amplified channel matrix; The magnification factor for the Pixel Shuffle operation is 2x or 3x.

Images