Channel information feedback method of large-scale MIMO system based on multi-scale feature fusion

Abstract

The invention discloses a multi-scale feature fusion-based channel information feedback method for a large-scale MIMO (Multiple Input Multiple Output) system, which comprises the following steps of: 1, constructing a multi-scale feature fusion-based encoder and decoder network, and combining a novel selective state space with a multi-head attention model and hybrid convolution, the channel state information (CSI) remote modeling capability is further enhanced through an input selectivity mechanism, and multi-scale channel features are extracted for CSI (de-compression). 2, constructing a convolution enhanced vector quantization model, and further improving the feedback precision under the same feedback overhead by using dimension reduction convolution; and 3, a channel feature enhancement (CFE) module is constructed, and channel information loss is further reduced by utilizing ascending channel convolution to assist in decompressing full connection. According to the codec, multi-scale CSI features are extracted by using an input selectivity mechanism, multi-head attention and convolution, and high-precision channel information feedback is completed by combining a convolution enhanced vector quantization model and a CFE module.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communications and deep learning technology, and in particular relates to a channel information feedback method for a large-scale MIMO system based on multi-scale feature fusion. Background Art

[0002] Massive MIMO technology is widely regarded as one of the key technologies to support multi-user scenarios in the next generation of wireless communication systems. In order to fully utilize the advantages of massive MIMO technology in frequency division duplexing (FDD) systems, each user device needs to feed back its estimated downlink channel state information (CSI) to the base station through the uplink channel, and the base station uses the CSI to make a precoding matrix. Therefore, the accuracy of CSI feedback directly affects the spatial diversity and multiplexing gain of the massive MIMO system. However, the huge antenna array of the FDD massive MIMO system leads to a large amount of CSI feedback data in the system, and directly feeding back the complete CSI will result in excessive uplink resource consumption. Therefore, robust CSI compression and quantization technology is essential.

[0003] In recent years, artificial intelligence (AI) and deep learning (DL) have attracted widespread attention as technologies for designing and optimizing wireless communication systems. Many researchers at home and abroad have conducted extensive research on the network layer, link layer, and physical layer. CSI feedback based on deep learning has also received increasing attention in the field of wireless communication research. Researchers have improved the accuracy of CSI feedback through innovative neural network design and the use of correlation. The CSI feedback network based on deep learning generally adopts a codec framework, and constructs different architectures including innovative convolutional neural networks and Transformers with multi-head attention mechanisms. It has been proven in many simulation scenarios that it has feedback accuracy that exceeds that of traditional compressed sensing algorithms.

[0004] However, convolution and various attention architectures have certain defects. For example, the inherent locality of the convolution receptive field makes it difficult to capture the global features of CSI, and the long-range modeling capability of Transformer increases the network computation and reasoning time. Therefore, introducing a new modeling framework and fully integrating multi-scale channel features is crucial to reducing the network computation and improving the accuracy of CSI feedback. Summary of the invention

[0005] The purpose of the present invention is to provide a channel information feedback method for a large-scale MIMO system based on multi-scale feature fusion. Aiming at the channel information feedback problem of the existing large-scale MIMO system, a variety of CSI encoding and decoding strategies are integrated, and the new structured state space model Mamba is combined with a multi-head attention model and a convolutional network. The CSI remote modeling capability is further enhanced through the input selectivity mechanism, and multi-scale channel features are used for CSI feedback and the quantization framework is optimized to improve the accuracy and computational efficiency of CSI feedback.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] The present invention discloses a channel information feedback method for a large-scale MIMO system based on multi-scale feature fusion, which includes an encoder and decoder network based on multi-scale feature fusion, a convolution enhanced vector quantization model and a CFE module, and the steps are as follows:

[0008] Step 1: Construct a simulation channel scenario, preprocess the generated channel matrix, and divide it into training set, validation set, and test set;

[0009] Step 2: construct a channel information feedback model based on multi-scale feature fusion, wherein the channel information feedback model includes an encoder based on multi-scale feature fusion, a convolution enhanced vector quantization model and a decoder, and constructs a channel feature enhancement (CFE) module at the decoder end to decompress the channel information;

[0010] Step 3: input the training set into the multi-scale feature fusion channel information feedback model, and train the neural network end-to-end in combination with the weighted loss function;

[0011] Step 4: Input the test data set into the trained neural network to detect the channel information feedback accuracy.

[0012] Furthermore, in the channel information feedback model based on multi-scale feature fusion described in step 2, both the encoder and the decoder are hierarchical structures, and each layer is composed of a cascaded Conformer module that integrates a multi-head attention mechanism and a hybrid convolution and a new structured state space model Mamba module, and the number of layers is set according to the complexity requirements. The encoder end cascades a double-layer Conformer module and a double-layer Mamba module to extract channel features, compress the fully connected layer F1 to output compressed codewords, and the CFE module further extracts codeword features. Finally, the decoder end cascades a double-layer Conformer module and a double-layer Mamba module to reconstruct the channel matrix.

[0013] The Conformer module includes a first feedforward layer, a Transformer network, a hybrid convolutional network, a second feedforward layer, and a layer normalization unit. The first feedforward layer and the second feedforward layer are macaron-style half-step residual connection structures, and the Transformer network and the hybrid convolutional network are residual structures;

[0014] The Mamba module includes a linear layer, a one-dimensional causal convolution layer, a selective state space, a layer normalization unit, and a SiLU activation function, wherein the linear layer changes the feature dimension, the one-dimensional causal convolution layer preprocesses the local features, the selective state space introduces a state transition matrix, and the output is obtained through the state equation and the output equation.

[0015] Furthermore, the convolution enhanced vector quantization model in step 2 includes three cascade modules: a convolution dimension reduction module, a vector quantization module, and a convolution dimension increase module;

[0016] The convolutional dimension reduction module consists of three 1×1 channel convolutions, one 1×3 dimension reduction convolution, a pooling layer, and an activation function, which transforms the codeword vector into a set of feature expansion vectors;

[0017] The quantization module has an internal codebook The codebook consists of K latent embedding vectors e of length D. i where the elements are real numbers Map the feature expansion vector to the potential embedding vector e of the codebook i , e i The corresponding codebook index forms a feedback bit stream;

[0018] The convolutional dimension-raising module consists of a 1×3 dimension-raising convolution, a 1×1 channel convolution and an activation function, and reconstructs the codeword through a reverse convolution operation.

[0019] Furthermore, the CFE module in step 2 includes two branches: the decompressed fully connected layer F2 and the up-channel convolution branch, and the dot product operation between the two obtains the decompressed CSI. The up-channel convolution branch is composed of three up-channel convolution layers with different convolution kernel sizes connected in series. The decompressed fully connected layer F2 performs the mirror operation of the compressed fully connected layer F1.

[0020] Furthermore, the input dimension of the Mamba module in step 2 is (b, L, D), where b is the batch size, L is the number of features, and D is the feature dimension; the input first passes through a layer normalization unit and an input linear layer to obtain an output (b, 8D, L), which is equally divided into two sequences s1 and s2 of dimension (b, 4D, L);

[0021] Input s1 into a one-dimensional causal convolution layer with a convolution kernel of 4 and a SiLU activation function to obtain a selective state space input sequence x of dimension (b, 4D, L);

[0022] The selective state space is dynamically updated by discretization using parameters (Δt, A, B, C), where the dimensions of the time change Δt, the state transition matrix A, the input-to-state transition matrix B, and the state-to-output transition matrix C are (b, 4D, d state ), (b,4D,L), (b,1,d state ,L),(b,1,d state ,L),d state is a network hyperparameter. s2 is multiplied by the SiLU activation function and the selective state space output y, and finally passes through the output linear layer to obtain an output with a dimension of (b, L, D);

[0023] Furthermore, the selective state space of the Mamba module utilizes an input selectivity mechanism to first update the potential state through the state equation, and then obtain the current output y(t) using the output equation.

[0024] The state equation is as follows:

[0025] h'(t) = Α h(t-1) + B x(t) (1)

[0026] Where h(t) and h(t-1) are the potential states of the current and previous inputs, respectively, h'(t) is the current potential state derivative, and x(t) is the current input;

[0027] The output equation is as follows:

[0028] y(t) = C h(t) (2)

[0029] The discretized state parameters are obtained by the zero-order hold (ZOH) rule and

[0030]

[0031] Among them, Δ is the time scale parameter, I is the unit matrix;

[0032] The discretized state equation is obtained as:

[0033]

[0034] Among them, h k and h k-1 are the current and previous discretized input potential states, respectively, x k is the current discretization input;

[0035] The discretized output equation is:

[0036] y k =Chk (5)

[0037] Finally, the output y is obtained through global convolution:

[0038]

[0039] in, is the structured convolution kernel.

[0040] Furthermore, the convolution enhanced vector quantization model described in step 2 transforms the input codeword vector H of length M into c Use the convolutional dimension reduction module DownConv(·) to transform into M / 2 feature expansion vectors q s ;

[0041] The vector quantization module VQ(·) is used for discretization, and the feature expansion vector q is obtained by nearest neighbor search. s Mapping to codebook The latent embedding vector e in i , the corresponding codebook index forms a bit stream, and the feature expansion vector q is reconstructed at the base station r =e i , where i = arg min m ||q s -e m ||, m∈{1,2,...,K}, and finally the codeword vector is restored through the convolutional dimension-upgrading module UpConv(·) The whole process is expressed as:

[0042]

[0043] Quantization model loss function L(H c ,e i )for:

[0044]

[0045] Where sg is the stop gradient operator and β is the weight coefficient.

[0046] The reconstructed compressed codeword is reconstructed through the serially connected CFE module and decoder to reconstruct the CSI. The total loss function L is:

[0047]

[0048] Where H a and are encoder input and decoder output respectively, ||·|| F is the Frobenius norm.

[0049] The present invention discloses a channel information feedback method for a large-scale MIMO system based on multi-scale feature fusion, which combines a novel selective state space with a multi-head attention model and a hybrid convolution, further enhances the CSI remote modeling capability through an input selective mechanism, and then makes full use of multi-scale channel features for CSI feedback, and combines the convolution enhanced vector quantization model and the CFE module of the present invention to obtain a high-precision and low-transmission-bit feedback effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of channel information feedback of the present invention;

[0051] Figure 2 This is a diagram of the channel information feedback network structure;

[0052] Figure 3 Convolution-enhanced vector quantization model structure diagram;

[0053] Figure 4 This is the CFE module structure diagram;

[0054] Figure 5 This is the Mamba module structure diagram;

[0055] Figure 6 The comparison chart of feedback accuracy and network complexity under different compression ratios (CR) in indoor environment (Indoor);

[0056] Figure 7 The figure is a comparison of feedback accuracy and network complexity under different compression ratios (CR) in outdoor environment (Outdoor). DETAILED DESCRIPTION

[0057] In order to better understand the purpose, structure and function of the present invention, the channel information feedback method of the large-scale MIMO system based on multi-scale feature fusion of the present invention is further described in detail below with reference to the accompanying drawings.

[0058] This embodiment provides a channel information feedback method for a large-scale MIMO system based on multi-scale feature fusion, and designs an encoder and decoder network based on multi-scale feature fusion, a convolution-enhanced vector quantization model, and a CFE module, so that a lighter neural network can complete high-precision channel information feedback at a lower transmission bit and computational complexity. The embodiment integrates multiple CSI encoding and decoding strategies such as Mamba, Transformer, and hybrid convolution to extract multi-scale CSI features, wherein the Mamba module can selectively focus on specific information by relying on the input selection mechanism, the convolution-enhanced vector quantization model alleviates the information loss caused by CSI compression through convolution, and the CFE module uses the channel-up convolution network to assist the anti-compression fully connected layer F2 to perform channel information decompression at the decoder end, thereby enhancing the scalability and flexibility of the system.

[0059] Specifically, if Figure 2 As shown in the figure, both the encoder and decoder networks are hierarchical structures, and each contains two modules: the Conformer module that integrates the multi-head attention mechanism and hybrid convolution, and the Mamba module, a new structured state space model. The Conformer module contains the first feedforward layer, the Transformer network, the hybrid convolution network, the second feedforward layer, and the layer normalization unit, which models the local and global features of CSI in a parameter-efficient way. The Mamba module filters irrelevant modeling information through the input selection mechanism, adaptively aggregates the long-distance dependencies of CSI features, and improves the network computational efficiency.

[0060] Specifically, if Figure 3 As shown in Figure 1, the convolution-enhanced vector quantization model consists of three cascaded modules: convolution dimension reduction module, vector quantization module, and convolution dimension increase module. The convolution dimension reduction module consists of three 1×1 channel convolutions, one 1×3 dimension reduction convolution, pooling layer, and activation function, which transforms the codeword vector into a feature expansion vector; the quantization module is a variational autoencoder with an intrinsic codebook, which maps each feature expansion vector to a potential embedding vector in the codebook, and the corresponding codebook index forms a feedback bit stream; the convolution dimension increase module uses a 1×3 dimension increase convolution, a 1×1 channel convolution, and an activation function to reconstruct the codeword through a reverse convolution operation.

[0061] Specifically, if Figure 4 As shown in the figure, the CFE module contains two branches: the decompressed fully connected layer F2 and the convolution branch, and the point multiplication operation between the two obtains the decompressed CSI. The convolution branch consists of three concatenated up-channel convolution layers with different convolution kernel sizes, and the decompressed fully connected layer F2 performs the mirror operation of the compressed fully connected layer F1. The convolution branch does not occupy the computing and storage resources of the user end, further reducing the channel information loss and enhancing the scalability and flexibility of the system.

[0062] Example 1: See Figure 1-Figure 5 , this implementation provides a channel information feedback method for a large-scale MIMO system based on multi-scale feature fusion, specifically including:

[0063] Step 1: Construct a simulation channel scenario, preprocess the generated channel matrix, and divide it into training set, validation set, and test set;

[0064] Step 2: construct a channel information feedback model based on multi-scale feature fusion, wherein the channel information feedback model includes an encoder based on multi-scale feature fusion, a convolution enhanced vector quantization model and a decoder, and constructs a channel feature enhancement (CFE) module at the decoder end to decompress the channel information;

[0065] Step 3: input the training set into the multi-scale feature fusion channel information feedback model, and train the neural network end-to-end in combination with the weighted loss function;

[0066] Step 4: Input the test data set into the trained neural network to detect the channel information feedback accuracy.

[0067] Specifically, the step 1 includes:

[0068] Step 1.1: Use the default settings in COST2100 to simulate and generate indoor micro-cell and outdoor rural scene channels in the 5.3 GHz and 300 MHz frequency bands. Number of antennas N t , number of subcarriers N c They are set to 32 and 1024 respectively, generating a total of 150,000 channel matrices;

[0069] Step 1.2: Divide the channel matrix into three parts, namely, 100,000 as training data set, 30,000 as validation data set, and 20,000 as test data set;

[0070] Step 1.3: Preprocess the generated channel matrix. Using the sparsity of the channel matrix, a sparse complex channel matrix in the angle delay domain is obtained through a two-dimensional discrete Fourier transform, where the first N a <N c The row is non-zero, take the first N a = 32 rows of concatenated real and imaginary parts as the input matrix of the neural network in Represents a real number.

[0071] Specifically, Figure 2 As shown, the step 2 includes:

[0072] Step 2.1: The encoder (user end) inputs the encoder through the serial multi-scale feature fusion, and uses the cascaded two-layer Conformer module and the two-layer Mamba module to extract multi-scale channel features. The output size of each layer is consistent with the input;

[0073] Step 2.2: Length 2N a N t The vector is sent to the encoder compression fully connected layer F1, and the compressed codeword is obtained according to a specific compression ratio;

[0074] Step 2.3: The compressed codeword is used to obtain a transmission bit stream through a convolution-enhanced vector quantization model;

[0075] Step 2.4: The decoder (base station end) uses the CFE module to extract codeword features, and its two branches recover the CSI dimension 2N through dot multiplication calculation a N t ;

[0076] Step 2.5: Recover the original CSI using the decoder (base station end) double-layer Conformer module and the double-layer novel structured state space model Mamba module;

[0077] Step 2.6: Recover the angle delay domain channel matrix by zero-padding in the frequency domain and inverse discrete Fourier transform.

[0078] Specifically, the double-layer Conformer module and the double-layer Mamba module constructed in step 2.1 and step 2.5 include:

[0079] Step 2.1.1: The input of the Conformer module passes through the first feedforward layer, the Transformer network, the hybrid convolutional network, the second feedforward layer, and the layer normalization unit in sequence. The first and second feedforward layers are macaron-style half-step residual connection structures, and the Transformer network and the hybrid convolutional network are residual structures. During training, 10% of the neurons in the Conformer module are randomly discarded to enhance the generalization ability of the model;

[0080] Step 2.1.2: The two linear layers are expanded by 4 times according to the expansion factor and then restored to the input dimension;

[0081] Step 2.1.2: The number of attention heads of the Transformer network is set to 8, and Q n , K n and V n The matrix gets the attention score matrix A n , using the normalized matrix Atten(A n ) and V n Calculate Z n , Z n Concatenate into the output matrix of the Transformer network to extract CSI global features;

[0082] Step 2.1.3: The hybrid convolutional network uses point convolution, depth convolution, batch normalization, layer normalization, gated linear unit (GLU) and Swish activation function to extract small-scale features, where the convolution kernel size of the point convolution is 1 and the depth convolution kernel size is 31;

[0083] Step 2.1.4: The input dimension of the two-layer new structured state space model Mamba module is (b, L, D), where b is the batch size, L is the number of features, and D is the feature dimension; the input first passes through the layer normalization unit and the input linear layer to obtain the output (b, 8D, L), which is divided into two sequences s1 and s2 of dimension (b, 4D, L);

[0084] Step 2.1.5: Input s1 into a one-dimensional causal convolution layer with a convolution kernel of 4 and a SiLU activation function to obtain a selective state space input sequence x of dimension (b, 4D, L);

[0085] Step 2.1.6: The selective state space is dynamically updated using the discretized parameters (Δt, A, B, C), where the dimensions of the time change Δt, the state transition matrix A, the input-to-state transition matrix B, and the state-to-output transition matrix C are (b, 4D, d state ), (b,4D,L), (b,1,d state ,L),(b,1,d state ,L), where d state is a network hyperparameter, set to 16. First, update the potential state through the state equation:

[0086] h'(t) = Α h(t-1) + B x(t) (1)

[0087] Where h(t) and h(t-1) are the potential states of the current and previous inputs, respectively, h'(t) is the current potential state derivative, and x(t) is the current input;

[0088] Step 2.1.7: Use the output equation to get the current output y(t):

[0089] y(t) = C h(t) (2)

[0090] Step 2.1.8: Obtain the discretized state parameters by the zero-order hold (ZOH) rule and

[0091] Among them, Δ is the time scale parameter, I is the unit matrix;

[0092] Step 2.1.9: Discretize state equations and output equations in the selective state space:

[0093]

[0094] Among them, h k and h k-1 are the current and previous discretized input potential states, respectively, x k is the current discretization input;

[0095] The selective state space output y is obtained through global convolution:

[0096]

[0097] in, is the structured convolution kernel.

[0098] Step 2.1.10: s2 is multiplied by y through the SiLU activation function, and finally passes through the second linear layer to obtain an output with dimension (b, L, D);

[0099] Step 2.1.11: Selective state space is trained in parallel in convolutional mode and autoregressive inference is performed in recurrent mode, combining the efficiency of training and inference.

[0100] Specifically, Figure 3 As shown, the convolution enhanced vector quantization model constructed in step 2.3 specifically includes:

[0101] Step 2.3.1: Codeword vector H of length M c Entering the convolution dimension reduction module, it is first reshaped into a single-channel matrix, and the channel dimension is transformed through three 1×1 channel convolutions, and then a 1×3 two-dimensional dimensionality reduction convolution is used to reduce the quantization size to obtain M / 2 feature expansion vectors q s As input to the vector quantization module;

[0102] Step 2.3.2: Construct channel quantized variational autoencoder codebook The codebook consists of K latent embedding vectors e of length D i where the elements are real numbers The quantization bit is fixed to 5, that is, K = 2 5 , D is set to 3. The feature expansion vector q s Send it to the variational autoencoder and use the nearest neighbor query to convert q s Mapping to codebook The latent embedding vector e in i , corresponding codebook index forms a bit stream;

[0103] Step 2.3.3: The bit stream is transmitted to the base station and the potential embedding vector q is reconstructed using the codebook r =e i , where i = argmin m ||q s -e m ||, m∈{1,2,...,K}, the quantized bitstream is matched with the latent embedding vector;

[0104] Step 2.3.4: Recover the codeword vector using the convolutional dimension-raising module It passes through 1×3 two-dimensional up-convolution and 1×1 channel convolution in sequence. The whole process is expressed as:

[0105]

[0106] Where UpConv(·), VQ(·), and DownConv(·) are the convolutional dimension reduction module, the vector quantization module, and the convolutional dimension increase module, respectively.

[0107] Step 2.3.5: Calculate the quantization model loss function L(H c ,e i )for:

[0108]

[0109] Where sg is the stop gradient operator and the weight coefficient β is set to 1.

[0110] Specifically, the step 3 includes:

[0111] Step 3.1: Calculate the total network loss function L as:

[0112]

[0113] Where H a and are encoder input and decoder output respectively, ||·|| F is the Frobenius norm.

[0114] Step 3.2: Perform end-to-end self-supervised learning on the model. The network is trained for 1000 iterations. The training adopts the cosine annealing scheme, the number of warm-up is set to 30 times, and the learning rate range is 5×10 -5 ~2×10 -4 ;

[0115] Step 3.3: Using the ADAM optimizer, the entire feedback process is measured by the mean square error MSE:

[0116]

[0117] The simulation results can be found in Figure 6-7 , the channel information feedback method of the large-scale MIMO system based on multi-scale feature fusion described in the present invention is compared with the existing channel information feedback network CsiConformer (including vector quantization) under different compression ratios (CR=8, 16, 32, 64) in terms of network complexity and feedback accuracy, where the feedback accuracy is measured by the normalized mean square error (NMSE) and the network computational complexity is expressed in floating point operations per second (FLOPs). The present invention combines a novel structured state space model Mamba, a multi-head attention model and a convolutional network to extract multi-scale CSI features, Figure 6-7The results show that the CSI encoding and decoding strategy based on multi-scale feature fusion can effectively extract CSI features. Combined with the convolution enhanced vector quantization model, it can achieve lower NMSE with fewer FLOPs under the same transmission bit, that is, achieve higher channel information feedback accuracy with low computational overhead. In addition, using the CFE module instead of the traditional fully connected anti-compression can further improve the feedback quality by relying on the computing resources of the base station, which has obvious advantages in outdoor scenarios.

Claims

1. A channel information feedback method for massive MIMO systems based on multi-scale feature fusion, It is characterized in that The method comprises the following steps: Step 1: Construct a simulation channel scenario, preprocess the generated channel matrix, and divide it into training set, validation set, and test set; Step 2: construct a channel information feedback model based on multi-scale feature fusion, wherein the channel information feedback model includes an encoder based on multi-scale feature fusion, a convolution enhanced vector quantization model and a decoder, and constructs a CFE module at the decoder end to decompress the channel information; Step 3: input the training set into the multi-scale feature fusion channel information feedback model, and train the neural network end-to-end in combination with the weighted loss function; Step 4: Input the test data set into the trained neural network to detect the channel information feedback accuracy.

2. The channel information feedback method for a large-scale MIMO system based on multi-scale feature fusion according to claim 1, It is characterized in that In the channel information feedback model based on multi-scale feature fusion, both the encoder and the decoder are hierarchical structures, each layer is composed of a series of Conformer modules that integrate multi-head attention mechanisms and mixed convolutions and a new structured state space model Mamba module, and the number of layers is set according to the complexity requirements; the encoder uses cascaded double-layer Conformer modules and double-layer Mamba modules to extract channel features, and then obtains compressed codewords by compressing the fully connected layer F1; the decoder extracts codeword features through the CFE module, and then inputs them into the cascaded double-layer Conformer modules and double-layer Mamba modules to reconstruct the channel matrix.

3. The channel information feedback method for a large-scale MIMO system based on multi-scale feature fusion according to claim 2, It is characterized in that The Conformer module includes a first feedforward layer, a Transformer network, a hybrid convolutional network, a second feedforward layer and a layer normalization unit; the first feedforward layer and the second feedforward layer are macaron-type half-step residual connection structures, and the Transformer network and the hybrid convolutional network are residual structures; the Mamba module includes a linear layer, a one-dimensional causal convolutional layer, a selective state space, a layer normalization unit and a SiLU activation function; wherein the linear layer changes the feature dimension, the one-dimensional causal convolutional layer preprocesses local features, and the selective state space introduces state transition matrices A, B, and C, and the output is obtained through the state equation and the output equation.

4. The channel information feedback method for a large-scale MIMO system based on multi-scale feature fusion according to claim 1, It is characterized in that The convolution enhanced vector quantization model includes three cascade modules: a convolution dimension reduction module, a vector quantization module, and a convolution dimension increase module; the convolution dimension reduction module transforms the compressed codeword into a feature expansion vector through three 1×1 channel convolutions, one 1×3 dimension reduction convolution, a pooling layer and an activation function; the vector quantization module is a module with an internal codebook. The codebook consists of K latent embedding vectors e of length D. i where the elements are real numbers Map each feature expansion vector to a potential embedding vector e in the codebook i , e i The corresponding codebook index forms a feedback bit stream; the convolutional dimension-raising module reconstructs the codeword using a 1×3 dimension-raising convolution, a 1×1 channel convolution and an activation function.

5. The channel information feedback method for a large-scale MIMO system based on multi-scale feature fusion according to claim 1, It is characterized in that The CFE module includes two branches: an inverse compression fully connected layer F2 and an ascending channel convolution branch, and a decompressed CSI is obtained by a dot product operation between the two. The ascending channel convolution branch is composed of three ascending channel convolution layers with different convolution kernel sizes connected in series, and the inverse compression fully connected layer F2 performs a mirror operation of the compression fully connected layer F1.

6. The channel information feedback method for a large-scale MIMO system based on multi-scale feature fusion according to claim 3, It is characterized in that The input dimension of the novel structured state space model Mamba module is (b, L, D), where b is the batch size, L is the number of features, and D is the feature dimension; the input first passes through the layer normalization unit and the input linear layer to obtain the output (b, 8D, L), which is divided into two sequences s with dimensions (b, 4D, L) 1 and 2 ; Will s 1 Input to a one-dimensional causal convolution layer with a convolution kernel of 4 and a SiLU activation function to obtain a selective state space input sequence x of dimension (b, 4D, L); The selective state space is dynamically updated by discretization using parameters (Δt, A, B, C), where the time change Δt, the state transition matrix A, and the transition matrix B input to the state; The dimensions of the state-to-output transition matrix C are (b, 4D, d state ), (b,4D,L), (b,1,d state ,L),(b,1,d state ,L),d state is the network hyperparameter; s 2 Through the SiLU activation function and the selective state space output y dot product, and finally through the output linear layer to obtain the output of dimension (b, L, D); The selective state space first updates the potential state through the state equation, and then uses the output equation to obtain the current output y(t). The state equation and output equation are: Where h(t) and h(t-1) are the potential states of the current and previous inputs, respectively, h'(t) is the current potential state derivative, and x(t) is the current input; The Mamba module obtains the discretized state parameters through the zero-order hold rule. and Among them, Δ is the time scale parameter, I is the unit matrix; The discretized state equation and output equation of the selective state space are: Among them, h k and h k-1 are the current and previous discretized input potential states, respectively, x k is the current discretization input; The selective state space output y is obtained by global convolution: in, is the structured convolution kernel.

7. The channel information feedback method for a large-scale MIMO system based on multi-scale feature fusion according to claim 4, It is characterized in that The convolution enhanced vector quantization model transforms the input codeword vector H of length M into c Use the convolutional dimension reduction module DownConv(·) to transform into M / 2 feature expansion vectors q s ; The vector quantization module VQ(·) is used for discretization, and the feature expansion vector q is obtained by nearest neighbor search. s Mapping to codebook The latent embedding vector e in i , the corresponding codebook index forms a bit stream, and the feature expansion vector q is reconstructed at the base station r =e i , where i = argmin m ||q s -e m ||, m∈{1,2,...,K}, and finally reconstruct the codeword through the convolutional dimension-upgrading module UpConv(·) The whole process is expressed as: Quantization model loss function L(H c ,e i )for: Where sg is the stop gradient operator and β is the weight coefficient; Reconstructed compressed codeword After the CFE module and decoder are connected in series to reconstruct the CSI, the total loss function L is: Where H a and are encoder input and decoder output respectively, ||·|| F is the Frobenius norm.

Images