Intelligent channel state information feedback method fusing physical model

Abstract

The invention discloses an intelligent channel state information feedback method fusing a physical model, and the method comprises the steps: generating a CSI data set under multiple scenes, and obtaining a multipath parameter tag corresponding to the CSI data set; a hierarchical perception encoder is combined with the global perception capability of large kernel convolution and the local feature extraction capability of small kernel convolution so as to realize enhancement of a multipath component region and extraction of multipath parameter features; recovering each multipath parameter by adopting a heterogeneous architecture decoder; a guiding fine tuning type two-stage training strategy is adopted, a whole feedback network constructed by a codec is firstly guided to learn a mapping relation between CSI and channel multipath parameters, and then fine tuning is performed through a physical multipath channel model to optimize CSI reconstruction performance. According to the method, efficient compression from high-dimensional CSI to low-dimensional multipath parameters and high-quality reconstruction of the CSI can be realized, the feedback overhead is reduced, the CSI reconstruction precision is improved, and the calculation complexity of the model is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication, in particular to a physical model integrated intelligent channel state information feedback method. BACKGROUND

[0002] With the continuous development of mobile communication technology, Massive Multiple-Input Multiple-Output (Massive MIMO) technology has become one of the core technologies of current communication systems. By configuring a large number of transmit antennas at the base station (BS) end, the channel capacity is improved by fully utilizing the spatial resources. The prerequisite for achieving the above advantages is that the base station end accurately obtains the downlink channel state information (CSI). Since the uplink and downlink frequency bands are different in the frequency division duplex (FDD) system, the channels do not have reciprocity, so the downlink CSI needs to be fed back to the BS by the user equipment (UE). However, the problem exists in the current massive MIMO scenario. The number of antennas increases sharply, and the feedback overhead increases rapidly. Traditional feedback schemes such as codebook-based or compressed sensing feedback methods face great challenges in practical applications.

[0003] In recent years, the rapid development of artificial intelligence (AI) technology has provided a new solution for CSI feedback. Existing methods treat the CSI channel as a two-dimensional image processing, and complete the end-to-end compression and reconstruction of CSI by constructing an autoencoder. However, this kind of method usually directly borrows advanced network architecture from the field of computer vision, and improves the performance by increasing the depth and complexity of the network, but ignores the rich physical meaning contained in the CSI data itself, that is, it is determined by the physical processes such as reflection and scattering of signals in a multipath propagation environment. This "black box" type of processing method leads to a lack of physical interpretability of the model. SUMMARY

[0004] The purpose of the present application is to provide a physical model integrated intelligent channel state information feedback method, which can realize the feedback and reconstruction of CSI and reduce the feedback overhead.

[0005] Technical solution: An intelligent channel state information feedback method fusing a physical model, which learns multipath parameters from a CSI channel matrix through constructing a CSI feedback neural network, wherein an encoder is deployed at a user end to compress a high-dimensional CSI channel matrix into a low-dimensional compressed code word containing parameter information; a decoder is deployed at a base station end to recover each path parameter from the compressed code word; finally, the base station end reconstructs downlink CSI through a physical multipath channel model using the recovered multipath parameters; comprising the following steps: S1, constructing a training input data set based on CSI sample data under multiple scenes, and constructing a training label data set based on the multipath parameter labels corresponding to the CSI sample data; S2, constructing a hierarchical perceptual encoder, using large kernel convolution for global perception to locate the multipath component area, and combining small kernel convolution for local feature enhancement to extract multipath parameter features and generate low-dimensional compressed code words; S3, constructing a heterogeneous architecture decoder composed of a main network and multiple subnetworks, the main network is used to extract multipath features in the compressed code word, and each subnetwork recovers each multipath parameter from the multipath features; S4, using a guided fine-tuning training strategy to train the CSI feedback neural network; using the multipath parameters output by the heterogeneous architecture decoder, calculating the downlink CSI according to the physical multipath channel model.

[0006] Further, the multipath parameters include the complex gain amplitude of each path , the complex gain phase of each path , the time delay of each path , and the corresponding base station antenna transmission angle of each path , wherein represents the index of each path; then the path parameter vector , , represents the total number of multipaths; the CSI is reconstructed by the multipath parameters through the physical multipath channel model, and the CSI matrix is obtained: , wherein, , represent the discrete Fourier transform matrix of the delay domain and the angle domain respectively, is the OFDM subcarrier number, is the number of base station antennas, represents the conjugate transpose, represents the multipath channel model, is the antenna steering vector, is the antenna spacing, is the signal wavelength, is the frequency domain response vector, is the subcarrier spacing, This indicates the matrix transpose.

[0007] Furthermore, the input to the hierarchical sensing encoder is the CSI matrix. The processing steps are as follows: First, a single-layer convolution is used to generate an intermediate feature map; then, an adaptive diamond-shaped small convolution kernel is generated using a two-branch large-kernel convolution structure to perform position-by-position convolution on the intermediate feature map; finally, the feature map after position-by-position convolution is straightened and output as compressed codewords through a fully connected layer. The fitting process of the hierarchical sensing encoder is as follows: , in, Represents encoder, These are the parameters of the encoder model.

[0008] Furthermore, the input to the heterogeneous architecture decoder is compressed codewords. The output is the recovered channel multipath parameters. The main network uses fully connected layers to increase the dimensionality of the input compressed codewords, reshapes them to the size of the CSI matrix through a reshaping layer, and then uses multiple convolutional layers for further feature extraction. The convolutional layers are stacked and connected through residual structures. The sub-network is a multi-branch fully connected layer structure, where each branch is used to recover a multipath parameter. Shallow fully connected branches are designed for parameters with relatively concentrated distributions, and deep fully connected branches are designed for parameters with relatively random distributions. The fitting process of the heterogeneous architecture decoder is as follows: , in, Represents decoder, These are the parameters of the decoder model.

[0009] Furthermore, in step S4, the mean square error between the standard channel parameter matrix and the prediction parameters of the heterogeneous architecture decoder is first used as the first loss function to guide the CSI feedback neural network to learn the mapping between CSI and multipath parameters. Subsequently, the normalized mean square error between the original CSI matrix and the CSI matrix reconstructed by the physical multipath channel model is used as the second loss function to fine-tune the entire CSI feedback neural network and optimize the CSI reconstruction performance.

[0010] Furthermore, the expression for the first loss function is as follows: , in, This represents the sample size of the training batch. For the standard parameter matrix, The parameter matrix predicted by the decoder. It is the Euclidean norm; The expression of the second loss function is as follows: , Wherein, is a standard channel matrix, represents the CSI matrix reconstructed by the multipath parameter output by the decoder .

[0011] Compared with the prior art, the present application has the following remarkable effects: 1. The present application takes the CSI channel matrix characterized by multipath parameters as input parameters, guides the neural network to learn the multipath parameter features from the downlink CSI, and further realizes the feedback reconstruction of the CSI; and takes the physical multipath channel model as the prior knowledge for network design and training, so that the reconstructed CSI after compression conforms to the physical law, and the physical interpretability of the model is enhanced; 2. The present application represents high-dimensional CSI with low-dimensional multipath parameters, and since the number of multipaths is much smaller than the number of antennas and carriers, the feedback overhead is greatly reduced; the network complexity of the hierarchical perceptual encoder and the heterogeneous architecture decoder designed in the present application is low, and the computational overhead is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 is a flowchart of the present application; Figure 2 is a core framework diagram of the intelligent CSI feedback method; Figure 3 is a neural network architecture diagram; Figure 4 is a flowchart of the guided fine-tuning training of the intelligent CSI feedback method. DETAILED DESCRIPTION

[0013] The present application will be further described in detail below in combination with the drawings and specific embodiments of the present application.

[0014] In the intelligent CSI feedback task, the existing work is limited to introducing advanced image compression reconstruction network structures in the field of computer vision, increasing the network complexity to improve the feedback performance. This image compression reconstruction network paradigm ignores the rich meaning of the physical process of signal multipath propagation contained in the CSI, and has poor physical interpretability. Therefore, it is necessary to design an intelligent channel state information feedback method fusing a physical model, guide the network to learn the physical characteristics of the CSI through the support of the physical model, and further reduce the feedback overhead.

[0015] As shown in Figure 1 , Figure 2 The intelligent channel state information feedback method fusing a physical model of the present application comprises the following steps: Step 1, based on CSI sample data under multiple scenes and its corresponding multipath parameter label respectively; In embodiments of the present application, the dataset consists of two parts: the CSI sample data is the angle-delay domain channel matrix as the training input dataset; and the multipath parameter label is the parameter set as the training label dataset, representing the total number of multipaths; and each CSI sample corresponds to a multipath parameter label one-to-one, while the CSI sample can be generated by the corresponding multipath parameter through the physical multipath channel model. The above multipath parameters include the complex gain amplitude of each path , the complex gain phase of each path , the time delay of each path , and the corresponding base station antenna transmission angle of each path , wherein represents the index of each path, so the path parameter vector is composed of 4 multipath parameters. The process of reconstructing the CSI by the multipath parameter through the physical multipath channel model to obtain the CSI matrix can be described as: (1) wherein, , represent the discrete Fourier transform (DFT) matrix of the delay domain and the angle domain respectively, is the orthogonal frequency division multiplexing system (OFDM) subcarrier number, is the number of base station antennas, represents the conjugate transpose, represents the physical multipath channel model, is the antenna steering vector, is the antenna spacing, is the signal wavelength, is the frequency domain response vector, is the subcarrier spacing, denotes the matrix transpose.

[0016] Step 2, design a hierarchical perception encoder, combine the advantages of large kernel convolution global perception and small kernel convolution local perception, use large kernel convolution to adaptively generate small kernel for convolution operation, enhance the multipath component region, extract multipath parameter features, and generate compressed code words; In embodiments of the present application, the input of the hierarchical perception encoder is the CSI matrix , and the output is the compressed code word The layered perception encoder utilizes the global perception function of large kernel convolution to determine the multipath component region, specifically, first, a single-layer convolution is used to generate an intermediate feature map; then an adaptive diamond-shaped small convolution kernel is generated by using a double-branch large kernel convolution structure, and the intermediate feature map is convolved position by position to enhance local feature extraction; finally, the feature map after position-by-position convolution is straightened, and compressed code words are output through a fully connected layer The activation function of the layered perception encoder is ReLU and LeakyReLU, and the fitting process is described as follows: (2) Wherein, represents the encoder, is the parameter of the encoder model.

[0017] Step 3, design a heterogeneous architecture decoder to recover each multipath parameter. The heterogeneous architecture decoder is composed of a main network and a subnetwork, the main network further extracts features from the compressed code words, and the subnetwork is designed according to the difficulty of recovering each multipath parameter.

[0018] In the embodiment of the present application, the input of the heterogeneous architecture decoder is the compressed code word , and the output is the recovered channel multipath parameter . The heterogeneous architecture decoder is composed of a main network and a subnetwork, the main network includes a fully connected layer, a reshaping layer, a convolution layer and a batch normalization layer, wherein the main network uses a single-layer fully connected layer to upgrade the dimension of the input compressed code word, reshapes it to the size of the CSI matrix through the reshaping layer, then uses multiple convolution layers for further feature extraction, and the convolution layer is connected through a residual structure stack; the subnetwork is designed as a multi-branch fully connected layer structure, each branch is used to recover one multipath parameter, and the depth of the fully connected layer of the branch network is designed according to the difficulty of recovering different multipath parameters, shallow fully connected branches are designed for parameters with relatively concentrated distribution, and deep fully connected branches are designed for parameters with relatively random distribution, the activation function of the network is ReLU and LeakyReLU, and the network fitting process is described as follows: (3) Wherein, represents the decoder, is the parameter of the decoder model.

[0019] Step 4, a guided fine-tuning training strategy is used to train the CSI feedback neural network, which is composed of a layered perception encoder network and a heterogeneous architecture decoder network; First, guide the entire CSI feedback neural network to learn the mapping relationship between CSI and multipath parameters; In the embodiment of the present application, the standard channel parameter matrix is used as the input of the heterogeneous architecture decoder, and the channel parameter matrix The mean square error between the and is taken as a cost function (first loss function) for training, the cost function is minimized, the Adam optimization algorithm is used to update the parameters of the heterogeneous architecture decoder, and the cost function is described as follows: (4) Wherein, represents the sample size of the training batch, is a standard parameter matrix, is a parameter matrix obtained by predicting the heterogeneous architecture decoder, is the Euclidean norm.

[0020] Secondly, the entire CSI feedback neural network is fine-tuned through a physical multipath channel model; The input CSI matrix and the output parameter The normalized mean square error between the CSI matrix obtained according to formula (1) is taken as a cost function (second loss function) for training, the cost function is minimized, the Adam optimization algorithm is used to update the parameters of the entire CSI feedback neural network, and the cost function is described as follows: (5) Wherein, is a standard channel matrix, represents the CSI matrix reconstructed by the multipath parameters output by the decoder .

[0021] Finally, the downlink CSI is calculated according to the physical multipath channel model described in formula (1) using the multipath parameters output by the heterogeneous architecture decoder.

[0022] The specific embodiments and drawings will be combined to further illustrate the detailed content of the present application.

[0023] In a large-scale MIMO system, a typical downlink transmission link is considered. The base station end is configured with a ULA, and the number of antennas is , and the user end is configured with a single antenna. The system uses an OFDM modulation scheme, the number of subcarriers is , the center frequency is 3.5 GHz, and the bandwidth is 50 MHz. The scene is an outdoor scene composed of two 40m wide vertical streets, the base station is 6 meters high and located on both sides of the street, and the user is randomly initialized in the street scene. According to the above settings, a data set is generated, and a single scene data set contains 20000 CSI samples and corresponding multipath parameter matrices, of which 16000 samples are used for training, 2000 samples are used for verification, and 2000 samples are used for testing. Specifically, the following steps are included: Step F1, as Figure 2This is a framework diagram of the intelligent CSI feedback method based on the fusion physical model according to an embodiment of the present invention. The CSI channel matrix can be derived from multipath parameters. Based on the characteristics of the representation, a neural network structure is designed to learn multipath parameters from the CSI channel matrix. The encoder is deployed at the user end to compress the high-dimensional CSI channel matrix into low-dimensional compressed codewords containing parameter information. The decoder is deployed at the base station to recover the path parameters from the compressed codewords. Finally, the base station can use the recovered multipath parameters to reconstruct the downlink CSI through the physical multipath channel model.

[0024] Step F2, as follows Figure 3 This diagram illustrates the neural network architecture of the intelligent CSI feedback method based on a fused physical model according to an embodiment of the present invention. The method designs a CSI feedback neural network model comprising a hierarchical perceptual encoder and a heterogeneous architecture decoder. The encoder is deployed at the user end, and the decoder is deployed at the base station end, to obtain reconstructed CSI samples. The input to the hierarchical perceptual encoder is the CSI matrix. The output is a compressed codeword. The system utilizes the global perception capability of large-kernel convolution to identify multipath component regions, and generates adaptive diamond-shaped small convolution kernels to enhance local feature extraction by applying them to the feature map. Finally, codewords are output through a fully connected layer. Heterogeneous architecture decoders extract compressed codewords Multipath parameters for channel recovery It consists of a main network and subnets. The main network includes convolutional layers and batch normalization layers, with the convolutional layers connected through residual structures. The subnets have network depths tailored to the recovery difficulty of different multipath parameters and consist of fully connected layers of varying numbers. The restoration design of shallow branch roads, The deep branches of the recovery design are used, and the activation functions of the network are ReLU and LeakyReLU.

[0025] Step F3, as follows Figure 4 This diagram illustrates the training process of the intelligent CSI feedback method based on a fused physical model according to an embodiment of the present invention. Training can be divided into two steps: bootstrapping and fine-tuning. The bootstrapping step uses a standard channel parameter matrix. The channel parameter matrix obtained from the decoder output The mean squared error between the two values ​​is used as the first loss function for training. To minimize this loss function, the Adam optimization algorithm is used to update the codec parameters. The cost function is described as follows: , in, This represents the sample size of the training batch. For the standard parameter matrix, The parameter matrix predicted by the decoder. is the Euclidean norm. The fine-tuning step uses the input CSI matrix and the output parameter The reconstructed CSI matrix The normalized mean square error between the reconstructed CSI matrix and the standard CSI matrix is taken as the second loss function, and the loss function is minimized to update the codec parameters using the Adam optimization algorithm, and the cost function is described as follows: , wherein, is the standard CSI matrix, represents the reconstructed CSI matrix of the multipath parameter output by the decoder , is the Euclidean norm.

[0026] In summary, according to the intelligent channel state information feedback method fusing a physical model proposed in the example of the present application, the characteristics that the CSI channel matrix can be characterized by the multipath parameters are used to guide the neural network to learn the multipath parameter features from the downlink CSI, and then the feedback reconstruction of the CSI is realized. The present application enhances the physical interpretability of the model by taking the physical multipath channel model as the prior knowledge of network design and training; at the same time, the method of representing high-dimensional CSI with low-dimensional multipath parameters greatly reduces the feedback overhead because the number of multipaths is much smaller than the number of antennas and carriers. In addition, compared with the defect of introducing advanced network structure in the image processing field to exchange performance with network complexity, the network complexity of the layered perception encoder and the heterogeneous architecture decoder designed in the present application is low, which has an advantage in computational overhead.

[0027] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms is not necessarily for the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, the skilled in the art can combine and combine the different embodiments or examples described in the present application and the features of the different embodiments or examples without contradiction.

[0028] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features limited by "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "N" is at least two, such as two, three, etc., unless otherwise specifically limited Any processes or methods described in the flowcharts or otherwise described herein can be understood as representing code modules, segments, or portions of code which include one or more executable instructions for implementing specific logic functions (or steps) of the application, and / or that the various processes or methods described in connection with the preferred embodiments of the application can be embodied in a computer readable medium (e.g., magnetic or optical), that is, any non-transitory medium of expression that can be capable of being employed as a computational medium for implementing the processes or methods described herein, and that can be accessed by a computer or an equivalent processing device.

Claims

1. A method for intelligent channel state information feedback that integrates a physical model, characterized in that, Multipath parameters are learned from the CSI channel matrix by constructing a CSI feedback neural network. An encoder is deployed at the user end to compress the high-dimensional CSI channel matrix into low-dimensional compressed codewords containing parameter information. A decoder is deployed at the base station to recover the path parameters from the compressed codewords. Finally, the base station uses the recovered multipath parameters to reconstruct the downlink CSI using a physical multipath channel model. The steps are as follows: S1, construct the training input dataset based on CSI sample data in multiple scenarios, and construct the training label dataset based on the multipath parameter labels corresponding to the CSI sample data; S2, construct a hierarchical perceptual encoder, use large kernel convolution for global perception to locate multipath component regions, and combine small kernel convolution for local feature enhancement, extract multipath parameter features and generate low-dimensional compressed codewords. S3, construct a heterogeneous architecture decoder consisting of a main network and multiple subnets. The main network is used to extract multipath features from compressed codewords, and each subnet recovers the multipath parameters from the multipath features. S4 employs a guided fine-tuning training strategy to train the CSI feedback neural network; it utilizes the multipath parameters output by the heterogeneous architecture decoder to calculate the downlink CSI based on the physical multipath channel model.

2. The intelligent channel state information feedback method based on the fused physical model according to claim 1, characterized in that, The multipath parameters include the complex gain amplitude of each path. Complex gain phase of each path Delay of each path The corresponding base station antenna transmission angle for each path ,in This represents the index of each path; therefore, the parameter vector for each path... , , Represents the total number of multipath paths; the CSI matrix is ​​obtained by reconstructing the CSI from the multipath parameters using a physical multipath channel model. : in, , These represent the discrete Fourier transform matrices in the time delay domain and the angle domain, respectively. The number of OFDM subcarriers, This refers to the number of base station antennas. Represents conjugate transpose. Represents a multipath channel model. This is the antenna steering vector. Antenna spacing, For the signal wavelength, It is the frequency domain response vector. For subcarrier spacing, This indicates the matrix transpose.

3. The intelligent channel state information feedback method based on the fused physical model according to claim 1, characterized in that, The input to the hierarchical sensing encoder is the CSI matrix. The processing steps are as follows: First, a single-layer convolution is used to generate an intermediate feature map; then, an adaptive diamond-shaped small convolution kernel is generated using a two-branch large-kernel convolution structure to perform position-by-position convolution on the intermediate feature map; finally, the feature map after position-by-position convolution is straightened and output as compressed codewords through a fully connected layer. The fitting process of the hierarchical sensing encoder is as follows: in, Represents encoder, These are the parameters of the encoder model.

4. The intelligent channel state information feedback method based on the fused physical model according to claim 3, characterized in that, The input to the heterogeneous architecture decoder is compressed codewords. The output is the recovered channel multipath parameters. The main network uses fully connected layers to increase the dimensionality of the input compressed codewords, reshapes them to the size of the CSI matrix through a reshaping layer, and then uses multiple convolutional layers for further feature extraction. The convolutional layers are stacked and connected through residual structures. The sub-network is a multi-branch fully connected layer structure, where each branch is used to recover a multipath parameter. Shallow fully connected branches are designed for parameters with relatively concentrated distributions, and deep fully connected branches are designed for parameters with relatively random distributions. The fitting process of the heterogeneous architecture decoder is as follows: in, Represents decoder, These are the parameters of the decoder model.

5. The intelligent channel state information feedback method based on the fused physical model according to claim 1, characterized in that, In step S4, the mean square error between the standard channel parameter matrix and the predicted parameters of the heterogeneous architecture decoder is first used as the first loss function to guide the CSI feedback neural network to learn the mapping between CSI and multipath parameters. Then, the normalized mean square error between the original CSI matrix and the CSI matrix reconstructed by the physical multipath channel model is used as the second loss function to fine-tune the entire CSI feedback neural network and optimize the CSI reconstruction performance.

6. The intelligent channel state information feedback method based on the fused physical model according to claim 1, characterized in that, The expression for the first loss function is as follows: in, This represents the sample size of the training batch. For the standard parameter matrix, The parameter matrix predicted by the decoder. It is the Euclidean norm; The expression for the second loss function is as follows: in, For the standard channel matrix, The CSI matrix representing the multipath parameter reconstruction of the decoder output. .

Images