WO2020220278A1

WO2020220278A1 - Channel estimation model training method and device

Info

Publication number: WO2020220278A1
Application number: PCT/CN2019/085230
Authority: WO
Inventors: 黄鸿基; 胡慧; 刘劲楠; 杨帆
Original assignee: 华为技术有限公司
Priority date: 2019-04-30
Filing date: 2019-04-30
Publication date: 2020-11-05
Also published as: CN113748614A; CN113748614B

Abstract

Embodiments of the present application provide a channel estimation model training method and device. The method comprises: converting a first channel matrix into codeword information, and reconstructing a channel matrix by using the codeword information, so as to obtain a second channel matrix; performing deep learning on a channel estimation model by using the first channel matrix and the second channel matrix, so as to obtain a trained channel estimation model, wherein the channel estimation model is constructed by using a deep neural network; obtaining a first signal transmitted by a terminal; performing channel estimation on the first signal by using the trained channel estimation model. Using the embodiments of the present application, errors of channel estimation can be reduced.

Description

Method and equipment for training channel estimation model

Technical field

This application relates to the field of communication technology, and in particular to a channel estimation model training method and device.

Background technique

The performance of a massive (Massive) input and output (multiple-input multiple output, MIMO) system largely depends on the availability of complete and effective channel state information (CSI) of the system, based on how the pilot symbols are used. Different, the current channel estimation algorithms can be roughly divided into the following three categories: pilot-assisted channel estimation, blind channel estimation and semi-blind channel estimation. The three channel estimation methods are introduced separately below:

1. Pilot-aided channel estimation adopts the strategy of mixing pilot and data frequency points in the symbol; for frequency domain estimation, first estimate the channel frequency response of each pilot point, and realize other pilots through interpolation calculations of different methods. Frequency channel frequency response estimation; for time domain estimation, first estimate the time domain separable gain, and then obtain the frequency domain channel transmission matrix for data detection through fast Fourier transform. Frequency domain estimation and time domain estimation generally adopt least square method or minimum mean square error method to realize channel estimation. The computational complexity of pilot-assisted channel estimation is relatively low, but the overhead from the pilot band reduces the spectral efficiency of the Massive MIMO system. 2. Blind channel estimation algorithms do not need to use training sequences or pilot signals for channel estimation. They mainly use some statistical characteristics of received and transmitted signals to achieve channel estimation. Such algorithms usually have relatively high computational complexity. 3. Semi-blind channel estimation mainly includes semi-blind estimation method based on subspace, semi-blind detection algorithm based on joint detection strategy and semi-blind estimation method assisted by adaptive filter. Among them, the main idea of the semi-blind estimation algorithm based on joint detection is to send fewer training sequences or pilots to obtain the initial value of the channel estimation, and based on this value, the decoder and the detector complete the channel estimation and iteratively. track. Because this method achieves better channel estimation performance under the premise of saving spectrum resources, it has received extensive attention from researchers. However, this type of method also has the problem of high computational complexity. It can be seen that pilot-assisted channel estimation, blind channel estimation, and semi-blind channel estimation either have the problem of low spectrum efficiency or high computational complexity. In view of this situation, the technology in this field proposes to integrate deep learning based on The entire Massive MIMO system is regarded as a black box and performs end-to-end learning to achieve unsupervised channel estimation. The structure of the deep learning network used in the unsupervised channel estimation process is shown in Figure 1. In specific implementation, the channel matrix is independently decomposed into a gain matrix and a steering vector for estimation, for example, each angle of arrival is fixed to obtain the corresponding received signal. Then the received signal and the angle of arrival are used as samples for training to realize the estimation of the steering vector through a deep neural network (DNN), and then the gain matrix is estimated by the same method. This method does not need to rely on pilots, so the spectrum efficiency is not reduced, but because its channel estimation is based on the gain matrix and steering vector estimation, the result of channel estimation has additional errors.

How to minimize the channel estimation error under the premise of ensuring relatively high spectrum efficiency and relatively low computational complexity is a technical problem being studied by those skilled in the art.

Summary of the invention

The embodiments of the present application disclose a channel estimation model training method and equipment, which can reduce channel estimation errors.

In a first aspect, an embodiment of the present application provides a channel estimation model training method, which includes: converting a first channel matrix into codeword information, and reconstructing a channel matrix using the codeword information to obtain a second channel matrix; The first channel matrix and the second channel matrix perform in-depth learning of the channel estimation model to obtain the channel estimation model after training, wherein the channel estimation model is constructed based on a deep neural network; and the first channel transmitted by the terminal is obtained. Signal; using the trained channel estimation model to perform channel estimation on the first signal.

By performing the above method, the first channel matrix is converted into codeword information, and then the matrix is reconstructed according to the codeword information to obtain the second channel matrix, and then the parameters of the channel estimation model are adjusted through deep learning to reduce the second channel matrix and the first channel matrix Therefore, the trained channel estimation model can be obtained; the subsequently trained channel estimation model can perform channel estimation according to the input transmission signal. Since the estimation of the intermediate steering vector and the gain matrix is not introduced in this process, the calculation pressure is significantly reduced; in addition, because this application only estimates the channel matrix, not the steering vector and gain matrix obtained by estimation. Channel estimation avoids information distortion in the intermediate links, so the error of the estimation result in the embodiment of the present application is smaller.

With reference to the first aspect, in a first possible implementation manner of the first aspect, the converting the first channel matrix into codeword information, and reconstructing the channel matrix using the codeword information to obtain the second channel matrix includes : Convert the real part and the imaginary part of the first channel matrix into two real vectors; convert the two real vectors into codeword information; use the codeword information to reconstruct the channel matrix to obtain the second channel matrix. .

With reference to the first aspect or any one of the foregoing possible implementation manners of the first aspect, in a second possible implementation manner of the first aspect, the use of the codeword information to reconstruct a channel matrix to obtain a second channel matrix , Including: extracting a second signal and a first white noise from the codeword information, where the second signal is a transmission signal; reconstructing a channel matrix according to the second signal and the first white noise to obtain the first Two-channel matrix.

With reference to the first aspect, or any one of the foregoing possible implementation manners of the first aspect, in a third possible implementation manner of the first aspect, the first channel matrix is converted into codeword information, and the Before reconstructing the channel matrix with codeword information to obtain the second channel matrix, the method further includes: generating the first channel matrix according to the third signal, the fourth signal, and the second white noise, wherein the third signal is sent by the network device Signal; the fourth signal is a signal obtained when the terminal receives the third signal, and the second white noise is white noise fed back by the terminal.

With reference to the first aspect, or any of the foregoing possible implementation manners of the first aspect, in a fourth possible implementation manner of the first aspect, the channel estimation model includes an encoding network and a decoding network, wherein the encoding The network includes a first convolutional layer and a first fully connected layer; the decoding network includes a second fully connected layer, a second convolutional layer, a maximum pooling layer, a first insertion layer, a second insertion layer, a deep connection module, Global pooling layer and third convolutional layer; the first channel matrix is used as the input of the first convolutional layer of the coding network, and the output of the first convolutional layer is used as the first global The input of the connection layer, the output of the first fully connected layer is used as the input of the second fully connected layer of the decoding network, and the output of the second fully connected layer is used as the input of the maximum pooling layer , The output of the maximum pooling layer is used as the input of the first insertion layer, the output of the first insertion layer is used as the input of the second insertion layer, and the output of the second insertion layer is used As the input of the deep connection module, the output of the deep connection module is used as the input of the full pooling layer, and the output of the full pooling layer is used as the input of the third convolutional layer, The third convolutional layer is used to generate the second channel matrix. The deep learning network with this structure has stronger learning ability and fast training speed, and the error of the channel matrix estimated by the channel estimation model at the training place is smaller.

With reference to the first aspect or any of the foregoing possible implementation manners of the first aspect, in a fifth possible implementation manner of the first aspect, the channel estimation model includes an encoding network and a decoding network, wherein the encoding The network includes a first convolutional layer and a first fully connected layer; the decoding network includes a second fully connected layer, a second convolutional layer, a maximum pooling layer, a third convolutional layer, a fourth convolutional layer, and a fifth Convolutional layer, sixth convolutional layer, and seventh convolutional layer; the first channel matrix is used as the input of the first convolutional layer of the coding network, and the output of the first convolutional layer is used as The input of the first fully connected layer, the output of the first fully connected layer is used as the input of the second fully connected layer of the decoding network, and the output of the second fully connected layer is used as the maximum The input of the pooling layer, the output of the maximum pooling layer is used as the input of the third convolutional layer, and the output of the third convolutional layer is used as the input of the fourth convolutional layer, so The output of the fourth convolutional layer is used as the input of the fifth convolutional layer, the output of the fifth convolutional layer is used as the input of the sixth convolutional layer, and the sixth convolutional layer The output of is used as the input of the seventh convolutional layer, and the seventh convolutional layer is used to generate the second channel matrix. The structure of this deep learning network still has stronger learning capabilities, faster training speed, smaller errors, and smaller computational complexity than existing technologies in the smaller massive MIMO system scenario.

With reference to the first aspect or any of the foregoing possible implementation manners of the first aspect, in a sixth possible implementation manner of the first aspect, the channel estimation model includes an encoding network and a decoding network, wherein the encoding The network includes a first convolutional layer and a first fully connected layer; the decoding network includes a second fully connected layer, a residual network, and a third fully connected layer; the first channel matrix is used as the first fully connected layer of the encoding network The input of a convolutional layer, the output of the first convolutional layer is used as the input of the first fully connected layer, and the output of the first fully connected layer is used as the second fully connected of the decoding network Layer input, the output of the second fully connected layer is used as the input of the residual network, the output of the residual network is used as the input of the third fully connected layer, the third fully connected The layer is used to generate the second channel matrix. The deep learning network of this structure ensures that the network can avoid high complexity on the premise of deepening the model depth through the method of cutting off the connection of the features.

With reference to the first aspect, or any of the foregoing possible implementation manners of the first aspect, in the seventh possible implementation manner of the first aspect, the channel estimation model adopts a compressed sensing mechanism. Therefore, the dimensionality of the deep learning network is reduced, thereby reducing the computational complexity.

With reference to the first aspect, or any of the foregoing possible implementation manners of the first aspect, in a seventh possible implementation manner of the first aspect, the first convolutional layer is configured to use the first channel matrix The real part and the imaginary part are converted into two real vectors; the first fully connected layer is used to convert the two real vectors into the codeword information.

In a second aspect, an embodiment of the present application provides a channel estimation model training device. The device includes a processor and a memory. The memory is used to store program instructions and model parameters. The processor is used to call the program instructions and model parameters. To perform the following operations: convert the first channel matrix into codeword information, and use the codeword information to reconstruct the channel matrix to obtain the second channel matrix; use the first channel matrix and the second channel matrix to estimate the channel The model performs deep learning to obtain a trained channel estimation model, where the channel estimation model is constructed based on a deep neural network; the first signal transmitted by the terminal is obtained; and the trained channel estimation model is used to calculate the first signal Channel estimation for a signal.

In the above equipment, the first channel matrix is converted into codeword information, and the matrix is reconstructed according to the codeword information to obtain the second channel matrix, and then the parameters of the channel estimation model are adjusted through deep learning to reduce the second channel matrix and the first channel matrix Therefore, the trained channel estimation model can be obtained; the subsequently trained channel estimation model can perform channel estimation according to the input transmission signal. Since the estimation of the intermediate steering vector and the gain matrix is not introduced in this process, the calculation pressure is significantly reduced; in addition, because this application only estimates the channel matrix, not the steering vector and gain matrix obtained by estimation. Channel estimation avoids information distortion in the intermediate links, so the error of the estimation result in the embodiment of the present application is smaller.

With reference to the second aspect, in a first possible implementation manner of the second aspect, the first channel matrix is converted into codeword information, and the codeword information is used to reconstruct the channel matrix to obtain the second channel matrix. The steps are: converting the real part and the imaginary part of the first channel matrix into two real vectors; converting the two real vectors into codeword information; using the codeword information to reconstruct the channel matrix to obtain the second channel matrix.

With reference to the second aspect, or any of the foregoing possible implementation manners of the second aspect, in a second possible implementation manner of the second aspect, the use of the codeword information to reconstruct a channel matrix to obtain a second channel matrix , Specifically: extracting a second signal and a first white noise from the codeword information, where the second signal is a transmission signal; reconstructing a channel matrix according to the second signal and the first white noise to obtain The second channel matrix.

With reference to the second aspect, or any of the foregoing possible implementation manners of the second aspect, in a third possible implementation manner of the second aspect, the first channel matrix is converted into codeword information, and the The codeword information reconstructs the channel matrix to obtain the second channel matrix. The processor is further configured to generate the first channel matrix according to the third signal, the fourth signal and the second white noise, wherein the third signal is a network A signal sent by the device; the fourth signal is a signal obtained when the terminal receives the third signal, and the second white noise is white noise fed back by the terminal.

With reference to the second aspect, or any one of the foregoing possible implementation manners of the second aspect, in a fourth possible implementation manner of the second aspect, the channel estimation model includes an encoding network and a decoding network, wherein the encoding The network includes a first convolutional layer and a first fully connected layer; the decoding network includes a second fully connected layer, a second convolutional layer, a maximum pooling layer, a first insertion layer, a second insertion layer, a deep connection module, Global pooling layer and third convolutional layer; the first channel matrix is used as the input of the first convolutional layer of the coding network, and the output of the first convolutional layer is used as the first global The input of the connection layer, the output of the first fully connected layer is used as the input of the second fully connected layer of the decoding network, and the output of the second fully connected layer is used as the input of the maximum pooling layer , The output of the maximum pooling layer is used as the input of the first insertion layer, the output of the first insertion layer is used as the input of the second insertion layer, and the output of the second insertion layer is used As the input of the deep connection module, the output of the deep connection module is used as the input of the full pooling layer, and the output of the full pooling layer is used as the input of the third convolutional layer, The third convolutional layer is used to generate the second channel matrix. The deep learning network with this structure has stronger learning ability and fast training speed, and the error of the channel matrix estimated by the channel estimation model at the training place is smaller.

With reference to the second aspect, or any of the foregoing possible implementation manners of the second aspect, in a fifth possible implementation manner of the second aspect, the channel estimation model includes an encoding network and a decoding network, wherein the encoding The network includes a first convolutional layer and a first fully connected layer; the decoding network includes a second fully connected layer, a second convolutional layer, a maximum pooling layer, a third convolutional layer, a fourth convolutional layer, and a fifth Convolutional layer, sixth convolutional layer, and seventh convolutional layer; the first channel matrix is used as the input of the first convolutional layer of the coding network, and the output of the first convolutional layer is used as The input of the first fully connected layer, the output of the first fully connected layer is used as the input of the second fully connected layer of the decoding network, and the output of the second fully connected layer is used as the maximum The input of the pooling layer, the output of the maximum pooling layer is used as the input of the third convolutional layer, and the output of the third convolutional layer is used as the input of the fourth convolutional layer, so The output of the fourth convolutional layer is used as the input of the fifth convolutional layer, the output of the fifth convolutional layer is used as the input of the sixth convolutional layer, and the sixth convolutional layer The output of is used as the input of the seventh convolutional layer, and the seventh convolutional layer is used to generate the second channel matrix. The structure of this deep learning network still has stronger learning capabilities, faster training speed, smaller errors, and smaller computational complexity than existing technologies in the smaller massive MIMO system scenario.

With reference to the second aspect, or any one of the foregoing possible implementation manners of the second aspect, in a sixth possible implementation manner of the second aspect, the channel estimation model includes an encoding network and a decoding network, wherein the encoding The network includes a first convolutional layer and a first fully connected layer; the decoding network includes a second fully connected layer, a residual network, and a third fully connected layer; the first channel matrix is used as the first fully connected layer of the encoding network The input of a convolutional layer, the output of the first convolutional layer is used as the input of the first fully connected layer, and the output of the first fully connected layer is used as the second fully connected of the decoding network Layer input, the output of the second fully connected layer is used as the input of the residual network, the output of the residual network is used as the input of the third fully connected layer, the third fully connected The layer is used to generate the second channel matrix. The deep learning network of this structure ensures that the network can avoid high complexity on the premise of deepening the model depth through the method of cutting off the connection of the features.

With reference to the second aspect, or any of the foregoing possible implementation manners of the second aspect, in the seventh possible implementation manner of the second aspect, the channel estimation model adopts a compressed sensing mechanism.

With reference to the first aspect, or any one of the foregoing possible implementation manners of the first aspect, in an eighth possible implementation manner of the first aspect, the first convolutional layer is configured to use the first channel matrix The real part and the imaginary part are converted into two real vectors; the first fully connected layer is used to convert the two real vectors into the codeword information.

In a third aspect, an embodiment of the present application provides a computer-readable storage medium having program instructions stored in the computer-readable storage medium. When the program instructions run on a processor, the first aspect or the first aspect is implemented. Any possible implementation of the aspect described in the method.

In a fourth aspect, the embodiments of the present application provide a computer program product, which, when the computer program product runs on a processor, implements the method described in the first aspect or any possible implementation of the first aspect.

By implementing the embodiment of this application, the first channel matrix is converted into codeword information, and then the second channel matrix is obtained by reconstructing the matrix according to the codeword information, and then adjusting the parameters of the channel estimation model through deep learning to reduce the second channel matrix and the first channel matrix. The relationship between the channel matrices can be used to obtain the trained channel estimation model; the subsequently trained channel estimation model can perform channel estimation according to the input transmission signal. Since the estimation of the intermediate steering vector and the gain matrix is not introduced in this process, the calculation pressure is significantly reduced; in addition, because this application only estimates the channel matrix, not the steering vector and gain matrix obtained by estimation. Channel estimation avoids information distortion in the intermediate links, so the error of the estimation result in the embodiment of the present application is smaller.

Description of the drawings

FIG. 1 is a schematic diagram of the structure of a deep learning network in the prior art provided by this application;

FIG. 2 is a schematic diagram of a scene of a communication system provided by an embodiment of the present application;

3 is a schematic flowchart of a channel estimation model training method provided by an embodiment of the present application;

FIG. 4 is a schematic structural diagram of a deep learning network provided by an embodiment of the present application;

FIG. 5 is a schematic structural diagram of a deep learning network provided by an embodiment of the present application;

Fig. 6 is a schematic structural diagram of a deep learning network provided by an embodiment of the present application;

Fig. 7 is a schematic structural diagram of a device provided by an embodiment of the present application.

Detailed ways

The embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

Please refer to FIG. 2, which is a schematic diagram of a scene of a communication system 200 provided by an embodiment of the present application. The communication system 200 includes a network device 201 and a terminal 202.

The network device 201 may be a base station, and the base station may be used to communicate with one or more terminals, and may also be used to communicate with one or more base stations with partial terminal functions (for example, communication between a macro base station and a micro base station). The base station can be the Base Transceiver Station (BTS) in the Time Division Synchronous Code Division Multiple Access (TD-SCDMA) system, or the Evolutional Node B (Evolutional Node B) in the LTE system , ENB), as well as base stations in 5G systems and New Air Interface (NR) systems. In addition, the base station may also be an access point (Access Point, AP), a transmission node (Trans TRP), a central unit (Central Unit, CU) or other network entities, and may include some or all of the functions of the above network entities . Optionally, considering that the network device 201 may face greater computing pressure, a server or a server cluster can be deployed for it to separately provide computing capabilities for the network device 201. At this time, the server or server cluster It can be regarded as a part of the network device 201.

The terminals 202 may be distributed in the entire wireless communication system 200, and may be stationary or mobile, and the number thereof is usually multiple. The terminal 202 may include handheld devices with wireless communication functions (for example, mobile phones, tablets, palmtop computers, etc.), vehicle-mounted devices (for example, automobiles, bicycles, electric vehicles, airplanes, ships, etc.), and wearable devices (for example, smart watches). (Such as iWatch, etc.), smart bracelets, pedometers, etc.), smart home equipment (for example, refrigerators, TVs, air conditioners, electric meters, etc.), smart robots, workshop equipment, other processing equipment that can be connected to wireless modems, and various Various forms of user equipment (User Equipment, UE), mobile station (Mobile station, MS), terminal (terminal), terminal equipment (Terminal Equipment), etc.

In the communication system 200, multiple transmitting antennas are deployed at the transmitting end (such as a network device), and multiple receiving antennas are deployed at the receiving end (such as a terminal), forming a Massive MIMO system.

Please refer to FIG. 3. FIG. 3 is a channel estimation model training method provided by an embodiment of the present application. The method includes but is not limited to the following steps:

Step S301: The network device generates a first channel matrix.

Specifically, the third signal is a signal sent by the network device in a specific direction; the fourth signal is a signal received by the terminal from the specific direction, and the second white noise is the terminal Feedback white noise.

For example, if the network device has N _t transmit antennas, N _s is the number of orthogonal frequency division multiplexing OFDM carriers, which is equal to the dimension of the sparse precoding vector. Then considering that most elements of the first channel matrix H are close to 0, especially in the delay domain, due to the delay of multipath arrival in a specific time slot, only the N _f row contains non-zero values. Therefore, the first channel matrix H can be reduced to a matrix of N _f ×N _t . For the first channel matrix H, it can be generated as follows:

In the first step, according to the received signal y _j at the jth subcarrier, the precoding vector v _j used for subcarrier power allocation at the jth subcarrier, the transmitted signal x _{j at} the jth subcarrier, and the jth subcarrier White noise z _j , determine the channel vector at the jth subcarrier

Among them, j takes a positive integer between 1 and N _f in turn, the transmitted signal x _j at the jth subcarrier is the signal sent by the network device in a specific direction, that is, the third signal; the received signal at the jth subcarrier y _j The signal received by the terminal from the specific direction, that is, the fourth signal, is fed back to the network device by the terminal after being acquired, and the precoding vector v _j at the jth subcarrier is based on the transmission power of each subcarrier The predefined quantity, the white noise z _j at the jth subcarrier, that is, the second white noise can also be fed back to the network device by the terminal, and the relationship between the above quantities is as shown in formula (1).

The second step, according to the channel vector at each subcarrier

Get the conjugate matrix

among them

The third step is to use formula (2) to compare the conjugate matrix

Perform dual DFT transformation to obtain the first channel matrix H.

The above three steps describe how to obtain the first channel matrix H of a batch of (patch) signaling. The embodiment of the present application needs to obtain the first channel matrix H of each patch signaling in multiple patch signaling through the same principle. For example, the obtained signaling matrices of M patch signaling are H ₁ , H ₂ , H ₃ , ..., _Hi , H _i+1 , ..., H _M. The acquired M first channel matrices need to be input into the channel estimation model as samples for training, so as to obtain an ideal channel estimation model.

The channel estimation model may be a deep learning network (also called a deep neural network), the deep learning network includes an encoder network and a decoder network, where the encoding network is used to obtain codeword information according to the first channel matrix , And the decoding network is used to reconstruct the matrix according to the codeword information; the first channel matrix H _{i of} each batch (patch) can be used as the input of the deep learning network, and correspondingly, the output of the deep learning network is the reconstruction matrix

The reconstruction matrix can be called the second channel matrix, and the parameter set θ={θ _en , θ _de } represents the data set of the encoding network and the decoding network, respectively. Optionally, in this embodiment of the present application, the decoding network may use a compressed sensing mechanism to reduce the dimension of the network, thereby reducing computational complexity. Optionally, an inception layer may be deployed in the decoding network, which includes convolution kernels of different sizes, reduces the complexity of the network through segmentation, and improves the performance of the network through splicing (such as learning ability, accuracy, etc.).

Three optional structures of the deep learning network are provided below to facilitate understanding.

The first one, as shown in Figure 4:

The coding network includes two neural network layers and the activation functions of these two neural network layers are both Rectified Linear Unit (ReLU). Each layer introduces a batch normalization mechanism. The first layer It is a convolutional layer (may be referred to as the first convolutional layer) with a filter size of 3×3 and a step size of 2. This layer is used to obtain the real and imaginary parts of the input first channel matrix H. The second layer is a fully connected layer (may be called the first fully connected layer), and its width is related to the compression rate of compressed sensing. Suppose the length and width of the network and the number of feature maps are a, b, and c respectively, and the compression rate is r, then the number of neurons in this layer is (a×b×c)/r. This layer converts the real and imaginary parts of the first channel matrix H into two real vectors, which reflect the characteristics of the channel (for example, antenna position, communication link fading coefficient, angle of arrival gain, etc.). The codeword information s is further generated according to these two real vectors.

Decoding network, including:

A fully connected layer (may be called a second fully connected layer): the width of this layer is consistent with the dimension of the first channel matrix H, that is, N _f ×N _t .

A convolutional layer (may be called the second convolutional layer): contains 8 3×3 filters, step size=2, s in Figure 4 represents the step size stride, using a zero padding mechanism. This layer generates 8 feature maps.

A max pooling layer: contains 8 3×3 filters, step size=2.

An inception layer (may be called the first insertion layer): 8 3×3 filters, step size=1; 8 5×5 filters, step size=1; one 3×3 filter Detector, max pooling layer with step=1. This layer adopts zero padding mechanism.

An insertion (inception) layer (can be called the second insertion layer): 8 3×3 filters, step size=1; 24 1×1 filters, step size=1, this layer uses zero padding mechanism . This layer generates 32 feature maps.

A deep connection (DepthConcat) module.

A global pooling layer (average pooling) layer: contains 8 3×3 filters with a step size=1.

A convolutional layer (can be called the third convolutional layer): Contains two 3×3 filters, step size = 2, using zero padding mechanism, this layer is used to reconstruct the second channel matrix

Optionally, all neural network layers of the decoding network select the rectified linear unit ReLU as the activation function.

The data processing flow in the above-mentioned deep neural network is: input the sample (the first channel matrix _Hi ) into the first convolutional layer of the Encoder of the encoding network, and the first convolutional layer passes the information flow to the first full Connection layer. The first fully connected layer passes the information flow to the fully connected layer at the decoder network Decoder; then, the fully connected layer passes the information flow to the max pool layer; the output of the max pool is passed to the first inception layer; the first The output of inception is passed to the second inception layer. The output of the second inception layer is passed to the DepthConcate module; then, the DepthConcat module passes the output to the average pooling layer, and passes the output of this layer to the last convolutional layer, which gets the output of the network, The second channel matrix

The second, as shown in Figure 5:

In the case where the massive MIMO system is relatively small (for example, the transmit antenna is between 0-64 and the number of users is between 0-16), the difference from the deep learning network shown in Figure 4 is that in Figure 5, Figure 4 The inception layer of the decoding network is replaced with a common convolutional layer. In order to reduce the complexity of the model, a small convolution kernel can be used to design a convolution layer (3×3). For example, the deep neural network includes an encoding network and a decoding network, wherein the encoding network includes a first convolutional layer and a first fully connected layer; the decoding network includes a second fully connected layer, a second convolutional layer, The largest pooling layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, the sixth convolutional layer, and the seventh convolutional layer; the first channel matrix is used as the first channel of the coding network The input of a convolutional layer, the output of the first convolutional layer is used as the input of the first fully connected layer, and the output of the first fully connected layer is used as the second fully connected of the decoding network The output of the second fully connected layer is used as the input of the maximum pooling layer, the output of the maximum pooling layer is used as the input of the third convolutional layer, and the third The output of the convolutional layer is used as the input of the fourth convolutional layer, the output of the fourth convolutional layer is used as the input of the fifth convolutional layer, and the output of the fifth convolutional layer is used As the input of the sixth convolution layer, the output of the sixth convolution layer is used as the input of the seventh convolution layer, and the seventh convolution layer is used to generate the second channel matrix . The experimental results show that although the performance is slightly lower than that of the embodiment shown in FIG. 4, it still has a better performance than other methods in a smaller-scale massive MIMO system scenario, and the computational complexity is lower. In addition, all neural network layers of the decoding network can use ReLU as the activation function.

The third type, as shown in Figure 6:

For larger-scale massive MIMO systems (for example, a large-scale millimeter wave scene usually has 256-512 transmitting antennas and 64-128 users), it is very complicated, so the network structure must be deepened. But generally speaking, the increase in the depth of the neural network will bring about a sharp increase in complexity. Therefore, as shown in Figure 6, compared with Figure 4, the coding network remains unchanged. A residual network is introduced to replace the decoding network with other structures except the first fully connected layer. On the first layer, a fully connected layer with a dimension of N _f ×N _t is designed. For example, the deep neural network includes an encoding network and a decoding network, wherein the encoding network includes a first convolutional layer and a first fully connected layer; the decoding network includes a second fully connected layer, a residual network, and a third Fully connected layer; the first channel matrix is used as the input of the first convolutional layer of the coding network, the output of the first convolutional layer is used as the input of the first fully connected layer, the The output of the first fully connected layer is used as the input of the second fully connected layer of the decoding network, the output of the second fully connected layer is used as the input of the residual network, and the output of the residual network Used as an input of the third fully connected layer, and the third fully connected layer is used to generate the second channel matrix. The method of cutting off connections through features ensures that the network can avoid high complexity while deepening the model depth. According to the article K. He, X. Zhang, S. Ren, and J. Sun, "Deep Residual Learning for Image Recognition," in Proc.2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, 2016 , pp.770-778. As shown in the results, this structure achieves better performance than the traditional convolutional network layer, so it can improve the performance of the application in a larger-scale massive MIMO scenario.

Step S302: The network device converts the first channel matrix into codeword information, and uses the codeword information to reconstruct the channel matrix to obtain a second channel matrix.

Specifically, the real part and the imaginary part of the first channel matrix are converted into two real vectors, and the two real vectors are converted into codeword information. Here, the real part and the imaginary part of the first channel matrix can be specifically converted by the encoding network. The imaginary part is converted into two real vectors, and the two real vectors are converted into codeword information.

Here, the decoding network can use the codeword information to reconstruct the channel matrix to obtain the second channel matrix. Optionally, the decoding network extracts the transmission signal and white noise from the codeword information, and the extracted transmission signal is called the first Second signal, the white noise extracted here is called the first white noise. It can be understood that when the parameter set of the decoding network changes due to training iterations, it will still extract the quantity used to characterize the second signal and the first white noise. One amount of white noise, but the specific values of the extracted second signal and first white noise may change. Optionally, in addition to extracting the second signal and the first white noise from the codeword information, the decoding network may also extract other characteristic information, such as channel fading coefficient, channel noise, etc., and what other information has been extracted? There are no restrictions. After the second signal and the first white noise are extracted, a channel matrix is reconstructed according to the second signal and the first white noise to obtain a second channel matrix.

Step S303: The network device uses the first channel matrix and the second channel matrix to perform in-depth learning of the channel estimation model to obtain a trained channel estimation model.

Specifically, a loss function can be introduced, which is used to constrain the deviation between the second channel matrix and the first channel matrix. If the deviation between the second channel matrix and the first channel matrix does not satisfy the constraint condition of the loss function, It is necessary to continue the iterative training. Optionally, the Stochastic Gradient Descent (SGD) algorithm can be used for iterative training. When the deviation between the second channel matrix and the first channel matrix meets the constraints of the loss function , The channel estimation model at this time is the channel estimation model after training. Optionally, the loss function l(Θ) used in the embodiment of the present application is as shown in formula (3).

In formula (3), M represents the total number of samples, specifically,

Represents the reconstruction matrix obtained from the m-th sample reconstruction, that is, the second channel matrix, and H _m represents the m-th first channel matrix among the M samples. The loss function l(Θ) is based on the idea of mean square error, which can minimize the error between the first channel matrix and the second channel matrix.

Step S304: The network device obtains the first signal transmitted by the terminal.

Specifically, the signal transmitted by the terminal is called the first signal, and accordingly, the network device receives the signal transmitted by the terminal.

Step S305: The network device uses the trained channel estimation model to perform channel estimation on the first signal.

It can be understood that the codeword information is generated according to the first channel matrix, the second signal and the first white noise are extracted from the codeword information, the channel matrix is reconstructed according to the second signal and the first white noise, and the The channel estimation model performs iterative training operations, which is equivalent to summarizing the relationship between the transmitted signal and the channel matrix. This relationship is described by the trained channel estimation model, so the first signal obtained is input to the trained channel The estimation model can then perform channel estimation. It should be noted that the second signal and the first signal here are essentially transmission signals, but we call the transmission signal used for training the second signal, and the transmission signal used in the actual estimation process as the first signal. In the same way, the first white noise and the second white noise here are essentially white noise, but we call the white noise used to calculate the training sample (ie the first channel matrix) the second white noise, which is extracted from the training sample The white noise is the first white noise.

In order to verify the performance of the above-trained channel estimation model, experiments were carried out on an outdoor massive MIMO system and an indoor massive MIMO system. The first deep learning network architecture described above was used. The specific parameter settings are as follows:

Outdoor massive MIMO system: equipped with 128 antennas, 32 single-antenna users, 100 propagation paths, subcarriers N _s =1024, N _f =32, and an area width of 400m;

Indoor massive MIMO system: equipped with 32 antennas, 16 single-antenna users, 6 propagation paths, subcarriers N _s ＝＝1024, N _f ＝32, and the area width is 20m.

At the same time, the normalized mean square error (NMSE) is used as the evaluation standard, and the NMSE calculation method is as shown in formula (4).

In formula (4), H is the actual channel matrix,

Is the channel matrix estimated by the channel estimation model.

Table 1-NMSE performance results

It can be seen from Table 1 that two methods based on deep learning (the traditional deep neural network method (conventional deep neural network, C-DNN) that decomposes the channel matrix into steering vectors and gain matrices for independent estimation, C-DNN and direct pair Channel matrix for channel estimation compressed sensing deep neural network method (compress sensing deep neural network, CS-DNN) and other non-deep learning methods (requires minimum absolute shrinkage and selection operation of pilots (least absolute shrinkage and selection operator, LASSO) and approximate message passing method (approximate message passing, AMP) achieve better performance. At the same time, the embodiment of the application has a lower NMSE than other methods under different compression ratios. Although the VTC-DNN method achieves a better NMSE in an indoor scene when the compression ratio is 0.03125, on the whole, the CS-DNN method in the embodiment of the present application achieves better channel estimation performance. In addition, it can be seen from the comparison of the average running time that the embodiment of the present application requires less running time and lower computational complexity.

In the method shown in Figure 3, the first channel matrix is converted into codeword information, and the matrix is reconstructed according to the codeword information to obtain the second channel matrix, and then the parameters of the channel estimation model are adjusted through deep learning to reduce the second channel matrix and The relationship between the first channel matrix is obtained, and the trained channel estimation model can be obtained; the channel estimation model after the training can perform channel estimation according to the input transmission signal. Since the estimation of the intermediate steering vector and the gain matrix is not introduced in this process, the calculation pressure is significantly reduced; in addition, because this application only estimates the channel matrix, not the steering vector and gain matrix obtained by estimation. Channel estimation avoids information distortion in the intermediate links, so the error of the estimation result in the embodiment of the present application is smaller.

The foregoing describes the method of the embodiment of the present application in detail, and the device of the embodiment of the present application is provided below.

Referring to FIG. 7, FIG. 7 shows a network device (ie, a channel estimation device) 700 provided by some embodiments of the present application. It is understandable that the network equipment 700 can be implemented as central office (CO) equipment, multi-dwelling unit (MDU), multi-merchant unit (MTU), digital subscriber line access multiplexer (DSLAM), multi-service access node (MSAN) ), Optical Network Unit (ONU), etc. As shown in FIG. 7, the network device 700 may include: one or more device processors 701, a memory 702, a transmitter 704, and a receiver 705. These components can be connected to the processor. among them:

The processor 701 may be implemented as one or more central processing unit (CPU) chips, cores (for example, multi-core processors), field programmable gate arrays (FPGA), application specific integrated circuits (ASIC), and/or digital signal processors (DSP), and/or can be part of one or more ASICs. The processor 701 may be configured to execute any of the solutions described in the above application embodiments, including data transmission methods. The processor 701 may be implemented by hardware or a combination of hardware and software.

The memory 702 is coupled with the processor 701, and is configured to store various software programs and/or groups of program instructions. Specifically, the memory 702 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic disk storage devices, flash memory devices, or other non-volatile solid-state storage devices. The memory 702 may also store a network communication program, which may be used to communicate with one or more additional devices, one or more terminal devices, and one or more devices.

The transmitter 704 can be used as an output device of the network device 700. For example, data can be transferred out of device 700. The receiver 705 can be used as an input device of the network device 700, for example, data can be transferred to the device 700. In addition, the transmitter 704 may include one or more optical transmitters, and/or one or more electrical transmitters. The receiver 705 may include one or more optical receivers, and/or one or more electrical receivers. The transmitter 704/receiver 705 can take the following forms: modem, modem bank, Ethernet card, universal serial bus (USB) interface card, serial interface, token ring card, optical fiber distributed data interface (FDDI) card, etc. Wait. Optionally, the network device 700 may not have a receiver and a transmitter, but a wired communication interface, which can communicate with other devices in a wired manner.

The processor 701 may be used to read and execute computer-readable program instructions and related model parameters. Specifically, the processor 701 may be used to call programs and model parameters stored in the memory 702, for example, the channel estimation model training method provided in one or more embodiments implements program instructions and model parameters on the network device side, and executes the program instructions And model parameters. Optionally, the processor 701 performs the following operations by calling program instructions and model parameters in the memory 702:

Convert the first channel matrix into codeword information, and use the codeword information to reconstruct the channel matrix to obtain the second channel matrix; use the first channel matrix and the second channel matrix to perform deep learning on the channel estimation model, Obtain a trained channel estimation model; wherein the channel estimation model is constructed based on a deep neural network; obtain the first signal transmitted by the terminal; use the trained channel estimation model to perform channel estimation on the first signal .

In the above network equipment, the first channel matrix is converted into codeword information, and the matrix is reconstructed according to the codeword information to obtain the second channel matrix, and then the parameters of the channel estimation model are adjusted through deep learning to reduce the second channel matrix and the first channel The relationship between the matrices can be used to obtain the trained channel estimation model; the subsequently trained channel estimation model can perform channel estimation according to the input transmission signal. Since the estimation of the intermediate steering vector and the gain matrix is not introduced in this process, the calculation pressure is significantly reduced; in addition, because this application only estimates the channel matrix, not the steering vector and gain matrix obtained by estimation. Channel estimation avoids information distortion in the intermediate links, so the error of the estimation result in the embodiment of the present application is smaller.

In a possible implementation manner, the converting the first channel matrix into codeword information, and reconstructing the channel matrix using the codeword information to obtain the second channel matrix is specifically: converting the real part of the first channel matrix And the imaginary part is converted into two real vectors; the two real vectors are converted into codeword information; the channel matrix is reconstructed using the codeword information to obtain a second channel matrix.

In a possible implementation manner, the reconstructing the channel matrix by using the codeword information to obtain the second channel matrix is specifically: extracting the second signal and the first white noise from the codeword information, and the second signal The second signal is a transmission signal; a channel matrix is reconstructed according to the second signal and the first white noise to obtain a second channel matrix.

In a possible implementation manner, before said converting the first channel matrix into codeword information and reconstructing the channel matrix using the codeword information to obtain the second channel matrix, the processor is further configured to Signal, fourth signal, and second white noise to generate the first channel matrix, where the third signal is a signal sent by a network device; the fourth signal is obtained when the terminal receives the third signal Signal, the second white noise is white noise fed back by the terminal.

In a possible implementation, the channel estimation model includes an encoding network and a decoding network, wherein the encoding network includes a first convolutional layer and a first fully connected layer; the decoding network includes a second fully connected layer , The second convolutional layer, the maximum pooling layer, the first insertion layer, the second insertion layer, the deep connection module, the global pooling layer, and the third convolutional layer; the first channel matrix is used as the coding network The input of the first convolutional layer, the output of the first convolutional layer is used as the input of the first fully connected layer, and the output of the first fully connected layer is used as the second of the decoding network The input of the fully connected layer, the output of the second fully connected layer is used as the input of the maximum pooling layer, and the output of the maximum pooling layer is used as the input of the first insertion layer. The output of an insertion layer is used as the input of the second insertion layer, the output of the second insertion layer is used as the input of the deep connection module, and the output of the deep connection module is used as the full pool The output of the full-pooling layer is used as the input of the third convolutional layer, and the third convolutional layer is used to generate the second channel matrix. The deep learning network with this structure has stronger learning ability and fast training speed, and the error of the channel matrix estimated by the channel estimation model at the training place is smaller.

In a possible implementation, the channel estimation model includes an encoding network and a decoding network, wherein the encoding network includes a first convolutional layer and a first fully connected layer; the decoding network includes a second fully connected layer , The second convolutional layer, the maximum pooling layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, the sixth convolutional layer, and the seventh convolutional layer; the first channel matrix is used for As the input of the first convolutional layer of the coding network, the output of the first convolutional layer is used as the input of the first fully connected layer, and the output of the first fully connected layer is used as the The input of the second fully connected layer of the decoding network, the output of the second fully connected layer is used as the input of the maximum pooling layer, and the output of the maximum pooling layer is used as the third convolutional layer The output of the third convolutional layer is used as the input of the fourth convolutional layer, the output of the fourth convolutional layer is used as the input of the fifth convolutional layer, and the first The output of the five convolutional layer is used as the input of the sixth convolutional layer, the output of the sixth convolutional layer is used as the input of the seventh convolutional layer, and the seventh convolutional layer is used to generate the The second channel matrix. The structure of this deep learning network still has stronger learning capabilities, faster training speed, smaller errors, and smaller computational complexity than existing technologies in the smaller massive MIMO system scenario.

In a possible implementation, the channel estimation model includes an encoding network and a decoding network, wherein the encoding network includes a first convolutional layer and a first fully connected layer; the decoding network includes a second fully connected layer , The residual network and the third fully connected layer; the first channel matrix is used as the input of the first convolutional layer of the coding network, and the output of the first convolutional layer is used as the first fully connected The input of the connection layer, the output of the first fully connected layer is used as the input of the second fully connected layer of the decoding network, and the output of the second fully connected layer is used as the input of the residual network, The output of the residual network is used as an input of the third fully connected layer, and the third fully connected layer is used to generate the second channel matrix. The deep learning network of this structure ensures that the network can avoid high complexity on the premise of deepening the model depth through the method of cutting off the connection of the features.

In a possible implementation manner, the channel estimation model adopts a compressed sensing mechanism.

In a possible implementation, the first convolutional layer is used to convert the real part and imaginary part of the first channel matrix into two real vectors; the first fully connected layer is used to convert the The two real vectors are converted into the codeword information.

It should be noted that the implementation of each operation may also correspond to the corresponding description of the method embodiment shown in FIG. 3.

An embodiment of the present application also provides a chip system. The chip system includes at least one processor, a memory, and an interface circuit. The memory, the transceiver, and the at least one processor are interconnected by wires, and the at least one memory Instructions are stored in the processor; when the instructions are executed by the processor, the method flow shown in FIG. 3 is realized.

An embodiment of the present application also provides a computer-readable storage medium, which stores instructions, and when it runs on a processor, the method flow shown in FIG. 3 is implemented.

The embodiment of the present application also provides a computer program product. When the computer program product runs on a processor, the method flow shown in FIG. 3 is realized.

In summary, the first channel matrix is converted into codeword information, and then the matrix is reconstructed according to the codeword information to obtain the second channel matrix, and then the parameters of the channel estimation model are adjusted through deep learning to reduce the second channel matrix and the first channel matrix Therefore, the trained channel estimation model can be obtained; the subsequently trained channel estimation model can perform channel estimation according to the input transmission signal. Since the estimation of the intermediate steering vector and the gain matrix is not introduced in this process, the calculation pressure is significantly reduced; in addition, because this application only estimates the channel matrix, not the steering vector and gain matrix obtained by estimation. Channel estimation avoids information distortion in the intermediate links, so the error of the estimation result in the embodiment of the present application is smaller.

A person of ordinary skill in the art can understand that all or part of the process in the above-mentioned embodiment method can be realized. The process can be completed by a computer program instructing relevant hardware. The program can be stored in a computer readable storage medium. , May include the processes of the foregoing method embodiments. The aforementioned storage media include: ROM or random storage RAM, magnetic disks or optical discs and other media that can store program codes.

Claims

A channel estimation method, characterized by comprising:

Converting the first channel matrix into codeword information, and reconstructing the channel matrix using the codeword information to obtain a second channel matrix;

Deep learning the channel estimation model by using the first channel matrix and the second channel matrix to obtain a channel estimation model after training, wherein the channel estimation model is constructed based on a deep neural network;

Acquiring the first signal transmitted by the terminal;

Using the trained channel estimation model to perform channel estimation on the first signal.
The method according to claim 1, wherein the converting the first channel matrix into codeword information and reconstructing the channel matrix using the codeword information to obtain the second channel matrix comprises:

Converting the real part and the imaginary part of the first channel matrix into two real vectors;

Converting the two real vectors into codeword information;

The channel matrix is reconstructed using the codeword information to obtain the second channel matrix.
The method according to claim 2, wherein the reconstructing a channel matrix using the codeword information to obtain a second channel matrix comprises:

Extracting a second signal and a first white noise from the codeword information, where the second signal is a transmission signal;

A channel matrix is reconstructed according to the second signal and the first white noise to obtain a second channel matrix.
The method according to any one of claims 1 to 3, wherein before the converting the first channel matrix into codeword information, and using the codeword information to reconstruct the channel matrix, the second channel matrix is obtained include:

The first channel matrix is generated according to the third signal, the fourth signal and the second white noise, where the third signal is a signal sent by a network device; the fourth signal is the terminal receiving the third signal The second white noise is the white noise fed back by the terminal.
The method according to any one of claims 1-4, wherein the channel estimation model includes an encoding network and a decoding network, wherein the encoding network includes a first convolutional layer and a first fully connected layer; The decoding network includes a second fully connected layer, a second convolutional layer, a maximum pooling layer, a first insertion layer, a second insertion layer, a deep connection module, a global pooling layer, and a third convolutional layer; the channel matrix Used as the input of the first convolutional layer of the coding network, the output of the first convolutional layer is used as the input of the first fully connected layer, and the output of the first fully connected layer is used as The input of the second fully connected layer of the decoding network, the output of the second fully connected layer is used as the input of the maximum pooling layer, and the output of the maximum pooling layer is used as the first insertion The input of the layer, the output of the first insertion layer is used as the input of the second insertion layer, the output of the second insertion layer is used as the input of the deep connection module, and the output of the deep connection module Used as the input of the full pooling layer, the output of the full pooling layer is used as the input of the third convolutional layer, and the third convolutional layer is used for generating the second channel matrix.
The method according to any one of claims 1-4, wherein the channel estimation model includes an encoding network and a decoding network, wherein the encoding network includes a first convolutional layer and a first fully connected layer; The decoding network includes a second fully connected layer, a second convolutional layer, a maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, and a seventh convolutional layer The channel matrix is used as the input of the first convolutional layer of the coding network, the output of the first convolutional layer is used as the input of the first fully connected layer, the first fully connected layer The output of is used as the input of the second fully connected layer of the decoding network, the output of the second fully connected layer is used as the input of the maximum pooling layer, and the output of the maximum pooling layer is used as The input of the third convolutional layer, the output of the third convolutional layer is used as the input of the fourth convolutional layer, and the output of the fourth convolutional layer is used as the fifth convolution The output of the fifth convolutional layer is used as the input of the sixth convolutional layer, and the output of the sixth convolutional layer is used as the input of the seventh convolutional layer. The seventh convolutional layer is used to generate the second channel matrix.
The method according to any one of claims 1-4, wherein the channel estimation model includes an encoding network and a decoding network, wherein the encoding network includes a first convolutional layer and a first fully connected layer; The decoding network includes a second fully connected layer, a residual network, and a third fully connected layer; the channel matrix is used as the input of the first convolutional layer of the coding network, and the output of the first convolutional layer is used As the input of the first fully connected layer, the output of the first fully connected layer is used as the input of the second fully connected layer of the decoding network, and the output of the second fully connected layer is used as the input of the second fully connected layer. The input of the residual network, the output of the residual network is used as the input of the third fully connected layer, and the third fully connected layer is used to generate the second channel matrix.
The method according to any one of claims 1-7, characterized in that:

The channel estimation model uses a compressed sensing mechanism.
The method according to any one of claims 5-7, characterized in that:

The first convolutional layer is used to convert the real part and the imaginary part of the first channel matrix into two real vectors; the first fully connected layer is used to convert the two real vectors into the code Word information.
A channel estimation device is characterized by comprising a processor and a memory, the memory is used to store program instructions and model parameters, and the processor is used to call the program instructions and model parameters to perform the following operations:

Converting the first channel matrix into codeword information, and reconstructing the channel matrix using the codeword information to obtain a second channel matrix;

Deep learning the channel estimation model by using the first channel matrix and the second channel matrix to obtain a channel estimation model after training, wherein the channel estimation model is constructed based on a deep neural network;

Acquiring the first signal transmitted by the terminal;

Using the trained channel estimation model to perform channel estimation on the first signal.
The device according to claim 10, wherein the converting the first channel matrix into codeword information, and using the codeword information to reconstruct the channel matrix to obtain the second channel matrix is specifically:

Converting the real part and the imaginary part of the first channel matrix into two real vectors;

Converting the two real vectors into codeword information;

The channel matrix is reconstructed using the codeword information to obtain the second channel matrix.
The device according to claim 11, wherein the reconstruction of the channel matrix by using the codeword information to obtain the second channel matrix is specifically:

Extracting a second signal and a first white noise from the codeword information, where the second signal is a transmission signal;

A channel matrix is reconstructed according to the second signal and the first white noise to obtain a second channel matrix.
The device according to any one of claims 10-12, wherein the first channel matrix is converted into codeword information, and the codeword information is used to reconstruct the channel matrix to obtain the second channel matrix. The processor is further configured to generate the first channel matrix according to a third signal, a fourth signal, and a second white noise, where the third signal is a signal sent by a network device; the fourth signal is the terminal A signal obtained when receiving the third signal, and the second white noise is white noise fed back by the terminal.
The device according to any one of claims 10-13, wherein the channel estimation model includes a coding network and a decoding network, wherein the coding network includes a first convolutional layer and a first fully connected layer; The decoding network includes a second fully connected layer, a second convolutional layer, a maximum pooling layer, a first insertion layer, a second insertion layer, a deep connection module, a global pooling layer, and a third convolutional layer; the channel matrix Used as the input of the first convolutional layer of the coding network, the output of the first convolutional layer is used as the input of the first fully connected layer, and the output of the first fully connected layer is used as The input of the second fully connected layer of the decoding network, the output of the second fully connected layer is used as the input of the maximum pooling layer, and the output of the maximum pooling layer is used as the first insertion The input of the layer, the output of the first insertion layer is used as the input of the second insertion layer, the output of the second insertion layer is used as the input of the deep connection module, and the output of the deep connection module Used as the input of the full-pooling layer, the output of the full-pooling layer is used as the input of the third convolutional layer, and the third convolutional layer is used for generating the second channel matrix.
The device according to any one of claims 10-13, wherein the channel estimation model includes a coding network and a decoding network, wherein the coding network includes a first convolutional layer and a first fully connected layer; The decoding network includes a second fully connected layer, a second convolutional layer, a maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, and a seventh convolutional layer The channel matrix is used as the input of the first convolutional layer of the coding network, the output of the first convolutional layer is used as the input of the first fully connected layer, the first fully connected layer The output of is used as the input of the second fully connected layer of the decoding network, the output of the second fully connected layer is used as the input of the maximum pooling layer, and the output of the maximum pooling layer is used as The input of the third convolutional layer, the output of the third convolutional layer is used as the input of the fourth convolutional layer, and the output of the fourth convolutional layer is used as the fifth convolution The output of the fifth convolutional layer is used as the input of the sixth convolutional layer, and the output of the sixth convolutional layer is used as the input of the seventh convolutional layer. The seventh convolutional layer is used to generate the second channel matrix.
The device according to any one of claims 10-13, wherein the channel estimation model includes a coding network and a decoding network, wherein the coding network includes a first convolutional layer and a first fully connected layer; The decoding network includes a second fully connected layer, a residual network, and a third fully connected layer; the channel matrix is used as the input of the first convolutional layer of the coding network, and the output of the first convolutional layer is used As the input of the first fully connected layer, the output of the first fully connected layer is used as the input of the second fully connected layer of the decoding network, and the output of the second fully connected layer is used as the input of the second fully connected layer. The input of the residual network, the output of the residual network is used as the input of the third fully connected layer, and the third fully connected layer is used to generate the second channel matrix.
The device according to any one of claims 10-16, characterized in that:

The channel estimation model uses a compressed sensing mechanism.
The device according to any one of claims 14-16, wherein the first convolutional layer is used to convert the real part and the imaginary part of the first channel matrix into two real vectors; A fully connected layer is used to convert the two real vectors into the codeword information.
A computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions run on a processor, the method according to any one of claims 1-9 is implemented.
A computer program product, characterized in that, when the computer program product runs on a processor, the method according to any one of claims 1-9 is implemented.