6G near-field non-stationary channel space-time joint extrapolation method and device based on deep learning

Abstract

The present invention discloses a 6G near-field non-stationary channel space-time joint extrapolation method based on deep learning, which obtains partial spatial domain channel state information at a historical moment; wherein the partial spatial domain channel state information is incomplete channel state information; the partial spatial domain channel state information is input into a space-time joint channel extrapolation model, and the complete spatial domain channel state information at a future target moment output by the space-time joint channel extrapolation model is obtained; wherein the space-time joint channel extrapolation model includes a LSTM module and a generator connected in series; the present invention determines the partial spatial domain channel state information at a future target moment according to the partial spatial domain channel state information at a historical moment through an LSTM module, and then generates the complete spatial domain channel state information at the future target moment through a generator, thereby realizing the extrapolation from the partial spatial domain channel state information at a historical moment to the complete spatial domain channel state information at a target moment, and reducing the pilot overhead and feedback overhead during system channel estimation.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication technology, and in particular to a 6G near-field non-stationary channel space-time joint extrapolation method based on deep learning. Background Art

[0002] The pilot signal is a part of the downlink power and shares the downlink power with other downlink channels. The pilot signal can be a signal used to measure the channel state between the base station and each user terminal, and the signal may not carry data.

[0003] As antenna scale increases and user mobility speeds increase, existing frequency division duplex (FDD) and time division duplex (TDD) systems will face problems such as feedback delay and processing delay, which will cause the acquired channel state information (CSI) to become outdated, seriously affecting the performance of adaptive wireless systems.

[0004] Moreover, with the increase in 6G system antenna scale, bandwidth and number of user connections, it is basically impractical to obtain CSI for each domain separately by sending pilot signals, which will generate huge pilot overhead. Summary of the invention

[0005] The purpose of the present invention is to provide a 6G near-field non-stationary channel space-time joint extrapolation method based on deep learning to reduce the dependence of channel state estimation on pilot signals.

[0006] The present invention adopts the following technical solution: a 6G near-field non-stationary channel space-time joint extrapolation method based on deep learning, comprising the following steps:

[0007] Acquire partial spatial channel state information at a historical moment; wherein the partial spatial channel state information is incomplete channel state information;

[0008] Part of the spatial channel state information is input into the space-time joint channel extrapolation model, and the complete spatial channel state information at the future target moment is output by the space-time joint channel extrapolation model; wherein the space-time joint channel extrapolation model includes a serially connected LSTM module and a generator.

[0009] Preferably, the LSTM module is used to generate partial spatial channel state information at a future target time according to partial spatial channel state information;

[0010] The LSTM module includes a first LSTM layer, a second LSTM layer, a third LSTM layer, and a fully connected layer connected in sequence;

[0011] Among them, Dropout layers are added to the first LSTM layer, the second LSTM layer, and the third LSTM layer.

[0012] Preferably, the generator is used to generate complete spatial channel state information at a future target time according to partial spatial channel state information at a future target time;

[0013] The generator consists of several convolutional layers, several deconvolutional layers, and several dilated convolutional layers.

[0014] Preferably, the generator includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a first dilated convolutional layer, a second dilated convolutional layer, a third dilated convolutional layer, a fourth convolutional layer, a first deconvolutional layer, a second deconvolutional layer, a fifth convolutional layer, a sixth convolutional layer and a seventh convolutional layer connected in sequence.

[0015] Preferably, the training method of the space-time joint channel extrapolation model comprises the following steps:

[0016] A corresponding first discriminator and a second discriminator are established according to the generator; the generator, the first discriminator and the second discriminator form a GAN network; the network structures of the first discriminator and the second discriminator are different, wherein the differentiated design of the first discriminator and the second discriminator can evaluate the authenticity of the spatial channel information generated by the generator from different angles and accuracies, which helps the GAN network to capture and learn complex non-stationary channel characteristics and enhance the overall system's ability to extrapolate near-field non-stationary channel state information;

[0017] The generator is trained based on a first training data set and a GAN network; wherein the first training data set includes partial spatial channel state information and real complete spatial channel state information at the same time.

[0018] Preferably, training the generator based on the first training data set and the GAN network includes:

[0019] Inputting partial spatial channel state information at the same time into the generator to obtain predicted complete spatial channel state information;

[0020] Inputting the predicted complete spatial domain channel state information and the real complete spatial domain channel state information into a first discriminator to obtain a first identification result; inputting the predicted complete spatial domain channel state information and the real complete spatial domain channel state information into a second discriminator to obtain a second identification result;

[0021] Combine the first identification result and the second identification result to obtain a total identification result;

[0022] The generator is optimized based on the total identification results.

[0023] Preferably, the total loss function of the generator optimized based on the total identification result is:

[0024]

[0025] Among them, LO (G, D) represents the total loss function, G represents the generator, D represents the discriminator network, the discriminator network includes the first discriminator and the second discriminator, α represents the weighted hyperparameter, L(G, D) represents the cross entropy loss function of the discriminator network, NMSE G Represents the NMSE loss function of the generator.

[0026] The input data of the generator and the discriminator come from the partial spatial channel state information and the complete spatial channel state information at a specific moment. This processing method helps the model capture and simulate the relationship between the spatial channel state information during the training process, thereby more accurately modeling the near-field spatial non-stationary characteristics.

[0027] Preferably, the training method of the space-time joint channel extrapolation model further includes the following steps:

[0028] Connect the input of the trained generator to the output of the LSTM module;

[0029] The LSTM module is trained based on a second training data set; wherein the second training data set includes partial spatial channel state information at a historical moment and complete spatial channel state information at a future target moment.

[0030] Preferably, the loss function when training the LSTM module based on the second training data set is:

[0031]

[0032] Among them, NMSE GL represents the NMSE loss of the LSTM module, h(t+Δt) represents the real complete spatial channel state information at the future target moment, t represents the historical moment, Δt represents the time interval between the historical moment and the future target moment, Represents the complete spatial channel state information of the future target time output by the generator, Represents the complete spatial channel state information of the future target time output by the LSTM module.

[0033] Another technical solution of the present invention: a 6G near-field non-stationary channel space-time joint extrapolation device based on deep learning, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the above method is implemented when the processor executes the computer program.

[0034] The beneficial effects of the present invention are as follows: the present invention determines partial spatial channel state information at a future target moment according to partial spatial channel state information at a historical moment through an LSTM module, and then generates complete spatial channel state information at the future target moment through a generator, thereby realizing the extrapolation from partial spatial channel state information at a historical moment to complete spatial channel state information at a target moment, reducing the dependence of CSI prediction on pilot signals, and reducing pilot overhead and feedback overhead during system channel estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Schematic diagram of application scenario of the method of the present invention;

[0036] Figure 2 It is the principle flow chart of the method of the present invention;

[0037] Figure 3 Schematic diagram of the network structure of the joint channel extrapolation model in the method of the present invention;

[0038] Figure 4 This is a comparison chart of the actual channel measurement results and space-time extrapolation results of users at different moving speeds in the simulation experiment;

[0039] Figure 5 This is a comparison chart of the normalized mean square error (NMSE) in the simulation experiment;

[0040] Figure 6 A comparison diagram of normalized mean square errors between the method of the present invention and a single-domain channel extrapolation method that does not consider spatial non-stationary characteristics;

[0041] Figure 7 It is a comparison diagram of the achievable rates of the method of the present invention and the single-domain channel extrapolation method without considering the spatial non-stationary characteristics for mobile users under different signal-to-noise ratio conditions in simulation experiments. DETAILED DESCRIPTION

[0042] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] In order to reduce the dependence of CSI calculation on pilot signals, channel extrapolation technology came into being. It can use known CSI to extrapolate unknown CSI by establishing a channel extrapolation model, thereby obtaining more accurate CSI without additional resource overhead.

[0044] Most existing channel extrapolation methods focus on the channel state information of a single domain and fail to effectively consider the relationship between multiple channel domains, thus limiting the application scope and prediction accuracy of the channel method. In addition, existing research often oversimplifies the application scenarios, usually assuming that the channel changes in time or space are stable, and fails to fully consider the non-stationary characteristics of the near-field space. This has a particularly significant impact on current high-speed mobile actual communication systems with large-scale antennas, which directly affects the communication system's effective use of spectrum resources and the reliability of data transmission.

[0045] Therefore, considering the complex channel environment of the actual communication system, the present invention proposes a 6G near-field non-stationary channel space-time joint extrapolation method based on deep learning, aiming to solve the timeliness and accuracy of CSI acquisition in the 6G system, further reduce the system channel estimation overhead, and improve system performance. By using deep learning technology to jointly model the time domain and spatial domain characteristics of the channel, high-precision extrapolation of non-stationary channels can be achieved, and the accuracy of CSI acquisition can be improved. At the same time, this method does not require additional channel estimation resources, can effectively reduce the system's energy consumption overhead, and improve the overall performance of the 6G system.

[0046] The present invention discloses a 6G near-field non-stationary channel space-time joint extrapolation method based on deep learning, comprising the following steps: obtaining partial spatial domain channel state information at a historical moment; wherein the partial spatial domain channel state information is incomplete channel state information; inputting the partial spatial domain channel state information into a space-time joint channel extrapolation model, and obtaining the complete spatial domain channel state information at a future target moment output by the space-time joint channel extrapolation model; wherein the space-time joint channel extrapolation model comprises a serially connected LSTM module and a generator.

[0047] The present invention determines partial spatial channel state information at a future target moment according to partial spatial channel state information at a historical moment through an LSTM module, and then generates complete spatial channel state information at the future target moment through a generator, thereby realizing extrapolation based on the complete spatial channel state information, reducing the dependence of CSI prediction on pilot signals, and reducing pilot overhead and feedback overhead during system channel estimation.

[0048] The present invention is applicable to 6G application scenarios combining near-field or near-far-field. The specific system model is as follows: Figure 1 As shown in the figure, considering the non-stationary characteristics of near-field space and spherical wave propagation, a geometry-based random model (GBSM) is used for channel modeling. The base station (BS) uses a UPA with M×N antennas, and the mobile user uses a single antenna with a moving speed of v(t). Both the BS and the mobile user are equipped with single-polarized antennas.

[0049] The research goal of this invention is to realize the space-time joint channel extrapolation of a single mobile user in the application scenario of 6G near-field non-stationary channels, so as to improve the accuracy of channel estimation and reduce the channel estimation overhead of 6G systems. The specific network process of this invention using the space-time joint channel extrapolation model to realize space-time channel extrapolation is as follows: Figure 2 As shown. The overall network model can be called a 2DGAN-LSTM network, which consists of an LSTM module and a GAN network with two discriminators.

[0050] In the method of the present invention, a large amount of space-time channel data is first generated by using a 6G channel simulator as offline training data. Secondly, the 2DGAN network is trained to realize spatial channel extrapolation. Finally, the trained 2DGAN generator is fixed to the output end of the LSTM network, and the 2DGAN-LSTM network is trained to realize space-time joint extrapolation. When using the trained 2DGAN-LSTM network for online extrapolation, the input layer of the network inputs part of the spatial channel state information at the historical moment. The output layer outputs the complete spatial channel state information h(t+Δt) at the current or future time. Δt represents the time domain range of network extrapolation. The network can extrapolate the channel state information of several future transmission frames by learning the patterns and rules of the user's historical channel state information.

[0051] Specifically, Figure 3 As shown in FIG. 1 , the LSTM module is used to generate partial spatial channel state information at a future target time according to partial spatial channel state information. The LSTM module includes a first LSTM layer, a second LSTM layer, a third LSTM layer, and a fully connected layer connected in sequence; wherein a Dropout layer is added to the first LSTM layer, the second LSTM layer, and the third LSTM layer.

[0052] The LSTM module is responsible for calculating the channel state information of some airspaces at historical moments Extrapolate partial spatial channel state information at current or future time

[0053] The LSTM layer (i.e., the first LSTM layer, the second LSTM layer, or the third LSTM layer) is responsible for processing the input sequence (i.e. The fully connected layer is responsible for converting the output of the LSTM layer into an extrapolated result. The Dropout layer added inside the LSTM layer can reduce the dependency between neurons by randomly setting the output of some neurons in the network to zero during the training process, thereby preventing the model from overfitting and improving the generalization ability of the model.

[0054] In one embodiment, the generator is used to generate the complete spatial channel state information at the future target time according to the partial spatial channel state information at the future target time. The generator includes a plurality of convolutional layers, a plurality of rice convolutional layers and a plurality of dilated convolutional layers.

[0055] The convolution layer, also known as the downsampling layer, compresses and refines the spatial channel state information through the sliding of the filter, realizes the dimension reduction and abstraction of the features, and learns and extracts the global features of the spatial channel state information. The deconvolution layer, also known as the upsampling layer, upsamples the feature map to increase the spatial size of the feature map, making the generated spatial channel state information clearer. The dilated convolution layer introduces holes in the convolution to increase the effective receptive field of the convolution kernel, so as to capture contextual information in a wider range of space.

[0056] Preferably, the generator includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a first dilated convolutional layer, a second dilated convolutional layer, a third dilated convolutional layer, a fourth convolutional layer, a first deconvolutional layer, a second deconvolutional layer, a fifth convolutional layer, a sixth convolutional layer and a seventh convolutional layer connected in sequence.

[0057] In summary, the generator G consists of 7 convolutional layers, 3 dilated convolutional layers, and 2 deconvolutional layers, and a rectified linear unit is set after each layer. Among them, the convolutional layer compresses and refines the spatial channel state information to achieve dimensionality reduction and abstraction of features. The deconvolution layer increases the spatial size of the feature map by upsampling the feature map, making the generated spatial channel state information clearer. The dilated convolutional layer captures spatial correlations of different scales by setting different dilation rates. In addition, batch normalization is used after each convolutional layer and deconvolution layer, which consists of activation functions Rule and Tanh to reduce internal covariate shift, thereby better modeling the near-field spatial non-stationary characteristics and solving the problem of inaccurate channel extrapolation of 6G non-stationary channels.

[0058] In the embodiment of the present invention, each module is optimized by batch training. The generator is first trained, and then the trained generator is connected to the output end of the LSTM module to train the LSTM module again.

[0059] First, a large number of space-time data sets are generated using the 6G channel simulator, which contains LOS path and NLOS path users and user data at different locations and different moving speeds. Then, the initial training parameters of each module are initialized.

[0060] In one embodiment, the training method of the space-time joint channel extrapolation model includes the following steps: establishing a corresponding first discriminator and a second discriminator according to a generator; the generator, the first discriminator and the second discriminator form a GAN network; the network structures of the first discriminator and the second discriminator are different; the generator is trained based on a first training data set and a GAN network; wherein the first training data set includes partial spatial channel state information and real complete spatial channel state information at the same time. The first training data set is a batch of data randomly extracted from the original data set, and then obtained after antenna sampling.

[0061] Two discriminators, discriminator D1 and discriminator D2, are introduced to identify whether the complete spatial channel state information is real or generated. By jointly optimizing the two discriminators and the entire network, the prediction accuracy and robustness of the network are further improved, and the space-time joint extrapolation of the 6G near-field non-stationary channel is achieved.

[0062] In an embodiment of the present invention, the first discriminator is composed of five convolutional layers and a fully connected layer connected in sequence, and the second discriminator is composed of four convolutional layers and a fully connected layer connected in sequence. The output results of the first discriminator and the second discriminator are spliced ​​and then connected to the same fully connected layer. It should be noted that a rectified linear unit is provided after each convolutional layer. By splicing the output results of the first discriminator (i.e., D1) and the second discriminator (i.e., D2), and then processing them through a fully connected layer, the joint judgment result of the two discriminators is output, thereby further improving the robustness of the space-time channel extrapolation network.

[0063] In an embodiment of the present invention, training a generator based on a first training data set and a GAN network includes: inputting partial spatial channel state information at the same time into the generator to obtain predicted complete spatial channel state information; inputting the predicted complete spatial channel state information and the true complete spatial channel state information into a first discriminator to obtain a first identification result; inputting the predicted complete spatial channel state information and the true complete spatial channel state information into a second discriminator to obtain a second identification result; combining the first identification result and the second identification result to obtain a total identification result; and optimizing the generator based on the total identification result.

[0064] The total loss function for optimizing the generator based on the total identification result is:

[0065]

[0066] Among them, L O (G, D) represents the total loss function, G represents the generator, D represents the discriminator network, the discriminator network includes the first discriminator and the second discriminator, α represents the weighted hyperparameter, L(G, D) represents the cross entropy loss function of the discriminator network, NMSE GRepresents the NMSE loss function of the generator.

[0067] NMSE G The NMSE loss function of the generator represented by is specifically:

[0068]

[0069] Among them, h(t) represents the real complete spatial channel state information at time t, Represents the generator according to The predicted complete spatial channel state information at time t is generated, Represents the partially complete spatial channel state information at time t.

[0070] The cross entropy loss function of the discriminator network is specifically:

[0071]

[0072] Wherein, D(h(t)) represents the identification result of the total discriminator output obtained by combining the first discriminator and the second discriminator, which usually represents the probability that the input h(t) is the true complete spatial channel state information; Indicates input is the probability of the true complete spatial channel state information.

[0073] That is, the generator NMSE loss function and the cross entropy loss functions of the two discriminators are jointly calculated iteratively to generate accurate complete spatial domain information.

[0074] In one embodiment, the training method of the space-time joint channel extrapolation model also includes the following steps: connecting the input end of the trained generator to the output end of the LSTM module; training the LSTM module based on a second training data set; wherein the second training data set includes partial spatial channel state information at historical moments and complete spatial channel state information at future target moments.

[0075] The training process can be iterated multiple times. In each iteration, a batch of data is extracted from all space-time data sets, and spatial sampling is performed according to the antenna selection method of the 2DGAN network. The 2DGAN-LSTM network refers to a network that connects the LSTM module, the generator, the first discriminator, and the second discriminator together. For convenience in the training stage, it is collectively referred to as the 2DGAN network.

[0076] The loss function when training the LSTM module based on the second training data set is:

[0077]

[0078] Among them, NMSE GLrepresents the NMSE loss of the LSTM module, h(t+Δt) represents the true complete spatial channel state information at the future target moment, t represents the historical moment, Δt represents the time interval between the historical moment and the future target moment, that is, the prediction step length, represents the complete spatial channel state information of the future target time output by the generator, that is, the complete spatial CSI after the generator completes. represents the complete spatial channel state information of the future target time output by the LSTM module, and ||·|| represents the Euclidean norm.

[0079] By introducing deep learning technology, this paper proposes an innovative method of space-time joint channel extrapolation network, which further reduces the system's pilot overhead and frequent feedback overhead compared to traditional channel estimation methods or single-domain channel extrapolation methods that do not consider the spatial non-stationary characteristics of the channel. At the same time, this method can accurately model and extrapolate the near-field non-stationary channel characteristics in 6G communications, improve the performance and reliability of the communication system, and has good practical application value.

[0080] The present invention conducts simulation experiments of users at different moving speeds, such as Figure 4 As shown, Figure 4 (a), (b), and (c) in the figure respectively show the comparison between the actual channel measurement results and the space-time extrapolation results when the spatial sampling interval is 2, the time domain prediction step is 1, and the user moving speed is 30m / s, 60m / s, and 80m / s respectively under the conditions that the system center frequency is 28GHz and the channel bandwidth is 2000MHZ. It can be clearly seen from these three comparison figures that the results obtained by channel extrapolation using the space-time joint channel extrapolation network proposed in the present invention are very close to the actual channel true measurement values.

[0081] Figure 5 The figure shows the normalized mean square error (NMSE) comparison of space-time channel extrapolation using the proposed network when the base station is equipped with a 16×16 UPA and the user moving speed is 50m / s. The horizontal axis in the figure represents the time domain prediction step, and the antenna compression ratios of 1 / 2, 1 / 4, and 1 / 8 represent the antenna sampling results when the spatial antenna sampling intervals are 2, 4, and 8, respectively. It can be seen from the figure that with the increase of the time domain extrapolation step, the spatial compression ratio of the spatial antenna increases, that is, the smaller the antenna sampling interval, the smaller the error.

[0082] Figure 6The figure shows the normalized mean square error comparison of the proposed space-time joint channel extrapolation network and the current single-domain channel extrapolation research that does not consider spatial non-stationary characteristics for users at different moving speeds. It can be seen from the figure that as the user's moving speed increases, the space-time channel extrapolation network proposed by the present invention is more robust to the user's moving speed and near-field spatial non-stationary characteristics.

[0083] Figure 7 The figure shows the achievable rate comparison of a mobile user with a moving speed of 50 m / s under different signal-to-noise ratio conditions when the system center frequency is 28 GHz and the channel bandwidth is 2000 MHZ. The figure shows the achievable rate curve of the user when the channel state information is obtained by channel extrapolation using the space-time joint channel extrapolation network proposed by the present invention and the existing single-domain channel extrapolation research that does not consider the spatial non-stationary characteristics. The comparison curves in the figure further show the advantages of the proposed space-time channel extrapolation network. In addition, combined with Figure 6 The simulation results show that the existing single-domain channel extrapolation research that does not consider the spatial non-stationary characteristics performs relatively well when the user's moving speed is low, but its performance will decrease significantly as the user's speed increases.

[0084] The present invention also discloses a 6G near-field non-stationary channel space-time joint extrapolation device based on deep learning, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The above method is implemented when the processor executes the computer program.

[0085] The present invention also discloses a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0086] The present invention also provides a computer program product. When the computer program product runs on a data storage device, the data storage device can implement the steps in the above-mentioned method embodiments when executing the computer program product.

[0087] Based on such understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may at least include: any entity or device capable of carrying the computer program code to a storage device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electric carrier signal, a telecommunication signal, and a software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk.

[0088] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0089] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the present invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

Claims

1. A 6G near-field non-stationary channel space-time joint extrapolation method based on deep learning, characterized in that: The following steps are involved: Acquire partial spatial channel state information at a historical moment; wherein the partial spatial channel state information is incomplete channel state information; Inputting the partial spatial channel state information into a space-time joint channel extrapolation model, and obtaining the complete spatial channel state information at a future target time output by the space-time joint channel extrapolation model; wherein the space-time joint channel extrapolation model includes a long short-term memory module (LSTM module) and a generator connected in series; The LSTM module is used to generate the partial spatial channel state information at the future target time according to the partial spatial channel state information; The LSTM module includes a first LSTM layer, a second LSTM layer, a third LSTM layer and a fully connected layer connected in sequence; Wherein, a Dropout layer is added to the first LSTM layer, the second LSTM layer and the third LSTM layer; The generator is used to generate the complete spatial channel state information at the future target time according to the partial spatial channel state information at the future target time; The generator includes a plurality of convolutional layers, a plurality of deconvolutional layers and a plurality of dilated convolutional layers; The generator includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a first dilated convolutional layer, a second dilated convolutional layer, a third dilated convolutional layer, a fourth convolutional layer, a first deconvolutional layer, a second deconvolutional layer, a fifth convolutional layer, a sixth convolutional layer and a seventh convolutional layer connected in sequence; The training method of the space-time joint channel extrapolation model comprises the following steps: A corresponding first discriminator and a second discriminator are established according to the generator; the generator, the first discriminator and the second discriminator form a generative adversarial network (GAN network); the network structures of the first discriminator and the second discriminator are different; The generator is trained based on a first training data set and the GAN network; wherein the first training data set includes partial spatial channel state information and real complete spatial channel state information at the same time; Training the generator based on the first training data set and the GAN network includes: Inputting partial spatial channel state information at the same time into the generator to obtain predicted complete spatial channel state information; Inputting the predicted complete spatial domain channel state information and the real complete spatial domain channel state information into a first discriminator to obtain a first identification result; inputting the predicted complete spatial domain channel state information and the real complete spatial domain channel state information into a second discriminator to obtain a second identification result; Combining the first identification result and the second identification result to obtain a total identification result; optimizing the generator based on the overall identification result; The total loss function for optimizing the generator based on the total identification result is: Among them, L O (G, D) represents the total loss function, G represents the generator, D represents the discriminator network, the discriminator network includes the first discriminator and the second discriminator, α represents the weighted hyperparameter, L(G, D) represents the cross entropy loss function of the discriminator network, NMSE G represents the NMSE loss function of the generator; The training method of the space-time joint channel extrapolation model also includes the following steps: Connecting the trained input end of the generator to the output end of the LSTM module; The LSTM module is trained based on a second training data set; wherein the second training data set includes partial airspace channel state information at a historical moment and complete airspace channel state information at a future target moment; The loss function when training the LSTM module based on the second training data set is: Among them, NMSE GL represents the NMSE loss of the LSTM module, h(t+Δt) represents the real complete spatial channel state information at the future target moment, t represents the historical moment, Δt represents the time interval between the historical moment and the future target moment, represents the complete spatial channel state information at the future target time output by the generator, Represents the complete spatial channel state information at the future target time output by the LSTM module.

2. A 6G near-field non-stationary channel space-time joint extrapolation device based on deep learning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to claim 1 is implemented.

Images