Contrastive learning-based 6g fully-decoupled network representative channel generation method and system

Abstract

Disclosed in the present invention are a contrastive learning-based 6G fully-decoupled network representative channel generation method and system. The method comprises: establishing a multipath channel model and a multiple-input multiple-output orthogonal frequency division multiplexing system transmission model; giving a representative channel generation target on the basis of the system transmission model; establishing a multipath channel information data set on the basis of the multipath channel model, and using a contrastive learning framework to train a Transformer-based encoder, wherein the encoder uses multipath channel information as an input, and outputs multipath channel representation information; training a Transformer-based decoder on the basis of the multipath channel information and the multipath channel representation information, wherein the decoder uses the multipath channel representation information as an input, and outputs the multipath channel information; in an actual deployment stage, inputting the multipath channel information, averaging the multipath channel representation information output by the encoder, then inputting the averaged multipath channel representation information into the decoder, and outputting a representative channel. The present invention can solve the problem of difficulties in selecting fixed transmission parameters in spectrum maps, and provides a feasible solution for transmission in 6G fully-decoupled access networks.

Description

A method and system for generating representative channels in 6G fully decoupled networks based on contrastive learning Technical Field

[0001] This invention belongs to the field of wireless mobile communication technology, and relates to contrastive learning and multiple-input multiple-output transmission. Specifically, it relates to a method and system for generating representative channels in a 6G fully decoupled network based on contrastive learning. Background Technology

[0002] In wireless communication, the channel describes signal propagation in the physical environment and determines transmission quality. In multiple-input multiple-output (MIMO) orthogonal frequency division multiplexing (OFDM) systems, the channel is typically represented in the spatial, frequency, and time domains. The spatial domain describes multipath signal scattering in the angular domain, involving the departure angle of the transmitter and the arrival angle of the receiver. The frequency domain describes multipath fading scattering in the delay domain, capturing the frequency-selective fading of the channel. The time domain reflects the multipath effects of environmental dynamics. Here, we refer to the channel at a fixed geographical location, thus excluding the Doppler effect. By studying the channel, its characteristics can be effectively utilized to facilitate MIMO transmissions, such as beamforming or spatial multiplexing. Channels exhibit correlation in the spatial, frequency, and time domains. When using uniform linear array or uniform planar array antennas, electromagnetic waves arrive at different antennas with a fixed phase difference. Furthermore, there is a fixed interval between different carriers, resulting in a fixed phase difference in the channel on different subcarriers. At a specific location, the channel exhibits temporal correlation because the dominant electromagnetic waves roughly follow the same propagation path. Correlation means that the channel can be represented in the form of higher information entropy, i.e., channel characterization, which is crucial for understanding the intrinsic properties of the channel.

[0003] Furthermore, in the next-generation mobile communication network (6G), Academician Yu Quan and others proposed a fully-decoupled RAN architecture for 6G, as presented in "A Fully-Decoupled RAN Architecture for 6G Inspired by Neurotransmission." In this fully decoupled network, traditional base stations are decoupled into control base stations, uplink base stations, and downlink base stations. Through hardware decoupling, uplink and downlink transmissions can be completely independent and flexibly adjusted to meet diverse and continuously changing service and coverage requirements of users. However, the fully decoupled network also brings new challenges to downlink MIMO transmission. Because the uplink and downlink base stations are hardware isolated, the existing feedback mechanism is no longer suitable for the fully decoupled RAN architecture. Therefore, the fully decoupled RAN uses a spectrum map approach to select transmission parameters, that is, using fixed transmission parameters for a fixed geographical location for MIMO transmission. However, how to select fixed transmission parameters remains a problem.

[0004] In summary, the current problems in wireless communication and 6G fully decoupled wireless access network architecture include: (1) Channel information can be represented as a channel representation with higher information entropy, but there is no good method for obtaining the channel representation. (2) In 6G fully decoupled networks, it is difficult to select fixed transmission parameters for building spectrum maps. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a method and system for generating representative channels in a 6G fully decoupled network based on contrastive learning. It uses contrastive learning to extract the representation information of multipath channels, and uses a decoder to restore the multipath channel representation to the multipath channel, thereby obtaining the representative channel and fixed transmission parameters for determining the spectrum map.

[0006] Technical Solution: To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for generating representative channels in a 6G fully decoupled network based on contrastive learning, comprising the following steps:

[0007] Step 1: Establish a multipath channel model and a transmission model for a multiple-input multiple-output orthogonal frequency division multiplexing system;

[0008] Step 2: Based on the transmission model of the multiple-input multiple-output orthogonal frequency division multiplexing system, give the representative channel generation target;

[0009] Step 3: Establish a multipath channel information dataset based on the multipath channel model, and train a Transformer-based encoder using a contrastive learning framework. The encoder takes multipath channel information as input and outputs multipath channel representation information.

[0010] Step 4: Based on the multipath channel information and multipath channel representation information, train a Transformer-based decoder. The decoder takes the multipath channel representation information as input and outputs the multipath channel information.

[0011] Step 5: In the actual deployment phase, input multipath channel information, the encoder outputs multipath channel representation information, and the multipath channel representation information, after being averaged over the time dimension, is input into the decoder to output a representative channel. Further, the multipath channel model is as follows:

[0012] Where t is the sampling time, e is the natural constant; N k N represents the number of subcarriers in the transmission model of the multiple-input multiple-output orthogonal frequency division multiplexing system. p B represents the number of multipaths in the multipath channel, and B represents the bandwidth of the transmission model of the multiple-input multiple-output orthogonal frequency division multiplexing system; b, u, and k are the sequence numbers of the base station, user, and subcarrier, respectively. Let θ represent the power, phase, and delay of the multipath channel between base station b and user u at time t.b,u,t , These represent the departure angle, zenith angle of arrival, and directional angle of arrival of the multipath between base station b and user u at time t, respectively; H b,u,t,k This represents the multipath channel on subcarrier k between base station b and user u at time t; It is the array response matrix of the user and the base station;

[0013] Multipath channel between base station b and user u at sampling time t U represents concatenating the matrix along a new dimension, and N... R N represents the number of physical antenna ports at the receiver. T Let N be the number of physical antenna ports at the transmitting end; a fixed base station b and user u are represented using the base station-user pair {b,u}, and N data are collected between base station b and user u. t There are multiple path channels, and the information of these multiple path channels is:

[0014] Furthermore, for users employing uniform linear antennas and base stations employing uniform planar array antennas, the array response matrix is ​​expressed as:

[0015] in It is the Kronecker product; T is the matrix transpose operation; a x (·),a y (·),a z (·) represent the array response matrices in the x, y, and z directions, respectively:

[0016] Where, N x N y N z These represent the number of physical antenna ports of the base station in the x, y, and z directions, respectively.

[0017] Furthermore, the transmission model of the multiple-input multiple-output orthogonal frequency division multiplexing system is as follows:

[0018] Where, x b,u,t,k y is the signal transmitted by base station b to user u on subcarrier k at time t; u,t,k W is the received signal of user u on subcarrier k at time t; b,u,t,k Let n be the precoding matrix of base station b to user u on subcarrier k at time t; b,i,t,k The additive complex white Gaussian noise emitted by base station b to user u on subcarrier k at time t.

[0019] Furthermore, the representative channel The target to be generated is:

[0020] in, Is with H b,u,t,k A matrix with the same dimension but not changing with t; N l It is the number of transport layers; It is the variance of the additive complex white Gaussian noise; ||·|| F It is the Frobenius norm; ||·||2 is the 2-norm; It is a singular vector extraction operation, which involves taking the top N vectors after performing singular value decomposition. l The singular vectors corresponding to the largest singular values.

[0021] Furthermore, the Transformer-based encoder decomposes the multipath channel information into multiple multipath channel information blocks, and maps these blocks to a grid of size d using a fully connected network. model Vectors, then add sine / cosine positional encoding to the vectors:

[0022] Where pos is the position encoding index and i is the vector element index. These are the positional encoding of a single element and the overall output; then the vector undergoes multiple self-attention mechanisms; these self-attention mechanisms map the input to a vector of size d through three independent fully connected layers. atten The vectors, the weights of the three fully connected layers are respectively represented by Ω. Θ ,Ω K ,Ω V The outputs of the three fully connected layers are represented by Θ, K, and V, respectively, for self-attention. This can be represented as: Θ = Ω Θ • input, K = Ω K input, V = Ω V ·input,

[0023] Wherein, input represents the input, softmax is the normalized exponential function; finally, the output after the self-attention mechanism is passed through average pooling and a fully connected layer to obtain the output of the Transformer-based encoder, which is the multipath channel representation information.

[0024] Furthermore, the contrastive learning framework utilizes positive and negative samples to achieve self-supervised learning, considering a positive sample of multipath channel information. and One negative sample of multipath channel information A contrastive learning training framework is used to train two Transformer-based encoders. This encoder aims to make the outputs of multipath channel information at different sampling times in positive samples as similar as possible, while minimizing the dissimilarity between the outputs of positive and negative samples. This is achieved using ε... pos ,ε neg It means that ε pos The loss function is:

[0025] Where exp is the natural exponential function; t′ is a different sampling time than t; and γ is the temperature parameter. neg The parameter θ neg According to ε pos The parameter θ pos Update: θ neg ←βθ neg +(1-β)θ pos ,

[0026] Where β represents the update factor. Each multipath channel information negative sample is maintained by a first-in-first-out queue, and each training ε neg Output a batch of channel characterization information into a queue, and remove the first batch of channel characterization information that entered the queue from the queue.

[0027] Furthermore, the Transformer-based decoder takes multipath channel representation information as input and outputs multipath channel information. Its purpose is to recover the input based on the output of the Transformer-based encoder. The loss function of the Transformer-based decoder is:

[0028] in For a Transformer-based decoder, F b,u,t N represents the multipath information channel characterization output by the encoder. u This represents the total number of users.

[0029] Based on the same inventive concept, this invention provides a representative channel generation system for a 6G fully decoupled network based on contrastive learning, comprising: a modeling module for establishing a multipath channel model and a multiple-input multiple-output (MIMO) orthogonal frequency division multiplexing (OFDM) system transmission model, and providing a representative channel generation target based on the OFDM system transmission model; an encoder training unit for establishing a multipath channel information dataset based on the multipath channel model, and training a Transformer-based encoder using a contrastive learning framework, the encoder taking multipath channel information as input and outputting multipath channel representation information; a decoder training unit for training a Transformer-based decoder based on the multipath channel information and multipath channel representation information, the decoder taking multipath channel representation information as input and outputting multipath channel information; and a representative channel generation unit for inputting multipath channel information during the actual deployment phase, the encoder outputting multipath channel representation information, and inputting the multipath channel representation information after averaging the multipath channel representation information over the time dimension into the decoder to output a representative channel.

[0030] Based on the same inventive concept, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the aforementioned method for generating a representative channel in a 6G fully decoupled network based on contrastive learning.

[0031] Beneficial Effects: Compared with existing technologies, this invention has the following advantages: 1. This invention provides a feasible representative channel generation method in 6G fully decoupled networks, avoiding the feedback defects caused by hardware isolation in 6G fully decoupled networks. 2. This invention proposes a contrastive learning method to obtain the representation of wireless channels. By utilizing the small-scale time-domain changes of the channel, a contrastive learning proxy task is designed to train an encoder capable of mapping channel information to channel representation information. 3. This invention uses deep learning to fit the relationship between channel representation and channel information, obtaining a decoder that can recover the channel from the channel representation. Through experimental comparison, the method of this invention can obtain a representative channel for the establishment of a spectrum map in a 6G fully decoupled access network. Attached Figure Description

[0032] Figure 1 is a flowchart of the representative channel generation method for 6G fully decoupled network based on contrastive learning used in the embodiments of the present invention.

[0033] Figure 2 is a flowchart of the comparative learning framework of an embodiment of the present invention.

[0034] Figure 3 is a schematic diagram of the neural network structure of the encoder and decoder according to an embodiment of the present invention.

[0035] Figure 4 is a bar chart comparing the performance of the representative channel and the original channel in an embodiment of the present invention.

[0036] Figure 5 is a line graph comparing the performance of a representative channel and the original channel in an embodiment of the present invention.

[0037] Figure 6 is a cumulative distribution function graph comparing the performance of the representative channel and the original channel in an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical methods, and advantages of this invention clearer, the embodiments of this invention are described in detail below with reference to the accompanying drawings. These embodiments are implemented based on the technical methods of this invention, providing detailed implementation methods and specific operating procedures. It should be understood that the specific examples described herein are merely illustrative of this invention, but the scope of protection of this invention is not limited to the following embodiments.

[0039] This invention discloses a novel method for generating representative channels in a 6G fully decoupled network based on contrastive learning. Figure 1 illustrates the process of generating representative channels in a 6G fully decoupled network based on contrastive learning. First, a multipath channel model and a multiple-input multiple-output (MIMO) orthogonal frequency division multiplexing (OFDM) system transmission model are established. Second, a representative channel generation target is given based on the OFDM system transmission model. Third, a multipath channel information dataset is established based on the multipath channel model, and a Transformer-based encoder is trained using a contrastive learning framework. The encoder takes multipath channel information as input and outputs multipath channel representation information. Then, a Transformer-based decoder is trained based on the multipath channel information and multipath channel representation information. The decoder takes multipath channel representation information as input and outputs multipath channel information. Finally, in the actual deployment stage, multipath channel information is input, the encoder outputs multipath channel representation information, and the averaged multipath channel representation information is input to the decoder to output the representative channel.

[0040] Specifically, this embodiment of the invention considers a single-user downlink scenario in a fully decoupled network.

[0041] The specific steps for channel generation are given below.

[0042] In this example, the transmitter's antenna is a dual-polarized uniform planar array antenna, with N antennas in the x, y, and z directions, respectively. x =8,N y =2,N z = 1 physical antenna port, with a total of N T =2N x N y N z = 32 physical antenna ports. The receiver's antenna is a single-polarization uniform linear antenna, with a total of N. R = 4 physical antenna ports. The frequency-selective fading channel is:

[0043] Where t is the sampling time, e is the natural constant; N k N represents the number of subcarriers in the transmission model of the multiple-input multiple-output orthogonal frequency division multiplexing system. p B represents the number of multipaths in the multipath channel, and B represents the bandwidth of the transmission model of the multiple-input multiple-output orthogonal frequency division multiplexing system; b, u, and k are the sequence numbers of the base station, user, and subcarrier, respectively. Let θ represent the power, phase, and delay of the multipath channel between base station b and user u at time t. b,u,t , These represent the departure angle, zenith angle of arrival, and directional angle of arrival of the multipath between base station b and user u at time t, respectively; H b,u,t,k This represents the multipath channel on subcarrier k between base station b and user u at time t; a UE (θ b,u,t ), It is the array response matrix of the user and the base station:

[0044] in It is the Kronecker product; T is the matrix transpose operation; a x (·),a y (·),a z (·) represent the array response matrices in the x, y, and z directions, respectively.

[0045] Multipath channel between base station b and user u at sampling time t U represents a matrix concatenated along a new dimension. A fixed base station b and user u are represented using a base station-user pair {b, u}. N data are collected between base station b and user u. t There are multiple path channels, and the information of these multiple path channels is:

[0046] The following are the specific steps for establishing a transmission model for a multiple-input multiple-output orthogonal frequency division multiplexing system:

[0047] Where, x b,u,t,k y is the signal transmitted by base station b to user u on subcarrier k at time t; u,t,k W is the received signal of user u on subcarrier k at time t; b,u,t,k Let n be the precoding matrix of base station b to user u on subcarrier k at time t; b,u,t,k The additive complex white Gaussian noise emitted by base station b to user u on subcarrier k at time t.

[0048] The following are representative channels. Target generation:

[0049] in, Is with H b,u,t,k A matrix with the same dimensions but not varying with t; N l It is the number of transport layers; It is the variance of the additive complex white Gaussian noise; ||·|| F It is the Frobenius norm; ||·||2 is the 2-norm; It is a singular vector extraction operation, which involves taking the top N vectors after performing singular value decomposition. l The singular vectors corresponding to the maximum singular values. Since the objective function is in the form of a nonlinear logarithmic sum, it is mathematically difficult to solve.

[0050] The following are the specific implementation steps of a Transformer-based encoder.

[0051] As shown in Figure 3, the Transformer-based encoder decomposes multipath channel information into multiple multipath channel information blocks, and maps these blocks to a grid of size d using a fully connected network. model Vectors, then add sine / cosine positional encoding to the vectors:

[0052] Where pos is the position encoding index and i is the vector element index. These are the positional encoding of a single element and the overall output, respectively. The vector then undergoes L self-attention mechanisms. These self-attention mechanisms map the input to a vector of size d through three independent fully connected layers. atten The vectors, the weights of the three fully connected layers are respectively represented by Ω. Θ ,Ω K ,Ω V The outputs of the three fully connected layers are represented by Θ, K, and V, respectively, and the self-attention layer... This can be represented as: Θ = Ω Θ • input, K = Ω K input, V = Ω V ·input,

[0053] Here, `input` represents the input, and `softmax` is the normalized exponential function. Finally, the output after the self-attention mechanism is passed through average pooling and a fully connected layer to obtain the output of the Transformer-based encoder, which represents the multipath channel information.

[0054] The specific implementation steps of the comparative learning framework are given below.

[0055] As shown in Figure 2, consider a positive sample of multipath channel information. and One negative sample of multipath channel information The contrastive learning framework trains two Transformer-based encoders that make the outputs of multipath channel information at different sampling times in positive samples as similar as possible, while making the outputs of positive and negative samples as dissimilar as possible, denoted by ε. pos ,ε neg It means that ε pos The loss function is designed as follows:

[0056] Where exp is the natural exponential function; t′ is a different sampling time than t; and γ is the temperature parameter. neg The parameter θ neg According to ε pos The parameter θ pos Update: θ neg ←βθ neg +(1-β)θ pos ,

[0057] Where β represents the update factor. Each multipath channel information negative sample is maintained by a first-in-first-out queue, and each training ε neg Output a batch of channel characterization information into a queue, and remove the first batch of channel characterization information that entered the queue from the queue.

[0058] The following describes the specific implementation steps of a Transformer-based decoder.

[0059] The Transformer-based decoder takes multipath channel representation information as input and outputs multipath channel information. Its purpose is to recover the input from the output of the Transformer-based encoder. The loss function of the Transformer-based decoder is:

[0060] in For a Transformer-based decoder, F b,u,t N represents the multipath information channel characterization output by the encoder. u This represents the total number of users.

[0061] To make this embodiment more intuitive and demonstrate the performance advantages of the method, this invention first generates multipath channel information in MATLAB 2023a. This invention sets the number of multipath paths N. p =10, number of subcarriers N k=128, bandwidth B=192MHz, and other channel parameters were determined by the ray tracing model. 50 channel data points were collected from 2 base stations and 3030 users, for a total of 303,000 multipath channel information points were collected. Secondly, the encoder and decoder were trained in a Python 3.8 and PyTorch 1.13.0 environment. The hardware platform used in this invention is a 12th generation Intel Core i7-12700H. The encoder's learning rate was designed to be 5e-5, and the encoder divides the input into 16*16 blocks, taking d... mdel =512, and undergoes L=4 self-attention mechanisms before finally being mapped to a vector of size 128 by a fully connected layer. The decoder and encoder use the same d mdel The learning rate and the size of the blocks are 1*1.

[0062] This invention uses throughput as a performance indicator to verify the method proposed in this invention:

[0063] 1. As shown in Figure 4, we will determine the transmission layer number N. l The performance of the representative channel was tested under different transmission settings, set to 1, 2, 3, and 4 respectively, and the average throughput of 3030 users was obtained. Figure 4 shows that directly averaging the original channel does not yield a good representative channel, and the throughput is lower than that of the original channel. However, the method proposed in this invention, which averages the multipath channel characterization information and then reconstructs the multipath channel using a decoder, achieves better performance than the original channel, demonstrating the effectiveness of our method. Furthermore, with N... l As the number of transmission layers increases, the performance improvement of this invention will decrease. This is because higher transmission layers require a more accurate description of multipath, which also shows that the representative channel of this invention provides an accurate description of multipath characteristics.

[0064] 2. As shown in Figures 5 and 6, this invention targets N. l When the value is 1, we can observe the throughput performance of each user. As can be seen from the figure, the representative channel obtained by this invention generally has better performance than the representative channel obtained by directly using the average of the original channels. This shows that the method of this invention has universality and can be extended to other scenarios.

[0065] Based on the same inventive concept, this invention discloses a representative channel generation system for a 6G fully decoupled network based on contrastive learning, comprising: a modeling module for establishing a multipath channel model and a multiple-input multiple-output (MIMO) orthogonal frequency division multiplexing (MFD) system transmission model, and providing a representative channel generation target based on the MFD system transmission model; an encoder training unit for establishing a multipath channel information dataset based on the multipath channel model, and training a Transformer-based encoder using a contrastive learning framework, the encoder taking multipath channel information as input and outputting multipath channel representation information; a decoder training unit for training a Transformer-based decoder based on the multipath channel information and multipath channel representation information, the decoder taking multipath channel representation information as input and outputting multipath channel information; and a representative channel generation unit for inputting multipath channel information during the actual deployment phase, the encoder outputting multipath channel representation information, and inputting the multipath channel representation information after averaging the multipath channel representation information over the time dimension into the decoder to output a representative channel.

[0066] Based on the same inventive concept, this invention discloses a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the aforementioned method for generating a representative channel in a 6G fully decoupled network based on contrastive learning.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A 6G full decoupling network representative channel generation method based on contrastive learning, characterized in that, The method comprises the following steps: Step 1: establishing a multipath channel model and a multi-input multi-output orthogonal frequency division multiplexing system transmission model; Step 2: giving a representative channel generation target according to the multi-input multi-output orthogonal frequency division multiplexing system transmission model; Step 3: establishing a multipath channel information dataset according to the multipath channel model, and training a Transformer-based encoder using a contrastive learning framework, the encoder taking multipath channel information as input and outputting multipath channel representation information; Step 4: training a Transformer-based decoder according to the multipath channel information and the multipath channel representation information, the decoder taking multipath channel representation information as input and outputting multipath channel information; Step 5: inputting multipath channel information in the actual deployment stage, the encoder outputting multipath channel representation information, the multipath channel representation information after being averaged in the time dimension being inputted into the decoder, and outputting a representative channel.

2. The 6G fully decoupled network representative channel generation method based on contrastive learning according to claim 1, characterized in that: The multipath channel model is: where t is the sampling time, e is the natural constant; N k N is the subcarrier number of the MIMO-OFDM system transmission model, N p N is the subcarrier number of the MIMO-OFDM system transmission model, N respectively denote the power, phase and delay of the multipath channel between base station b and user u at time t; are respectively the angle of departure, the zenith angle of arrival and the direction angle of arrival of the multipath between the base station b and the user u at time t; H b,u,t,k denotes the multipath channel on the subcarrier k between the base station b and the user u at time t; a UE (θ b,u,t ), is an array response matrix of users and base stations; Multipath channel between base station b and user u at time t sample ∪ denotes concatenation along a new dimension, N R is the number of receive-side physical antenna ports, T is the number of transmit-side physical antenna ports; a base station b and a user u are denoted by a base station-user pair {b, u}, and N t is the number of multipath channels between base station b and user u, and the multipath channel information is 3. The method of claim 2, wherein the method is a method of generating a representative channel for a 6G fully decoupled network based on contrastive learning. For a user with a uniform linear antenna and a base station with a uniform planar array antenna, the array response matrix is represented as: wherein is the Kronecker product; T is the matrix transpose operation; a x (·),a y (·),a z (·) are the array response matrices in x, y, z directions, respectively: where N x ,N y ,N z are the number of physical antenna ports of the base station in the x, y, z directions, respectively.

4. The method of claim 1, wherein the method is based on a contrastive learning-based 6G fully decoupled network representative channel generation method. The multiple-input multiple-output orthogonal frequency division multiplexing system transmission model is: where x b,u,t,k is the signal transmitted by base station b to user u on subcarrier k at time t; y u,t,k is the received signal of user u on subcarrier k at time t; W b,u,t,k is the precoding matrix of base station b to user u on subcarrier k at time t; H b,u,t,k is the multipath channel on subcarrier k between base station b and user u at time t; n b,u,t,k is the additive complex Gaussian white noise on subcarrier k of base station b to user u at time t; N b is the total number of base stations.

5. The 6G fully decoupled network representative channel generation method based on contrastive learning according to claim 1, characterized in that: The representative channel The generation target is: wherein is the H b,u,t,k a matrix with the same dimensions but not varying with t; H b,u,t,k is the multipath channel on subcarrier k between base station b and user u at time t; N l is the number of transmission layers; is the variance of the additive complex white noise; || · || F is the Frobenius norm; || · ||2is the 2-norm; is to take singular vector operation, that is, to take singular vector corresponding to the first N l maximum singular value after singular value decomposition is performed on 6. The 6G fully decoupled network representative channel generation method based on contrastive learning according to claim 1, characterized in that: The Transformer-based encoder decomposes the multi-path channel information into multiple multi-path channel information blocks, and maps the multi-path channel information blocks into vectors of size d model The vectors are then added with sine / cosine positional encodings: where pos is the position encoding number, i is the vector element number, are position encoding single elements and overall outputs, respectively; then the vector undergoes multiple self-attention mechanisms; the self-attention mechanisms map the input to a vector of size d atten through three independent fully connected layers, the weights of which are denoted by Ω Θ , Ω K , Ω V , respectively, and the outputs of the three fully connected layers are denoted by Θ, K, V, respectively, and the self-attention is expressed as: Θ = Ω Θ • input, K = Ω K • input, V = Ω V • input, wherein input represents input, softmax is a normalized exponential function; finally, the output after the self-attention mechanism is subjected to average pooling and a fully connected layer to obtain the output of the Transformer-based encoder, which is multipath channel representation information.

7. The 6G fully decoupled network representative channel generation method based on contrastive learning according to claim 2, characterized in that: The contrast learning framework realizes self-supervised learning by using positive and negative samples, considers one multi-path channel information positive sample and multi-path channel information negative samples The contrast learning training framework is used for training two encoders based on the Transformer, which makes the outputs of the multi-path channel information at different sampling times in the positive sample as similar as possible, and the outputs of the positive sample and the negative sample as dissimilar as possible, respectively represented by ε pos ,ε neg and ε pos The loss function is: where exp is the natural exponential function; t' is another sampling time instant different from t, γ is a temperature parameter; ε neg the parameter θ neg according to the parameter θ pos of ε pos , update: θ neg ← βθ neg + (1 - β)θ pos , wherein β represents an update factor; The multipath channel information negative samples are maintained by a first-in-first-out queue, and ε neg The channel representation information of one batch is output into the queue, and the channel representation information of one batch which enters the queue first is allowed to exit the queue.

8. The 6G fully decoupled network representative channel generation method based on contrastive learning according to claim 1, characterized in that: The input of the Transformer-based decoder is multipath channel representation information, and the output is multipath channel information. The loss function of the Transformer-based decoder is: wherein F is a Transformer-based decoder, b,u,t H is a multipath information channel representation output by the encoder, b,u,t H is a multipath channel between a base station b and a user u; N b N is the total number of base stations, u N is the total number of users, t N is the number of time instants.

9. A 6G fully decoupled network representative channel generation system based on contrastive learning, characterized in that, It comprises: a modeling module for establishing a multipath channel model and a multi-input multi-output orthogonal frequency division multiplexing system transmission model, and giving a representative channel generation target according to the multi-input multi-output orthogonal frequency division multiplexing system transmission model; an encoder training unit for establishing a multipath channel information dataset according to the multipath channel model, and training a Transformer-based encoder using a contrastive learning framework, the encoder taking multipath channel information as input and outputting multipath channel representation information; a decoder training unit for training a Transformer-based decoder according to the multipath channel information and the multipath channel representation information, the decoder taking multipath channel representation information as input and outputting multipath channel information; and a representative channel generation unit for inputting multipath channel information in the actual deployment stage, the encoder outputting multipath channel representation information, the multipath channel representation information after being averaged in the time dimension being inputted into the decoder, and outputting a representative channel.

10. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the steps of the contrastive learning-based 6G full decoupling network representative channel generation method according to any one of claims 1-8.

Images