Robust joint user scheduling and precoding method and system based on neural network

By employing a robust joint user scheduling and precoding method based on neural networks, the performance degradation caused by channel aging is resolved, achieving efficient user rate enhancement and low computational complexity, making it suitable for mobile communication systems.

CN118984176APending Publication Date: 2024-11-19SOUTHEAST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410926421.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

Existing user scheduling and precoding algorithms suffer from poor performance in mobile communication scenarios due to channel aging, failing to achieve optimal performance and exhibiting high computational complexity.

Method used

A robust joint user scheduling and precoding method based on neural networks is adopted. Channel state information and statistical information are collected by the base station to establish a receiving channel model. The model is transformed into low-dimensional features using a neural network with permutation equivariance properties. A robust joint optimization scheme is designed by combining precoding and user scheduling neural networks.

Benefits of technology

It improves user rate performance, reduces computational complexity, and has good generalization and robustness, effectively mitigating the impact of channel aging.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118984176A_ABST
    Figure CN118984176A_ABST
Patent Text Reader

Abstract

The invention provides a robust united user scheduling and precoding method and system based on a neural network, and a base station carries out united user scheduling and precoding during downlink transmission according to channel state information obtained from an uplink channel and statistical prior information obtained through long-term observation. And designing a robust united user scheduling and precoding scheme by taking maximization of user expectation and rate as a target. In order to simplify solving, a joint optimization problem is converted into two sub-problems for solving: solving a robust precoding sub-problem, and for any user scheduling scheme, an optimal precoding vector under the scheme can be obtained; and a robust precoding algorithm is given, and a robust user scheduling sub-problem is solved. According to the method, a deep learning robust joint user scheduling and precoding algorithm meeting permutation equivariability is designed by utilizing a permutation equivariability deep neural network structure. According to the algorithm, the user performance and the rate performance are improved, the calculation complexity is low, and good robustness is achieved for a scene with channel aging.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of communications, and in particular relates to a robust joint user scheduling and precoding method and system based on a neural network. Background Art

[0002] With the development of the information age, the demand for wireless communication data transmission has greatly increased. Deploying large-scale antenna arrays at base stations can make full use of space resources and increase the system channel capacity exponentially without increasing spectrum resources and antenna transmission power. Therefore, it is regarded as the core technology of mobile communications.

[0003] User scheduling refers to selecting appropriate users for data transmission in a multi-user communication system to maximize system performance. Precoding refers to encoding data at the transmitter in a multi-antenna system to achieve spatial signal interference suppression. User scheduling and precoding are performed at different layers of the communication system, so these two technologies are usually studied separately. However, the performance of user scheduling depends on the precoding criteria, and precoding is performed on the scheduled user set. Therefore, independent design of the two technologies cannot achieve optimal performance, and a joint design of the two technologies is required.

[0004] Currently, most user scheduling and precoding algorithms are designed based on accurate channel state information. However, it is challenging to obtain accurate channel state information at the transmitter. Usually, in a static channel environment, relatively accurate channel state information can be obtained by using channel estimation methods. In mobile scenarios, the speed of channel change is affected by the relative movement speed of the device and the base station. Channel aging is the change process that the channel undergoes from uplink channel estimation to downlink transmission time, and it is also one of the important reasons for imperfect channel information in a mobile environment. Therefore, it is necessary to design robust user scheduling and precoding algorithms to counter the impact of channel aging. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a robust joint user scheduling and precoding method and system based on neural network, which has good robustness in the scenario of channel aging, can improve user and rate performance, and has low computational complexity.

[0006] Technical solution: To achieve the above-mentioned invention object, a robust joint user scheduling and precoding method based on a neural network of the present invention comprises the following steps:

[0007] The base station collects channel state information with channel aging and statistical channel information obtained through long-term observation through the uplink channel, models the downlink channel to obtain a receiving channel model affected by the channel aging effect, and further obtains a receiving signal model;

[0008] Based on the received signal model, the robust joint user scheduling and precoding optimization problem is established with the base station's transmit power as a constraint to maximize the user's expectation and rate.

[0009] The joint optimization problem is transformed into two sub-problems: the robust precoding sub-problem of maximizing user expectations and rates; and the robust user scheduling sub-problem of maximizing user expectations and rates given a robust precoding algorithm.

[0010] The robust precoding problem is transformed into a robust weighted mean square error minimization problem, and the low-dimensional features of the closed-form solution of the transformed problem are obtained by using a precoding neural network with permutation equivariant properties, and the precoding matrix is ​​further obtained; the precoding neural network input includes Noise Power And the maximum transmission power P of the base station T , the output low-dimensional feature is the receiving factor vector and the weight vector in is the number of user terminals selected by the base station through user scheduling, M is the number of antennas of the base station, β k is the time correlation coefficient, is the channel state information of user k obtained by uplink estimation, is the statistical channel information of user k obtained through long-term observation;

[0011] Under a given robust precoding algorithm, a user scheduling neural network with permutation equivariant properties is used to obtain a user scheduling indication vector to obtain a scheduled user set; the user scheduling neural network input includes and P T , output user scheduling indication vector Where K is the number of all user terminals covered by the base station;

[0012] By combining the precoding neural network and the user scheduling neural network, a robust joint user scheduling and precoding scheme based on neural network is obtained.

[0013] Preferably, the channel model of the downlink channel is expressed as:

[0014]

[0015] Among them, h k is the channel vector from the downlink base station to user k, and the right side of the equal sign represents the deterministic part and the random part respectively. is a matrix composed of discrete Fourier transform matrices, is a random vector whose elements are independent and obeys a complex Gaussian distribution with variance 1, and ⊙ is the Hadamard product.

[0016] Preferably, the joint optimization problem is expressed as:

[0017]

[0018] in represents the expected rate of user k, represents the matrix composed of all user precoding vectors, Represents the collection of all users.

[0019] As a preferred method, the joint optimization problem is transformed into the following two sub-problems:

[0020]

[0021] For each indicator vector η, there is a corresponding optimal precoding matrix Optimal precoding matrix By solving We can take it as an input of η and an output of ; in obtaining Based on the mapping, we further obtain an optimization problem that is only related to η Solution The problem is to obtain a robust user scheduling scheme based on the robust precoding algorithm.

[0022] Preferably, the robust minimization weighted mean square error problem is expressed as:

[0023]

[0024] where ε k represents the mean square error of the receiver at user k, represents the set of scheduled users; the closed-form solution to this problem is:

[0025]

[0026] Where (i) represents the i-th iteration, I M represents the unit matrix of dimension M, and As shown below:

[0027]

[0028]

[0029] Where diag{·} means converting the vector into a diagonal matrix, [·] m represents the mth element of a vector, [·] mm Represents the elements of the matrix with m rows and m columns.

[0030] Preferably, the robust precoding neural network includes a preprocessing layer and a feature extraction module. The preprocessing layer concatenates the data into of the form, where and Represent the real part and the imaginary part respectively, 1 represents the all-1 vector, the feature extraction module includes a multi-layer set attention module SAB, a linear layer and an output layer; the data X input SAB expression is:

[0031] SAB(X)=LayerNorm(S+rFF(S))

[0032] Where LayerNorm is the layer normalization function, rFF is the row-wise feedback layer, which processes each user's data independently and identically, and neither is affected by the order of input arrangement; the expression of S is:

[0033] S = LayerNorm(X+Attention(X,N h ))

[0034] in N h Indicates the number of multiple heads, and Cat means splicing the matrix in the row direction. Where Q, K, and V represent the query, key, and value matrices respectively, and Q = K = V = X. are all model parameters of the neural network, and d represents the normalization parameter. After the data passes through multiple layers of SAB, the linear layer and activation function are integrated and output to obtain the characteristic parameters The loss function is the inverse of user expectation and rate, and the precoding neural network is trained in an unsupervised manner.

[0035] Preferably, the robust user scheduling neural network includes a preprocessing layer, a feature extraction module and a quantization module; the preprocessing layer splices the data into The feature extraction module includes a multi-layer set attention module, a linear layer and an output layer; the probability vector of user scheduling is obtained through the integration output of the linear layer and the activation function. The quantization module is used to quantize the probability value between 0 and 1 to 0 or 1 to obtain the final scheduling variable η; the robust precoding network is used to greedily schedule with the goal of maximizing the user and rate to generate the scheduling label, which is used for supervised learning and training of the robust user scheduling network.

[0036] Preferably, the combining the precoding neural network and the user scheduling neural network to obtain a robust joint user scheduling and precoding scheme comprises:

[0037] In the offline training phase, the robust precoding network is trained to obtain the mapping Solve the precoding matrix P according to the closed form; then use the robust precoding network to greedily schedule the label of the scheduling network with the goal of maximizing the user and rate; train the robust user scheduling network to obtain the mapping

[0038]

[0039] In the online use phase, the channel data is input into the user scheduling network to obtain η; the channel data of the scheduled user is input into the precoding network to obtain P.

[0040] Based on the same inventive concept, the present invention provides a neural network-based robust joint user scheduling and precoding system, comprising:

[0041] A joint optimization problem modeling module is used for the base station to collect channel state information with channel aging through the uplink channel, as well as statistical channel information obtained by long-term observation, model the downlink channel to obtain a receiving channel model affected by the channel aging effect, and further obtain a receiving signal model; and based on the receiving signal model, with the base station's transmit power as a constraint, establish a robust joint user scheduling and precoding optimization problem for the criterion of maximizing user expectations and rates;

[0042] A problem conversion module is used to convert the joint optimization problem into two sub-problems: a robust precoding sub-problem of maximizing user expectations and rates; and a robust user scheduling sub-problem of maximizing user expectations and rates given a robust precoding algorithm;

[0043] A neural network building module is used to transform the robust precoding problem into a robust weighted mean square error minimization problem, and use a permutation equivariant precoding neural network to obtain the low-dimensional features of the closed-form solution of the transformed precoding problem, and further obtain the precoding matrix; the precoding neural network input Noise Power And the maximum transmission power P of the base station T , the output low-dimensional feature is the receiving factor vector and the weight vector Where K is the number of user terminals selected by the base station through user scheduling, M is the number of antennas of the base station, β k is the time correlation coefficient, is the channel state information of user k obtained by uplink estimation, is the statistical channel information of user k obtained through long-term observation; and under a given robust precoding algorithm, a user scheduling neural network with permutation equivariant properties is used to obtain a user scheduling indication vector to obtain a scheduled user set; the user scheduling neural network input and PT , output user scheduling indication vector Where K is the number of all user terminals covered by the base station;

[0044] The joint user scheduling and precoding module is used to combine the obtained precoding neural network and the user scheduling neural network to obtain a robust joint user scheduling and precoding scheme based on neural network.

[0045] 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 neural network-based robust joint user scheduling and precoding method.

[0046] Beneficial effects: The present invention provides a method for robust joint user scheduling and precoding based on a neural network. The method uses a posterior channel model based on a joint correlation channel model, and designs a robust joint user scheduling and precoding scheme for the criterion of maximizing user expectations and rates. The optimization problem is converted into two related sub-problems: 1. Robust precoding sub-problem, for any user scheduling scheme, the optimal precoding matrix under the scheme is obtained; 2. Robust user scheduling problem under a given robust precoding scheme. Combined with the permutation equivariant property of the problem itself, a robust joint user scheduling and precoding scheme is proposed using a neural network that satisfies this property. By introducing a neural network to replace the iterative process, a mapping is established with outdated channel state information and statistical channel information as input and low-dimensional features as output, achieving high performance and low energy consumption, and improving system performance. Compared with the prior art, the present invention adopts a hybrid model-driven and data-driven approach, which can alleviate the performance loss caused by model mismatch. In addition, the combination of neural networks can further reduce the complexity of the calculation, and the permutation equivariant property of the network also makes the algorithm have good generalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of a multi-user communication system;

[0048] Figure 2 It is the overall flow chart of the method of the present invention;

[0049] Figure 3 The figure is a detailed method flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0051] The embodiment of the present invention discloses a robust joint user scheduling and precoding method based on a neural network, Figure 1 The wireless communication transmission system is a MU-MISO system, in which the base station is equipped with M antennas, and the cell coverage area contains K single-antenna terminals. All terminals use a collection The base station is selected by user scheduling. user terminals, the scheduled terminals are used as a set Therefore, the channel state information collected by the uplink channel is aged channel state information, that is, there is a gap between the time when the channel state information is collected in the uplink channel and the time when the downlink is transmitted.

[0052] Figure 2 The flowchart of the scheme is shown in Figure 1. The method first models the channel of downlink transmission, obtains the expression of the user's received signal, and designs the optimization problem of robust joint user scheduling and precoding that maximizes the user's expectation and rate; the above joint optimization problem is transformed to obtain the sub-problem of robust precoding design and the sub-problem of robust user scheduling given the precoding scheme; for sub-problem 1, by finding the lower bound of the objective function of the robust precoding problem, the problem is transformed into a problem of minimizing the weighted mean square error. This problem has a closed-form solution, and a stable solution can be found through iteration. In order to simplify the process, a deep neural network with permutation equivariant properties is used to replace the iterative process, and a network with channel state information as input and low-dimensional feature parameters in the closed-form solution as output is obtained. The precoding vector can be further restored using the output of the network. For sub-problem 2, the user scheduling problem can be further solved given the robust precoding network, and the label of the user scheduling network is obtained by greedy means. Using the label, a user scheduling network that satisfies the permutation equivariant property is further obtained through supervised training. Thus, a robust joint user scheduling and precoding scheme based on a neural network is obtained. The details are as follows:

[0053] Step 1: The base station collects aging channel state information and statistical priors obtained from long-term observations through the uplink channel, and models the downlink channel according to the a posteriori channel model based on the joint correlation channel model, thereby obtaining a received signal model affected by the channel aging effect.

[0054] Taking into account the impact of channel aging, the posterior channel model is expressed as follows:

[0055]

[0056] in, is the channel matrix estimated by the base station uplink channel, is the downlink channel matrix from the base station to user k, is a matrix composed of discrete Fourier transform matrices, is a sparse vector with non-negative elements, that is, the statistical channel information is obtained through long-term observation. is a random vector with independent elements and a complex Gaussian distribution with variance 1, β k The correlation coefficient related to the mobile speed and the aging time introduced by the first-order Markov model to model the channel time variation is ⊙, and ⊙ is the Hadamard product. For a detailed description of the specific a posteriori channel model, please refer to [A.-A.Lu, X.Gao, W.Zhong, C.Xiao, and X.Meng, "Robust transmission for massive MIMO downlink with imperfect CSI," IEEE Trans.Commun., vol.67, no.8, pp.5362–5376].

[0057] Based on this, the received signal of user k is expressed as:

[0058]

[0059] Among them, η k ∈{0,1} is the scheduling instruction of user k, p k is the precoding vector of user k, whose dimension is M×1, d k To transmit a signal to user k, z k is additive noise, subject to distributed, is the noise power.

[0060] Step 2: Based on the received signal model and taking the base station's transmit power as a constraint, a robust joint user scheduling and precoding optimization problem is established for the criterion of maximizing user expectations and rates.

[0061] The expected rate of user k at the receiving end is expressed as:

[0062]

[0063] in and They represent the user scheduling indication vector and , respectively, the matrix composed of the precoding of all users. Indicates the channel h k Find the expected value, the same below. The maximum transmission power P of the base station T Under the constraints of , with the goal of maximizing user expectations and rates, the joint optimization problem can be expressed as:

[0064]

[0065] Step 3, transform the joint optimization problem into two sub-problems: 1. The robust precoding sub-problem of maximizing user expectations and rates; 2. Given a robust precoding algorithm, solve the robust user scheduling sub-problem of maximizing user expectations and rates.

[0066] The optimization variables of the above problem include discrete variables η and continuous variables P A , is a mixed integer nonlinear programming problem. According to the optimization problem It can be observed that for each indicator vector η there is a corresponding optimal precoding matrix It is expressed as follows:

[0067]

[0068] Optimal precoding matrix It can be solved by Here, it is regarded as an input of η and an output of The mapping of Based on the mapping, we can further obtain an optimization problem that is only related to η

[0069]

[0070] Solution The problem can be solved by obtaining a robust user scheduling scheme based on the robust precoding algorithm.

[0071] Step 4: Solve the robust precoding problem. The robust precoding problem is transformed into a robust weighted mean square error minimization problem, and a permutation-equivariant neural network is used to obtain the low-dimensional features of the closed-form solution of the transformed problem, and further obtain the precoding matrix.

[0072] The robust precoding subproblem Transformed into a robust minimization weighted mean square error problem. It is easy to see that when η k = 0, the p of the unscheduled user k = 0. Therefore, It can be simplified as

[0073]

[0074] in represents the precoding matrix for scheduling users. Both represent the set of scheduled users Considering the non-convex nature of the problem, by finding the lower bound of the problem, the problem can be further transformed into the following robust weighted minimization of mean square error problem.

[0075]

[0076] in and denote the receiving factor vector and weight vector respectively, Represents the mean square error of the receiver at user k. Combining the Lagrange multiplier method to solve the optimization problem, the following iterative closed-form solution can be obtained:

[0077]

[0078] Where (i) represents the i-th iteration. In the above closed-form expression and As shown below

[0079]

[0080] in Combined with neural network to extract low-dimensional features in the closed-form expression of the above pre-coded vector Simplify the computational complexity of the algorithm. Construct a neural network with the following input and output:

[0081]

[0082] in The robust precoding neural network is divided into a preprocessing layer and a feature extraction module. The preprocessing layer concatenates the data into The feature extraction module consists of a multi-layer set attention module (SAB), a linear layer and an output layer, and has the property of permutation equivariance as a whole. The data X input SAB expression is

[0083] SAB(X)=LayerNorm(S+rFF(S))

[0084] Among them, LayerNorm is the layer normalization function, which retains the size relationship between different features of users. rFF is the row-wise feedback layer, which processes the data of each user independently and identically. Both are not affected by the order of input arrangement. The expression of S is:

[0085] S = LayerNorm(X+Attention(X,N h ))

[0086] in N h Indicates the number of multiple heads, and Cat means splicing the matrix in the row direction. Where Q, K, and V represent the Query, Key, and Value matrices, respectively. In this embodiment, Q = K = V = X. are all model parameters of the neural network, and d is a normalization parameter. After the data passes through multiple layers of SAB, the characteristic parameters are obtained by integrating the linear layer and the activation function. The output characteristic parameters are input into the precoding recovery module to further recover the user's precoding matrix P. The user channel can be generated through the Quadriga simulation generator, and the user's statistical channel state information can be further obtained. The network is trained in an unsupervised manner using the inverse of the user's expectation and rate as the loss function.

[0087] Step 5: Solve the robust user scheduling problem. Under a given robust precoding algorithm, a permutation-equivariant neural network is used to obtain the user scheduling indication vector and obtain the scheduled user set.

[0088] Similarly, the following robust user scheduling network is constructed, with the user scheduling indication vector as the feature:

[0089]

[0090] in The neural network of robust user scheduling is divided into a preprocessing layer, a feature extraction module, and a quantization module. The preprocessing layer concatenates the data into The feature extraction module consists of a multi-layer set attention module (SAB), a linear layer and an output layer, and has the property of permutation equivariance as a whole. After the data passes through the multi-layer SAB, the linear layer and the activation function are integrated to obtain the probability vector of user scheduling. The quantization module is used to further sort the probability values ​​between 0 and 1 into 0 or 1 according to the number of user scheduling, and the final scheduling vector η is obtained. The robust precoding network is used to greedily schedule the maximum number of users and the rate to generate the scheduling label, which is used for supervised learning and training of the robust user scheduling network here.

[0091] In summary, a robust joint user scheduling and precoding structure can be obtained, such as Figure 3 The joint solution can be expressed as follows:

[0092] Offline training phase

[0093] a. Train the robust precoding network to obtain the mapping

[0094] b. Solve the precoding matrix P according to the closed form;

[0095] c. Using the robust precoding network, greedy scheduling is performed to maximize the number of users and the rate to obtain the label of the scheduling network;

[0096] d. Train the robust user scheduling network to obtain the mapping Online use stage

[0097] a. Input the channel data into the user scheduling network to obtain η;

[0098] b. Input the channel data of the scheduled user into the precoding network to obtain P.

[0099] Based on the same inventive concept, an embodiment of the present invention discloses a robust joint user scheduling and precoding system based on a neural network, comprising: a joint optimization problem modeling module, which is used for a base station to collect channel state information with channel aging through an uplink channel, as well as statistical channel information obtained by long-term observation, model a downlink channel to obtain a receiving channel model affected by the channel aging effect, and further obtain a receiving signal model; and based on the receiving signal model, taking the transmission power of the base station as a constraint, establishing a robust joint user scheduling and precoding optimization problem for the criterion of maximizing user expectations and rates; a problem conversion module, which is used to convert the joint optimization problem into two sub-problems: a robust precoding sub-problem of maximizing user expectations and rates; and given a robust precoding algorithm, solving a robust user scheduling sub-problem of maximizing user expectations and rates; a neural network construction module, which is used to convert the robust precoding problem into a robust minimization weighted mean square error problem, and use a precoding neural network with permutation equivariant properties to obtain low-dimensional features of a closed-form solution of the converted precoding problem, and further obtain a precoding matrix; and a joint user scheduling and precoding module, which is used to combine the obtained precoding neural network and user scheduling neural network to obtain a robust joint user scheduling and precoding scheme based on a neural network. The specific implementation process of each module can be found in the above method embodiment and will not be repeated here.

[0100] Based on the same inventive concept, an embodiment of the present invention discloses a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the neural network-based robust joint user scheduling and precoding method.

[0101] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A robust joint user scheduling and precoding method based on neural network, characterized in that: The following steps are involved: The base station collects channel state information with channel aging and statistical channel information obtained through long-term observation through the uplink channel, models the downlink channel to obtain a receiving channel model affected by the channel aging effect, and further obtains a receiving signal model; Based on the received signal model, the robust joint user scheduling and precoding optimization problem is established with the base station's transmit power as a constraint to maximize the user's expectation and rate. The joint optimization problem is transformed into two sub-problems: the robust precoding sub-problem of maximizing user expectations and rates; and the robust user scheduling sub-problem of maximizing user expectations and rates given a robust precoding algorithm. The robust precoding problem is transformed into a robust weighted mean square error minimization problem, and the low-dimensional features of the closed-form solution of the transformed problem are obtained by using a precoding neural network with permutation equivariant properties, and the precoding matrix is ​​further obtained; the precoding neural network input includes Noise Power And the maximum transmission power P of the base station T , the output low-dimensional feature is the receiving factor vector and the weight vector in is the number of user terminals selected by the base station through user scheduling, M is the number of antennas of the base station, β k is the time correlation coefficient, is the channel state information of user k obtained by uplink estimation, is the statistical channel information of user k obtained through long-term observation; Under a given robust precoding algorithm, a user scheduling neural network with permutation equivariant properties is used to obtain a user scheduling indication vector to obtain a scheduled user set; the user scheduling neural network input includes and P T , output user scheduling indication vector Where K is the number of all user terminals covered by the base station; By combining the precoding neural network and the user scheduling neural network, a robust joint user scheduling and precoding scheme based on neural network is obtained.

2. The method for robust joint user scheduling and precoding based on a neural network according to claim 1, characterized in that: The channel model of the downlink channel is expressed as: Among them, h k is the channel vector from the downlink base station to user k, and the right side of the equal sign represents the deterministic part and the random part respectively. is a matrix composed of discrete Fourier transform matrices, is a random vector whose elements are independent and obeys a complex Gaussian distribution with variance 1, and ⊙ is the Hadamard product.

3. The method for robust joint user scheduling and precoding based on neural network according to claim 1, characterized in that: The joint optimization problem is expressed as: in represents the expected rate of user k, represents the matrix composed of all user precoding vectors, Represents the collection of all users.

4. The method for robust joint user scheduling and precoding based on neural network according to claim 3, characterized in that: The joint optimization problem is transformed into the following two sub-problems: For each indicator vector η, there is a corresponding optimal precoding matrix Optimal precoding matrix By solving We can take it as an input of η and an output of ; in obtaining Based on the mapping, we further obtain an optimization problem that is only related to η Solution The problem is to obtain a robust user scheduling scheme based on the robust precoding algorithm.

5. The method for robust joint user scheduling and precoding based on neural network according to claim 4, characterized in that: The robust minimization weighted mean square error problem is expressed as: where ε k represents the mean square error of the receiver at user k, represents the set of scheduled users; the closed-form solution to this problem is: Where (i) represents the i-th iteration, I M represents the unit matrix of dimension M, and As shown below: where h k is the downlink channel vector from the base station to user k, is a matrix composed of discrete Fourier transform matrices, diag{·} means converting a vector into a diagonal matrix, [·] m represents the mth element of a vector, [·] mm Represents the elements of the matrix with m rows and m columns.

6. The method for robust joint user scheduling and precoding based on neural network according to claim 1, characterized in that: The robust precoding neural network includes a preprocessing layer and a feature extraction module. The preprocessing layer concatenates the data into of the form, where and Represent the real part and the imaginary part respectively, 1 represents the all-1 vector, and the feature extraction module includes a multi-layer set attention module SAB, a linear layer and an output layer; the data X input SAB expression is: SAB(X)=LayerNorm(S+rFF(S)) Where LayerNorm is the layer normalization function, rFF is the row-wise feedback layer, and processes each user's data independently and identically. Both are not affected by the order of input arrangement. The expression of S is: S=LayerNorm(X+Attention(X,N h )) in N h Indicates the number of multiple heads, Cat means splicing the matrix in the row direction, Where Q, K, and V represent the query, key, and value matrices respectively, Q = K = V = X, W, are all model parameters of the neural network, d represents the normalization parameter, and after the data passes through multiple layers of SAB, the linear layer and the activation function are integrated and output to obtain the characteristic parameters The loss function is the inverse of user expectation and rate, and the precoding neural network is trained in an unsupervised manner.

7. The method for robust joint user scheduling and precoding based on neural network according to claim 1, characterized in that: The robust user scheduling neural network includes a preprocessing layer, a feature extraction module, and a quantization module; the preprocessing layer concatenates the data into of the form, where and Represent the real part and the imaginary part respectively, and 1 represents the all-1 vector; the feature extraction module includes a multi-layer set attention module, a linear layer and an output layer; the probability vector of user scheduling is obtained through the integration output of the linear layer and the activation function The quantization module is used to quantize the probability value between 0 and 1 to 0 or 1 to obtain the final scheduling variable η; the robust precoding network is used to greedily schedule with the goal of maximizing the user and rate to generate the scheduling label, which is used for supervised learning and training of the robust user scheduling network.

8. The method for robust joint user scheduling and precoding based on neural network according to claim 1, characterized in that: The combining of the precoding neural network and the user scheduling neural network to obtain a robust joint user scheduling and precoding scheme includes: In the offline training phase, the robust precoding network is trained to obtain the mapping Solve the precoding matrix P according to the closed form; then use the robust precoding network to greedily schedule the label of the scheduling network with the goal of maximizing the user and rate; train the robust user scheduling network to obtain the mapping In the online use phase, the channel data is input into the user scheduling network to obtain η; the channel data of the scheduled user is input into the precoding network to obtain P.

9. A robust joint user scheduling and precoding system based on neural network, characterized in that: include: A joint optimization problem modeling module is used for the base station to collect channel state information with channel aging through the uplink channel, as well as statistical channel information obtained by long-term observation, model the downlink channel to obtain a receiving channel model affected by the channel aging effect, and further obtain a receiving signal model; and based on the receiving signal model, with the base station's transmit power as a constraint, establish a robust joint user scheduling and precoding optimization problem for the criterion of maximizing user expectations and rates; A problem conversion module is used to convert the joint optimization problem into two sub-problems: a robust precoding sub-problem of maximizing user expectations and rates; and a robust user scheduling sub-problem of maximizing user expectations and rates given a robust precoding algorithm; A neural network building module is used to transform the robust precoding problem into a robust weighted mean square error minimization problem, and use a permutation equivariant precoding neural network to obtain the low-dimensional features of the closed-form solution of the transformed precoding problem, and further obtain the precoding matrix; the precoding neural network input Noise Power And the maximum transmission power P of the base station T , the output low-dimensional feature is the receiving factor vector and the weight vector in is the number of user terminals selected by the base station through user scheduling, M is the number of antennas of the base station, β k is the time correlation coefficient, is the channel state information of user k obtained by uplink estimation, is the statistical channel information of user k obtained through long-term observation; and under a given robust precoding algorithm, a user scheduling indication vector is obtained by using a user scheduling neural network with permutation equivariant properties to obtain a scheduled user set; the user scheduling neural network input and P T , output user scheduling indication vector Where K is the number of all user terminals covered by the base station; The joint user scheduling and precoding module is used to combine the obtained precoding neural network and the user scheduling neural network to obtain a robust joint user scheduling and precoding scheme based on neural network.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by the processor, the steps of the robust joint user scheduling and precoding method based on neural network according to any one of claims 1-8 are implemented.