Multi-cell large-scale MIMO user scheduling method using position information

Abstract

The invention discloses a multi-cell large-scale MIMO (Multiple Input Multiple Output) user scheduling method using position information. A base station and a time-frequency resource block to which a user belongs are determined directly according to the position of the user. Firstly, the multi-cell user scheduling problem is converted into a sequence-to-sequence learning task, an input sequence is composed of position coordinates of users to be scheduled, and an output sequence is a corresponding scheduling scheme. The method comprises the following steps of: firstly, establishing an end-to-end mapping mechanism from user position information to a scheduling scheme, then, constructing an MUGFormer (Multi-cell User Grouping Former) network based on a Transform architecture in a central scheduler, and establishing an end-to-end mapping mechanism from the user position information to the scheduling scheme by adopting a data-driven strategy and combining with a supervised learning framework; according to the invention, high system performance can be realized with low complexity, and the problem of user scheduling of a multi-cell large-scale MIMO system is effectively solved.

Description

Technical Field

[0001] The present invention relates to user scheduling and time-frequency resource allocation, and particularly to a multi-cell massive MIMO user scheduling method using location information, belonging to the field of wireless communication technology. Background Art

[0002] With the gradual commercial implementation of the 5th Generation Mobile Communications Systems (5G), the research focus in the communication field has gradually shifted to the 6th Generation Mobile Communications Systems (6G). Among the many key performance indicators of 6G, ultra-high data transmission rate occupies an important position, aiming to fully meet the needs of users for massive data transmission in diverse scenarios. Ultra-dense connection is also a widely recognized key performance indicator for 6G. It is expected that the 6G wireless communication system can achieve a device connection density of 10 8 per square kilometer, providing strong support for the in-depth development of the Internet of Everything. By significantly increasing the number of antennas on the base station side, the massive multiple-input multiple-output (MIMO) technology can support hundreds of users simultaneously. This technology allows the base station to serve multiple users on the same time-frequency resource block, effectively improving the system capacity and showing great potential in meeting the requirements of ultra-high data transmission rate and ultra-dense connection in the 6G communication system. However, there is co-channel interference among the users scheduled to the same time-frequency resource block (RB), and the degree of interference among users is closely related to the system sum rate, which directly affects the overall data transmission performance of the system. By means of intelligent user scheduling and resource allocation technology, the interference among the users sharing the same resource block can be effectively reduced, thereby significantly improving the system capacity. Most of the existing multi-cell user scheduling methods are model-driven, rely on ideal assumptions, and have high computational complexity, making it difficult to scale to the typical scenarios of future 6G massive connections and extremely low latency. In the user scheduling problem, how to achieve high system performance with low computational complexity is a core challenge. Based on this, the present invention realizes fast and efficient multi-cell user scheduling through a data-driven strategy, combined with user location information, thereby reducing the dependence on real-time channel state information and reducing the system overhead. Summary of the Invention

[0003] Object of the Invention: The object of the present invention is to provide a multi-cell massive MIMO user scheduling method using location information to solve the problems of strong interference in multi-cells and high complexity of resource allocation in a dense user scenario in the multi-cell transmission scenario of massive MIMO, and to reduce the computational complexity and improve the system performance.

[0004] Technical solution: To achieve the above object, the present invention adopts the following technical solution:

[0005] A multi-cell massive MIMO user scheduling method using location information, comprising: modeling multi-cell user scheduling as a sequence-to-sequence learning task, where the input sequence is the location coordinates of the users to be scheduled, and the output sequence is the corresponding scheduling scheme; the MUGFormer (Multi-cell User Grouping Former) network based on the Transformer architecture to achieve an end-to-end mapping from user location information to the scheduling scheme; the MUGFormer network includes an embedding module, a user feature aggregation module, and a scheduling decision module; among them, there is an independent embedding module before the user feature aggregation module and the scheduling decision module respectively, which map the location coordinates of the users and the scheduling results to a high-dimensional space; the user feature aggregation module uses the multi-head self-attention mechanism to integrate all the information of the users to be scheduled and generate a new weighted feature representation for each user; the scheduling decision module combines the user features output by the user feature aggregation module and the partially generated user scheduling results to determine the scheduling scheme for the next user.

[0006] Further, the training of the MUGFormer network adopts a supervised learning method, and each sample contains a group of users with different locations and their corresponding optimal multi-cell user scheduling schemes; during the training process, the user location information is used as the input, and the corresponding optimal scheduling scheme is used as the label, and the network parameters are iteratively updated by minimizing the cross-entropy loss between the output and the label; reaching the preset period or the environment changes will trigger the model update mechanism.

[0007] Further, the embedding module before the user feature aggregation module maps the location coordinates of each user to a d-dimensional sequence for capturing user features; the input of the embedding module before the scheduling decision module is the multi-cell user scheduling scheme or the scheduling start marker x0, and the output is a d-dimensional embedding sequence. For the k-th user, when k = 1, the embedding sequence is determined by the start marker x0; when k > 1, the embedding sequence is generated based on the optimal scheduling scheme of the (k - 1)-th user during the training phase and based on the predicted scheduling result of the (k - 1)-th user during the execution phase.

[0008] Further, the user feature aggregation module of the MUGFormer network is composed of multiple feature fusion units connected in series. Except for the first feature fusion unit taking the output of the embedding module as the input, the rest of the feature fusion units take the output of the previous unit as the input.

[0009] Furthermore, in the feature fusion unit, data sequentially passes through a multi-head self-attention layer, a residual connection and normalization layer, a feed-forward network, and another residual connection and normalization layer. Among them, the multi-head self-attention layer generates new weighted features for each user by fusing the features of other users to be scheduled through the self-attention mechanism. The residual connection and normalization layer adds the input and output of the previous layer and performs normalization processing. The feed-forward network is composed of a multi-layer perceptron to enhance the non-linear mapping ability of the model.

[0010] Furthermore, the scheduling decision module of the MUGFormer network consists of multiple serially connected decision units and a linear layer. Among them, except for the first decision unit that takes the output of the user feature aggregation module and the output of the embedding module as inputs, the remaining decision units take the output of the user feature aggregation module and the output of the previous decision unit as inputs. The output of the last decision unit is mapped through the linear layer to generate the probability distribution of different scheduling results for users.

[0011] Furthermore, in the decision unit, data sequentially passes through a masked self-attention layer, a residual connection and normalization layer, an encoder-decoder attention layer, a residual connection and normalization layer, a feed-forward network, and a residual connection and normalization layer. Among them, the masked self-attention layer fuses the information of the partial user scheduling results that have been obtained through the masked self-attention mechanism. The encoder-decoder attention layer obtains the key vector and value vector from the output of the user feature aggregation module, obtains the query vector from the output of the previous layer, and then obtains a new feature sequence through the self-attention mechanism.

[0012] Furthermore, during online scheduling, each base station transmits the location coordinates of the users to be scheduled to the central scheduler. The central scheduler inputs the location coordinates of all users into the MUGFormer network, obtains the multi-cell user scheduling scheme, determines the base station and time-frequency resource block corresponding to each user, and then transmits the results to each base station.

[0013] Furthermore, during the training phase of the MUGFormer network, each base station obtains the statistical channel information of users at different locations through channel estimation methods, records their location coordinates, and then transmits this information to the central scheduler. The central scheduler generates a dataset based on this data. Among them, each sample contains a group of users at different locations and the corresponding optimal scheduling scheme for this group of users. The optimal multi-cell user scheduling scheme is obtained through an exhaustive search method or an iterative clustering method.

[0014] The present invention also provides a computer program product, including computer programs / instructions, and when the computer programs / instructions are executed by a processor, the steps of the multi-cell large-scale MIMO user scheduling method using location information are implemented.

[0015] Beneficial effects: The multi-cell user scheduling method using location information proposed by the present invention is based on the Transformer architecture and adopts a data-driven strategy, realizing an end-to-end mapping from user location to scheduling scheme, getting rid of the dependence on specific random channel models in traditional methods, while reducing the demand for real-time channel state information and significantly reducing system overhead. The simulation results show that in small and medium-sized user scenarios, the multi-cell user scheduling scheme obtained by the present invention is exactly the same as the optimal scheduling scheme; in large-scale user scenarios, the system sum rate of the obtained results approaches the optimal scheduling result, and the running time of the algorithm is less than 1% of that of traditional methods. Brief Description of the Drawings

[0016] Figure 1 It is a diagram of the multi-cell user scheduling method using location information according to an embodiment of the present invention.

[0017] Figure 2 It is a diagram of the MUGFormer network architecture in an embodiment of the present invention.

[0018] Figure 3 It is a diagram of the algorithm sum rate performance of the simulation experiment in an embodiment of the present invention. Detailed Embodiments

[0019] The technical solution of the present invention will be described in detail below with reference to the drawings and specific embodiments.

[0020] As Figure 1 shown, a multi-cell large-scale MIMO user scheduling method using location information disclosed in an embodiment of the present invention models multi-cell user scheduling as a sequence-to-sequence learning task, where the input sequence is the location coordinates of the users to be scheduled, and the output sequence is the corresponding scheduling scheme; and a data-driven strategy is adopted to construct a MUGFormer network based on the Transformer architecture, combined with a supervised learning framework, to realize an end-to-end mapping from user location information to scheduling scheme. The architecture design of the MUGFormer network is based on the Transformer framework, including an embedding module, a user feature aggregation module, and a scheduling decision module. Among them, there is an independent embedding module before each of the user feature aggregation module and the scheduling decision module, which respectively map the location coordinates of the users and the scheduling results to a high-dimensional space. The user feature aggregation module uses the multi-head self-attention mechanism to integrate all the information of the users to be scheduled and generate a new weighted feature representation for each user. The scheduling decision module determines the scheduling scheme for the next user in combination with the user features output by the user feature aggregation module and the partially generated user scheduling results.

[0021] The training of the MUGFormer network adopts a supervised learning method. The central scheduler generates sample data, and each sample contains a group of users at different locations and their corresponding optimal multi-cell user scheduling schemes. During the training process, the user location information is used as the input, and the corresponding optimal scheduling scheme is used as the label. The network parameters are iteratively updated by minimizing the cross-entropy loss between the output and the label. Reaching the preset period or a change in the environment will trigger the model update mechanism.

[0022] Next, in combination with a specific system model, the detailed implementation process of the method according to the embodiments of the present invention will be exemplarily described.

[0023] I. System Model

[0024] Consider a multi-cell large-scale MIMO system composed of N base stations and K users. Each base station is equipped with a uniform linear array of M antennas. The antenna array spacing is d = 0.5λ, where λ is the carrier wavelength. Each user is equipped with a single antenna. After OFDM modulation, the system has L time-frequency resource blocks to be allocated. It is assumed that the channel coefficients in each time-frequency resource block remain unchanged.

[0025] On the time-frequency resource block l, the channel between base station n and user k can be expressed as

[0026]

[0027] where Q n,k is the number of paths between base station n and user k, a n,k,q is the channel gain of the q-th path, f c is the carrier frequency, Δf is the frequency interval between adjacent time-frequency resource blocks, is the antenna array steering vector with respect to the direction cosine Ω, and are the delay and direction cosine of the q-th path, respectively. and are the delay set and direction cosine set of the multipaths between base station n and user k, respectively.

[0028] Define the matrix V as where are direction cosines uniformly sampled in the range of -1, 1, and is the number of sampled direction cosines, where Then, V is a semi-unitary matrix, i.e., VV H = I. g n,k,l can be approximated as

[0029]

[0030] where the x-th element of

[0031]

[0032] Set And In this embodiment, V is referred to as the beam matrix, and is referred to as the beam-domain channel.

[0033] For the beam-domain channel, the statistical channel covariance matrix Its (x, y)-th element is calculated as

[0034]

[0035] As can be seen from Equation (3), different elements of the beam-domain channel represent the channel gains of different paths. When using a large antenna, the base station can separate signals from different angles. Based on the uncorrelated scattering assumption, each element of is an independent complex random variable with zero mean and different variances, then Equation (a) holds, and Γ n,k is a diagonal matrix. In addition, the x-th diagonal element of Γ n,k [Γ n,k x represents the channel energy of user k in the x-th sampling direction of base station n.

[0036] As the number of links and user density in wireless communication networks continue to increase, the overhead of traditional pilot-based channel estimation methods for obtaining channel state information (CSI) becomes huge and may no longer be applicable in future ultra-high rate and ultra-low latency applications. At the same time, the rich location-specific channel data and powerful data mining capabilities in wireless networks provide the possibility for the development of environment-aware communication, thus promoting more efficient CSI acquisition. For example, a channel knowledge map is a database for a specific area, and the corresponding channel information can be indexed through the positions of the receiver and transmitter. It is based on historical channel data, establishes a mapping relationship between location and channel information, and enables rapid estimation of channel information at any location within the area. The statistical channel covariance can be estimated from the current environment and the positions of the base station and users

[0037]

[0038] where f Θ (·) is a function with parameter Θ, and the parameter Θ is related to the environment, is the position of base station n, is the position of user k.

[0039] Consider the uplink transmission of a multi-cell massive MIMO system. Define the vector to represent the multi-cell user scheduling scheme. The k-th element x in the vector k ​∈[1,2,…,NL] represents the scheduling scheme of user k, that is, user k uses time-frequency resource block l k = x k %L and base station n k = x k / / L communication, where % represents the remainder calculation and / / represents the integer division calculation. Considering the uplink transmission of a multi-cell massive MIMO system and assuming that each user transmits signals with equal power p. Under linear detection, the received signal of base station n k regarding user k can be expressed

[0040]

[0041] where is the received vector of base station n k on time-frequency resource block l k regarding user k, and satisfies is additive Gaussian noise, is the data signal of user k. Then, the signal-to-interference-plus-noise ratio of user k can be calculated as

[0042]

[0043] where Since V is a semi-unitary matrix, substituting equation (2) into the expression of the signal-to-interference-plus-noise ratio gives equation (b). The uplink ergodic achievable rate of user k is given by

[0044]

[0045] To reduce the computational complexity of the Monte Carlo averaging in the ergodic rate, this embodiment gives an approximate solution. Specifically, by taking the average of the numerator and denominator of the SINR respectively and combining the generalized Rayleigh entropy property, we obtain the approximate rate

[0046]

[0047] where Serial number Then, the sum rate of the system can be expressed as

[0048]

[0049] This equation shows that the sum rate of the system only depends on the statistical channel information and the multi-cell user scheduling scheme.

[0050] II. Problem statement

[0051] The multi-cell user scheduling problem aims to achieve the pairing of users and base stations and RB allocation at the same time to maximize the system and rate. The optimal multi-cell user scheduling solution is

[0052]

[0053] where x opt The kth element of represents the optimal scheduling scheme for user k, and the constraint condition ensures that all users are scheduled. According to equation (11), the calculation of the system sum rate depends entirely on the user grouping scheme and the statistical channel covariance. In addition, from equation (5), it can be seen that the statistical channel covariance is directly related to the environment and the location of the base station and the user. Assuming that the environment changes relatively slowly, that is, the function f Θ The mapping relationship between the position represented by (·) and the statistical channel covariance is stable and effective in the long term and will not change. Based on this assumption, it is possible to use position information to schedule users in multiple cells.

[0054] The multi-cell user scheduling method using location information is different from the traditional multi-cell user grouping method using statistical channel state information (CSI). It does not need to rely on statistical CSI for complex iterative solutions, but uses a neural network to find a mapping relationship. When the user location is input, a multi-cell user scheduling plan is output, and the plan is highly consistent with the optimal user scheduling plan obtained using statistical CSI. Specifically, each specific area containing N base stations is equipped with a dedicated neural network. The neural network inputs the location information of all users within the coverage area. Output predicted user grouping scheme The kth element represents the predicted scheduling result of user k. In this framework, our goal is to find an optimal mapping function that minimizes the predicted user scheduling solution and the optimal user grouping solution x opt The above optimization problem can be expressed as

[0055]

[0056] Among them, G Φ (A) means mapping A to function (neural network), Φ represents the neural network parameters, is the quantized x opt and The loss function of the difference between them is used. In this embodiment, cross entropy loss is used.

[0057] 3. Algorithm Design

[0058] Regarding the multi-cell user scheduling problem and the characteristics of the Massive MIMO system, in this embodiment, a MUGFormer network is designed based on the Transformer architecture to solve the multi-cell user scheduling problem using location information. The framework of the MUGFormer network is as shown in Figure 2 . The MUGFormer consists of an embedding module, a user feature aggregation module, and a scheduling decision module. The embedding module maps the two-dimensional location information and the one-dimensional scheduling result to a high-dimensional space respectively, so as to more accurately present the similarities and differences of user features at different positions and the features of different user sets. The user feature aggregation module is responsible for integrating all user information and providing comprehensive decision support for multi-cell user scheduling. The scheduling decision module then cleverly uses the user information output by the user feature aggregation module and the generated scheduling result to predict the scheduling scheme of the next user. Next, each module will be introduced in detail.

[0059] 1. Embedding module

[0060] Since the input and output tags of the multi-cell user problem are different, the parameters of the embedding modules before the user feature aggregation module and the scheduling decision module are independent, and are denoted as embedding module 1 and embedding module 2 respectively. Embedding module 1 maps the position coordinates of each user to a high-dimensional sequence where d represents the dimension of the vector after embedding. If the Euclidean distance ||z k -z k' ||2 between the embedding vectors of two users k and k' is small, it indicates that the features between these two users have a high degree of similarity. The input of embedding module 2 is the one-dimensional user set number or the scheduling start tag x0, where x0 is usually set to 0. This embedding layer maps the input single-user scheduling result to a d-dimensional space to obtain a sequence If the Euclidean distance between the sequences corresponding to the scheduling results of two users is smaller, the similarity of the corresponding user set features is higher.

[0061] 2. User feature aggregation module

[0062] The user feature aggregation module consists of S e feature fusion units and is connected in series. Among them, the first feature fusion unit takes the output of embedding module 1 as the input, and the subsequent feature fusion units take the output of the previous feature fusion unit as the input. Finally, the output of the S e th feature fusion unit is used as the output of user feature aggregation. For the convenience of writing, in this embodiment, S e =1 is taken to introduce the specific execution process. The feature fusion unit consists of a multi-head self-attention layer, a residual connection and normalization, and a feed-forward network.

[0063] The multi-head self-attention layer does not rely on external information, but realizes feature extraction through information interaction and integration within the input sequence {z k}. Specifically, for each user, the self-attention mechanism calculates the correlation between it and all users to be scheduled, and generates a weighted feature representation based on these correlations

[0064]

[0065] where the k-th column of is the final output of user k through the multi-head attention layer is the weight matrix to be learned, S is the number of attention heads, the Concat(·) function represents concatenating the matrices by column, and the matrix is the output of the input at the s-th attention head, and the specific calculation is

[0066]

[0067] where is the query matrix of Z at the s-th attention head, is the key matrix of Z at the s-th attention head, is the value matrix of Z at the s-th attention head, is the weight matrix, and the softmax(·) function represents performing the softmax operation on each column of the matrix.

[0068] Both the multi-head self-attention layer and the feed-forward network adopt a residual connection and layer normalization module, which directly passes the input of the previous layer to the subsequent layer, thus effectively alleviating the problems of gradient disappearance and gradient explosion in the training of deep networks. Specifically, the output of the multi-head self-attention layer is added to the input to form a residual connection, that is

[0069]

[0070] where the k-th column of Z' is z' k . Subsequently, the features of each user are normalized to obtain

[0071]

[0072] where LN· represents the layer normalization operation, is the k-th column of, and its i-th element is

[0073]

[0074] where

[0075]

[0076] represent the mean and variance respectively.

[0077] The feed - forward network introduces a non - linear activation function to enhance the non - linear mapping ability of the model, enabling it to learn more complex feature representations. Specifically, the feed - forward module consists of a two - layer multi - layer perceptron. The first layer uses the ReLU non - linear activation function, while the second layer does not contain a non - linear activation function. For the input The output of the feed - forward network is

[0078]

[0079] where and are the weight matrix and bias matrix of the first layer respectively, while and are the weight matrix and bias matrix of the second layer respectively. The dimension of the output of the first layer is d f , the activation function ReLU(x)=max(0,x), and the dimension of the output of the second layer is d. Then, a residual connection and layer normalization are performed again to obtain the output of the feature fusion unit

[0080]

[0081] So far, a feature fusion unit has completed the feature extraction and transformation of the input sequence. When the number of feature fusion units S e = 1 in the user feature aggregation module, the output of the user feature aggregation module is Z final .

[0082] 3. Scheduling Decision Module

[0083] The scheduling decision module consists of S d serially connected decision units and a linear layer. The first decision unit takes the output Z final of the user feature aggregation module and the output of the embedding module 2 as inputs, while the subsequent decision units take Z final and the output of the previous decision unit as inputs. For the sake of convenient writing, in this embodiment, S d = 1 is taken for detailed introduction.

[0084] It should be noted that the embedding vector obtained by user k through the embedding module 2 is obtained from the scheduling information of the (k - 1) - th user. Specifically, when k = 1, the embedding vector is obtained from the scheduling start token x0; when k > 1, the embedding vector is obtained from the optimal user scheduling result during the training phase, and from the prediction result during the inference phase.

[0085] The decision-making unit consists of a masked self-attention layer, a residual connection and normalization module, an encoder-decoder attention layer, and a feed-forward network. Among them, the masked self-attention introduces a masking mechanism to ensure that the generation of the current output is not affected by future information. The designs of the residual connection and normalization as well as the feed-forward network are consistent with those in the user feature aggregation module, aiming to enhance the stability and non-linear expression ability of the model. The encoder-decoder attention layer can make full use of the global features of all users extracted by the user feature aggregation module, thereby generating more accurate user scheduling predictions. The following will focus on the designs of the masked self-attention layer and the encoder-decoder attention layer in the decision-making unit.

[0086] The masked self-attention layer is basically the same as the multi-head self-attention layer in structure. The difference is that a masking matrix is added when calculating the attention score to ensure that the generation of the current output is not affected by future information. Specifically, the output of the masked self-attention layer is

[0087]

[0088] Among them, The k-th column of is the final output of user k after passing through the masked self-attention layer is the weight matrix to be learned, and the matrix is the output of the input at the s-th attention head, and the specific calculation is

[0089]

[0090] Among them, and are the query matrix, key matrix, and value matrix of C at the s-th attention head respectively, is the weight matrix, and the masking matrix is an upper triangular matrix, and the elements below the diagonal are negative infinity; other elements are 0.

[0091] The encoder-decoder self-attention layer allows the decision-making unit to focus on the output of the user feature aggregation module, thereby integrating all user global information. Specifically, the query matrix comes from the output of the previous layer of this layer while the key matrix and value matrix come from the output Z of the user feature aggregation module final . The output of the encoder-decoder self-attention layer is

[0092]

[0093] Among them, is the weight matrix to be learned, and They are the query matrix, key matrix, and value matrix of the s-th attention head respectively.

[0094] Subsequently, After another layer of residual connection and normalization, we get Then we input it into the feed-forward network to obtain Finally, after another residual connection and normalization process, we get the final output of the decision-making unit Finally, through a linear layer, we obtain the original output of the MUGFormer network

[0095] P = W P C final + B P #(24)

[0096] Where and are the weight and bias to be learned respectively. The k-th column of P is the original output p of user k through the MUGFormer network k . Further, by applying the softmax function, P is transformed into a probability matrix Where the k-th column of represents the scheduling probability of user k. The multi-cell user scheduling result predicted by MUGFormer is Where

[0097] IV. Implementation Effects

[0098] To enable those skilled in the art to better understand the solution of the present invention, the following presents the performance display and computational complexity comparison of the multi-cell user scheduling method using location information in this embodiment under specific system configurations. The simulation environment is generated by the QuaDRiGa platform. Among them, the base station is equipped with M = 128 antennas, and each user is equipped with a single antenna. The number of training and test samples of the MUGFormer network are 20,000 and 8,000 respectively. Each sample contains a group of users with different positions and their corresponding optimal user scheduling schemes, where the optimal user scheduling scheme is obtained by using a graph clustering-based multi-cell user scheduling algorithm (Iterative Clustering Algorithm, ICA). In addition, to better compare the performance of the MUGFormer network, the DGCA (Density-cut based Graph Clustering Algorithm) algorithm and the SC-MS (Spectral Clustering-aided Multi-band user Scheduling) algorithm are introduced as baseline algorithms.

[0099] Figure 3 Shows the sum-rate performance of different user scheduling algorithms in the test set under the system configurations of (N = 3, K = 100, L = 6), (K = 3, K = 300, L = 18), and (N = 3, K = 500, L = 18). Among them, the ICA curve is the target result, and the closer the MUGFormer curve is to the ICA curve, the better. When the number of users K = 100, the MUGFormer curve completely coincides with the ICA curve; when the number of users K = 300, the MUGFormer curve approaches the ICA curve; when the number of users K = 500, the gap between the MUGFormer curve and the ICA curve increases, but the sum-rate of MUGFormer is more than 98% of that of ICA. When the number of users is small, the multi-cell user scheduling based on MUGFormer is exactly the same as the target scheme, and the obtained sum-rate can reach 100% of the target result. As the number of users increases, the system sum-rate obtained by MUGFormer decreases compared with the target result but still reaches more than 98% of the target value.

[0100] Table 1 shows the running time of each algorithm under different system configurations. The running time of each algorithm increases with the increase in the number of users. Specifically, under all system configurations, the multi-cell user scheduling algorithm based on MUGFormer shows the shortest running time, its performance is close to that of the ICA algorithm, but the required time is less than 1% of that of the ICA algorithm.

[0101] Table 1 Running Time Table

[0102]

[0103] *(N, L, K) represent the number of base stations, the number of time-frequency resource blocks, and the number of users, respectively.

[0104] The embodiments of the present invention also disclose a computer program product, including computer programs / instructions, and when the computer programs / instructions are executed by a processor, the steps of the multi-cell massive MIMO user scheduling method using location information are implemented. The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to the processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing devices, so that when the program codes are executed by the processor or controller, the steps of the method of the present invention are implemented. The program codes can be executed entirely on the machine, partially on the machine, executed partially on the machine as an independent software package and partially on a remote machine, or executed entirely on a remote machine or server. Where the present invention is not described in detail, it is all well-known techniques in the art.

Claims

1. A multi-cell large-scale MIMO user scheduling method using location information, characterized in that Including: Modeling multi-cell user scheduling as a sequence-to-sequence learning task, where the input sequence is the location coordinates of the users to be scheduled, and the output sequence is the corresponding scheduling scheme; the MUGFormer network based on the Transformer architecture to achieve an end-to-end mapping from user location information to the scheduling scheme; The MUGFormer network includes an embedding module, a user feature aggregation module, and a scheduling decision module; among them, there is an independent embedding module before the user feature aggregation module and the scheduling decision module, respectively mapping the location coordinates of the users and the scheduling results to a high-dimensional space; the user feature aggregation module uses the multi-head self-attention mechanism to integrate all the information of the users to be scheduled and generate a new weighted feature representation for each user; the scheduling decision module combines the user features output by the user feature aggregation module and the partially generated user scheduling results to determine the scheduling scheme for the next user.

2. A multi-cell massive MIMO user scheduling method using location information according to claim 1, characterized in that The training of the MUGFormer network adopts a supervised learning method, and each sample contains a set of users at different locations and their corresponding optimal multi-cell user scheduling schemes; during the training process, the user location information is used as the input, and the corresponding optimal scheduling scheme is used as the label, and the network parameters are iteratively updated by minimizing the cross-entropy loss between the output and the label; reaching the preset period or the environment changes will trigger the model update mechanism.

3. A multi-cell massive MIMO user scheduling method using location information according to claim 1, characterized in that The embedding module before the user feature aggregation module maps the location coordinates of each user to a d-dimensional sequence for capturing user features; the input of the embedding module before the scheduling decision module is the multi-cell user scheduling scheme or the scheduling start token x0, and the output is a d-dimensional embedding sequence. For the k-th user, when k = 1, the embedding sequence is determined by the start token x0; when k > 1, the embedding sequence is generated based on the optimal scheduling scheme of the (k - 1)-th user during the training phase and based on the predicted scheduling result of the (k - 1)-th user during the execution phase.

4. A multi-cell large-scale MIMO user scheduling method using location information according to claim 1, characterized in that The user feature aggregation module of the MUGFormer network is composed of multiple feature fusion units connected in series. Except for the first feature fusion unit taking the output of the embedding module as the input, the remaining feature fusion units all take the output of the previous unit as the input.

5. A multi-cell massive MIMO user scheduling method using location information according to claim 4, characterized in that In the feature fusion unit, the data sequentially passes through the multi-head self-attention layer, the residual connection and normalization layer, the feed-forward network, and the residual connection and normalization layer again; among them, the multi-head self-attention layer generates a new weighted feature for each user that integrates the features of other users to be scheduled by means of the self-attention mechanism; the residual connection and normalization layer adds the input and output of the previous layer and performs normalization processing; the feed-forward network is composed of a multi-layer perceptron to enhance the non-linear mapping ability of the model.

6. A multi-cell massive MIMO user scheduling method using location information according to claim 1, characterized in that, The scheduling decision module of the MUGFormer network consists of multiple serially connected decision units and a linear layer; among them, except for the first decision unit taking the output of the user feature aggregation module and the output of the embedding module as the input, the remaining decision units take the output of the user feature aggregation module and the output of the previous decision unit as the input; the output of the last decision unit is mapped through the linear layer to generate the probability distribution of different scheduling results for the users.

7. A multi-cell large-scale MIMO user scheduling method using location information according to claim 6, characterized in that, In the decision-making unit, data sequentially passes through a masked self-attention layer, a residual connection and normalization layer, an encoder-decoder attention layer, a residual connection and normalization layer, a feed-forward network, and a residual connection and normalization layer; among them, the masked self-attention layer uses the masked self-attention mechanism to fuse the partially obtained user scheduling result information; the encoder-decoder attention layer obtains the key vector and value vector from the output of the user feature aggregation module, then obtains the query vector from the output of the previous layer, and then obtains a new feature sequence through the self-attention mechanism.

8. A multi-cell large-scale MIMO user scheduling method using location information according to claim 1, characterized in that, During online scheduling, each base station transmits the position coordinates of the users to be scheduled to the central scheduler; the central scheduler inputs the position coordinates of all users into the MUGFormer network, obtains the multi-cell user scheduling scheme, determines the base station and time-frequency resource block corresponding to each user, and then transmits the result to each base station.

9. A multi-cell large-scale MIMO user scheduling method using location information according to claim 1, characterized in that, In the training stage of the MUGFormer network, each base station obtains the statistical channel information of users at different positions in a limited manner through the channel estimation method and records their position coordinates, and then transmits this information to the central scheduler; the central scheduler generates a data set based on these data, where each sample contains a group of users at different positions and the corresponding optimal scheduling scheme for this group of users; the optimal multi-cell user scheduling scheme is obtained through an exhaustive search method or an iterative clustering method.

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 a multi-cell massive MIMO user scheduling method using location information according to any one of claims 1-9 are implemented.

Images