A compression matrix design method for uplink cell-free access system

Abstract

The application provides a compression matrix design method for an uplink non-cellular access system, and belongs to the technical field of wireless communication, comprising the following steps: pre-generating a plurality of non-cellular MIMO system scenes; obtaining a plurality of sub-tasks by task sampling from the pre-generated non-cellular MIMO system scenes, constructing a joint task from the plurality of sub-tasks, and modeling the non-cellular MIMO system scenes corresponding to the plurality of sub-tasks into a heterogeneous graph model; performing joint training on the joint task based on meta-learning parameters, generating compression matrices corresponding to the plurality of sub-tasks, and obtaining optimal parameters corresponding to the joint task, wherein the joint training process adopts a message passing algorithm; calculating a difference vector between the optimal parameters and the meta-learning parameters to complete the update of the meta-learning parameters, and performing the next joint training to finally obtain optimal meta-learning parameters; and deploying the optimal meta-learning parameters to a non-cellular MIMO system complying with a standard block fading model and performing data transmission. The application is suitable for dynamic communication scenes.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to a compression matrix design method for uplink non-cellular access systems. Background Technology

[0002] In recent years, the rapid development of mobile internet and the Internet of Things (IoT) has placed higher demands on the data transmission rate, network coverage, and system capacity of wireless communication systems. Although traditional cellular network architecture has evolved over many years, it still faces significant challenges in handling large-scale device access and high-traffic service demands. To overcome the inherent limitations of traditional architecture, cellular-free massive MIMO (Multiple-Input Multiple-Output) technology has gained widespread attention as an innovative solution. This technology provides more uniform network coverage and higher spectral efficiency by distributing a large number of access points and connecting them to a central processing unit via fronthaul links for coordinated signal processing. This architectural innovation effectively solves the coverage blind spots and capacity bottlenecks in traditional cellular networks, providing a new technical path to achieve the performance goals of next-generation communication systems. However, in cellular-free MIMO systems, the fronthaul link between access points and the central processing unit is typically limited by capacity. When faced with the concurrent access demands of massive numbers of devices, this limitation prevents the system from effectively and promptly processing a large number of parallel uplink signals, resulting in reduced system spectral efficiency. Therefore, in order to address the capacity limitation of the fronthaul link between the access point and the central processing unit, each access point needs to compress the signal before transmitting the uplink signal to the central processing unit to reduce the fronthaul overhead.

[0003] To ensure information integrity during signal compression, existing research typically employs signal dimensionality reduction methods based on compression matrices. However, in non-cellular MIMO systems, the distributed architecture of multiple access points working collaboratively makes the joint optimization of compression matrices particularly complex. The widespread application of deep learning has provided an effective solution to this problem. Existing compression matrix design methods based on Deep Neural Networks (DNNs) can perceive the effective information contained in the uplink signal at each access point and design compression matrices for each access point with the goal of improving system capacity. This compression matrix can store the effective information in the uplink signal in a low-dimensional form, thereby reducing fronthaul communication overhead. However, existing deep neural network-based compression matrix design methods typically rely on fixed model parameters for training. In actual deployment, when system parameters change, such as a change in the number of users, the original model parameters become inapplicable. This scalability limitation makes such deep neural network-based compression matrix design methods difficult to apply to dynamic communication systems. Summary of the Invention

[0004] The purpose of this invention is to propose a compression matrix design method for uplink non-cellular access systems, which is applicable to dynamic communication scenarios.

[0005] This invention is achieved through the following technical solution: A compression matrix design method for uplink non-cellular access systems includes the following steps: Step S1: Pre-generate multiple non-cellular MIMO system scenarios. Each non-cellular MIMO system scenario contains the same number of access points, and each non-cellular MIMO system scenario contains a different number of users, different user location distributions, and different rate requirements. Step S2: Sample tasks from the pre-generated non-cellular MIMO system scenario to obtain several sub-tasks, and construct a joint task from these sub-tasks. Each sub-task contains different numbers of users, user distribution and rate requirements. Based on graph neural networks, model the non-cellular MIMO system scenario corresponding to each sub-task as a heterogeneous graph model. Step S3: Perform joint training on the joint task based on the learning parameters, generate the compression matrix corresponding to each subtask, and obtain the optimal parameters corresponding to the joint task. The joint training process adopts the message passing algorithm, and the meta-learning parameters are randomly generated during the first joint training. Step S4: Calculate the difference vector between the optimal parameters and the meta-learning parameters, and update the meta-learning parameters with a fixed step size and the difference vector. Then proceed to step S1 to perform the next joint training to continuously optimize the meta-learning parameters. Step S5: Deploy the optimized meta-learning parameters on a cellular-free MIMO system that follows the standard block fading model and transmit the data.

[0006] Furthermore, in step S2, the nodes in the heterogeneous graph model are users and access points, the edges are the connection relationships between access points and users, the user's characteristic is the rate requirement, the access point's characteristic is the compression matrix to be optimized, and the edge's characteristic is the uplink channel matrix between the connected access point and the user.

[0007] Furthermore, the message passing algorithm in step S3 includes feature definition and message passing, specifically as follows: Feature definition: For each subtask in the joint task, initialize the compression matrix of each corresponding access point using random numbers to obtain the initialized compression matrix. And based on the channel parameters between the user and the access point. and user speed requirements Complete the feature definition of nodes and edges in the corresponding heterogeneous graph model, where, l Indicates the access point number. Indicates the first Sub-tasks k Indicates the first k One user; Message passing includes message aggregation and feature updates, specifically: Message aggregation: The first in heterogeneous graph models l Each access point obtains information from all its neighboring nodes; this process is represented as follows: ,in, Indicates the first The corresponding subtask in the first l The access point is at the _ t Information obtained during the optimization round For pooling functions, For the first l The set of all neighboring nodes of an access point For parameters The parameterized function of the driver; Feature update: 1st l Each access point updates its features for the next optimization round based on the acquired messages and the features at the current optimization round. This process is represented as follows: ,in, For the first The corresponding subtask in the first l The access point t Features during round optimization For parameters The parameterized function of the driver.

[0008] Furthermore, in step S3, in each round of optimization, each access point sequentially completes message passing and performs... T Round optimization to obtain the first The final compression matrix under each subtask Based on the known channel state information between each access point and the user and the final compression matrix Complete the coordinated processing of the uplink signal. After coordinated processing, the first... k The signal-to-interference-plus-noise ratio (SIR) for each user during transmission is: .

[0009] Furthermore, in step S3, after obtaining the final compression matrix of all subtasks, according to the formula... Calculate the performance loss of the current graph neural network on the joint task. Repeat joint training until Convergence is achieved, yielding the optimal parameters for the joint task. Where Ω represents the number of subtasks. For the first Number of users for each sub-task For the first User speed requirements for each sub-task λ This is the penalty coefficient.

[0010] Furthermore, in step S4, according to the formula Complete the update of meta-learning parameters. The step size is fixed.

[0011] Furthermore, step S5 specifically includes the following steps: Step S51: At the beginning of each channel block, each access point estimates the uplink channel parameters of the user and transmits them to the central processing unit through the fronthaul link. The central processing unit completes the construction of the heterogeneous graph model of the non-cellular MIMO system scenario corresponding to the current channel block based on the information uploaded by all access points. Step S52: The central processing unit calls the offline optimized meta-learning parameters to initialize the graph neural network in the heterogeneous graph model, and based on the current heterogeneous graph model, completes the feature definition and message passing process of each access point in parallel to optimize the compression matrix corresponding to each access point. Step S53: The central processing unit transmits the optimized compression matrix back to the corresponding access points. During the remaining time of the channel block, each access point completes the reception of user uplink data and compresses the user uplink data based on the optimized compression matrix before transmitting it to the central processing unit. The central processing unit uses the minimum mean square error algorithm to complete the collaborative processing of each user's uplink data according to the optimized compression matrix and the system channel state information.

[0012] Furthermore, in each training session, several subtasks are sampled from different non-cellular MIMO system scenarios to construct a joint task.

[0013] The present invention has the following beneficial effects: 1. This invention pre-generates multiple non-cellular MIMO system scenarios. Before each joint training, several sub-tasks are sampled from the pre-generated non-cellular MIMO system scenarios, and a joint task is constructed from these sub-tasks. Then, the joint task is jointly trained based on the learning parameters to generate a compression matrix corresponding to each sub-task and obtain the optimal parameters corresponding to the joint task. Next, the difference vector between the optimal parameters and the meta-learning parameters is calculated, and the meta-learning parameters are updated with a fixed step size and the difference vector. The joint training is repeated multiple times to continuously optimize the meta-learning parameters. Finally, the optimized meta-learning parameters are deployed in a non-cellular MIMO system following a standard block number fall model for data transmission. The compression matrix designed in this way can adapt to dynamic changes in the system. Even if the number of users in the scenario changes, it can still quickly adapt to the new communication environment and meet the high-quality, high-speed service requirements of dynamic systems. In this process, the non-cellular MIMO system is modeled as a heterogeneous graph structure. Combined with the message passing algorithm, the model failure problem caused by dynamic changes in the number of users is effectively solved. The introduction of meta-learning technology, through the division of task scenarios and joint training, improves the model's adaptability and generalization ability in dynamic scenarios. Attached Figure Description

[0014] The present invention will now be described in further detail with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart of the present invention.

[0016] Figure 2 The schematic diagram of the heterogeneous graph model is used to model the scenario of the non-cellular MIMO system of the present invention.

[0017] Figure 3 This is a flowchart illustrating the online optimization and actual transmission stages of the present invention.

[0018] Figure 4 This is a performance comparison chart of the present invention and the comparison algorithm on different compression dimensions. Detailed Implementation

[0019] like Figure 1 As shown, the compression matrix design method for uplink non-cellular access systems includes the following steps: Step S1: Pre-generate multiple non-cellular MIMO system scenarios. Each non-cellular MIMO system scenario contains the same number of access points, and each non-cellular MIMO system scenario contains a different number of users, different user location distributions, and different rate requirements to simulate a real dynamic communication environment. In the same type of non-cellular MIMO system scenario, the number of users is the same, but the location distribution and rate requirements of the users are different. Specifically, in this embodiment, the number of access points in each non-cellular MIMO system scenario is... L Both have 4 antennas. M Both are 12, the equivalent number of antennas after compression. N Both are 6, number of users K ∈{8,12,16,20}, each user is randomly distributed within the service area.

[0020] Step S2: Sample tasks from the pre-generated non-cellular MIMO system scenario to obtain several sub-tasks, and construct a joint task from these sub-tasks. Each sub-task contains different numbers of users, user distribution and rate requirements. Based on graph neural networks, model the non-cellular MIMO system scenario corresponding to each sub-task as a heterogeneous graph model. like Figure 2 As shown, the construction of the heterogeneous graph model includes node construction and edge construction. Node construction involves creating two different types of nodes in the heterogeneous graph model: access points and users. The user's characteristics include its respective rate requirements, and the access point's characteristics are the compression matrix to be optimized. Edge construction involves constructing the connection relationships between access points and users as edges in the heterogeneous graph model. The edge's characteristics are the uplink channel matrix between the connected access point and user. Each subtask in the joint task will be modeled as a unique heterogeneous graph model based on its specific scenario distribution.

[0021] In the heterogeneous graph model, the features of users and edges serve as input information for the compression matrix optimization and remain constant throughout the optimization process. The access point features contain the parameters of the compression matrix to be optimized. To effectively process the system input information, a message-passing algorithm is employed in the compression matrix optimization process.

[0022] Step S3: Perform joint training on the joint task based on the meta-learning parameters, generate the compression matrix corresponding to each sub-task, and obtain the optimal parameters corresponding to the joint task. The joint training process adopts a message passing algorithm, and the first set of meta-learning parameters during joint training... Randomly generated; The message passing algorithm includes feature definition and message passing, specifically: Feature definition: For each subtask in the joint task, initialize the compression matrix of each corresponding access point using random numbers to obtain the initialized compression matrix. And based on the channel parameters between the user and the access point. and user speed requirements Complete the feature definition of nodes and edges in the corresponding heterogeneous graph model, where, l Indicates the access point number. Indicates the first Sub-tasks k Indicates the first k One user; Message passing includes sequential message aggregation and feature updates, specifically: Message aggregation: The first in heterogeneous graph models l Each access point obtains information from all its neighboring nodes; this process is represented as follows: ,in, Indicates the first The corresponding subtask in the first l The access point is at the _ t Information obtained during the optimization round For pooling functions, For the first l The set of all neighboring nodes of an access point For parameters The parameterized function of the driver; Feature update: 1st l Each access point updates its features for the next optimization round based on the acquired messages and the features at the current optimization round. This process is represented as follows: ,in, For the first The corresponding subtask in the first l The access point t Features during round optimization For parameters The parameterized function driven by the meta-learning parameters includes parameters. and parameters ; In each round of optimization, each access point sequentially completes message aggregation and feature updates, and then performs... T Round optimization (i.e., the process of duplicate message aggregation and feature update) T (time), to obtain the first The final compression matrix under each subtask The central processing unit, based on the least mean square error algorithm, uses the known channel state information between each access point and the user, along with the final compression matrix, to compress the data. Complete the coordinated processing of the uplink signal. After coordinated processing, the first... k The signal-to-interference-plus-noise ratio (SIR) for each user during transmission is: The process of acquiring channel state information is an existing technology. After obtaining the final compression matrix of all subtasks, according to the formula... Computing the performance loss of graph neural networks on joint tasks Repeat joint training until Convergence is achieved, yielding the optimal parameters of the graph neural network for this joint task. Where Ω represents the number of subtasks. For the first Number of users for each sub-task For the first User speed requirements for each sub-task λ This is the penalty coefficient.

[0023] To ensure broad adaptability during training, several subtasks are sampled from different non-cellular MIMO system scenarios to construct a joint task in each training session. Each subtask contains different numbers of users, location distributions, and rate requirements, enabling the model to learn diverse scenario characteristics.

[0024] Step S4: Calculate the difference vector between the optimal parameters and the meta-learning parameters, and update the meta-learning parameters with a fixed step size and the difference vector. Then proceed to step S1 to perform the next joint training to continuously optimize the meta-learning parameters. Specifically, according to the formula Complete the update of meta-learning parameters. The step size is fixed.

[0025] After multiple rounds of task sampling, joint training, and parameter optimization, the meta-learning parameter θ will contain prior information from different scenarios during training. When facing new scenarios, the model can quickly adapt to unknown system parameters, thereby reducing latency in the compression matrix optimization process and improving the final performance of the optimized compression matrix.

[0026] Step S5: Deploy the optimized meta-learning parameters on a cellular-free MIMO system that follows the standard block fading model and transmit the data; Specifically, the steps include the following: Step S51: At the beginning of each channel block, each access point estimates the uplink channel parameters of the user and transmits them to the central processing unit through the fronthaul link. The central processing unit completes the construction of the heterogeneous graph model of the non-cellular MIMO system scenario corresponding to the current channel block based on the information uploaded by all access points. In this embodiment, the non-cellular MIMO system following the standard block fading model includes four access points with 12 antennas each and one central processing unit. Each access point is deployed in a square grid pattern. The number of users per antenna changes dynamically between different channel blocks and the users are randomly distributed. To reduce the communication overhead of the fronthaul link, the antenna dimension of the access points needs to be reduced to 6 through a compression matrix.

[0027] Step S52: The central processing unit calls the offline optimized meta-learning parameters to initialize the graph neural network in the heterogeneous graph model, and based on the current heterogeneous graph model, completes the feature definition and message passing process of each access point in parallel to optimize the compression matrix corresponding to each access point. During the optimization process, the central processing unit will update the model parameters multiple times based on the current scenario. Since the meta-learning parameters fully extract the key features of the compression matrix design in the non-cellular MIMO system during training and contain prior information from multiple different scenarios, the model can still achieve excellent performance with a limited number of updates, even with extremely strict latency constraints in the actual online optimization process.

[0028] Step S53: The central processing unit transmits the optimized compression matrix back to the corresponding access points. During the remaining time of the current channel block, each access point completes the reception of user uplink data and compresses the user uplink data based on the optimized compression matrix before transmitting it to the central processing unit to reduce the communication overhead of the fronthaul link. The central processing unit uses the minimum mean square error algorithm to complete the collaborative processing of the user uplink data according to the optimized compression matrix and system channel state information.

[0029] A flowchart illustrating the online optimization and actual transmission stages is shown below. Figure 3 As shown.

[0030] Figure 4 The graph shows a performance comparison between the present invention and the comparison algorithm (antenna selection algorithm) at different compression dimensions. It can be seen that, compared with the antenna selection algorithm, the present invention exhibits good compression performance at lower compression dimensions, and can effectively preserve the original information in the compressed low-dimensional representation. At the same time, at higher compression dimensions, the present invention can achieve performance close to that of no compression, and can effectively reduce the communication overhead between the access point and the central processing unit while ensuring near-lossless performance.

[0031] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the present invention. All equivalent changes and modifications made in accordance with the scope of the patent application and the contents of the specification of the present invention should still fall within the scope of the patent of the present invention.

Claims

1. A compression matrix design method for uplink non-cellular access systems, characterized in that: Includes the following steps: Step S1: Pre-generate multiple non-cellular MIMO system scenarios. Each non-cellular MIMO system scenario contains the same number of access points, and each non-cellular MIMO system scenario contains a different number of users, different user location distributions, and different rate requirements. Step S2: Sample tasks from the pre-generated non-cellular MIMO system scenario to obtain several sub-tasks, and construct a joint task from these sub-tasks. Each sub-task contains different numbers of users, user distribution and rate requirements. Based on graph neural networks, model the non-cellular MIMO system scenario corresponding to each sub-task as a heterogeneous graph model. Step S3: Perform joint training on the joint task based on the learning parameters, generate the compression matrix corresponding to each subtask, and obtain the optimal parameters corresponding to the joint task. The joint training process adopts the message passing algorithm. Step S4: Calculate the difference vector between the optimal parameters and the meta-learning parameters, and update the meta-learning parameters with a fixed step size and the difference vector. Then proceed to step S1 to perform the next joint training to continuously optimize the meta-learning parameters. The meta-learning parameters are randomly generated during the first joint training. Step S5: Deploy the optimized meta-learning parameters on a cellular-free MIMO system that follows the standard block fading model and transmit the data; In step S2, the nodes in the heterogeneous graph model are users and access points, the edges are the connection relationships between access points and users, the user's characteristic is the rate requirement, the access point's characteristic is the compression matrix to be optimized, and the edge's characteristic is the uplink channel matrix between the connected access point and the user. The message passing algorithm in step S3 includes feature definition and message passing, specifically: Feature definition: For each subtask in the joint task, initialize the compression matrix of each corresponding access point using random numbers to obtain the initialized compression matrix. And based on the channel parameters between the user and the access point. and user speed requirements Complete the feature definition of nodes and edges in the corresponding heterogeneous graph model, where l represents the access point label. Indicates the first There are several subtasks, where k represents the k-th user; Message passing includes message aggregation and feature updates, specifically: Message aggregation: In a heterogeneous graph model, the l-th access point obtains information from all its neighboring nodes. This process is represented as follows: ,in, Indicates the first The information obtained by the l-th access point in each subtask during the t-th round of optimization. For pooling functions, Let L be the set of all neighboring nodes of the L-th access point. For parameters The parameterized function of the driver; Feature Update: The l-th access point updates its features for the next optimization round based on the acquired messages and the features under the current optimization round. This process is represented as follows: ,in, For the first The features of the l-th access point in each subtask during the t-th round of optimization. For parameters The parameterized function of the driver; In step S3, in each round of optimization, each access point sequentially completes message passing and performs T rounds of optimization to obtain the result. The final compression matrix under each subtask Based on the known channel state information between each access point and the user and the final compression matrix After completing the coordinated processing of the uplink signal, the signal-to-interference-plus-noise ratio (SIR) for the k-th user during transmission is: In step S3, after obtaining the final compression matrix of all subtasks, the formula is used... Computing the performance loss of graph neural networks on joint tasks Repeat joint training until Convergence is achieved, yielding the optimal parameters for the joint task. Where Ω represents the number of subtasks. For the first Number of users for each sub-task For the first The user rate requirement for each sub-task, where λ is the penalty coefficient.

2. The compression matrix design method for uplink non-cellular access systems according to claim 1, characterized in that: In step S4, according to the formula Complete the update of meta-learning parameters. The step size is fixed.

3. The compression matrix design method for uplink non-cellular access systems according to claim 2, characterized in that: Step S5 specifically includes the following steps: Step S51: At the beginning of each channel block, each access point estimates the uplink channel parameters of the user and transmits them to the central processing unit through the fronthaul link. The central processing unit completes the construction of the heterogeneous graph model of the non-cellular MIMO system scenario corresponding to the current channel block based on the information uploaded by all access points. Step S52: The central processing unit calls the offline optimized meta-learning parameters to initialize the graph neural network in the heterogeneous graph model, and based on the current heterogeneous graph model, completes the feature definition and message passing process of each access point in parallel to optimize the compression matrix corresponding to each access point. Step S53: The central processing unit transmits the optimized compression matrix back to the corresponding access points. During the remaining time of the channel block, each access point completes the reception of user uplink data and compresses the user uplink data based on the optimized compression matrix before transmitting it to the central processing unit. The central processing unit uses the minimum mean square error algorithm to complete the collaborative processing of each user's uplink data according to the optimized compression matrix and the system channel state information.

4. A compression matrix design method for uplink non-cellular access systems according to any one of claims 1 to 3, characterized in that: In each training session, several subtasks are sampled from different non-cellular MIMO system scenarios to construct a joint task.

Images