A pilot and power control method for low-altitude communication system

Abstract

The application belongs to the field of wireless communication, non-cell low-altitude communication and artificial intelligence, and particularly relates to a low-altitude communication system pilot and power control method. The application firstly constructs a local graph observation by taking accessible user equipment as nodes and pilot sharing relationship as edges for each unmanned aerial access point. Secondly, the relative service importance of the accessible user equipment is mined based on the local graph observation by using a topology-aware graph. Then, a shared topology-aware graph network decision backbone is trained by using multiple unmanned aerial asynchronous historical experiences at a central computing node, and the shared decision backbone is inherited at each edge unmanned aerial side and locally adapted through low-rank edge individualization fine-tuning. Finally, each unmanned aerial outputs a pilot authentication decision and a downlink power control decision based on the local observation, so as to realize distributed resource management. The application can reduce the online training parameter quantity and communication overhead of the edge unmanned aerial, and improve the network throughput of the non-cell low-altitude communication system.

Description

Technical Field

[0001] This invention belongs to the fields of wireless communication, low-altitude communication, non-cellular communication and deep reinforcement learning, and specifically relates to a pilot and power control method for a low-altitude communication system. Background Technology

[0002] With the development of sixth-generation mobile communication systems and low-altitude economic applications, low-altitude communication networks composed of UAV access points can provide fast and flexible wireless coverage for temporary hotspots, emergency rescue, remote areas, and dynamic business areas. Cellular-free low-altitude communication systems, through the collaborative service of user equipment via multiple UAV access points, can break through the limitations of traditional cellular boundaries and dynamically reconfigure service areas according to business needs, thus being considered an important development direction for low-altitude communication networks.

[0003] However, radio resource management in non-cellular low-altitude communication systems still faces significant challenges. On the one hand, the mobility of both UAV access points and user equipment (UAVs) leads to dynamic changes in the set of accessible UAVs, link channel conditions, and interference relationships over time. On the other hand, in time-division duplex systems, the number of available orthogonal pilots is limited, and multiple UAVs may reuse the same pilot, causing pilot pollution, reducing downlink channel estimation accuracy, and further affecting downlink precoding, pilot access, and power control performance. Furthermore, if traditional centralized optimization methods are used, each UAV access point needs to frequently exchange global instantaneous channel information and solve a hybrid discrete-continuous resource allocation problem in each time slot, easily resulting in high communication overhead and online computation latency. Therefore, how to achieve low-overhead pilot access and power control under dynamic network topology and limited pilot resources is a key issue in resource management of non-cellular low-altitude communication systems.

[0004] In recent years, to improve the efficiency of wireless resource management, artificial intelligence technology has been gradually introduced into the resource management problems of non-cellular low-altitude communication systems. Existing advanced technologies mainly include:

[0005] (1) Resource management technology based on deep reinforcement learning. This type of technology constructs a state space, action space, and reward function, and uses deep neural networks to learn the mapping relationship from the wireless network state to resource management decisions, thereby avoiding the complex iterative solution in each time slot of traditional optimization methods. This type of method has the advantages of fast inference speed and suitability for online decision-making. However, ordinary deep neural networks usually represent the network state as a fixed-dimensional vector, which is difficult to adapt to the dynamic changes in the number of users that can access the network within the coverage area of ​​the UAV access point, and it is also difficult to explicitly characterize the pilot sharing relationship and interference relationship between user equipment.

[0006] (2) Resource management technology based on topology-aware graphs. This type of technology uses nodes and edges to model the topology of wireless networks, and can capture the spatial topological relationships between UAV access points, user equipment, and wireless links, thereby improving the decision-making performance of tasks such as user association, pilot allocation, power control, or spectrum management. However, existing topology-aware graph resource management methods usually still rely on global network information or completely local training. Methods that rely on global network information need to frequently collect global instantaneous channel state and topology information, resulting in high communication overhead; methods that rely on completely local training are difficult to utilize long-term shared experience across UAV access points, and require updating more model parameters on the resource-constrained edge UAV side, resulting in a large edge training burden. Summary of the Invention

[0007] To address the aforementioned problems, this invention proposes a pilot and power control method for low-altitude communication systems. The innovation of this invention lies in the following: For each UAV access point, a local graph observation is constructed using its accessible user equipment as nodes and pilot sharing relationships as edges. A topology-aware graph network is used to mine the relative service importance of accessible user equipment, thereby mitigating the impact of pilot pollution on downlink transmission performance. Simultaneously, the long-term shared policy learning of the central computing node is combined with the short-term local personalized decision-making of the edge UAV access points. The central computing node trains the shared topology-aware graph network decision backbone using asynchronous historical experience from multiple UAV access points, while each edge UAV access point performs local adaptation through low-rank personalized fine-tuning. This achieves distributed pilot access and power control while reducing communication overhead and the number of edge training parameters.

[0008] The technical solution adopted in this invention is as follows:

[0009] A pilot and power control method for a low-altitude communication system, used in a time-division duplex downlink non-cellular low-altitude communication system composed of multiple UAV access points and multiple user equipment, includes the following steps:

[0010] S1. In each time slot, each user equipment randomly selects an uplink pilot sequence and sends a pilot signal. Each UAV access point obtains the local downlink channel estimate of the user equipment that can be accessed based on the received uplink pilot signal and channel reciprocity.

[0011] S2. Using the accessible user equipment at the UAV access point as nodes and the pilot sharing relationship between accessible user equipment that selects the same pilot as edges, construct a local graph observation for each UAV access point, and use the UAV service status, UAV physical information, user equipment service status, user equipment physical information, pilot sharing topology and edge interaction information as input to the topology-aware graph.

[0012] S3. Utilize the topology-aware map to perform message passing and feature aggregation on the local map observations to obtain the aggregated features of each accessible user equipment, so as to characterize the relative service importance of each user equipment under imperfect channel estimation and pilot pollution conditions.

[0013] S4. At the central computing node, asynchronous historical experience uploaded by multiple UAV access points is used to train a shared topology-aware graph network decision backbone, which includes graph backbone, actor backbone and commentator backbone.

[0014] S5. Each UAV access point periodically inherits the shared topology-aware graph network decision backbone, and adapts the shared topology-aware graph network decision backbone locally through low-rank edge personalized fine-tuning.

[0015] S6. Each UAV access point outputs the pilot authentication importance score, total power factor, and user equipment power allocation factor based on the adapted local decision network, jointly determines the pilot authentication and downlink power control scheme, and executes downlink transmission accordingly.

[0016] Furthermore, in step S1, the non-cellular low-altitude communication system includes One drone access point and Individual user equipment, drone access point In the time slot The set of accessible user equipment is denoted as Drone access point With user equipment The channel between them is represented as

[0017] ,

[0018] in and Representing large-scale fading and small-scale fading respectively; UAV access point In the The received uplink pilot signal on each pilot is:

[0019] ,

[0020] For using the first One pilot can be accessed by user equipment Its downlink channel estimation is:

[0021] ,

[0022] in Indicates the length of the pilot sequence. Indicates user equipment Selected pilot index, Indicates user equipment The uplink pilot transmit power, This indicates received noise.

[0023] Furthermore, in step S2, the local map observation includes the status of UAV service resources. UAV physical information Accessible user equipment status set The set of physical information of accessible user equipment And local topology and edge information used to describe pilot sharing relationships and pairwise interaction characteristics. .

[0024] Furthermore, in step S3, a graph attention network is used to perform message passing on the local graph observations, and the aggregated features are represented as follows:

[0025] ,

[0026] in , Indicates user equipment The features obtained by aggregating the pilot sharing information of neighboring user equipment.

[0027] Furthermore, in step S4, the drone access point In the time slot The local experience is represented as:

[0028] ,

[0029] in , and These represent local map observation information, local actions, and local rewards, respectively. After asynchronous backhaul delay alignment, the central computing node uses the sum of the local rewards of all UAVs as its central reward, represented as:

[0030] ,

[0031] Given the maximum latency slots for central nodes and edge drones, the graph backbone in the shared topology-aware graph network decision-making is used to capture pilot sharing relationships, the actor backbone is used to determine pilot authentication and power control decisions, and the critic backbone is used to evaluate state-action pairs and guide policy learning.

[0032] Furthermore, in step S5, the low-rank edge personalization parameters are implemented by freezing the high-dimensional linear layer in the shared decision backbone and learning low-rank additive updates. If we represent any frozen linear layer parameters in the shared decision backbone, then the UAV access point... The personalized parameters have been updated as follows:

[0033] ,

[0034] in and It is a low-rank trainable matrix;

[0035] In addition, each drone access point maintains its local network parameters according to an asynchronous backbone sharing mechanism:

[0036] ,

[0037] in This indicates shared core decision-making parameters. Indicates edge-personalized network parameters. Indicates drone access point The backhaul delay to the central computing node.

[0038] Furthermore, in step S6, the pilot authentication score... and total power factor We obtain it from the following formula:

[0039] ,

[0040] For drone access points Pilot authentication, and through thresholding The binary pilot authentication decision is obtained, namely:

[0041] ,

[0042] For drone access points For power control that can be accessed by users, the first step is to calculate the user equipment that can be accessed. The power importance mask, namely:

[0043] ;

[0044] Then calculate each accessible user equipment The normalized power allocation factor is:

[0045] ,

[0046] in Final drone access point For user equipment Downlink power is:

[0047] .

[0048] The beneficial effects of this invention include:

[0049] (1) The present invention can avoid frequent exchange of global instantaneous channel information in each time slot, enabling each UAV access point to realize distributed pilot authentication and power control based on local estimated channel;

[0050] (2) This invention explicitly characterizes the pilot sharing relationship and the local interaction relationship between accessible user equipment through topology sensing maps, which can improve the ability to characterize the importance of user services under pilot pollution conditions;

[0051] (3) This invention combines central training of shared decision backbone with edge low-rank personalized adaptation, which can adapt to the local wireless environment of different UAV access points while retaining shared policy knowledge;

[0052] (4) The present invention reduces the communication overhead between the central computing node and the UAV access point through an asynchronous backbone sharing mechanism, and reduces the amount of training parameters that need to be updated on the edge UAV side. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the overall framework of the pilot authentication and power control method for a non-cellular low-altitude communication system based on edge personalized topology sensing map of the present invention. The central computing node trains a decision backbone network with shared topology sensing map from asynchronous historical experience, and each edge UAV access point performs edge personalized adaptation using local instantaneous observation.

[0054] Figure 2 This is a schematic diagram of the backbone actor network in the edge-personalized topology-aware graph of the present invention.

[0055] Figure 3 This is a schematic diagram of the backbone commentator network in the edge-personalized topology-aware graph of this invention.

[0056] Figure 4 This diagram illustrates a comparison of the network total throughput convergence performance of the present invention and the benchmark method during the training phase.

[0057] Figure 5 This is a schematic diagram comparing the cumulative distribution function of total network throughput during the deployment phase of the present invention and the benchmark method. Detailed Implementation

[0058] The present invention will be further described below with reference to embodiments.

[0059] This invention addresses the challenges of pilot pollution, downlink transmission interference, and high communication overhead associated with centralized resource management in a time-division duplex non-cellular low-altitude communication system comprised of multiple UAV access points and user equipment. First, for each UAV access point, a local graph observation is constructed, using accessible user equipment as nodes and pilot sharing relationships as edges. Second, a topology-aware graph is used to mine the relative service importance of accessible user equipment based on local graph observations, and selective pilot authentication and service mitigation are employed to alleviate the impact of pilot pollution on downlink transmission performance. Then, a shared topology-aware graph network decision backbone is trained at the central computing node using asynchronous historical experience from multiple UAVs. Each edge UAV inherits this shared decision backbone and performs local adaptation through low-rank edge-specific fine-tuning. Finally, each UAV outputs pilot authentication and downlink power control decisions based on its local observations, achieving distributed resource management. This invention reduces the number of online training parameters and communication overhead for edge UAVs and improves the network throughput of the non-cellular low-altitude communication system.

[0060] This invention considers a time-division duplex non-cellular low-altitude downlink communication system, the system comprising: One drone access point and Each drone access point has a maximum communication coverage area and can only serve user equipment located within its current air-to-ground coverage area; these are called accessible user equipment. Drone Access Point In the time slot The set of accessible user equipment is denoted as The number of user devices that can be connected is Due to the movement of drone access points and user equipment, Its composition changes dynamically over time.

[0061] Drone access point With user equipment Channel modeling between them is

[0062]

[0063] in and These represent large-scale fading and small-scale fading, respectively.

[0064] This invention utilizes the channel reciprocity in a time-division duplex system to estimate the downlink channel from the uplink pilot. Let... The pilot pool represents the number of mutually orthogonal pilot sequences. ,in Indicates length is The Each user equipment (UE) selects a pilot signal randomly and independently before transmission to reduce communication overhead. Indicates user equipment The selected pilot index will result in the uniquely heated pilot indicator vector for each user equipment being:

[0065]

[0066] in ,when hour .

[0067] At the start of each time slot, each user equipment sends its selected uplink pilot signal, and the UAV access point... Receive from its set of accessible user equipment Pilot signals. Unmanned aerial vehicle (UAV) access point. In the The received uplink pilot signal on each pilot is

[0068]

[0069] in Indicates the uplink pilot transmit power. This indicates received noise. For Use the first One pilot can be accessed by user equipment Its downlink channel can be estimated using the minimum mean square error method.

[0070]

[0071] Based on estimated downlink channel Each drone access point uses maximum-ratio transmission to avoid additional inter-drone communication. Drone access point To user equipment The maximum ratio of precoding coefficients is

[0072]

[0073] Since user equipment sharing the same pilot signal can pollute channel estimation, continuing to serve user equipment with severely polluted channel estimation will lead to inaccurate precoding and degraded downlink transmission performance. Therefore, this invention first defines the UAV access point. Binary pilot discrimination vector

[0074]

[0075] in Indicates the use of the first The pilot signal allows access to user equipment via drone access points. It is considered to have high service importance. This indicates that the pilot signal was not accessed. Based on this access result, the user equipment... At the drone access point The service indicator variable at the location is

[0076]

[0077] in Indicates an indicator function, when At that time, user equipment The selected pilot signal is accessed by the UAV. Authentication successful, subsequent downlink transmissions allowed; otherwise .

[0078] Therefore, drone access point The transmitted signal is

[0079]

[0080] in Indicates satisfaction Data signals, Indicates drone access point To user equipment Candidate transmit power. User equipment. The received signal is

[0081]

[0082] in This indicates received noise. Correspondingly, the user equipment... The signal-to-interference-plus-noise ratio is

[0083]

[0084] in Let represent the noise variance. Indicates system bandwidth, then user equipment throughput and drone access points The local throughput metrics are respectively

[0085]

[0086] Therefore, this invention requires joint optimization of the binary pilot authentication variables at each UAV access point. and downlink transmit power To maximize total network throughput:

[0087]

[0088] in Indicates drone access point The maximum downlink transmit power. Due to and pass The problem is a mixed-integer nonlinear optimization problem in a dynamic network environment, characterized by mutual coupling and the inclusion of both discrete and continuous variables in the decision variables.

[0089] This invention employs a pilot authentication and power control method based on edge-personalized topology-aware graph networks.

[0090] 1) Method Introduction

[0091] To address the challenges of high communication overhead in centralized resource management, the difficulty for edge UAV access points to independently obtain global instantaneous channel information, and the dynamic changes in the local wireless environment in low-altitude non-cellular communication systems, this invention proposes a distributed intelligent decision-making framework based on an edge-personalized topology-aware graph network. This framework separates long-term shared resource control strategy learning from short-term device-specific decision-making processes: a central computing node learns reusable pilot access and power control strategy knowledge from the historical communication environments and interaction experiences of multiple UAV access points, encoding this knowledge into a shared topology-aware graph network decision backbone; each edge UAV access point performs edge-personalized adaptation of the shared decision backbone based on its own local observations and service feedback. Through this approach, each UAV access point can execute distributed pilot access and downlink power control decisions relying solely on local observation information, while still utilizing the long-term shared strategy knowledge learned by the central computing node.

[0092] like Figure 1 As shown, specifically, the distributed intelligent decision-making framework includes a central computing agent and multiple independent edge drone agents. The central computing agent collects asynchronous historical experience from multiple drone access points and trains a shared topology-aware graph network decision backbone based on this asynchronous historical experience. The shared decision backbone extracts reusable pilot access and power control decision knowledge under different network conditions, thereby providing a unified policy initialization for each edge drone access point. After receiving the shared decision backbone, each edge drone agent performs edge-specific fine-tuning using its own local instantaneous observation information and local service feedback in its current time slot, enabling the inherited shared policy to adapt to the local channel conditions, distribution of accessible user equipment, pilot sharing status, and quality of service requirements of the corresponding drone access point.

[0093] Furthermore, this invention introduces a topology-aware map into the shared decision backbone to learn the relative service importance of each accessible user equipment (UAE) based on locally estimated channel conditions and pilot sharing relationships among accessible UAEs. This allows each UAV access point to more accurately select the pilot groups to access and the UAEs to serve, even in the presence of pilot pollution and local interference, and to further perform targeted downlink power control. Through this framework, this invention avoids frequent online exchange of global instantaneous information between UAV access points, reduces the computational burden on the edge side by updating only local personalized parameters, and improves the system's adaptability to diverse low-altitude communication environments.

[0094] 2) Local map observation and topology-aware map feature aggregation

[0095] To enable each UAV access point to perform pilot authentication and power control based solely on local information, this invention constructs a local graph observation for each UAV access point. Specifically, each accessible user equipment (UE) is treated as a node, and edges are constructed between UEs using the same pilot. This graph is used to explicitly characterize pilot sharing relationships and potential service contention relationships caused by pilot multiplexing.

[0096] This map observation consists of local observation information and estimated channel information, and is organized into five categories of information: (1) UAV service resource status This includes throughput of the previous time slot, quality of service penalty, power usage, number of accessible and served user devices, and access statistics; (2) physical information of unmanned aerial vehicles. (3) Accessible user equipment status set This includes local service status, service quality mismatch, historical power and pilot information, and current and historical estimated channel status; (4) a set of physical information of accessible user equipment. This includes user equipment location, quality of service requirements, and selected pilot indications; (5) local topology and side information. It is used to describe pilot sharing relationships and pairwise interaction characteristics, including estimating potential interference indications, user equipment distances, quality of service gaps, and throughput gaps.

[0097] This invention employs a graph attention network to aggregate pilot signals and share information about neighboring user equipment. Let... This indicates the message passing process between accessible user devices, specifically the drone access point. The aggregation characteristics are

[0098]

[0099] in , Indicates that it can be connected to user equipment. Features are aggregated from neighboring user equipment. These aggregated features can characterize the relative service importance of different accessible user equipment under imperfect channel estimation conditions.

[0100] 3) Joint pilot authentication and power control decision mapping

[0101] This invention designs a decision network capable of simultaneously handling discrete pilot authentication and continuous power control output. First, the UAV is physically embedded... UAV observation status and aggregated user equipment representation The data is spliced ​​together to obtain continuous pilot authentication scores. and total power factor ,Right now

[0102]

[0103] in , Indicates the first The continuous access fraction of each pilot. and These represent the corresponding network modules. (UAV access point) The binary pilot authentication decision is obtained through thresholding:

[0104]

[0105] and This represents the total power factor obtained through Sigmoid activation.

[0106] Then, this invention employs a continuous power importance mask to softly adjust the power allocation of accessible user equipment based on the pilot authentication score. For accessible user equipment... The continuous importance mask is

[0107]

[0108] Unlike a hard ruling, It is a continuous importance weight obtained by mapping the pilot authentication score, which can ensure continuous gradients during training and suppress subsequent power control of low-importance user devices.

[0109] For each accessible user device Its normalized power factor is derived from , and The concatenation representation is obtained as follows:

[0110]

[0111] in Ultimately, user equipment The downlink transmit power is

[0112]

[0113] This can suppress unnecessary power allocation to user equipment with low service importance.

[0114] 4) The center trains and shares decision-making backbone personnel.

[0115] This invention models the central computing node as a central agent and each UAV access point as an edge agent. Although pilot authentication and power control decisions are executed locally by each UAV access point, different UAV access points still share reusable long-term pilot authentication and power control rules. Therefore, this invention utilizes the computing power of the central computing node to train a shared topology-aware graph decision backbone network from the historical experience of multiple UAV access points.

[0116] During training, the central agent asynchronously collects historical local experience from drone access points. In the time slot Local experience is represented as

[0117]

[0118] in This indicates local map observation information. Indicates local action, This indicates the corresponding local reward. Due to varying backhaul delays between different drone access points and the central computing node, let... Indicates drone access point The transmission delay to the central computing node, then the time slot The experience generated In the time slot Reaching the central intelligent agent. Let... Indicates the maximum return latency, then in the time slot The central agent can generate time alignment experience for all drone access points.

[0119] The shared topology-aware graph decision backbone network comprises a graph backbone, an actor backbone, and a critic backbone. The graph backbone is used to capture pilot sharing relationships, the actor backbone is used to determine pilot authentication and power control decisions, and the critic backbone is used to evaluate state-action pairs and guide policy learning. The actor design is as follows: Figure 2 As shown, the critic's design is as follows Figure 3 As shown. Since the central computing node can obtain the historical experience of each UAV access point, this invention defines the central reward as the sum of all time-aligned local rewards:

[0120]

[0121] The center's incentives enable shared decision-making backbones to learn long-term strategic knowledge at the network level and contribute to improving the network's total throughput target.

[0122] 5) Edge Personalization Adaptation

[0123] While the shared decision backbone trained centrally can capture long-term shared pilot authentication and power control strategies, directly applying the same backbone to all UAV access points may not be suitable for the local channel conditions, distribution of accessible user equipment, pilot sharing status, and quality of service requirements of different UAV access points. On the other hand, completely retraining the entire backbone at each UAV access point would place an excessively high computational and energy burden on resource-constrained edge devices.

[0124] Therefore, this invention employs a lightweight, device-specific edge personalization mechanism. Its basic idea is to freeze the high-dimensional linear layers in the shared decision backbone and learn low-rank additive updates for each UAV access point. Let... If we represent any frozen linear layer parameters in the shared decision backbone, then the UAV access point... The corresponding personalized parameters are

[0125]

[0126] in and This is a low-rank trainable matrix. (UAV access point) Personalized backbone computing for

[0127]

[0128] in Indicates drone access point Local observations, Indicates drone access point The corresponding trainable network module.

[0129] The local reward design for drone access points is as follows

[0130]

[0131] in This indicates a local service quality penalty, used to reflect the impact on drone access points. The quality of service of accessible user equipment, and This is a weighting factor used to balance throughput and quality of service penalties. This local reward incentivizes edge drone access points to increase their achievable throughput and meet the quality of service requirements of each user device.

[0132] 6) Asynchronous backbone network sharing mechanism

[0133] To reduce communication overhead between edge drone access points and the central computing node, this invention employs an asynchronous backbone network sharing mechanism that considers heterogeneous backhaul latency. The shared decision backbone network primarily captures long-term policy knowledge; therefore, the central agent broadcasts the shared decision backbone at fixed intervals, rather than transmitting it synchronously in every time slot. Simultaneously, device-specific edge-personalized parameters are retained locally at each drone access point and can be updated in each time slot.

[0134] Drone access point The parameters consist of the most recently received shared decision backbone parameters and local edge personalized parameters, and can be represented as follows:

[0135]

[0136] in This indicates shared core decision-making parameters. Indicates drone access point Edge-specific network parameters, and These represent the update slot and the latest update slot, respectively.

[0137] Example:

[0138] In this embodiment, a square service area with a width and length of 2000 m is considered. The system is configured with... There are four drone access points with initial horizontal positions of (500, 500) m, (250, 250) m, (250, 750) m, (750, 250) m, and (750, 750) m, corresponding to heights of 100 m, 200 m, 150 m, 300 m, and 150 m, respectively. Each drone access point moves randomly at a speed of 2 m / s and changes its direction of movement upon reaching the area boundary. The system considers... A number of user equipments are randomly generated within the service area, and each user equipment uses a random walk movement model with a speed of 2 m / s. The number of uplink orthogonal pilots is set to... The carrier frequency is set to 2.5 GHz, and the system bandwidth is set to 10 MHz. The latency from the UAV access point to the central computing node... The time slots are set to 0, 1, 2, 3, and 4 respectively. The maximum communication coverage radius of each UAV access point is 1000 m, and the maximum transmit power is 43 dBm.

[0139] In implementation, each user equipment (UE) first randomly selects a pilot signal and transmits an uplink pilot signal in each time slot, allowing the UAV access point to obtain local downlink channel estimates for the accessible UEs. Subsequently, each UAV access point constructs a local graph observation based on the accessible UEs, pilot sharing relationships, local service status, UE physical information, and edge interaction information, and obtains the results through a graph attention network. Then, the central computing node trains a shared decision backbone based on asynchronous historical experience and periodically broadcasts it to each UAV access point. After receiving the shared decision backbone, each UAV access point only updates its low-rank edge personalized parameters to adapt to its local wireless environment.

[0140] During the online execution phase, no parameters need to be updated at any of the drone access points. By inputting local map observations into the decision network, pilot authentication results and power allocation results can be obtained, thereby completing the corresponding downlink transmission.

[0141] To verify the effectiveness of this invention, it can be compared with centralized graph neural networks, distributed graph neural networks, distributed fully connected networks, and stochastic methods. Specifically, centralized graph neural networks allow all UAV access points to share a centrally trained decision network without considering central-edge latency; distributed graph neural networks allow each UAV access point to independently train its own local graph neural network; distributed fully connected networks do not use graph neural networks to model local pilot sharing relationships; and stochastic methods randomly determine pilot authentication and power control decisions.

[0142] Figure 4 This paper demonstrates the convergence performance of different methods in terms of total network throughput during the training phase. Centralized graph neural networks benefit from the shared policy obtained through central training in the early stages of training, exhibiting good initial performance. However, due to their unified decision representation, they struggle to adapt to the heterogeneous local wireless environments of different UAV access points, thus limiting subsequent performance improvements. Distributed graph neural networks can gradually improve performance through local training at each UAV access point, but their learning speed is slow due to the lack of long-term shared policy knowledge across UAV access points, resulting in relatively low final convergence performance. In contrast, this invention utilizes both the shared policy knowledge obtained from training at the central computing node and the device-specific learning capabilities of each UAV access point. This allows the shared decision backbone to provide good policy initialization and adapt to the local wireless environment through edge-personalized fine-tuning, thus outperforming the comparative methods in both training convergence speed and final convergence performance. Furthermore, the performance gap between distributed graph neural networks and distributed fully connected networks indicates that graph neural networks can more effectively learn pilot sharing relationships among accessible user equipment and the relative service importance of user equipment. Fully connected networks, lacking explicit modeling capabilities of the local graph structure, struggle to learn effective resource management strategies from complex local observations, thus their performance approaches that of stochastic methods.

[0143] Figure 5 The cumulative distribution function of total network throughput for different methods during the deployment and validation phase is shown. Compared to the benchmark methods, the cumulative distribution function curve corresponding to this invention shifts to the right overall, indicating that this invention achieves a higher probability of higher total network throughput during the deployment phase. Specifically, the average throughput achieved by this invention during the deployment phase is 380.01 Mbps, while the average throughputs of centralized graph neural networks and distributed graph neural networks are 308.59 Mbps and 220.98 Mbps, respectively. Compared to centralized graph neural networks and distributed graph neural networks, the average throughput of this invention is improved by 23.1% and 71.9%, respectively. These results demonstrate that this invention, by combining a centrally shared decision backbone with edge-specific fine-tuning, can maintain effective online decision-making capabilities without continuously updating all network parameters during the deployment phase, thus making it suitable for practical non-cellular low-altitude communication scenarios with low online computational overhead.

[0144] Table 1 shows a comparison of the number of trainable parameters for different methods:

[0145] Table 1. Number of trainable parameters for different methods

[0146]

[0147] In this invention, each edge UAV access point only needs to update 162,457 trainable parameters, while in a distributed graph neural network, each edge UAV access point needs to update 231,084 trainable parameters. Therefore, this invention reduces the number of edge trainable parameters by approximately 29.7% compared to a distributed graph neural network. This reduction in parameters mainly stems from a low-rank device-specific personalized fine-tuning mechanism, where most parameters in the shared decision backbone are frozen at the edge UAV access point, and only a small number of low-rank personalized parameters are updated. Thus, this invention reduces the edge training and computational burden while retaining the topology-aware graph network's ability to represent the importance of local pilot shared topology and user equipment services.

Claims

1. A pilot and power control method for a low-altitude communication system, used in a time-division duplex downlink non-cellular low-altitude communication system composed of multiple UAV access points and multiple user equipment, characterized in that, Includes the following steps: S1. In each time slot, each user equipment randomly selects an uplink pilot sequence and sends a pilot signal. Each UAV access point obtains the local downlink channel estimate of the user equipment that can be accessed based on the received uplink pilot signal and channel reciprocity. S2. Using the accessible user equipment at the UAV access point as nodes and the pilot sharing relationship between accessible user equipment that selects the same pilot as edges, construct a local graph observation for each UAV access point, and use the UAV service status, UAV physical information, user equipment service status, user equipment physical information, pilot sharing topology and edge interaction information as input to the topology-aware graph. S3. Utilize the topology-aware map to perform message passing and feature aggregation on the local map observations to obtain the aggregated features of each accessible user equipment, so as to characterize the relative service importance of each user equipment under imperfect channel estimation and pilot pollution conditions. S4. At the central computing node, asynchronous historical experience uploaded by multiple UAV access points is used to train a shared topology-aware graph network decision backbone, which includes graph backbone, actor backbone and commentator backbone. S5. Each UAV access point periodically inherits the shared topology-aware graph network decision backbone, and adapts the shared topology-aware graph network decision backbone locally through low-rank edge personalized fine-tuning. S6. Each UAV access point outputs the pilot authentication importance score, total power factor, and user equipment power allocation factor based on the adapted local decision network, jointly determines the pilot authentication and downlink power control scheme, and executes downlink transmission accordingly.

2. The pilot and power control method for a low-altitude communication system according to claim 1, characterized in that, In step S1, the non-cellular low-altitude communication system includes One drone access point and Individual user equipment, drone access point In the time slot The set of accessible user equipment is denoted as Drone access point With user equipment The channel between them is represented as ， in and Representing large-scale fading and small-scale fading respectively; UAV access point In the The received uplink pilot signal on each pilot is: ， For using the first One pilot can be accessed by user equipment Its downlink channel estimation is: ， in Indicates the length of the pilot sequence. Indicates user equipment Selected pilot index, Indicates user equipment The uplink pilot transmit power, This indicates received noise.

3. The pilot and power control method for a low-altitude communication system according to claim 2, characterized in that, In step S2, the local map observation includes the status of UAV service resources. UAV physical information Accessible user equipment status set The set of physical information of accessible user equipment And local topology and edge information used to describe pilot sharing relationships and pairwise interaction characteristics. .

4. The pilot and power control method for a low-altitude communication system according to claim 3, characterized in that, In step S3, a graph attention network is used to perform message passing on local graph observations, and the aggregated features are represented as follows: ， in , Indicates user equipment The features obtained by aggregating the pilot sharing information of neighboring user equipment.

5. The pilot and power control method for a low-altitude communication system according to claim 1, characterized in that, In step S4, the drone access point In the time slot The local experience is represented as: ， in , and These represent local map observation information, local actions, and local rewards, respectively. After asynchronous backhaul delay alignment, the central computing node uses the sum of the local rewards of all UAVs as its central reward, represented as: ， Given the maximum latency slots for central nodes and edge drones, the graph backbone in the shared topology-aware graph network decision-making is used to capture pilot sharing relationships, the actor backbone is used to determine pilot authentication and power control decisions, and the critic backbone is used to evaluate state-action pairs and guide policy learning.

6. The pilot and power control method for a low-altitude communication system according to claim 1, characterized in that, In step S5, the low-rank edge personalization parameters are implemented by freezing the high-dimensional linear layer in the shared decision backbone and learning low-rank additive updates. If we represent any frozen linear layer parameters in the shared decision backbone, then the UAV access point... The personalized parameters have been updated as follows: ， in and It is a low-rank trainable matrix; In addition, each drone access point maintains its local network parameters according to an asynchronous backbone sharing mechanism: ， in This indicates shared core decision-making parameters. Indicates edge-personalized network parameters. Indicates drone access point The backhaul delay to the central computing node.

7. The pilot and power control method for a low-altitude communication system according to claim 1, characterized in that, In step S6, the pilot authentication score is calculated. and total power factor We obtain it from the following formula: ， For drone access points Pilot authentication, and through thresholding The binary pilot authentication decision is obtained, namely: ， For drone access points For power control that can be accessed by users, the first step is to calculate the user equipment that can be accessed. The power importance mask, namely: ； Then calculate each accessible user equipment The normalized power allocation factor is: ， in Final drone access point For user equipment Downlink power is: 。

Images