Unmanned aerial vehicle small base station RAN slicing method based on segmentation federated learning

Abstract

An unmanned aerial vehicle small base station RAN slicing method based on segmentation federal learning comprises the following steps: in a service-oriented RAN slicing framework, using a flexible index of resource sharing to measure slice performance isolation quality; the isolation quality maximization problem is modeled as a joint optimization problem of shared resource reservation and resource slicing; solving the problem by adopting a multi-machine cooperation-oriented method; a segmentation federated learning SFL method based on clustering is constructed, the positions of the unmanned aerial vehicle and the ground base station are used for clustering, and the base station serves as a cluster head and is responsible for receiving unmanned aerial vehicle model parameters serving as members; the model of the member is divided into an air model and a ground model, and local iteration is completed through gradient interaction; transform enhanced air model aggregation is adopted, Transform is integrated on a ground base station serving as a cluster head, and a multi-head attention mechanism is used for optimizing air model aggregation so as to enhance generalization of a global model.

Description

Technical Field

[0001] The present invention is an application of artificial intelligence technology in the field of communication technology, specifically a UAV small base station RAN slicing method based on segmented federated learning. Background Art

[0002] With the advent of the 6G era, drone small base stations have gradually become an indispensable part of modern communication networks. Due to the advantages of flexible deployment and line-of-sight communication, drone small base stations can quickly cover network blank areas in remote or emergency scenarios, provide customized services for applications such as the Internet of Things (IoT), smart cities, and unmanned driving, and expand the functions of traditional base stations. In the 6G network, drone small base stations not only need to cope with the needs of massive device access and high-speed data transmission, but also need to provide ultra-low latency and ultra-high reliability communication guarantees, becoming an indispensable element in future networks.

[0003] Considering the differences in the tasks that drones perform and the fact that different types of tasks have very different requirements for network resources, the introduction of Radio Access Networks (RAN) slicing technology has become an inevitable choice. RAN slicing can dynamically divide the communication, computing and other physical resources in heterogeneous networks according to changes in network conditions. After the task arrives at the node, it is processed by the corresponding slice to provide differentiated Quality of Service (QoS) guarantees. [1] . With the widespread deployment of small base stations for drones, the application of RAN slicing technology has been extended to aerial networks. Due to the dynamic changes in flight trajectory, coverage, and service requirements, and the fact that drones are limited by battery life, storage capacity, and computing power, RAN slicing technology takes into account both lightweight and adaptability. Slice resource allocation and optimization need to consider the real-time location of drones, network load, and resource constraints, so as to provide efficient, reliable, and low-latency customized services for a variety of applications. For drone networks, the AirSlice framework [2] Through flexible resource slicing and QoS strategies, SoftFloat provides differentiated service support for drone network services. [3] It is a software-defined 5G network slicing framework that supports QoS-aware drone communications. The framework enables differentiated service provision and intelligent processing of data traffic through intelligent resource management.

[0004] Existing drone RAN slices mainly include stand-alone independent decision-making [4] and reliance on ground base stations [5] However, both of them are difficult to support the autonomous and large-scale movement of a large number of drones.[6] and Split Learning (SL) [7] It provides an effective solution for resource management of distributed UAV networks. FL achieves collaborative learning of multiple UAVs without sharing data, improving overall performance. SL reduces the computing and storage burden of UAVs by splitting model training between UAVs and ground base stations. However, their integration with UAV systems still faces many challenges, including:

[0005] 1) Dual optimization of joint training cost and performance isolation. During aerial training, the burden of data interaction and model update is heavy, which reduces the sustainability of the UAV network. [8] , training frequency reduction [9] or client scheduling

[10] It can reduce the training cost, but it will lead to a decrease in model accuracy. In addition, resource sharing is an important indicator for measuring the quality of performance isolation. Most existing studies adopt a fixed resource sharing strategy, which means that the scope of resource sharing is predetermined and cannot be flexibly adjusted according to the needs and changes of actual tasks. Although this strategy performs well in some scenarios, it cannot achieve optimal resource utilization when the UAV mission load fluctuates greatly or the environment changes frequently. How to achieve dual optimization of training cost and performance isolation while ensuring model accuracy is a key issue that needs to be urgently solved in the dynamic deployment of UAV systems.

[0006] 2) Environmental adaptability and generalization of airborne models. Drones that perform tasks on the move place higher demands on the generalization and robustness of the model. Traditional FL methods have difficulty coping with the data heterogeneity and differences caused by mobility. What’s more, due to the limited coverage, the data collected by drones has strong local characteristics. Traditional attention mechanisms have difficulty optimizing models by extracting local key information. GSM-FedAVG

[11] It is a global shared model-federated average algorithm that improves the environmental adaptability and generalization of the model by sharing the model among all clients.

[12] By introducing regularization terms to cope with the heterogeneity of the system and data, the stability and generalization of the model are enhanced. Ma et al. used the attention mechanism to focus on local key tasks and improve the path planning accuracy and robustness of the multi-UAV system.

[13] How to strengthen the model’s perception of the local environment and integration of global information needs further exploration. Summary of the invention

[0007] In response to the above challenges and limitations, this paper proposes a split federated learning (SFL) method based on Transformer (model) multi-head attention. This technical solution supports drone RAN slicing in multi-machine collaboration scenarios, with the aim of maximizing performance isolation and reducing training costs.

[0008] The main contributions of this invention include three aspects:

[0009] First, taking into account both performance isolation and training cost, a flexible indicator that considers resource sharing is used to measure the quality of slice performance isolation. The isolation quality maximization problem is modeled as a joint optimization problem of shared resource reservation and resource slicing. A multi-machine collaboration-oriented method is used to solve the problem.

[0010] Second, a clustering-based segmented federated learning (SFL) method is constructed. The locations of drones and ground base stations are used for clustering. The base station (as the cluster head) is responsible for receiving the model parameters of the drones (as members). To reduce energy consumption, the member models are divided into aerial models and ground models, and local iterations are completed through gradient interaction.

[0011] Third, Transformer-enhanced aerial model aggregation. Transformer is integrated on the ground base station as the cluster head, and the multi-head attention mechanism is used to optimize aerial model aggregation to enhance the generalization of the global model.

[0012] Simulation results show that the proposed scheme outperforms the typical FL method in terms of prediction accuracy and slice performance isolation. In addition, the proposed model segmentation scheme significantly reduces the training cost without sacrificing prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 represents a RAN slicing framework for resource sharing;

[0014] Figure 2 It represents the RAN structure of the drone small base station - SFL;

[0015] Figure 3 Represents the aggregation of model segmentation, local iteration, and Transformer enhancement;

[0016] Figure 4(a) to Figure 4(c) They represent the changes of prediction errors of slices 1, slice 2 and slice 3 with the communication rounds;

[0017] Figure 5(a) to Figure 5(c) Represent the probability density functions of the prediction errors of the three slices, slice 1, slice 2 and slice 3 respectively;

[0018] Figure 6It indicates the impact of the total number of RBs on performance isolation;

[0019] Figure 7(a) to Figure 7(c) They represent the changes of prediction errors of slices 1, slice 2 and slice 3 with model segmentation and no segmentation as the communication rounds increase.

[0020] Figure 8 Represents the impact of model segmentation on training cost. DETAILED DESCRIPTION

[0021] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0022] The first section summarizes the technical solution of the present invention.

[0023] Section 2 introduces the RAN slicing framework, performance isolation metrics, and models the optimization problem.

[0024] Section 3 designs a clustering-based segmentation federated learning (SFL) method to support collaborative training in large-scale scenarios.

[0025] Section 4 presents the experimental design and result analysis for performance evaluation.

[0026] Section V concludes the study and discusses future prospects.

[0027] 1 Overview

[0028] The present invention proposes a split federated learning (SFL) method based on multi-head attention to support resource slicing in multi-machine collaboration scenarios and achieve dual optimization of performance isolation and energy consumption.

[0029] The present invention adopts a resource sharing flexibility index to measure the slice performance isolation quality. The isolation quality maximization problem is modeled as a joint optimization problem of shared resource reservation and resource slicing. A SFL-based method is developed to solve this problem.

[0030] To improve scalability, the UAV positions are used for clustering before executing SFL. The ground base station acts as a cluster head to receive the model parameters of the member UAVs.

[0031] To reduce energy consumption, the onboard model is partitioned and partially migrated to the cluster head.

[0032] Local iteration is completed between the open-space models through interactive gradients.

[0033] The Transformer is integrated on the ground base station as the cluster head, and the multi-head attention is used to improve the generalization of the aggregation model.

[0034] Simulation results show that the proposed scheme outperforms the benchmark federated learning (FL) method in terms of slice performance isolation and training cost.

[0035] 2 System Model

[0036] As a small base station in the air, drones can be deployed over a large area according to the mission and are not restricted by geographical location. The drones and ground base stations are connected via 5G NR.

[14] The protocol completes the handshake and inquiry. When a drone needs to leave the area to perform other tasks, it will issue a notification to release the connection. For ease of reading, the main symbols and variables used in the following are summarized in Table 1.

[0037] Table 1. Definition of variables

[0038]

[0039] 2.1 Service-Oriented RAN Slicing Framework

[0040] The present invention considers a resource slicing method for resource sharing. Physical resources are pooled and divided into multiple slices and a shared area. According to resource requirements, each drone dynamically determines the allocation of shared areas and slice resources. Figure 1 Assume that time is dynamically divided into multiple slice windows, each containing a certain number of discrete time slots. Assume that the set of slice windows is represented as The time slot set and window length of slice window a are expressed as and T (a) .

[0041] Assume that the total number of RBs (resource blocks) held by drone k is B k , these resources are virtualized as M k service slices, and the slice set is recorded as At the beginning of each window, the RBs of each service slice and shared area are reallocated and maintained until the end of the window. At the end of each window, drone k predicts the resource requirements of each slice and the size of the shared area in the next window. The arriving tasks are scheduled to different slices according to their attributes. Each slice allocates resources to tasks centered on the time slot. At the beginning of window a, the number of RBs allocated to slice m and shared area on drone k is expressed as and The sum of the two cannot exceed the total number of RBs held by the drone, satisfying

[0042]

[0043] In the case of insufficient slice resources, RBs in the shared area can be temporarily occupied. The setting of resource sharing area helps to alleviate the slice performance violations caused by the sharp fluctuation of short-term resource demand. When the resource pressure in slice window a is high (low), the shared area should be appropriately expanded (reduced) to buffer the fluctuation of resource demand. The size of the shared area cannot exceed B k 15%

[0044] 2.2 Slice Performance Isolation Metrics

[0045] Performance isolation ensures that the resource usage of each slice does not interfere with the service provision of other slices, which is a necessary condition for the coexistence of multiple slices.

[15]

[16] The resource allocation strategy at the beginning of each slice window determines the performance isolation quality of each slice within the window.

[0046] Assume that in the time slot The number of RBs required for slice m on drone k is The number of RBs required for slice m on drone k in time slot t is The number of occupied RBs in the shared area is The performance isolation quality at this moment is defined as

[0047]

[0048] From formula (2), we can see that and The smaller the gap ( The higher the prediction accuracy of ), the higher the performance isolation quality. Slice m does not need to be occupied It can meet current needs and achieve the highest level of isolation quality like All in the shared area Even if RBs are occupied, it is not enough to meet the service requirements of slice m, and performance isolation is destroyed. like Slice m needs to temporarily occupy shared RBs to meet current needs, and performance isolation is weakened.

[0049] 2.3 Problem Modeling and Solution Framework

[0050] Assume that under the slice window a, the slice resource allocation vector on drone k is Based on formula (2), the average isolation quality of the slices on drone k is calculated as Apply it to For all windows in the long-term slice performance isolation optimization problem, the problem is modeled as

[0051]

[0052] question The essence is to allocate each slice and RB in the shared area To maximize performance isolation quality.

[0053] Assume that in window a, the number of RBs required for each slice on drone k is predicted to be The number of RBs allocated to slice m on drone k is calculated as

[0054]

[0055] Substituting equation (3) into equation (1), we can obtain This completes the problem It can be seen that is the starting point for problem solving. The following Section 3 proposes the SFL method for multi-machine collaboration scenarios to solve

[0056] 3 SFL-based solutions

[0057] This section designs a SFL method to enhance the model performance. Figure 2 The presented workflow consists of opportunistic clustering, local iteration towards a segmentation model, and aggregation of a multi-head attention model.

[0058] 3.1 Opportunistic Federated Clustering

[0059] When a large number of drones are observed hovering in the jurisdiction, the MEC controller uses weighted k-means

[17] Start clustering. Clustering is started at an opportune moment. The MEC controller broadcasts a notification to each drone and asks whether to participate in this round of clustering. The ground base station serves as the cluster head and the drones as members of the cluster.

[0060] The weighted coefficient of the drone position is incorporated into the Euclidean distance calculation of the k-means algorithm. The position of drone k and base station i is represented as g k and g i , the discrete degree of the position is used to determine the weighting coefficient. The weight of the drone position is quantified in the form of standard deviation, that is,

[0061]

[0062] in, is the average value of the drone position. Based on this, the weighted Euclidean distance between drone k and base station i (center of mass) is calculated as

[0063]

[0064] By minimizing The MEC controller determines the connection between the drone and the base station.

[0065] The clustering process based on weighted k-means is summarized as Algorithm 1. Assume that the set of base stations randomly selected as the initial centroid and the cardinality are and I k According to equations (4) and (5), the distance between other drones and this I k The weighted Euclidean distance of each centroid is classified into the class with the closest weighted Euclidean distance. is the set of drones in cluster i. Clustering helps improve the scalability of collaborative training and reduce communication costs.

[0066] Algorithm 1: Opportunistic clustering

[0067] enter:

[0068]

[0069] Output:

[0070] 1) Randomly select I k Base stations are used as the initial centroid;

[0071] 2)

[0072] 3) Repeat

[0073] 4) for each do

[0074] 5)

[0075] 6)

[0076] 7) end for

[0077] 3.2 Model segmentation and local iteration

[0078] Bidirectional Long Short-Term Memory (Bi-LSTM)

[18] As a lightweight model deployed on drones To predict resource requirements. Assume that in time slot The resource prediction value of the forward LSTM is The resource prediction value of the reverse LSTM is Then, the resource prediction value of Bi-LSTM is expressed as

[0079]

[0080] in, represents the combination of two hidden states. Assume that the model carried by drone k is represented by w i,k , then the number of RBs required for slice m in window a on drone k is predicted to be

[0081]

[0082] based on and Loss Function It is used to measure the prediction error of the number of RBs required for slice m on drone k. i,k The loss function is expressed as

[0083]

[0084] In SFL, i,k It is divided into an aerial model and a ground model, denoted as w i,k,b ′ and w i,k,p ′; w i,k,b ′ is kept locally, w i,k,p ′ is transferred to the base station i as the cluster head; the gradient of the aerial model and the ground model is back-propagated through wireless transmission; let and Represents w i,k,b ′ and w i,k,p ′’s loss function, and The accumulation of

[0085]

[0086] Figure 3 The local iteration process after the segmentation of the model on drone k is shown. The data carried by drone k and the newly collected data are used to i,k,b ′Forward propagation, the intermediate features are passed to base station i. Base station i performs forward and backward propagation to update w i,k,p ′ and returns the gradient. After receiving the gradient, through back propagation, the underlying model of drone k is updated to

[0087]

[0088] Among them, μ k is the learning rate of drone k. The above complete forward and backward propagation process is called a local iteration.

[0089] 3.3 Transformer-enhanced segmentation model aggregation

[0090] Received i,k,bAfter that, base station i starts federated aggregation to update the global model. Figure 3 As shown in the figure, in this process, the Transformer encoder (TE) and decoder (TD)

[20] It is integrated on the ground base station to enhance the generalization of the global model.

[0091] w i,k,b The input sequence is encoded by TE into an input sequence. After multiple layers of stacking, the input sequence generates a vector Q about Query, Key and Value. k , K k 、V k

[21] In federated aggregation, Transformer’s multi-head attention mechanism handles W i,b (Q k ) and w i,k,b (K k ). Through formula (12), the number of drones in cluster i and base station The attention weights of the underlying aggregation model parameters are expressed as and in

[0092]

[0093] After updating the local model, The model of the drone, namely w i,k,b , are uploaded to base station i. These models are aggregated by base station i into

[0094]

[0095] Then, TD decodes and outputs the aggregated underlying model parameter vector. Base station i converts W b The model aggregation implementation details are summarized in Algorithm 2. The base station i receives the w of the drones in the cluster. i,k,b .w i,k,b First encoded by TE as w i,k,b (Line 3), and then aggregated into W through the multi-head attention mechanism i,b (Lines 4-6), which is finally decoded by TD as W i,b (Line 7). The global underlying model parameters after base station aggregation in are expressed as W b , it was sent to each drone for a new round of training.

[0096] Algorithm 2: Serial segmentation model aggregation algorithm based on multi-head attention

[0097] Input: w i,k,b ; W i,b

[0098] Output: W b

[0099] 1) for each do

[0100] 2) for each do

[0101] 3) w i,k,b Encoded by TE as w i,k,b ;

[0102] 4) w i,k,b Q is generated by multi-layer stacking k , K k 、V k ;

[0103] 5)

[0104] 6)

[0105] 7)W i,b Decoded as W by TD i,b ;

[0106] 8) end for

[0107] 9)

[0108] 10) Repeat steps 3-10 to get W b ;

[0109] 11) end for

[0110] 3.4 Analysis of Joint Training Cost

[0111] This section analyzes the impact of model partitioning on communication and computational costs in collaborative training from a theoretical perspective.

[0112] The communication cost refers to the amount of data exchanged between UAV k and base station i during the joint training process. The number of parameters of the model carried by UAV k is expressed as |w i,k |. In traditional FL mode, all |w i,k |Need to upload to base station i. FL involves two-way transmission of the model, and the data volume is increased to 2|w i,k|. After the model is segmented, only the gradient of the back propagation needs to be transmitted. Assume that the amount of data transmitted each time is L. Since the segmentation model includes forward and backward propagation, the amount of data after all gradients are superimposed is 4L. In other words, as long as 4L < 2|w i,k |, the communication cost of the proposed SFL is less than that of FL.

[0113] The computational cost reflects the time cost of the training process. After the model is split, only w i,k,b Need to be updated. The time cost function is expressed as if only It holds true that the time cost of the proposed segmentation method must be less than that of the traditional FL. Table 2 lists the communication cost and time cost of UAV k and base station i in a training cycle.

[0114] Table 2 Analysis of segmentation-based FL training cost

[0115]

[0116] 4 Experimental design and results analysis

[0117] The present invention carried out a simulation experiment on a high-performance server. The server configuration included an Intel Core i9-14900K processor, 64GB DDR5 5200MHz memory, a 4TB PCIe 4.0 solid-state drive, an ASUS PRIME Z790-PWIFID5 motherboard, and two Gigabyte RTX 4090 24GB WindForce graphics cards.

[0118] On the UAV side, a Bi-LSTM model consisting of two bidirectional LSTM layers (each with 128 hidden neurons) was deployed, followed by a linear layer to map features to prediction results. Transformer encoder-decoders were deployed on each UAV and ground base station for feature merging and parameter aggregation. To simulate collaboration between nodes, the data and models of each UAV and ground base station were encapsulated in independent Docker containers. These containers run on the same physical server. The containers communicate through a virtual network, simulating the message passing and collaboration process in a real network environment. This setting enables us to flexibly adjust the number of UAVs and ground base stations, simulate a large-scale, distributed multi-UAV network environment, and evaluate the scalability and effectiveness of the proposed method in practical applications. In order to evaluate the impact of different strategies on the overall performance, the proposed scheme is represented as Transformer-enhanced FL (Proposed-1 without SL, Proposed-2 with SL).

[0119] For a comprehensive and objective comparison, the three methods listed in Table 3 are selected as baseline methods (without SL).

[0120] Table 3

[0121] Method naming <![CDATA[Ordinary weight aggregation

[22] > <![CDATA[FedAvg

[23] ]]> <![CDATA[Traditional FL

[24] > Baseline-1 √ Baseline-2 √ Baseline-3 √

[0122] The method proposed by Thammawichai et al.

[25] It is used to calculate the energy consumption during parameter sending and receiving; the computational cost is reflected by the time cost.

[26] The communication data from Trentino, Italy, was selected, including data services, voice calls, etc.

[27] , simulate the arrival of slice tasks to evaluate the performance of the proposed method in different business scenarios. The detailed parameter settings are shown in Table 4.

[0123] Table 4 Experimental parameter settings

[0124]

[0125]

[0126] 4.1 Analysis of resource demand forecast accuracy

[0127] This group of experiments uses MSE to measure the error between the predicted value and the true value of RB demand within a slice window. Figure 4(a) to Figure 4(c) By comparing the prediction errors of different methods, it is shown how the prediction error of the RB required for each slice changes with the communication round. As the communication round increases, the prediction error of each slice shows a downward trend and converges to an interval. The prediction error of the method proposed in the present invention on each slice is lower than that of the baseline method, indicating that it can accurately predict resource requirements in most time slots. In addition, compared with the baseline method, the proposed method uses the Transformer's multi-head attention mechanism to focus on nodes that contribute more to the global model update, thereby accelerating the model convergence and indirectly reducing communication and computational overhead. Compared with Baseline-1, 2, and 3, Proposed-1 can effectively process heterogeneous data by combining Transformer, and achieves performance gains of 8.67%-23.78%, 7.19%-14.46%, and 2.6%-8.67% on the three slices, respectively.

[0128] Figure 5(a) to Figure 4(c)The probability density function (PDF) of the prediction error is shown. For slice 1, the probability that the prediction error of Proposed-1 is less than 0.2 is about 49.75%, while that of Baseline-1, 2, and 3 is about 31.06%-34.09%; for slice 2, the probability that the prediction error of Proposed-1 is less than 0.1 is about 89.65%, while that of Baseline-1, 2, and 3 is about 76.26%-85.23%; for slice 3, the probability that the prediction error of Proposed-1 is less than 0.1 is about 60.38%, while that of Baseline-1, 2, and 3 is about 50.12%-55.12%. This shows that Proposed-1 can accurately capture the changes in resource requirements.

[0129] 4.2 Impact of Resource Pressure on Slice Performance Isolation

[0130] This group of experiments investigates the impact of the total number of RBs held by drones on the performance isolation of slices. The available RBs of all drones are set to the same number. Figure 6 As shown in the figure, in the initial stage, the performance isolation quality of each method began to improve rapidly as the resource pressure was relieved. Among all the compared methods, the performance isolation quality of the Proposed-1 method was the highest. Under the same learning method, the performance isolation quality of the proposed method was higher than that of the baseline method. This phenomenon shows that the model aggregation scheme based on the Transformer multi-head attention mechanism promotes the improvement of model performance. When the resources of the drone are sufficient, the performance isolation quality gradually approaches the ideal state. Compared with Baseline-1, 2, and 3, Proposed-1 has 15.39%, 10.66%, and 5.58% performance improvements in slice performance isolation quality, respectively. This is due to the accurate prediction of slice resource requirements and the reasonable setting of shared areas.

[0131] 4.3 Impact of Model Segmentation on Collaborative Training

[0132] This group of experiments selected Proposed-2 based on model segmentation and Proposed-1 without segmentation to investigate the impact of the proposed model segmentation on the accuracy of resource demand prediction in collaborative training. All local models were set with the same segmentation point (see Figure 3 ).like Figure 7(a) to Figure 7(c) As shown, with the increase of communication rounds, the prediction error of each slice after the segmentation model is almost consistent with the trend shown by the model without segmentation, indicating that the adopted model segmentation hardly sacrifices the prediction accuracy. Figure 8The training costs of the segmented model and the non-segmented model are compared, including communication cost and computational cost. Model segmentation helps reduce the amount of data transmitted and the time cost during training. Compared with FL, SFL significantly reduces the training cost without sacrificing prediction accuracy.

[0133] 5 Conclusion

[0134] Aiming at UAV RAN slicing in large-scale multi-machine collaborative training scenarios, the present invention proposes a SFL method based on clustering and Tansformer multi-head attention to improve model performance in a low-cost manner. Experimental results show that the proposed method is superior to the traditional FL method in terms of slicing performance isolation quality and training cost. In particular, the proposed model segmentation scheme significantly reduces the training cost without sacrificing prediction accuracy. The proposed method is not only suitable for large-scale UAV scenarios, but can also be flexibly extended to applications such as target detection and crowd intelligence perception on UAV platforms.

[0135] References

[0136] [1]Puligheddu C, Ashdown J, Chiasserini CF, et al. SEM-O-RAN: Semantic and flexible O-RAN slicing for NextG edge-assisted mobile systems[C] / / IEEEInfocom 2023-IEEEConference on Computer Communications.IEEE,2023:1-10.

[0137] [2] Yuan Z, Muntean G M. AirSlice: A network slicing framework for UAVcommunications[J]. IEEE Communications Magazine, 2020, 58(11): 62-68.

[0138] [3]Basu D,Ghosh U,Datta R.SoftFloat:softwarized 5G-driven networkslicing framework for QoS aware UAV communications[C] / / Proceedings of the 5thInternational ACM Mobicom Workshop on Drone Assisted Wireless Communicationsfor 5G and Beyond.2022:13-18.

[0139] [4]Le L,Nguyen T N,Suo K,et al.5G network slicing and drone-assistedapplications:A deep reinforcement learning approach[C] / / Proceedings of the5th International ACM Mobicom Workshop on Drone Assisted WirelessCommunications for 5G and Beyond.2022:109-114.

[0140] [5]Coelho A,Fontes H,Campos R,et al.Placement and allocation ofcommunications resources in slicing-aware flying networks[C] / / 2022 17thWireless On-Demand Network Systems and Services Conference(WONS).IEEE,2022:1-8.

[0141] [6]Qu Y,Dai H,Zhuang Y,et al.Decentralized federated learning for UAVnetworks:Architecture,challenges,and opportunities[J].IEEE Network,2021,35(6):156-162.

[0142] [7]Thapa C,Arachchige P C M,Camtepe S,et al.Splitfed:When federatedlearning meets split learning[C] / / Proceedings of the AAAI Conference onArtificial Intelligence.2022,36(8):8485-8493.

[0143] [8]Lu Y,Liu Z,Huang Y.Parameters compressed mechanism in federatedlearning for edge computing[C] / / 2021 8th IEEE International Conference onCyber Security and Cloud Computing(CSCloud) / 2021 7th IEEE InternationalConference on Edge Computing and Scalable Cloud(EdgeCom).IEEE,2021:161-166.

[0144] [9]Wang H,Zhao H,Li B.Bridging multi-task learning and meta-learning:Towards efficient training and effective adaptation[C] / / Internationalconference on machine learning.PMLR,2021:10991-11002.

[0145]

[10] Lu Y,Huang X,Zhang K,et al.Blockchain empowered asynchronousfederated learning for secure data sharing in internet of vehicles[J].IEEETransactions on Vehicular Technology,2020,69(4):4298-4311.

[0146]

[11] Tanmoy O B,Al Mamun M,Hasan S,et al.Enhancing Federated Learningwith Globally Shared Model:A Modified FedAVG Approach(GSM-FedAVG)[C] / / 20236th International Conference on Electrical Information and CommunicationTechnology(EICT).IEEE,2023:1-6.

[0147]

[12] Zhao Y.Comparison of Federated Learning Algorithms for ImageClassification[C] / / 20232nd International Conference on Data Analytics,Computing and Artificial Intelligence(ICDACAI).IEEE,2023:613-615.

[0148]

[13] Ma Z,Gong H,Wang X.Cooperative Optimal Control of UAVs Based onBrain-Attention Mechanism Neural Network[C] / / 2023 9th InternationalConference on Automation,Robotics and Applications(ICARA).IEEE,2023:263-267.

[0149]

[14] Lin X,Li J,Baldemair R,et al.5G new radio:Unveiling theessentials of the next generation wireless access technology[J].IEEECommunications Standards Magazine,2019,3(3):30-37.

[0150]

[15] Cho J W,Yang P,Quek T Q S,et al.Service-aware resource allocationdesign of uav ran slicing[C] / / 2020 International Conference on Informationand Communication Technology Convergence(ICTC).IEEE,2020:801-805.

[0151]

[16] He M,Wu H,Zhou C,et al.Resource Slicing with Cross-CellCoordination in Satellite-Terrestrial Integrated Networks[C] / / ICC 2024-IEEEInternational Conference on Communications.IEEE,2024:2501-2506.

[0152]

[17] Kumar H H,Karthik V R,Nair M K.Federated k-means clustering:Anovel edge ai based approach for privacy preservation[C] / / 2020 IEEEInternational Conference on Cloud Computing in Emerging Markets(CCEM).IEEE,2020:52-56.

[0153]

[18] Xu P,Wang C,Ye J,et al.State-of-charge estimation and healthprognosis for lithium-ion batteries based on temperature-compensated Bi-LSTMnetwork and integrated attention mechanism[J].IEEE Transactions on IndustrialElectronics,2023.

[0154]

[19] Liao Y,Xu Y,Xu H,et al.Mergesfl:Split federated learning withfeature merging and batch size regulation[C] / / 2024 IEEE 40th InternationalConference on Data Engineering(ICDE).IEEE,2024:2054-2067.

[0155]

[20] Ai S,Song J,Cai G.Sequence-to-sequence remaining useful lifeprediction of the highly maneuverable unmanned aerial vehicle:A multilevelfusion transformer network solution[J].Mathematics,2022,10(10):1733.

[0156]

[21] Hebert L,Golab L,Poupart P,et al.Fedformer:Contextual federationwith attention in reinforcement learning[J].arXiv preprint arXiv:2205.13697,2022.

[0157]

[22] Li Z,Lin T,Shang X,et al.Revisiting weighted aggregation infederated learning with neural networks[C] / / International Conference onMachine Learning.PMLR,2023:19767-19788.

[0158]

[23] Collins L,Hassani H,Mokhtari A,et al.Fedavg with fine tuning:Local updates lead to representation learning[J].Advances in NeuralInformation Processing Systems,2022,35:10572-10586.

[0159]

[24] Mammen P M.Federated learning:Opportunities and challenges[J].arXiv preprint arXiv:2101.05428,2021.

[0160]

[25] Thammawichai M,Baliyarasimhuni S P,Kerrigan E C,et al.Optimizingcommunication and computation for multi-UAV information gatheringapplications[J].IEEE Transactions on Aerospace and Electronic Systems,2017,54(2):601-615.

[0161]

[26] Tian Y,Wan Y,Lyu L,et al.FedBERT:When federated learning meetspre-training[J].ACM Transactions on Intelligent Systems and Technology(TIST),2022,13(4):1-26.

[0162]

[27] Barlacchi G,De Nadai M,Larcher R,et al.A multi-source dataset ofurban life in the city of Milan and the Province of Trentino[J].Scientificdata,2015,2(1):1-15.

Claims

1. A UAV small base station RAN slicing method based on segmentation federated learning, characterized in that In the service-oriented RAN slicing framework, flexible indicators of resource sharing are used to measure the quality of slice performance isolation; Step (i) the isolation quality maximization problem is modeled as a joint optimization problem of shared resource reservation and resource slicing; Step (ii) uses the clustering-based SFL method to solve the joint optimization problem; In the service-oriented RAN slicing framework, physical resources are pooled and divided into multiple slices and a shared area. Based on resource demand, each drone small base station dynamically determines the allocation of shared areas and slice resources based on the slice resource demand and shared area size predicted by the airborne model. For drone k, time is dynamically divided into multiple slice windows, each of which contains multiple discrete time slots t; the set of slice windows is expressed as The time slot set and window length of slice window a are expressed as and T (a) ; Assume that the total number of resource blocks RB held by drone k is B k , which are virtualized as M k slices, and the slice set is recorded as At the beginning of each slice window, the RBs of each slice and shared area are reallocated and maintained until the end of the slice window; At the end of each slice window, the UAV k onboard model predicts the resource demand and shared area size of each slice in the next slice window; the arriving tasks are scheduled to different slices according to their attributes; each slice allocates resources to tasks based on the time slot; At the beginning of slice window a, the number of RBs allocated to slice m and shared area on drone k are expressed as and The sum of the two does not exceed B k ,satisfy When slice resources are insufficient, the RBs in the shared area are temporarily occupied. When the resource pressure in the slice window a is high or low, the shared area is expanded or reduced accordingly to buffer the fluctuation of resource demand; In the step (a), it is assumed that in the time slot The number of RBs required for slice m on drone k is The number of RBs required for slice m on drone k in time slot t is The number of occupied RBs in the shared area is The performance isolation quality at this moment is defined as According to formula (2), and The smaller the gap, The higher the prediction accuracy, the higher the performance isolation quality; like Slice m does not occupy RBs in the shared area, which meets the current service requirements, and the isolation quality reaches the highest level. like Slice m occupies all of the shared area Even if RBs are not available, service requirements cannot be met and performance isolation is destroyed. like Slice m temporarily occupies RBs in the shared area to meet current service requirements, and performance isolation is weakened. Assume that in slice window a, the slice resource allocation vector on drone k is Based on formula (2), the average isolation quality of the slices on drone k is calculated as Apply it to For all windows in the long-term slice performance isolation optimization problem, the problem is modeled as question The essence of,is to allocate RBs in each slice and shared area to maximize,performance isolation quality; Assuming that in slice window a, the number of RBs required for each slice on drone k is predicted to be represents the predicted number of RBs of slice m on drone k in window a; the number of RBs allocated to slice m on drone k is calculated as Substituting formula (3) into formula (1), we get This completes the problem The solution of It is the starting point for problem solving; The steps of step (ii) include: 2.1) Opportunistic federated clustering: When a large number of drones hovering in the area under its jurisdiction are observed, the MEC controller starts clustering based on weighted k-means. Clustering is started at an opportune moment. The MEC controller broadcasts notifications to each drone and asks whether to participate in this round of clustering. The ground base station serves as the cluster head, and the drones serve as members within the cluster. Using the weighted k-means clustering algorithm, the set and number of base stations randomly selected as the initial centroid are and I k ,The weighted Euclidean distances of other UAVs from these centroids are classified into the category with the closest weighted Euclidean distance; 2.2) Model segmentation and local iteration In drone The bidirectional long short-term memory network Bi-LSTM model is deployed as a prediction model to predict resource requirements Assume that in the time slot The resource prediction value of the forward LSTM is The resource prediction value of the reverse LSTM is Then, the resource prediction value of the Bi-LSTM model is expressed as in, Represents the combination of two hidden states; Assume that the prediction model carried by drone k is represented as w i,k , then the number of RBs required for slice m in window a on drone k is predicted to be based on and Loss Function It is used to measure the prediction error of the number of RBs required for slice m on drone k; i,k The loss function is expressed as Prediction Model Segmentation: w i,k It is divided into an aerial model and a ground model, denoted as w i,k,b ′ and w i,k,p ′; w i,k,b ′ is kept locally, w i,k,p ′ is transferred to the base station i as the cluster head; the gradient of the aerial model and the ground model is back-propagated through wireless transmission; let and Represents w i,k,b ′ and w i,k,p ′’s loss function, and The accumulation of Local iteration: The data carried by drone k and the newly collected data are used to analyze w i,k,b ′Forward propagation, the intermediate features are passed to base station i; base station i performs forward and backward propagation to update w i,k,p ′ and returns the gradient; after receiving the gradient, through back propagation, the underlying model of drone k is updated to Among them, μ k is the learning rate of drone k; 2.3) Transformer-enhanced segmentation model aggregation Represents the set of drones in cluster i, by traversing The base station i receives the w of the drones in the cluster. i,k,b ;w i,k,b The encoder TE of the Transformer integrated on the ground base station is encoded as w i,k,b , and then aggregated into W through the multi-head attention mechanism i,b , and finally decoded into W by the Transformer decoder TD integrated in the ground base station i,b ; The global underlying model parameters after base station aggregation in are expressed as W b , it was distributed to each drone for a new round of training.

2. According to the method for slicing UAV small base station RAN based on segmented federated learning in claim 1, it is characterized in that In the service-oriented RAN slicing framework, the size of the shared area is set to no more than B k 15% of.

3. The method for slicing UAV small base station RAN based on segmented federated learning according to claim 1 is characterized in that In step 2.1), the weighted coefficient of the drone position is incorporated into the Euclidean distance calculation of the k-means algorithm; the positions of the drone k and the base station i are represented as g k and g i , the discrete degree of the position is used to determine the weighting coefficient; the weight of the drone position is quantified in the form of standard deviation, that is, in, is the average value of the drone positions, K represents the number of drones; The weighted Euclidean distance between drone k and base station i as the centroid is calculated as By minimizing The MEC controller determines the connection between the drone and the base station.

4. The method for slicing a UAV small base station RAN based on segmented federated learning according to claim 3 is characterized in that In step 2.3), we receive w i,k,b After that, base station i starts federated aggregation to update the global model. During this process, w i,k,b It is encoded by TE into an input sequence, which is stacked in multiple layers to generate a vector Q about Query, Key and Value. k , K k 、V k ; In federated aggregation, Transformer’s multi-head attention mechanism handles W i,b (Q k ) and w i,k,b (K k ); UAVs within a cluster and base station The attention weights of the underlying model parameters are expressed as and in After updating the local model, Model of drone w i,k,b are uploaded to base station i, and these models are aggregated by base station i into Next, TD decodes and outputs the aggregated underlying model parameter vector; base station i converts W b Send it to the drones in the cluster.

Images