Source-domain-independent MDTL-driven unmanned aerial vehicle RAN slicing method

Abstract

The invention provides a source domain-independent MDTL-driven unmanned aerial vehicle RAN slicing method, and aims at improving the slicing performance isolation effect and improving the model decision-making efficiency. Firstly, a window adaptive resource slicing framework giving consideration to performance isolation and decision efficiency is developed. Under the framework, an isolation quality maximization problem in a single-machine autonomous mobile scene is modeled as a joint optimization problem of a slice window and resource configuration. Then, designing a resource sharing perception adjusting strategy, and dynamically determining the division of the slice window; and finally, in order to support resource allocation in dynamic deployment of the unmanned aerial vehicle, a knowledge circulation mode of'movement-target domain selection-fine adjustment 'is provided, and a source domain irrelevant MDTL method is designed. Transform and a multi-head attention mechanism thereof are integrated on each unmanned aerial vehicle, so that generalization of feature merging is improved.

Description

Technical Field

[0001] The present invention is an application of artificial intelligence technology in the field of communication technology, specifically a UAV RAN slicing method driven by source domain-independent multi-target domain transfer learning. Background Art

[0002] With the rapid development of post-5G and 6G networks, drone small base stations, as a new type of network infrastructure, have become an important part of modern communication systems. Drone small base stations can not only quickly fill network gaps through flexible deployment, but also provide customized services for applications such as unmanned driving, Internet of Things (IoT), and smart cities, further expanding the functions of traditional ground base stations.

[0003] In this context, Radio Access Networks (RAN) slicing technology, which has the ability to customize the allocation of virtual network resources, has emerged. The RAN slicing technology in traditional 5G networks divides physical network resources into independent logical networks to provide differentiated services for different types of business needs. [1] With the widespread deployment of drone small base stations, the application of RAN slicing technology has been extended to air networks. Due to the dynamic changes in flight trajectory, coverage, and service requirements, as well as the limitations of drones in battery life, storage capacity, and computing power, RAN slicing technology takes into account both lightweight and adaptability. Existing work includes classifying drone applications according to QoS requirements. [2] , build a logical drone network for specific QoS requirements [3] , Evaluate the priority of slices based on the order in which resources are used [4] wait.

[0004] Transfer Learning (TL) [5] It provides an effective method for adaptive resource management of distributed UAV networks. TL helps a single UAV quickly adapt to new environments through cross-scenario or cross-network knowledge transfer, avoids frequent model retraining, and saves energy and time costs. However, the combination of TL and UAV systems still faces many challenges, including:

[0005] 1) Knowledge accumulation and continuous optimization of onboard models. In dynamic deployment, onboard models need to be frequently trained / fine-tuned to adapt to new scenarios and tasks. This not only consumes limited storage and computing resources, but also blocks the recycling of knowledge. Traditional TL methods lack an efficient knowledge transfer mechanism, which results in the inability of existing knowledge to be used for subsequent model optimization, limiting the sustainable optimization capabilities of the system in long-term dynamic deployment. To solve this problem, it is urgent to build a flexible knowledge circulation mechanism to achieve continuous optimization of the model and efficient use of resources. Yu et al. proposed a data-free knowledge distillation method (Data-Free Knowledge Distillation, DFKD) to ensure that model knowledge can be continuously updated and accumulated. [6] Li et al. proposed a UAV target tracking and path planning framework based on Mixed Noise Deep Deterministic Policy Gradient (MN-DDPG) and TL. By accumulating existing knowledge to optimize the performance of future tasks, it achieves continuous accumulation and optimization of knowledge, which helps to improve long-term performance. [7] .

[0006] 2) Dual optimization of isolation quality and online decision efficiency. The efficiency of online decision making is closely related to the length of the slice window. A shorter slice window can speed up resource updates, but it increases control overhead. While extending the slice window can reduce control overhead, the performance isolation of the slice may be destroyed in task-intensive scenarios. Most existing slice window divisions use a static mode. [8] , which cannot be applied to resource-constrained and highly dynamic drone networks. How to optimize decision-making efficiency while ensuring isolation quality is a key issue that needs to be solved in drone network resource management. Summary of the invention

[0007] In response to the above challenges and limitations, this paper proposes a source-free multi-target domain transfer learning (MDTL) driven UAV small base station resource slicing method. The goal is to improve the slicing performance isolation effect and reduce the decision cost. This method supports single-machine autonomous mobility scenarios, and its main contributions include:

[0008] First, a window-adaptive resource slicing framework is designed to balance performance isolation and decision efficiency. In this framework, a shared resource-aware slice window adaptation strategy is used to measure the quality of slice performance isolation. The occupancy of shared resources is used to determine the division of slice windows. The isolation quality maximization problem is modeled as a joint optimization problem of slice windows and resource configuration.

[0009] Second, focusing on the knowledge cycle mode of "mobile-target domain selection-fine-tuning", a source domain-independent MDTL method for a single machine is proposed. Unconstrained by the movement trajectory and range, the drone carries the model of the previous mission area (source domain) to the next mission area (target domain), and quickly fine-tunes the model with the help of multiple target domain ground base stations. The Transformer integrated on the drone is used to optimize feature merging to improve the generalization of onboard model fine-tuning.

[0010] Simulation results show that the proposed method is superior to typical TL methods in terms of model accuracy prediction and performance isolation effect. Based on the same performance isolation effect, the decision cost of the proposed method is also lower than that of typical TL methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 Dynamic RAN slicing framework with representation window adaptation;

[0012] Figure 2 It represents the RAN structure of the UAV small base station L;

[0013] Figure 3 Represents the MDTL architecture;

[0014] Figure 4(a) to Figure 4(c) They represent the changes of prediction errors of three slices, slice 1 to slice 3, with the communication rounds;

[0015] Figure 5(a) to Figure 5(c) Represent the probability density functions of the prediction errors of the three slices from slice 1 to slice 3 respectively;

[0016] Figure 6 It represents the impact of the total number of RBs in the static window on the performance isolation;

[0017] Figure 7(a) to Figure 7(c) They respectively represent the impact of different window partitioning strategies on the slice performance isolation effect, where:

[0018] Figure 7(a) shows the task traffic fluctuation of the static window partitioning strategy;

[0019] Figure 7(b) shows the task traffic fluctuation of the dynamic window partitioning strategy;

[0020] Figure 7(c) shows the average isolation quality of different slice window partitioning strategies;

[0021] Figure 8 Represents the impact of different slicing window strategies on decision efficiency. DETAILED DESCRIPTION

[0022] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0023] The first section summarizes the technical solution of the present invention.

[0024] Section 2 introduces the RAN slicing framework, performance isolation metrics, and models the optimization problem.

[0025] Section 3 introduces a shared resource-aware slice window adaptation strategy to decide the partitioning of slice windows.

[0026] Section 4 introduces the construction of a source domain-independent MDTL framework to support problem solving in single-machine mobility.

[0027] Section V presents the experimental design and results analysis for performance evaluation.

[0028] Section VI concludes the study and discusses future prospects.

[0029] 1 Overview

[0030] This paper proposes a source-free multi-target domain transfer learning (MDTL) driven UAV radio access network (RAN) slicing method. The purpose is to improve the slicing performance isolation effect and improve the efficiency of model decision-making.

[0031] Firstly, a window-adaptive resource slicing framework that takes into account both performance isolation and decision efficiency is developed. In this framework, the isolation quality maximization problem in a single-machine autonomous mobility scenario is modeled as a joint optimization problem of slicing windows and resource allocation.

[0032] Then, a resource sharing-aware regulation strategy is designed to dynamically determine the partitioning of slice windows.

[0033] Finally, in order to support resource allocation in the dynamic deployment of UAVs, a knowledge cycle model of "mobile-target domain selection-fine-tuning" is proposed, and a source domain-independent multi-target domain transfer learning MDTL method is designed. Transformer and its multi-head attention mechanism are integrated on each UAV to improve the generalization of feature merging.

[0034] Simulation results show that the proposed scheme outperforms the benchmark TL method in terms of slice performance isolation and decision efficiency.

[0035] 2 System Model

[0036] As an aerial small base station [9] , drones can be deployed over a large area according to the mission and are not restricted by region. The drones and ground base stations communicate via 5G NR

[10] (New Radio wireless access standard) protocol to complete the handshake and inquiry. When a drone needs to leave the area to perform other tasks, it will issue a notice 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] Consider a window-adaptive resource slicing method. Physical resources are pooled and divided into multiple slices and a resource sharing area. According to resource requirements, each drone dynamically determines the slice window length and resource allocation. 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 The time slot set and window length of slice window a are expressed as and

[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 are reallocated and remain so until the end of the window. At the end of each window, UAV 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 the shared area on UAV k is expressed as and The sum of the two cannot exceed the total number of RBs held by the drone, satisfying

[0042]

[0043] When slice resources are insufficient, RBs in the shared area can be temporarily occupied.

[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.

[11]

[12] 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 occupied RBs in the shared area is

[0047]

[0048] The performance isolation quality at this moment is defined as

[0049]

[0050] From formula (2) and (3), 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.

[0051] 2.3 Problem Modeling and Solution Framework

[0052] Assume that under the slice window a, the slice resource allocation vector on drone k is Based on formula (3), 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

[0053]

[0054] question The essence is to allocate each slice And determine the slice window length To maximize performance isolation quality. It is the key to problem solving.

[0055] Next, a resource sharing zone-aware slice window adaptation strategy is proposed to solve

[0056] The setting of resource sharing area helps to alleviate the slice performance violation caused by the sharp fluctuation of short-term resource demand. When the resource pressure in slice window a is high (low), the occupancy of shared RB increases (decreases) to buffer the fluctuation of resource demand. Therefore, the occupancy of shared RB can be used as a reference for slice window adjustment. The size of the shared area cannot exceed B k Assume that the step size of the slice window adjustment is denoted as ΔT k , ΔT k The setting cannot exceed Given and The duration of the slice window a is adaptively adjusted to

[0057]

[0058] This enables resource allocation to respond to fluctuations in resource demand in a timely manner. When the shared RB occupancy increases, the slice window interval becomes shorter to speed up slice resource updates. When the shared RB occupancy decreases, the slice window interval is expanded to reduce decision costs.

[0059] Assuming that in slice 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

[0060]

[0061] Substituting equation (5) into equation (2), we can get Will Substituting into formula (4), we can obtain This completes the problem It can be seen that is the starting point for solving the problem. The following Section 3 proposes the MDTL method for the single-machine mobile scenario to solve

[0062] 3 MDTL-based solutions

[0063] When a drone is performing a mission, each time it arrives at a mission area, the model it carries usually needs to be retrained to adapt to the new environment. This process not only makes knowledge unrecyclable, but also increases communication and computing costs.

[13] A multi-objective domain transfer learning (MDTL) method is proposed to achieve sustainable optimization of models on the move.

[0064] After a drone completes its mission in the previous mission area (source domain), it carries the source domain model to the next mission area. After arriving in a new area, it may be covered by multiple ground base stations (target domain)

[14] .by Figure 2 The proposed solution is explained as an example. Since it is covered by both base stations 1 and 2, the source domain model carried by UAV 3 will be delivered to both base stations. In TL, the historical data of base stations 1 and 2 are used to fine-tune the source domain model carried by UAV 3.

[0065] In the proposed MDTL framework, the pre-trained source domain model (not the source domain data) is used for TL. Figure 2 For example, the main steps include:

[0066] ① Multi-target domain data selection: The source domain model carried by UAV 3 is transmitted to the connected ground base station. The MEC controller selects a batch of target domain base stations for the UAV based on the similarity of data distribution. The selected base stations select fine-tuning data based on a certain ratio value.

[0067] ② Multi-target domain knowledge transfer: Based on the selected data, base stations 1 and 2 assist UAV 3 to complete MDTL to fine-tune the source domain model.

[0068] ③Multi-target domain feature merging: Transformer encoder (TE) and decoder (TD)

[15] It is integrated on UAV 3. After the fine-tuned models are transmitted back to UAV 3, TE and TD merge the features of the models to update the source domain model.

[0069] 3.1 Multi-target domain data selection

[0070] The data stored on the adjacent base station (target domain data) is used to fine-tune the UAV model. In this section, a multi-target domain data selection strategy is designed to alleviate the negative impact of the differences in different target domain data on fine-tuning.

[0071] In MDTL, the source domain and target domain i correspond to drone k and base station i respectively. Assume x i,j represents the jth input sequence data of target domain i, y i,j is the jth label, N i is the number of samples in the target domain i. Assume that the source domain model is o k ′, the model of the j-th input sequence data of target domain i is f i (x i,j ;o k ′). The loss function of the source domain and target domain models is expressed as

[0072]

[0073] in, is the loss function for a single sample.

[0074] KL divergence

[16] The criterion for selecting the target domain (i.e., the base station involved in model fine-tuning) is as follows. The smaller (larger) its value means that the distribution of historical data provided by the drone and the base station is more similar (the greater the difference). Based on this criterion, similar data is used for model fine-tuning as much as possible to reduce the risk of overfitting. Assume that the data distribution of drone k and base station i is as follows P k and P i

[0075]

[0076]

[0077] In formula (7), h k,j and h i,j is the hidden representation of the model of drone k and base station i. The KL divergence of base station i connected to drone k is calculated as

[0078]

[0079] Among them, H(P k ) is P k Information entropy of H(P i ,P k ) is P i and P k The cross entropy of Represents the set of all possible distributions.

[0080] The set and cardinality of ground base stations covering UAV k are expressed as and I k ′. The target domain base station set selected to participate in model fine-tuning is determined as

[0081]

[0082] Where θ is the KL divergence threshold. Select and P k Similar historical data fine-tuning k ′, the proportion of this part of data is

[0083]

[0084] in, It is negatively correlated with the amount of selected data.

[0085] 3.2 Knowledge Transfer from Multiple Target Domains

[0086] Bidirectional Long Short-Term Memory (Bi-LSTM)

[17] As a lightweight model, it is deployed on each UAV to predict resource requirements. 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

[0087]

[0088] in, Represents the combination of two hidden states.

[0089] The number of resource blocks required for slice m in window a on drone k is predicted to be

[0090]

[0091] Loss Function It is used to measure the prediction error of the number of RBs required for slice m on drone k. The aggregate loss function of slices on drone k is as follows

[0092]

[0093] The execution process of MDTL is as follows Figure 3 Using the gradient descent method, the local model parameters of drone k in window a are updated as

[0094]

[0095] Among them, α is the learning rate, Represents the loss function relative to o k Based on equations (8) and (9), the goal of fine-tuning base station i is to minimize the following loss function

[0096]

[0097] In formula (15), β is a regulation factor used to control the influence of KL divergence on weight. k In the target domain i is fine-tuned as

[0098]

[0099] Among them, η i is the learning rate of base station i, Represents the gradient of the loss function. For base station i, the proportion is ε iThe local data of w is used to fine-tune the model. k,i Will be used for the target task of base station i to improve o k Adaptability in target domain i. After that, The base station in delivers the fine-tuned model to drone k.

[0100] 3.3 Transformer-enhanced multi-target domain feature merging

[0101] After receiving multiple target domain TL fine-tuned models, the Transformer architecture is used for feature merging to improve the adaptability and generalization of the model. Relying on TE and TD, the features of multiple target domain models are dynamically weighed and integrated by the Transformer multi-head attention mechanism to optimize feature representation. TE is responsible for encoding the fine-tuned local model parameters into vector form, and TD decodes and outputs the encoded information after multi-head attention aggregation.

[0102] like Figure 3 As shown, w k,i 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. i , K i and V i

[18] During the fine-tuning process, w k and w k,i is taken as the query Q i and keyword K i The input is fed into the Transformer, where the multi-head attention mechanism calculates the importance of each position in the input sequence and is used to aggregate the encoded vector. The attention weight is calculated as

[0103]

[0104] Where d represents the dimension of the vector. According to formula (17), The ground base station w k,i The attention weight is expressed as Substituting equation (17) into equation (18), the model of the target domain base station is aggregated as

[0105]

[0106] After aggregation, TD is used to decode w after drone k is processed by Transformer. k And output w k The implementation details of feature merging are summarized in Algorithm 1. The algorithm traverses The ground base station in the target domain i is fine-tuned wk,i First encoded by TE as w k,i (see row 2), and then aggregated into w by the multi-head attention mechanism k (see lines 3-5), which is finally decoded by TD as (See row 6), thus achieving joint fine-tuning in multiple target domains.

[0107] Algorithm 1: Multi-target domain feature merging

[0108] Input: w k,i

[0109] Output: w k

[0110]

[0111] 4 Experimental design and results analysis

[0112] 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.

[0113] On the drone side, a Bi-LSTM model consisting of two bidirectional LSTM layers (each with 128 hidden neurons) was deployed, and then the features were mapped to prediction results through a linear layer. Transformer encoders and decoders were deployed on each drone and ground base station for feature merging and parameter aggregation. In order to simulate the collaboration between nodes, the data and models of each drone 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 drones and ground base stations, simulate a large-scale, distributed multi-drone 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 divided into the following two types:

[0114] 1) Proposed-1: Transformer-enhanced MDTL for static windows.

[0115] 2) Proposed-2: Transformer-enhanced MDTL for dynamic windows.

[0116] In order to make a comprehensive and objective comparison, two TL-based benchmark methods were selected, including:

[0117] 1) Baseline-1: Single target domain TL for static windows

[19] .

[0118] 2) Baseline-2: MDTL for static windows.

[0119] The communication data from Trentino, Italy, was selected, including data services, voice calls and other services.

[20] Different types of business data are used to 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 2.

[0120] Table 2 Experimental parameter settings

[0121]

[0122]

[0123] 4.1 Analysis of resource demand forecast accuracy

[0124] 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, the prediction error of the RB required for each slice is shown as the number of communication rounds changes. As the number of communication rounds increases, the prediction error of each slice shows a downward trend and converges to an interval. The prediction error of the proposed method 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 and 2, Proposed-1 can more effectively integrate multi-target domain data, improve model accuracy and generalization, and achieve performance gains of 10.31%-25.54%, 29.32%-55.18%, and 7.51%-27.46% on three slices, respectively.

[0125] Figure 5(a) to Figure 5(c)The probability density function (PDF) of the prediction error of the three slices is shown. For slice 1, the probability that the prediction error of Proposed-1 is less than 0.2 is about 72.6%, while Baseline-1 and 2 are only 30.05% and 30.68%; for slice 2, the probability that the prediction error of Proposed-1 is less than 0.1 is about 92.3%; for slice 3, the probability that the prediction error of Proposed-1 is less than 0.1 is about 69.88%, while Baseline-1 and 2 are only 33% and 46.12%. This shows that Proposed-1 can accurately capture the changes in resource requirements.

[0126] 4.2 Impact of Resource Pressure on Slice Performance Isolation

[0127] This group of experiments investigates the impact of the total number of RBs held by drones in a static window 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 them, the performance isolation quality of the Proposed-1 method was higher than that of other methods. Among all the compared methods, the performance isolation quality of the proposed method was the highest. 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 and 2, Proposed-1 has 18.55% and 9.62% performance improvements in slice performance isolation quality, respectively. This is due to the accurate prediction of slice resource requirements.

[0128] 4.3 Impact of Slice Window Strategy on Performance

[0129] This set of experiments investigates the impact of different slice window division strategies on performance isolation. We intercepted and visualized the task arrival data of a cluster (containing 4 drones) in a period of time in the experimental scenario. The static window strategy divides the timeline evenly into multiple slice windows (see Figure 7(a)). In contrast, the dynamic window strategy divides the timeline unevenly into multiple slice windows (see Figure 7(b)). As can be seen from Figure 7(c), in the selected period, the average isolation quality of Proposed-2 using the dynamic strategy is better than that of Proposed-1 using the static strategy, which proves the flexibility, adaptability and effectiveness of the dynamic strategy in dealing with task fluctuations.

[0130] 4.4 Online decision-making efficiency analysis

[0131] We observe the impact of different window partitioning strategies (static or dynamic windows) on the decision efficiency of collaborative training over a long time scale. Decision efficiency refers to the number of times resource reallocation is triggered within a given period of time. The number here is equivalent to the number of slice windows, which is determined by the window control strategy. Figure 8 As shown in the figure, under the same performance isolation effect, the long-term decision cost of the dynamic window is lower than that of the static window. Although the decision cost of the dynamic window may be higher than that of the static window in a specific period of time (such as when the task traffic in the coverage area fluctuates greatly), from the perspective of long-term optimization (such as within a day), the decision cost of the proposed dynamic window division strategy is lower than that of the static window. Compared with Proposed-1 using a static window, Proposed-2 has a performance improvement of 9.33%-15.83% in decision cost. Combined with the results of Figure 7, it can be seen that the dynamic window strategy can enhance the flexibility of RAN slices and balance control overhead and service provision.

[0132] 5 Conclusion

[0133] For drone RAN slicing in a single-machine autonomous mobility scenario, a source-domain-independent MDTL method is designed, which follows the knowledge cycle mode of "mobile-target domain selection-fine-tuning" and takes into account both performance isolation and energy consumption optimization. Experimental results show that the proposed method is superior to traditional TL methods in terms of slice performance isolation quality and decision efficiency. Based on the same performance isolation effect, the long-term decision cost of the proposed dynamic window strategy is also significantly reduced. The proposed method is not only suitable for large-scale drone scenarios, but can also be flexibly extended to applications such as target detection and crowd intelligence perception on drone platforms.

[0134] References

[0135] [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-IEEE Conference on Computer Communications.IEEE,2023:1-10.

[0136] [2]Bellone L,Galkin B,Traversi E,et al.Deep reinforcement learningfor combined coverage and resource allocation in uav-aided ran-slicing[C] / / 2023 19th International Conference on Distributed Computing in Smart Systemsand the Internet of Things(DCOSS-IoT).IEEE,2023:669-675.

[0137] [3]Yang P,Xi X,Guo K,et al.Proactive UAV network slicing for URLLCand mobile broadband service multiplexing[J].IEEE Journal on Selected Areasin Communications,2021,39(10):3225-3244.

[0138] [4]Ampririt P,Qafzezi E,Bylykbashi K,et al.A fuzzy-based scheme forslice priority assessment in 5g wireless networks[C] / / InternationalConference on Advanced Information Networking and Applications.Cham:SpringerInternational Publishing,2022:651-661.

[0139] [5]Wang M,Lin Y,Tian Q,et al.Transfer learning promotes 6G wirelesscommunications:Recent advances and future challenges[J].IEEE Transactions onReliability,2021,70(2):790-807.

[0140] [6]Yu G.Data-free knowledge distillation for privacy-preservingefficient uav networks[C] / / 2022 6th International Conference on Robotics andAutomation Sciences(ICRAS).IEEE,2022:52-56.

[0141] [7]Li B,Yang Z,Chen D,et al.Maneuvering target tracking of UAV basedon MN-DDPG and transfer learning[J].Defence Technology,2021,17(2):457-466.

[0142] [8]Li J,Shi W,Yang P,et al.Ahierarchical soft RAN slicing frameworkfor differentiated service provisioning[J].IEEE Wireless Communications,2020,27(6):90-97.

[0143] [9]Shen H,Ye Q,Zhuang W,et al.Drone-small-cell-assisted resourceslicing for 5G uplink radio access networks[J].IEEE Transactions on VehicularTechnology,2021,70(7):7071-7086.

[0144]

[10] Chang S Y,Park K,Kim J,et al.Securing UAV flying base station formobile networking:A review[J].Future Internet,2023,15(5):176.

[0145]

[11] 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.

[0146]

[12] 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.

[0147]

[13] Liu J,Zeng D,Li L,et al.Source-Free Domain Adaptation forMillimeter Wave Radar Based Human Activity Recognition[C] / / ICASSP 2024-2024IEEE International Conference on Acoustics,Speech and Signal Processing(ICASSP).IEEE,2024:7120-7124.

[0148]

[14] Lu Y,Huang H,Zeng B,et al.Multi-Source and Multi-Target DomainAdaptation Based on Dynamic Generator With Attention[J].IEEE Transactions onMultimedia,2024.

[0149]

[15] 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.

[0150]

[16] Cao Z,Zhou Y,Yang A,et al.Deep transfer learning mechanism forfine-grained cross-domain sentiment classification[J].Connection Science,2021,33(4):911-928.

[0151]

[17] 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.

[0152]

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

[0153]

[19] Zhuang F,Qi Z,Duan K,et al.Acomprehensive survey on transferlearning[J].Proceedings of the IEEE,2020,109(1):43-76.

[0154]

[20] 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 source domain-independent MDTL-driven UAV RAN slicing method, characterized by Includes steps: (i) In the window-adaptive resource slicing framework, the isolation quality maximization problem in the scenario of autonomous mobility of a single UAV small base station is modeled as a joint optimization problem of slicing window and resource allocation; In the following text, the drone small base station is referred to as "drone"; (ii) Adopting a resource sharing-aware adjustment strategy to dynamically determine the division of slice windows; (3) Adopting the knowledge cycle mode of movement-target domain selection-fine-tuning and the source domain-independent multi-target domain transfer learning method MDTL to optimize the drone-borne model; In step (a), The drone small base station and the ground base station complete the handshake and inquiry through the 5G NR protocol; when a drone needs to leave the area to perform other tasks, it will issue a notification to release the connection; The window-adaptive resource slicing framework is a service-oriented RAN slicing framework, in which physical resources are pooled and divided into multiple slices and a resource sharing area; based on resource demand, each drone dynamically determines the slice window length and resource allocation based on the slice resource demand and sharing area size predicted by the onboard model; For UAV 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 set of time slots of slice window a is expressed as The window length of the slice window a is expressed as Assume that the total number of resource blocks RB held by drone k is B k , which are virtualized as M k service slices, and the slice set is recorded as At the beginning of each 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 onboard model of UAV k predicts the resource demand of each slice and the size of the shared area 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 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, RBs in the shared area are temporarily occupied; The resource allocation strategy at the beginning of each slice window determines the performance isolation quality of each slice within the window; Assume that in the time slot The number of RBs required for slice m on drone k 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 (3), and The smaller the gap, The higher the prediction accuracy, the higher the performance isolation quality; like Slice m does not occupy That is, it meets current needs and the isolation quality reaches the highest level like All in the shared area Even if RBs are occupied, the service requirements of slice m are not 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 under the slice window a, the slice resource allocation vector on drone k is Based on formula (3), 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 is to allocate each slice And determine the slice window length Isolate quality to maximize performance; determine It is the key to problem solving; In step (2), a resource sharing zone-aware slice window adaptation strategy is proposed to solve The occupancy of the shared RB is used as a reference for slice window adjustment. The size of the shared area is not greater than B k ; Assume that the step size of the slice window adjustment is denoted as ΔT k , ΔT k The setting is not greater than Given and The duration of the slice window a is adaptively adjusted to Assuming that in slice 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 Substituting equation (5) into equation (2), we get Will Substituting into formula (4), we obtain This completes the problem The solution of It is the starting point for problem solving; In step (3), for the single-machine mobile scenario, the source domain-independent MDTL method is used to solve When a drone is performing a mission, each time it arrives at a mission area, the model it carries usually needs to be retrained to adapt to the new environment. In this case, the source domain-independent MDTL method is used to achieve sustainable optimization of the model on the move. After a drone completes its mission in the previous mission area, i.e., the source domain, it moves to the next mission area with the source domain model. After arriving at the mission area, it is likely to be covered by multiple ground base stations, i.e., the target domain. Since it is covered by multiple ground base stations at the same time, the source domain model carried by the drone will be delivered to these base stations. In transfer learning (TL), the historical data of these base stations are used to fine-tune the source domain model carried by the drone. Under the MDTL framework, the pre-trained source domain model is used for TL. The steps include: ① Multi-target domain data selection: The source domain model carried by the drone is transmitted to the connected base station; the MEC controller selects a batch of target domain base stations for the drone based on the similarity of data distribution; the selected base station selects fine-tuning data based on a certain ratio value; ②Multi-target domain knowledge transfer: Based on the selected data, these base stations assist the drones to complete MDTL to fine-tune the source domain model; ③ Multi-target domain feature merging: Transformer’s encoder TE and decoder TD are integrated on the drone; when the fine-tuned models are sent back to the drone, TE and TD merge the model features to update the source domain model.

2. The source domain-independent MDTL-driven UAV RAN slicing method according to claim 1 is characterized by: The shared area size should not exceed B k 15%; ΔT k The setting cannot exceed 10% of.

3. The source domain-independent MDTL-driven UAV RAN slicing method according to claim 1, characterized in that In step (iii), the data stored on the adjacent base station, i.e., the target domain data, is used to fine-tune the UAV-mounted model. A multi-target domain data selection strategy is adopted to alleviate the negative impact of the differences in different target domain data on fine-tuning: In MDTL, the source domain and target domain i correspond to drone k and base station i, respectively; Assume x i,j represents the jth input sequence data of target domain i, y i,j is the jth label, N i is the number of samples in the target domain i; assuming the source domain model is o k ′, the model of the j-th input sequence data of target domain i is f i (x i,j ;o k ′); the loss function of the source domain and target domain model is expressed as Among them, l(·) is the loss function of a single sample; KL divergence is used as the criterion for selecting the target domain, i.e., the base stations involved in model fine-tuning. The smaller or larger its value is, the more similar or different the distribution of historical data provided by the drone is to that provided by the base station. Based on this criterion, similar data is used for model fine-tuning as much as possible to reduce the risk of overfitting. Assume that the data distribution of drone k and base station i is as follows: k and In formula (7), h k,j and h i,j is the hidden representation of the model of drone k and base station i; the KL divergence of base station i connected to drone k is calculated as Among them, H(P k ) is P k Information entropy of H(P i ,P k ) is P i and P k The cross entropy of Represents the set of all possible distributions; The set and cardinality of base stations covering UAV k are expressed as and I k ′, the target domain base station set selected to participate in model fine-tuning is determined as Where θ is the KL divergence threshold; base station Select and P k Similar historical data fine-tuning k ′, the proportion of this part of data is in, It is negatively correlated with the amount of selected data.

4. The source domain-independent MDTL-driven UAV RAN slicing method according to claim 3 is characterized in that In step (iii), the bidirectional long short-term memory network Bi-LSTM is deployed on each drone as a 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 Bi-LSTM is expressed as in, Represents the combination of two hidden states; The number of resource blocks required for slice m in window a on drone k is predicted to be Loss Function It is used to measure the prediction error of the number of RBs required for slice m on drone k; the aggregate loss function of slices on drone k is as follows Using the gradient descent method, the local model parameters of drone k in slice window a are updated as Among them, α is the learning rate, Represents the loss function relative to o k ′; Based on equations (8) and (9), the goal of fine-tuning base station i is to minimize the following loss function In formula (15), β is a regulating factor used to control the influence of KL divergence on weight; k In the target domain i is fine-tuned as Among them, η i is the learning rate of base station i, represents the gradient of the loss function; For base station i, the proportion is ε i The local data of w is used to fine-tune the model; during the iteration process, w k,i will be used for the target task of base station i; then, The base station in delivers the fine-tuned model to drone k.

5. The method for UAV RAN slicing driven by source domain-independent multi-target domain transfer learning according to claim 4 is characterized in that In step (iii), after receiving multiple target domain TL fine-tuned models, Transformer is used for feature merging; Relying on TE and TD, the features of multiple target domain models are dynamically weighed and integrated by the Transformer multi-head attention mechanism; TE is responsible for encoding the fine-tuned local model parameters into vector form, and TD decodes and outputs the encoded information after multi-head attention aggregation; w k,i 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. i , K i and V i ; During model fine-tuning, w k and w k,i is taken as query Q i and keyword K i The input is sent to the Transformer, where the multi-head attention mechanism calculates the importance of each position in the input sequence and is used to aggregate the encoded vector; Base Station The attention weight is calculated as Where d represents the dimension of the vector; According to formula (17), The ground base station w k,i The attention weight is expressed as Substituting equation (17) into equation (18), the model of the target domain base station is aggregated as After aggregation, TD is used to decode w after drone k is processed by Transformer. k And output w k ; The execution method of feature merging is: traversal The ground base station in the target domain i is fine-tuned w k,i First encoded by TE as w k,i , and then aggregated into w by the multi-head attention mechanism k , and finally decoded by TD as w k , thus achieving joint fine-tuning in multiple target domains.

Images