Unmanned aerial vehicle small base station cluster ran slicing method based on hierarchical federated learning

Abstract

The present application is aimed at the scenario that the unmanned aerial vehicle small base station cluster provides differentiated services for ground users, and proposes a kind of unmanned aerial vehicle small base station cluster RAN slicing method based on hierarchical federated learning Based on the dynamic RAN slicing framework of hierarchical federated learning, to improve the slicing performance isolation effect and reduce the communication cost of collaborative training process. First, considering the dynamic deployment of unmanned aerial vehicles and the characteristics of insufficient data, data augmentation is used to promote local model training. Then, according to the position and data distribution information, a clustering strategy for unmanned aerial vehicle cluster is designed to support edge model aggregation, so that the member unmanned aerial vehicles of the edge have more opportunities to participate in federated learning and reduce the communication cost. Finally, the edge and global model aggregation scheme based on attention mechanism is explored to enhance the generalization ability of the global federated learning model. Simulation results show that compared with typical distributed machine learning methods, the present application can improve the slicing performance isolation effect and reduce the communication cost of unmanned aerial vehicle collaborative training.

Description

Technical Field

[0001] This invention belongs to the field of computer communication technology, specifically a method for RAN slicing of unmanned aerial vehicle (UAV) small base station clusters based on hierarchical federated learning. Background Technology

[0002] The fifth-generation mobile communication system (5G) supports diverse application scenarios, enabling billions of Internet of Things (IoT) devices to connect to the network. [1] This places higher demands on resource provision and Quality of Service (QoS) assurance. While terrestrial 5G radio access networks (RANs) can generally meet the service needs of most applications, they become inadequate in situations lacking infrastructure or during temporary hotspots such as holiday gatherings or major sporting events. In these situations, drones equipped with wireless transceivers and edge computing modules form aerial small cells. [2] This can expand network coverage and alleviate resource allocation pressure, meeting the service needs of terminal devices in sudden or temporary situations. Due to the limitations of a single drone, future drone small base stations will be deployed in clusters to provide services to ground users. [3] This approach has become an important part of post-5G and 6G RAN.

[0003] RAN slicing is one of the key enabling technologies in 5G networks and scalable network architectures, and can provide customized services based on customer needs. [4] RAN slicing divides a shared physical wireless network into multiple isolated logical networks, allowing for dynamic and elastic allocation of network resources to users. [5] In the development of communication networks, a natural step is to "extend" RAN slices to UAV small base stations, enabling them to support diverse services. Machine learning methods can express resource demand data at a higher level and in a more abstract way, becoming an important tool for solving complex decision-making problems in RAN slicing. In terms of slice resource allocation, machine learning methods can accurately predict users' slice resource needs based on changes in network conditions. [6] It can also achieve resource allocation across multiple time scales. [7] and adaptive resource scheduling [8] However, most existing RAN slicing solutions are designed for terrestrial networks.

[0004] Unlike terrestrial base stations, UAV small base stations have limited computing and storage capabilities, making it difficult to train complex models over extended periods. Furthermore, due to their dynamic deployment characteristics, UAV small base stations struggle to collect sufficient data within a limited timeframe, hindering the improvement of local model quality. In recent years, researchers have explored UAV-assisted RAN slicing, including classifying UAV applications based on QoS requirements. [9] Customized logical drone networks for specific QoS requirements

[10] Prioritize slices based on the order in which resources are used.

[11] Optimize the spectral efficiency of slices

[12] This provides differentiated QoS support for ground users. However, how to design an effective RAN slicing architecture for UAV small cell clusters requires further research.

[0005] As a typical distributed collaborative training framework, Federated Learning (FL) supports cross-domain model aggregation. The FL framework opens up new avenues for improving the overall model quality of UAV small base station clusters. Intuitively, after each UAV small base station completes its local model training, it only needs to upload its model parameters to the central controller for global model aggregation, which can significantly improve model quality.

[13] In fact, the efficient integration of federated learning and drone swarms still faces many challenges, mainly including the following three issues:

[0006] (1) The FL model parameters need to be exchanged between the UAV and the central controller, resulting in high communication costs.

[14] A key issue to consider is how to reduce the communication cost of UAV swarms while ensuring training effectiveness. Existing methods mainly focus on reducing data transmission (e.g., using gradient compression).

[15] ) and client scheduling (e.g., selecting a subset of devices to participate in each training round)

[16]

[17] This can be addressed from various angles, such as... The former usually incurs some gradient loss, leading to a decrease in accuracy. Although researchers have proposed some methods (such as adaptive threshold gradient compression algorithms)...

[18] This avoids a decrease in accuracy, but is more complex. The latter requires more training rounds.

[19] This will reduce the model convergence speed and makes it difficult to directly apply to dynamically deployed and resource-limited drone small base stations.

[0007] (2) Although the deployment locations of the drones in the cluster are close, there may still be some differences in the data sampled by different drones.

[20] This will lead to a decrease in the convergence speed of the training process.

[21] Weight bias occurs when aggregating with the model.

[22] Despite data sharing strategies

[23] and minimizing the KL divergence of marginal data

[24] Methods such as data variability can mitigate the negative impact of data discrepancies, but these methods rely on large amounts of data transfer and complex computations. FL client clustering

[25] It is also a possible solution, but existing methods mainly consider fixed clients on the Internet, which is fundamentally different from dynamically deployed and resource-constrained drones.

[0008] (3) UAV small base stations are typically deployed dynamically on demand, making it difficult to collect sufficient data in a short period. Coupled with limited computing and storage capabilities, model quality is difficult to improve. Poor performance of the local UAV model will inevitably affect the performance of the global model. Furthermore, there is system heterogeneity among different UAV small base stations, namely differences in computing, storage, and communication capabilities.

[26] This leads to differences in the models. If the traditional federal average is used...

[27] The generalization ability of the model will be weakened by this method. Summary of the Invention

[0009] To address the aforementioned technical issues, this invention proposes a RAN slicing method for UAV small base station clusters based on hierarchical FL, aiming to improve slicing performance isolation and reduce the communication cost of model collaborative training.

[0010] This invention provides a hierarchical federated learning-based RAN slicing method for unmanned aerial vehicle (UAV) small cell clusters. In the RAN scenario of UAV small cell clusters, the physical resources on each UAV small cell are virtualized into multiple slices according to the slice prediction model, and each slice supports a type of customized service; the UAV small cell is referred to as a UAV.

[0011] The slice prediction model is a federated learning (FL) model, hereinafter referred to as the model; a learning-based approach is used for collaborative training and optimization of the model, including the following steps:

[0012] 1) Local data augmentation and model training: When the drones reach a new deployment location, each drone uses data augmentation strategies to promote local model training;

[0013] 2) Clustering for edge model aggregation: The ground base station clusters the UAV clusters based on the data distribution and location information submitted by each UAV, and elects a cluster head;

[0014] The drone swarm is divided into multiple clusters. Members within a cluster upload their local model parameters to the cluster head; the cluster head then distributes the updated global model parameters from the base station to its members.

[0015] 3) Hierarchical model aggregation: After the cluster head aggregates the model parameters of the members within the cluster, it forms an edge aggregation model, and then uploads the model parameters to the ground base station; the ground base station aggregates the edge aggregation model parameters to establish a global model;

[0016] The cluster head distributes the updated global model parameters to the cluster members; the drones in the cluster update their local models based on the received global model parameters.

[0017] The main contributions of this invention include:

[0018] First, in response to problem (1), this invention designs a collaborative training framework based on hierarchical federated learning. In addition to autonomous local training, the UAV can exchange parameters with ground base stations using relays, thereby reducing communication costs and optimizing communication efficiency.

[0019] Second, in response to problem (2), this invention designs a drone swarm clustering strategy based on geographical location and data distribution. The drone acting as the cluster head is responsible for receiving the model parameters of the members within the cluster and performing edge model aggregation. The edge aggregation scheme also allows drones that are far from the ground base station to have more opportunities to participate in the collaborative optimization of local models.

[0020] Third, regarding problem (3), this invention takes into account the characteristics of dynamic deployment and insufficient data, and promotes local training through data augmentation to improve the performance of local models. For edge and global model aggregation, it explores a model aggregation scheme based on attention mechanism, which works in conjunction with the hierarchical FL framework to improve the generalization ability of the global model, so that each UAV small base station can obtain a stable performance improvement.

[0021] Simulation results show that the proposed scheme can better adapt to the dynamics of the network and outperforms existing typical distributed machine learning methods in terms of slice performance isolation and communication cost. Attached Figure Description

[0022] Figure 1 It is a RAN structure for drone small base station clusters.

[0023] Figure 2 This is a diagram illustrating service-oriented resource allocation.

[0024] Figure 3 This is a schematic diagram of a hierarchical collaborative training framework.

[0025] Figure 4 This is a schematic diagram of hierarchical model aggregation based on the attention mechanism.

[0026] Figure 5 This is a diagram illustrating the impact of the number of resource blocks on performance isolation.

[0027] Figure 6 This is a diagram showing the time percentage for performance isolation.

[0028] Figure 7 This is a schematic diagram illustrating the resource demand forecasting error of a slice.

[0029] Figures 8(a) to 8(f) This is a schematic diagram showing the variation of prediction error for three slices with communication rounds and the probability density function of the error, where:

[0030] Figure 8(a) shows the variation of prediction error for slice 1 with the number of communication rounds;

[0031] Figure 8(b) shows the variation of prediction error in slice 2 with the number of communication rounds;

[0032] Figure 8(c) shows the variation of prediction error for slice 3 with the number of communication rounds;

[0033] Figure 8(d) shows the probability density function of the prediction error for slice 1;

[0034] Figure 8(e) shows the probability density function of the prediction error in slice 2;

[0035] Figure 8(f) shows the probability density function of the prediction error for slice 3.

[0036] Figure 9 This is a schematic diagram illustrating the impact of cluster size on prediction error. Detailed Implementation

[0037] The present invention will now be described in conjunction with the accompanying drawings and specific embodiments. 1. Overview of the Invention

[0039] This invention addresses the scenario of UAV small base station clusters providing differentiated services to ground users, proposing a dynamic RAN slicing framework based on hierarchical federated learning to improve slice performance isolation and reduce communication costs during collaborative training. Considering the dynamic deployment of UAVs and insufficient data, this invention promotes local model training through data augmentation. Based on location and data distribution information, this invention designs a clustering strategy for UAV clusters, supporting edge model aggregation, allowing edge member UAVs to participate in federated learning more often and reducing communication costs. This invention explores edge and global model aggregation schemes based on attention mechanisms to enhance the generalization ability of the global federated learning model. Simulation results show that, compared with typical distributed machine learning methods, the proposed scheme can improve slice performance isolation and reduce communication costs during UAV collaborative training.

[0040] 2 System Model and Problem Description

[0041] 2.1 RAN Structure of UAV Small Base Station Cluster

[0042] like Figure 1As shown, a replicable RAN scenario is considered for a cluster of unmanned aerial vehicles (UAVs) using small base stations. UAVs can be dynamically deployed over a wide area according to preset trajectories, providing differentiated services to ground equipment. Ground base stations connect to MEC servers, providing service support to adjacent UAV clusters. When a UAV enters the coverage area of ​​a ground base station, it can communicate via a handshake protocol.

[28] Establish a connection with the base station; if the drone is far from the base station or cannot interact directly with it, it can relay information to the ground base station via a relay, thus having more opportunities to participate in local collaborative optimization. Due to its mobility, when the drone needs to go to other locations to perform tasks, it releases the connection by issuing a notification. The ground base station periodically refreshes the drone connection list and deletes information on invalid associated drones.

[0043] In the aforementioned RAN architecture, the physical resources of each UAV small cell are virtualized into multiple slices. Each slice supports a type of customized service. The UAV employs a learning-based method for model training to determine how to dynamically allocate resources to the slices, thereby providing differentiated QoS guarantees for terminal devices within its coverage area. Due to the limited capacity of UAV small cells, when the number of associated terminals exceeds the maximum capacity, no new devices are allowed to connect.

[0044] With the assistance of ground-based base stations, drone small base station clusters are clustered and undergo hierarchical model collaborative optimization to enhance resource slicing performance. For example, if Figure 1 UAV 3 is selected as the cluster head to perform edge aggregation. The local models of its member UAVs 4 and 5 can be aggregated using this cluster head for edge FL model aggregation. Cluster head UAV 3 further uploads the edge-aggregated model to the ground base station for global FL model aggregation. The cluster head distributes the updated global model parameters to member UAVs 4 and 5 within the cluster. Member UAVs 4 and 5 update their local models based on the received global model parameters. For ease of reading, the main symbols and variables used in this paper are summarized in Table 1.

[0045] Table 1 System Parameters

[0046]

[0047]

[0048] 2.2 Service-Oriented Dynamic Resource Slicing Framework

[0049] Resource slicing needs to be adaptively adjusted according to fluctuations in resource demand for various services. This section considers a service-oriented resource demand-aware resource slicing scheme, in which the controller on the UAV dynamically allocates resources to slices based on the resource demand of services.

[0050] by Figure 2 Let's take an example to illustrate service-oriented resource allocation. Assume that time is divided into multiple slice windows, and each slice window is further divided into multiple discrete time slots. The set of time slots contained in slice window 'a' is represented by T, where the number of elements in the set is denoted as T. a The drone predicts the number of resource blocks required for each slice in the next window a+1 during the operation of the current slice window a. At the start of window a+1, the resources for each slice are reallocated based on the predicted resource requirements. Figure 2 In Within window 'a', the number of resource blocks allocated to each slice remains constant. At the start of each time slot 't' within the slice window, the local controller on the UAV allocates resource blocks for the received tasks. To cope with sudden, instantaneous changes in resource demand, each UAV reserves a portion of resource blocks that can be shared by all slices. In the event of resource shortage, slices can temporarily occupy these shared resource blocks.

[0051] 2.3 Problem Modeling and Evaluation Methods

[0052] Performance isolation is a crucial prerequisite for the coexistence of resource slices, ensuring that overload in one slice does not affect other slices. The resource allocation strategy at the beginning of each slice window largely determines the effectiveness of performance isolation among slices in the next window.

[0053] Assume that drone k has B resource blocks. k These resource blocks are arranged into M k There are slices, and the set of slices is denoted as . In slice window a, assume that the predicted resource demand for slice m on drone k in window a+1 is... In the time slot The actual resource allocation is Then there is The allocation of slice resources needs to consider the following two situations:

[0054] (1) This means that the actual number of resource blocks required in time slot t is greater than the predicted value within slice window a. In this case, slice m needs to temporarily occupy the reserved shared resource blocks.

[0055] (2) This means that the required number of resource blocks is less than the predicted value, implying that the resources of slice m can meet the current demand.

[0056] Suppose that the shared resource block reserved by drone k is represented as B′. k Regarding the above situation (1), if, If all shared resource blocks are insufficient to handle sudden resource demands, then slice performance isolation cannot be guaranteed. Let binary variables... This means that each slice on UAV k in time slot t can maintain performance isolation; otherwise, it is 0.

[0057]

[0058] As shown in equation (1), within window a+1, a higher proportion of slice performance isolation time indicates a more effective resource allocation scheme. Therefore, the corresponding slice performance isolation optimization problem can be described as follows:

[0059]

[0060] During the operation of window a, the predicted number of resource blocks for slice m on drone k in the next window a+1 is... and real resource needs The smaller the gap, The higher the probability of a value being 1, the better. Therefore, by minimizing the mean square error (MSE) between the predicted and actual values, the problem... It can be transformed into the following optimization problem

[0061]

[0062] question The essence of this study is how to reduce prediction errors by optimizing model training under the constraint of total resources, so as to maximize the time ratio of slice performance isolation. Due to the characteristics of drones, such as dynamic deployment, limited sample size, and limited computing and storage capabilities, it is difficult to achieve satisfactory model results by relying solely on local model training; at the same time, drones themselves are also difficult to run complex algorithms. Therefore, subsequent studies will explore how to design a hierarchical collaborative model training method to enhance the resource slicing performance of drone swarms.

[0063] 3. A hierarchical collaborative training method for UAV swarms

[0064] 3.1 Collaborative Training Framework

[0065] This section designs a hierarchical federated learning framework based on edge aggregation, aiming to improve the quality of UAV swarm models while reducing the communication cost of collaborative training.

[0066] like Figure 3 As shown, the main steps of the collaborative training framework include:

[0067] (1) Local data augmentation and model training: Upon reaching a new deployment location, each UAV utilizes data augmentation strategies to facilitate local model training, as detailed in Section 3.2.

[0068] (2) Clustering for edge-oriented model aggregation: The base station clusters the drone cluster based on the data distribution and location information submitted by each drone and elects a cluster head. For example, Figure 3 The UAV swarm is divided into two clusters. Members within a cluster upload their local model parameters to the cluster head; the cluster head then distributes the updated global model parameters from the base station to its members. See Section 3.3 for details.

[0069] (3) Hierarchical model aggregation: After the cluster head aggregates the model parameters of the members within the cluster, it forms an edge aggregation model and further uploads the model parameters to the base station; the base station aggregates the edge aggregation model parameters to establish a globally shared model. Both edge and global model aggregation adopt an attention-based scheme. See Section 3.4 for details.

[0070] 3.2 Local Data Augmentation and Model Training

[0071] In slice window a, drone k utilizes global model parameters w (a) Initialize its own model and train it locally based on the collected data. Let... Let represent the actual resource demand vector of slice m on drone k within slice window a. Furthermore, the actual resource demand matrix for all slices on drone k is:

[0072]

[0073] The number of resource blocks required for drone k to predict slice m in window a+1 is

[0074]

[0075] Where f(·) represents the prediction model for drone k. These are the model parameters.

[0076] When a drone first enters and hovers over the target area, it is often difficult to obtain enough data in a short period of time to improve model quality. To address this, local data augmentation methods are used to facilitate model training. Specifically, a window slicing method is employed.

[29] Data augmentation involves using a sliding window to continuously sample data across a sequence, generating multiple shorter subsequences. For resource demand data of slice m on drone k, if the subsequence length is set to s, then T-s+1 time series can be generated, which is...

[0077]

[0078] Each subsequence represents a fragment of the original data, which can be represented as:

[0079]

[0080] Combined with formula (7), The corresponding set of "input-output pairs" can be represented as Where y k，m，n For the corresponding x k，m，n The output, N = T - s + 1, is the number of samples. Therefore, equation (3) can be restated as follows:

[0081]

[0082] Next, the loss function is introduced. The prediction error measures the number of resource blocks required for slice m on drone k. The loss function for all resource slices on drone k can be expressed as:

[0083]

[0084] Finally, the local model parameters of UAV k are updated using gradient descent.

[0085]

[0086] Where η is the learning rate. The loss function is represented by w. (a) The gradient.

[0087] 3.3 Clustering for Edge Model Aggregation

[0088] This section designs a clustering strategy for edge model aggregation, aiming to enable UAVs located far from ground base stations to participate in local model aggregation, reduce the communication cost of uploading UAV models, and provide support for edge and global model aggregation.

[0089] The base station clusters the drone swarm based on the drone data distribution and location. To reduce communication costs, drone k does not need to send raw data to the base station; it only needs to send its own data distribution and location information. Assume the drone's location information... And the data distribution Θ={θ1,...,θ K Incorporating k-means

[30] In this process, the cluster division is determined. As shown in Algorithm 1, the cluster division is first determined based on the location information of the UAV cluster. Perform clustering; then use based on The obtained centroids are used as initial centroids, and then clustering is performed based on Θ. This allows drones with similar geographical locations and data distributions to be grouped into the same cluster.

[0090] Algorithm 1 considers both drone location and data distribution to optimize edge aggregation while reducing communication energy consumption during drone model uploads, especially for drones located far from the base station. If clustering is based solely on data distribution, some member drones may be farther from the cluster head than from the base station. In this case, the overall communication cost of collaborative training might even be higher than the communication cost of not clustering (i.e., all drones interact directly with the base station).

[0091]

[0092]

[0093] Within each cluster, the base station elects a cluster head based on the remaining energy of the drone and the distance between the drone and the base station. Assume the distance between drone i and the base station in cluster c is d. i The remaining energy of drone i is ξ i Therefore, the normalization formula (11) is used to score the drone i.

[0094]

[0095] Where, ξ (max) d represents the upper limit of the drone's energy. (max) This represents the upper limit of the distance between the drone and the base station. The drone with the highest score in cluster c is selected as the cluster head by the base station, and its index is [index missing].

[0096]

[0097] 3.4 Attention-based hierarchical model aggregation

[0098] The cluster head is responsible for receiving model parameters from members within the cluster and performing federated average (FL) aggregation of edge models. Afterward, the cluster head uploads the parameters of the aggregated edge models to the base station for building the global model. It is worth noting that individual UAVs may have differences in data volume and model quality, and the contribution of each cluster's edge aggregation model to building the global model will also vary. If a traditional federated averaging algorithm is used, the generalization performance of the global model will be weakened. To address this issue, this section proposes an attention-based mechanism...

[31] A hierarchical model aggregation framework for intra-cluster FL aggregation and inter-cluster model FL aggregation.

[0099] Within slice window a+1, assume the edge aggregation model parameters of cluster c are expressed as follows: The parameters of the edge aggregation model for C clusters are expressed as follows: Suppose that cluster c contains I drones, and the set is... Here The local model parameters of UAV i are represented as follows: The local model parameters of I drones are represented as follows:

[0100] To achieve attention-based FL aggregation, it is necessary to compute... and w (a) O (a+1) The correlation between them. and For example, let For query, For the corresponding keywords, the relevance between the query and the keywords can be calculated using the attention scoring function according to equation (13).

[0101]

[0102] Here, V, λ, and φ all represent learnable parameter matrices. Through correlation calculation, we can obtain... and O (a+1) The corresponding two-component values ​​[s1, s2, ..., s] I ] and [s′1, s′2, ..., s′ C The scores are normalized using the Softmax function to obtain the attention distribution γ = [γ1, γ2, ..., γ] of the parameters of the UAV local model within cluster c and the edge aggregation model of C clusters. I ] and β = [β1, β2,...,β C ],in

[0103]

[0104] By minimizing and The expected distance between them can be used to obtain an edge aggregation model that is relatively close to the local model in parameter space. Then, the attention score output by equation (13) is used as a weight to minimize the value. and The distance between them, i.e.

[0105]

[0106] Where σ(·,·) represents the Euclidean distance between the two sets of parameters. Taking the derivative of the objective function in equation (15), we obtain the gradient.

[0107]

[0108] Finally, the gradient descent algorithm is executed to update the parameters of the edge aggregation model. Therefore, the update formula for the global model parameters can be expressed as:

[0109]

[0110] Where ε represents and w (a+1) The step size moved in the opposite gradient direction in each iteration. Afterwards, the base station sends the global model parameters to the cluster head, which then distributes them to members within the cluster for local model updates. Details of the hierarchical model aggregation implementation are shown in Algorithm 2, where the calculation of the attention distribution is crucial for obtaining the weighted distance between parameters; the specific process corresponds to lines 3-4.

[0111]

[0112] 4. Simulation Experiments and Result Analysis

[0113] This section presents a series of simulation experiments to evaluate the performance of the proposed method. For ease of comparison and analysis, the proposed method is named HierFed+DG+Att, where: HierFed represents the hierarchical federated architecture proposed in Section 3, DG represents the clustering strategy used in Section 3.3, and Att represents the model aggregation scheme based on the attention mechanism used in Section 3.4. To facilitate the evaluation of the effectiveness of the DG strategy, a simplified version is presented, named HierFed+Att. Furthermore, the following two benchmark methods are selected to evaluate their performance:

[0114] 1)FedAvg

[32] Model aggregation is performed using the average value of the local model parameters of the UAV;

[0115] 2) LSTM

[33] The model is deployed on each drone for distributed training.

[0116] To ensure fairness, all methods employed the same neural network architecture. Each drone used an LSTM model containing two LSTM layers (each with 64 hidden neurons) and one linear layer that maps features to predictions. The dataset used was communication data from the province of Trentito, Italy, including data services and voice calls.

[34] This paper uses different types of business data to simulate the resource demand fluctuations of data slices. The experimental parameter settings are shown in Table 2.

[0117] Table 2 Experimental Parameter Settings

[0118]

[0119] 4.1 Time percentage for slice performance isolation

[0120] This experiment investigated the impact of varying the total number of physical resource blocks for drones on the slicing performance isolation effect. All 10 drones had the same number of physical resource blocks. Figure 5The impact of varying the total number of UAV resource blocks on slice performance isolation is presented. When the number of resource blocks is 50, the slice performance isolation of each method begins to improve rapidly. Specifically, the HierFed+DG+Att method shows a higher time ratio for slice performance isolation compared to the other two benchmark methods. When the number of resource blocks is 150, the advantage of the proposed method in terms of performance isolation time ratio becomes even more pronounced.

[0121] Increasing the number of resource blocks typically helps improve the time commitment for slice performance isolation, but the actual effect is largely determined by the resource allocation strategy. In many cases, if the resource allocation for a slice does not match the actual resource requirements, the overall performance isolation effect will be compromised.

[0122] Next, we will observe the convergence and optimality of the proposed method. Figure 6 The variation of slice performance isolation time percentage with the slice window is presented when the number of resource blocks is 250. Starting from the third slice window, the slice performance isolation time percentage begins to rise rapidly until it stabilizes. This is because initially, the drone does not have enough data to improve model quality. As the amount of data collected by the drone increases, the model performance also improves. In terms of convergence speed, the proposed method and FedAvg are both faster than LSTM. The fully distributed LSTM has weaker accuracy in predicting resource requirements than the other two methods, leading to an unreasonable resource allocation strategy and reduced performance isolation effect. After iterative convergence, compared to FedAvg and LSTM, Fed+DG+Att achieves performance improvements of 8.4% and 16.5% respectively in slice performance isolation time percentage, which is attributed to accurate prediction of slice resource requirements.

[0123] 4.2 Analysis of the accuracy of resource demand forecasting

[0124] This group of experiments uses MSE to measure the error between predicted and actual resource demand. The Hier-Fed+Att method is introduced as a benchmark to verify the potential benefits of DG. Figure 7The overall prediction error of the resources required for each slice is shown. HierFed+DG+Att shares knowledge through model aggregation, which is more robust than LSTM, achieving performance gains of 15.67%, 32.64%, and 43.00% on the three slices, respectively. The proposed method has two main advantages over FedAvg: 1) the DG strategy mitigates the decline in model accuracy caused by data heterogeneity; 2) the attention-based model aggregation scheme considers the contribution of different models, resulting in a stronger generalization ability of the global model. This method achieves performance gains of 13.96%, 32.98%, and 14.64% on the three slices, respectively. Compared to HierFed+Att, HierFed+DG+Att considers the similarity of data distribution during edge aggregation, enhancing the edge FL aggregation effect, achieving performance gains of 13.34%, 26.43%, and 12.58% on the three slices, respectively.

[0125] Figures 8(a)-(c) show the variation of the prediction error of the resource blocks required for each slice with the number of communication rounds. Overall, the prediction error of each slice decreases with increasing communication rounds and converges within a certain range. The prediction error of HierFed+DG+Att on each slice is always lower than that of HierFed+Att, indicating that introducing DG can improve the model's prediction performance. Furthermore, compared to the baseline method, the proposed method achieves the same prediction accuracy using fewer communication rounds. This result demonstrates that the proposed method can reduce UAV energy consumption without compromising performance.

[0126] To facilitate observation of the differences in model prediction performance, this paper presents the probability density function (PDF) of the prediction error, as shown in Figures 8(d)-(f). For slice 1, the proposed method achieves a prediction error of less than 0.3 in 68% of cases, while HierFed+Att only achieves the same result in 20% of cases. For slice 2, the two methods achieve prediction errors of less than 0.2 in approximately 99% and 94% of cases, respectively, which are quite close. For slice 3, the proposed method achieves a prediction error of less than 0.2 in 87% of cases, while HierFed+Att achieves it in approximately 52%.

[0127] 4.3 The impact of the number of clusters on performance

[0128] This experiment examines the impact of variations in the number of clusters, C, on the predictive performance of resource requirements for three slices. C=1 means no clustering. Figure 9 As can be seen, the prediction error is highest when C=1 and lowest when C=4. Without clustering, all data are mixed together to produce a global model, increasing data heterogeneity. Therefore, the prediction error is always higher without clustering than with clustering.

[0129] The number of clusters also affects the communication energy consumption during collaborative training, including transmission and reception energy consumption. This group of experiments uses the model in reference

[35] to quantify the communication energy consumption. Table 3 shows the communication energy consumption under different numbers of clusters. Although the communication energy consumption of HierFed+DG+Att is slightly higher than that of HierFed+Att, the improvement in prediction performance is more significant. The communication loss of both hierarchical federation methods is lower than that of FedAvg. Under the HierFed structure, UAVs far from the base station can upload their own parameter information through the cluster head, while under the FedAvg method, all UAVs can only communicate directly with the base station. The results of this group confirm the effectiveness of the method proposed in this invention in reducing the overall communication cost of UAV clusters.

[0130] Table 3. Impact of Cluster Number on Communication Energy Consumption

[0131]

[0132] Note: The communication power consumption under the FedAvg scheme is 0.76KJ.

[0133] 5. Conclusion

[0134] For RAN slicing in UAV swarms, this invention proposes a collaborative training method based on hierarchical federated learning. First, a data augmentation strategy is employed to facilitate local model training among UAVs. Then, the UAV swarm is clustered according to the distribution of UAV sampled data and their geographical location, mitigating the negative impact of local data discrepancies on edge FL model aggregation. The cluster head performs edge FL model aggregation, giving UAVs farther from the base station more opportunities to participate in federated learning and reducing communication costs. Finally, to construct the edge aggregation model and the global model, an attention-based FL model aggregation scheme is designed, improving the generalization ability of the global model. Simulation results show that, compared with two benchmark methods, the proposed method improves the performance isolation effect of slicing and has lower communication costs.

[0135] References:

[0136] [1]Xia T, Wang MM, Jiang C, et al.Practical machine-type communi-cation for energy Internet of Things: An introduction[J]. IEEE CommunicationsStandards Magazine, 2019, 3(1): 48-59.

[0137] [2]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.

[0138] [3] Chen Xinying, Sheng Min, Li Bo. A review of UAV communication for 6G[J]. Journal of Electronics and Information Technology, 2022, 44(03): 781-789.

[0139] [4] Zhang S. An overview of network slicing for 5G[J]. IEEE WirelessCommunications, 2019, 26(3): 111-117.

[0140] [5]Shen x, Gao J, wu w, et al.AI-assisted network-slicing based next-generation wireless networks[J]. IEEE Open Journal of Vehicular Technology, 2020, 1: 45-66.

[0141] [6] Gutterman C, Grinshpun E, Sharma S, et al. RAN resource usage prediction for a 5G slice broker[C] / / Proceedings of the twentieth ACMinternational symposium on mobile ad hoc networking and computing. 2019: 231-240.

[0142] [7] Qi Weijing, Song Qingyang, Guo Lei. Dual-time-scale RAN slice resource allocation for software-defined multimodal vehicle networking [J]. Journal of Communications, 2022, 43(04): 60-70.

[0143] [8]Nassar A,Yilmaz Y.Deep reinforcement learning for adaptive networkslicing in 5G for intelligent vehicular systems and smart cities[J].IEEEInternet of Things Journal,2021,9(1):222-235.

[0144] [9]Yuan Z,Muntean G M.Airslice:A network slicing framework for uavcommunications[J].IEEE Communications Magazine,2020,58(11):62-68.

[0145]

[10] Yang P,Xi x,Guo K,et al.Proactive UAV network slicing for URLLCand mobile broadband service multiplexing[J].IEEE Joumal on Selected Areas inCommunications,2021,39(10):3225-3244.

[0146]

[11] 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.Springer,Cham,2022:651-661.

[0147]

[12] Hossain M A,Ansari N.Network slicing for NOMA-enabled edgecomputing[J].IEEE Transactions on Cloud Computing,2021.

[0148]

[13] Yang H,Zhao J,Xiong Z,et al.Privacy-preserving federated learningfor UAV-enabled networks:Learning-based joint scheduling and re-sourcemanagement[J].IEEE Journal on Selected Areas in Commu-nications,2021,39(10):3144-3159.

[0149]

[14] Lian Z,Wang w,Su C.COFEL:communication-efficient and optimizedfederated learning with local differential privacy[C] / / ICC 2021-IEEEInternational Conference on Communications.IEEE,2021:1-6.

[0150]

[15] Lu Y,Liu Z,Huang Y.Parameters Compressed Mechanism in FederatedLearning for Edge Computing[C] / / 2021 8th IEEE Interna-tional Conference onCyber Security and Cloud Computmg(CSCloud) / 2021 7th IEEE InternationalConference on Edge Com-puting and Scalable Cloud(EdgeCom).IEEE,2021:161-166.

[0151]

[16] Yang H H,Liu Z,Quek T Q S,et al.Scheduling policies for federatedlearning in wireless networks[J].IEEE transactions on communications,2019,68(1):317-333.

[0152]

[17] Lu Y, Huang

[0153]

[18] Tang Lun, Wang Zhiping, Pu Hao, et al. Research on efficient federated learning communication mechanism based on adaptive gradient compression [J]. Journal of Electronics and Information Technology, 2022, 44: 1-8.

[0154]

[19] Luo S, Chen

[0155]

[20] Tao Meixia, Wang Dong, Sun Rui, Zhang Naifu. User scheduling strategy based on time division multiple access in federated learning [J]. Journal of Communications, 2021, 42(06): 16-29.

[0156]

[21] Xiao J, Du C, Duan Z, et al. A Novel Server-side Aggregation Strategy for Federated Learning in Non-IID situations [C] / / 2021 20tb Interna-tional Symposium on Parallel and Distributed Computing (ISPDC). IEEE, 2021: 17-24.

[0157]

[22] Yeganeh Y,Farshad A,Navab N,et al.Inverse distance aggregationfor federated learning with non-iid data[M] / / Domain Adaptation andRepresentation Transfer,and Distributed and Collaborative Learning.Springer,Cham,2020:150-159.

[0158]

[23] Zhao Y,Li M,Lai L,et al.Federated learning with non-iid data[J].arXiv preprint arXiv:1806.00582,2018.

[0159]

[24] Deng Y,Lyu F,Ren J,et al.SHARE:Shaping data distribution at edgefor commiunication-efricient hierarchical federated learning[C] / / 2021 IEEE41st International Conference on Distributed Computing Sys-tems(ICDCS).IEEE,2021:24-34.

[0160]

[25] Briggs C,Fan Z,Andras P.Federated learning with hierarchicalclustering of local updates to improve training on non-IID da-ta[C] / / 2020International Joint Confefence on Neural Networks(IJCNN).IEEE,2020:1-9.

[0161]

[26] Li T,Sahu A K,Zaheer M,et al.Federated optimization in hetero-geneous networks[J].Proceedings of Machine Learning and Systems,2020,2:429-450.

[0162]

[27] Nilsson A,Smith S,ulm G,et al.A performance evaluation offederated learning algorithms[C] / / Proceedings of the second workshop ondistributed infrastructures for deep learning.2018:1-8.

[0163]

[28] Song Q,Zeng Y,Xu J,et al.A survey of prototype and experiment forUAV communications[J].Science China Information Sciences,2021,64(4):1-21.

[0164]

[29] Wen Q,Sun L,Yang F,et al. Time series data augmentation for deeplearning:A survey[J].arXiv preprint arXiv:2002.12478,2020.

[0165]

[30] Vouros A,Langdell S,Croucher M,et al.An empirical comparisonbetween stochastic and deterministic centroid initialisation for K-meansvariations[J].Machine Learning,2021,110(8):1975-2003.

[0166]

[31] Ji S,Pan S,Long G,et al.Learning private neural language modelingwith attentive aggregation[C] / / 2019 International joint conference on neuralnetworks(IJCNN).IEEE,2019:1-8.

[0167]

[32] McMahan B,Moore E,Ramage D,et al.Communication-efficient learningof deep networks from decenttalized data[C] / / Artificial intel-ligence andstatistics.PMLR,2017:1273-1282.

[0168]

[33] Yu Y,Si X,Hu C,et al.Areview of recurrent neural networks:LSTMcells and network architectures[J].Neural computatioin,2019,31(7):1235-1270.

[0169]

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

[0170]

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

Claims

1. A hierarchical federated learning-based RAN slicing method for UAV small base station clusters, characterized in that... In the RAN scenario of UAV small base station cluster radio access network, the physical resources on each UAV small base station are virtualized into multiple slices according to the slice prediction model, and each slice supports a type of customized service. The term "drone small base station" is abbreviated as "drone". The slice prediction model is a federated learning (FL) model, hereinafter referred to as the model; The model co-training optimization is performed using a learning-based method, and the steps include: 1) Local data augmentation and model training: Once the drones reach a new deployment location, each drone utilizes data augmentation strategies to facilitate local model training; 2) Clustering for edge model aggregation: The ground base station clusters the drone clusters based on the data distribution and location information submitted by each drone and elects a cluster head; The drone swarm is divided into multiple clusters. Members within a cluster upload their local model parameters to the cluster head; the cluster head then distributes the updated global model parameters from the base station to its members. 3) Hierarchical model aggregation: After the cluster head aggregates the model parameters of the members within the cluster, it forms an edge aggregation model, and then uploads the model parameters to the ground base station; the ground base station aggregates the edge aggregation model parameters to establish a global model; The cluster head distributes the updated global model parameters to the cluster members; the drones in the cluster update their local models based on the received global model parameters. The drone's controller dynamically allocates resources to slices based on the service's resource requirements, using the following method: Assume that time is divided into multiple slice windows, and each slice window is further divided into multiple discrete time slots. Represents a slice window The set of included time slots, the number of elements in which the set is represented as ; Drone in the current slice window Predict the next window during operation The number of resource blocks required for each slice within the window; Initially, based on resource demand forecasts Reallocate resources for each slice; In the window Within this range, the number of resource blocks allocated to each slice remains constant, and the number of resource blocks allocated to each time slot within the slice window remains constant. Initially, the controller on the drone allocates resource blocks for the received tasks; Each drone reserves a portion of resource blocks that can be shared by all slices; in the event of insufficient resources, the slice temporarily occupies these shared resource blocks; Assuming a drone The number of resource blocks is These resource blocks are arranged into There are slices, and the set of slices is denoted as . In the slice window Assuming drones upper slice In the window The resource demand forecast is In the time slot The actual resource allocation is Then there is ; There are two scenarios for slice resource allocation: Scenario 1: This represents a time slot. The required number of actual resource blocks is greater than the number of slice windows. The predicted value within the slice It is necessary to temporarily occupy the reserved shared resource block; Scenario 2: This means that the required number of resource blocks is less than the predicted value, so the slice is... Resources are sufficient to meet current needs; Assuming a drone The reserved shared resource block is represented as For case 1, if If all shared resource blocks are insufficient to meet sudden resource demands, then slice performance isolation cannot be guaranteed. Let the binary variable =1 represents in the time slot drones Each slice on the slice can maintain performance isolation; otherwise, the isolation value is 0. ； From equation (1), in the window Within the slice performance isolation timeframe, the higher the proportion of time spent on the slice, the more effective the resource allocation scheme. Therefore, the corresponding slice performance isolation optimization problem can be described as follows: ； In the window During operation, the drone slices In the next window Predicted number of resource blocks and real resource needs The smaller the gap, The higher the probability of a value being 1, the better; therefore, by minimizing the mean square error between the predicted and actual values, the problem... It is transformed into the following optimization problem , ； question The essence is how to reduce prediction errors by optimizing model training under the constraint of total resources, so as to maximize the time ratio of slice performance isolation. In equations (2) and (3), T is the number of discrete time slots within a slice window, and M is the number of slices; In step 1), in the slice window drones Using global model parameters Initialize its own model and train it locally based on the collected data; make Representing drones upper slice In the slice window The actual resource demand vector, then the drone The actual resource requirement matrix for all slices is as follows: ， drones Predicted slices In the window The required number of resource blocks is ， in, Representing drones Predictive models, These are model parameters; The method of using local data augmentation to facilitate model training is as follows: Data augmentation is achieved using a window slicing method, which involves continuously sampling across the sequence using a sliding window to generate multiple shorter subsequences; this method is suitable for UAVs. upper slice Regarding resource requirement data, if the subsequence length is set to Then it will produce A time series, that is ， Each subsequence represents a fragment of the original data, denoted as... ， The corresponding set of "input-output pairs" is represented as follows: ,in For the output corresponding to s, The number of samples; Equation (3) is restated as ， Next, the loss function is introduced. Measuring drones upper slice Prediction error of the required number of resource blocks; drones The loss function for all resource slices is expressed as: ， Finally, the gradient descent method is used to update the drone. Local model parameters ， in, For learning rate, Represents the loss function relative to The gradient; In step 2), the ground base station clusters the drone swarm based on the distribution and location of the drone data: drones It only sends its own data distribution and location information; Location information of the drone and data distribution Integrating k-means into the clustering process to determine cluster partitioning, the initial step is based on the drone cluster location information. Perform clustering; then use based on The obtained centroid is the initial centroid, and then based on the data distribution... Perform clustering; Within each cluster, the data distribution is as follows: the base station elects a cluster head based on the remaining energy of the drone and the distance between the drone and the base station; assuming cluster... Chinese drones The distance to the base station is drones The remaining energy is Therefore, the normalization formula (11) is used for the UAV. Rate ， in, Represents the upper limit of the drone's energy. This represents the upper limit of the distance between the drone and the base station; cluster The drone with the highest score was selected as the cluster head by the base station; the index of this drone is... ； In step 3), a hierarchical model aggregation framework based on attention mechanism is adopted for intra-cluster model aggregation and inter-cluster model aggregation; In the slice window Inside, assuming cluster The edge aggregation model parameters are expressed as , The parameters of the edge aggregation model for each cluster are represented as follows: Assuming clusters Inside A collection of drones ;Will Chinese drones The local model parameters are represented as , The local model parameters of each drone are represented as follows: ; calculate and ,as well as and The correlation between the two is calculated using the same method; by and For example, let For query, For the corresponding keywords, the relevance between the query and the keywords is calculated using the attention scoring function according to equation (13). ， in, and Both represent learnable parameter matrices; By calculating the correlation, we can obtain the relationship with... and Corresponding two component values and ; The scores are normalized using the Softmax function to obtain clusters. The local model of the drone and Attention distribution of parameters in the edge aggregation model for each cluster and ,in ， By minimizing and The expected distance between them is used to obtain an edge aggregation model that is closer to the local model in parameter space; Then, the attention score output by equation (13) is used as a weight to minimize the value. and The distance between them, i.e. ， in, The gradient is obtained by taking the derivative of the objective function in equation (15) to represent the Euclidean distance between the two sets of parameters. ， Finally, the gradient descent algorithm is executed to update the parameters of the edge aggregation model; Therefore, the update formula for the global model parameters is expressed as: ， in, represent and The step size in the opposite gradient direction during each iteration; Afterwards, the ground base station sends the global model parameters to the cluster head, which then distributes them to members within the cluster for local model updates of the UAV.

2. The RAN slicing method for UAV small base station clusters based on hierarchical federated learning according to claim 1, characterized in that, When the number of terminals associated with a drone exceeds the maximum capacity, no new devices are allowed to connect.

Images