Heterogeneous client autonomous collaboration method based on twin channel model in air-ground clustering federated learning scene

Abstract

The invention discloses a heterogeneous client autonomous collaboration method based on a twin channel model in an air-ground clustering federated learning scene, and belongs to the technical field of machine training. Dividing a client cluster by adopting an AP algorithm and preferably selecting a part of unmanned aerial vehicles from the unmanned aerial vehicles as cluster center nodes; an HT-UCB algorithm is adopted in the cluster, a multi-dimensional reward function is designed, and clients participating in global training are dynamically selected; the cluster center nodes and the air-ground heterogeneous clients selected in the cluster execute federated learning, and finally, the center node aggregates all cluster center node parameters to update a global model and perform loop iteration until the model converges; according to the invention, the identification precision and convergence rate of the overall model and the overall utilization rate of resources in the cooperative inspection task of the unmanned aerial vehicle and the unmanned vehicle are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of machine training technology, specifically to a heterogeneous client autonomous collaboration method based on a twin channel model in an air-ground clustered federated learning scenario. Background Technology

[0002] With the rapid development of low-altitude intelligent network technology, drones and unmanned vehicles (UAVs) have demonstrated unique advantages through collaborative work. Drones, leveraging their aerial mobility, can quickly cover vast areas and acquire wide-area image data from a high-altitude perspective in real time; while UAVs, flexibly traversing the ground, use onboard high-precision sensors to collect detailed information about specific areas, compensating for the limitations of drones in acquiring microscopic data. With their advanced imaging capabilities, drones and UAVs, collecting image data in real time, have been widely applied in various fields such as inspection, search and rescue, assisted driving, and military operations. These application scenarios have extremely high real-time requirements for data analysis, which conventional methods cannot meet, necessitating customized and secure machine learning (ML) models.

[0003] However, traditional machine learning model training heavily relies on uploading raw datasets collected by drones and unmanned vehicles to a central server for processing. This approach exposes multiple hidden dangers in the dynamic and heterogeneous scenarios of low-altitude intelligent networks: First, the direct transmission of wide-area images from drones and local high-resolution data from unmanned vehicles may lead to the leakage of sensitive geographical information due to the openness of wireless channels, and even face the risk of malicious node hijacking or gradient inversion attacks; second, the real-time transmission of massive amounts of heterogeneous data, such as tilted-view video and multispectral images, will consume a large amount of communication bandwidth, which can easily cause channel saturation during drone swarm collaboration or network congestion. The high mobility of UAV nodes leads to frequent dynamic reconfiguration of network topology, and the channel state changes rapidly and non-stationarily due to the complex spatiotemporal environment. Link interference and collisions occur frequently in scenarios with limited spectrum resources, further exacerbating communication latency jitter and packet loss rate. At the same time, the heterogeneity of multidimensional resources (such as computing power, storage, energy, and communication bandwidth) varies significantly among terminal nodes, making it difficult to achieve resource collaborative optimization under dynamic task loads. This results in a serious imbalance between local computing bottlenecks and global model convergence efficiency, severely restricting the real-time performance and robustness of the air-ground collaborative intelligent system.

[0004] In low-altitude intelligent networks, how to jointly train heterogeneous data formed by UAV top-down perspective and unmanned vehicle (UAV) eye-level perspective is a common and challenging problem. UAVs perform wide-area image scanning of target areas from a low-altitude top-down perspective, while UAVs acquire local high-resolution images through onboard telephoto lenses and macro cameras. The difference in perspective leads to geometric heterogeneity. Traditional clustering methods rely solely on data feature similarity for partitioning, making it difficult to create stable and efficient client clusters. Traditional client selection strategies often employ random selection or static mechanisms based on a single indicator such as device battery level, which can easily lead to parameter shifts, wasted resources, or training bias. Therefore, it is essential to design a heterogeneous client autonomous collaboration method based on a twin channel model in an air-to-ground clustered federated learning scenario. Summary of the Invention

[0005] The purpose of this invention is to provide a heterogeneous client autonomous collaboration method based on a twin channel model in an air-ground clustered federated learning scenario, so as to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a heterogeneous client autonomous collaboration method based on a twin channel model in an air-ground clustered federated learning scenario, comprising the following steps:

[0007] During the collaborative inspection task, where UAVs and unmanned vehicles (UAVs) collect information separately, the wide-area image acquisition by UAVs and the local high-resolution image and point cloud data acquisition by UAVs are synchronized by the GPS module. The sensor sampling is triggered according to a preset period based on GPS time, and a timestamp is embedded in the data frame to ensure synchronized acquisition. Both transmit GPS position, environmental perception data and channel status information (CSI) back to the base station in real time. The base station inputs the collected information into the digital twin online channel model to predict the environmental and communication status in real time.

[0008] The central node first distributes the global model parameters to each drone and unmanned vehicle;

[0009] The central node calculates the data similarity between each pair of clients based on the high-level feature vectors of the images of unmanned vehicles and drones after model processing, as well as the point cloud data of the unmanned vehicles. It also calculates the geographical similarity based on GPS coordinates and constructs a global similarity matrix using the AP (nearest neighbor propagation) algorithm. The drone and unmanned vehicle groups are then clustered. By updating their responsibility and attribution values, the client with the largest sum of responsibility and attribution values ​​is selected as the cluster center. Non-cluster center clients are assigned to the cluster center that maximizes the sum of their responsibility and attribution values.

[0010] Images acquired by autonomous vehicles are processed by STN-CNN to transform the view from a head-on perspective to a top-down perspective, achieving dimensional unification between autonomous vehicle and drone data. Building upon this, data similarity focuses on uncovering the heterogeneous information remaining in the feature distributions of autonomous vehicle and drone data, fully considering the heterogeneity in feature distribution between wide-area drone image data and local high-resolution autonomous vehicle image data; geographic similarity considers the differences in the spatial range and scene focus of the two data acquisition methods.

[0011] The central node calculates the data similarity s between every two clients based on the feature similarity of each client. data (i,j), used to evaluate whether the data distributions among clients are similar, is represented as:

[0012] s data ij = w image · image i j+w pointcloud · pointcloud ij)

[0013]

[0014] Among them, s image (i,j) is the image similarity, f i and f j s are the high-level feature vectors extracted from the image data of clients i and j by the model, where · represents the dot product and ||·|| represents the magnitude of the vector. pointcloud (i,j) represents the point cloud data similarity of the autonomous vehicle, p i p j It is the point cloud data extracted by PointNet from client i and j. The autonomous vehicle extracts point cloud features using PointNet. image and w pointcloud These are the weighting coefficients;

[0015] The cloud calculates the geographical similarity s between any two clients based on the 3D geolocation of each client. geo This is used to assess whether geographical publishing is similar between clients, and is represented as:

[0016]

[0017] Among them, (x i ,y i ,z i ), (x j ,y j ,z j ) represents the three-dimensional geographic coordinates of client i and j, w x w yw z ε is the position weight, and ε is the position correction coefficient.

[0018] The cloud platform generates a global similarity matrix S(i,j) through weighted fusion based on data similarity and geographic location similarity, denoted as:

[0019] S(i,j)=α·s data (i,j)+β·s geo (i,j)

[0020] Drone nodes are set with higher initial preference values, while unmanned vehicle nodes are set with low or negative values ​​to ensure that only drones compete for cluster centers.

[0021] The responsibility value r(i,j) is updated to quantify the superiority of node i in choosing j as the cluster center compared to other candidates, expressed as:

[0022] r(i,j)←S(i,j)-max j′≠j [a(i,j′)+S(i,j′)]

[0023] The attribution value a(i,j) is updated, quantifying the cumulative support of node j as a cluster center. The attribution value of node j itself is equal to the positive responsibility value of all other nodes to it, expressed as:

[0024]

[0025] To restrict the affiliation value to non-positive numbers and prevent node i from excessively supporting multiple cluster centers,

[0026]

[0027] For each node j, if a(j,j)+r(j,j)>0, then j is the cluster center. If j is an unmanned vehicle (R(j)=0), even if the score is high, it is forcibly excluded. Non-cluster center node i belongs to the cluster center k that maximizes a(j,j)+r(j,j).

[0028] If the central node performs multiple calculations and remains unchanged after several iterations, or if the algorithm executes beyond the set number of iterations, then the clustering of UAVs and unmanned vehicles ends. If the battery level E(j) of the cluster-center UAV is ≤ E... min Then, unmanned vehicles and drones are re-clustered to form new clusters and a suitable cluster center is selected.

[0029] Considering that traditional client selection strategies often employ random selection or static mechanisms based on a single metric (such as device battery level), lacking multi-dimensional collaborative optimization of data distribution similarity, resources, and performance, if there are substandard clients, such as drones that are nearly out of power but are forcibly selected due to static screening mechanisms, their unstable communication links may lead to parameter transmission interruptions or even cluster core collapse. Alternatively, camera clients carrying noisy data or adversarial examples may be incorrectly classified into the same cluster due to the non-independent and identically distributed nature of their data distribution and that of drones, causing parameter shifts during the model aggregation stage, resulting in resource waste or training bias.

[0030] Therefore, after the central node completes the client clustering based on multidimensional similarity, each cluster starts the client scheduling process using HT-UCB (Heterogeneous Terminal Perception Upper Confidence Bound Algorithm) in the initial stage of federated learning training. First, for each cluster, before the start of each round of training, the system selects each client in the cluster to participate in a local model training task once in batches to obtain its initial performance data. This process ensures that each client is selected independently once in the initialization stage, avoiding interference between batches, and the selection order only has a short-term impact on the subsequent cumulative calculation of reward values. Then, the initial reward value of each client is calculated.

[0031] Because drones possess high-altitude, wide-area perception capabilities, they can quickly collect large-scale image data. However, their high mobility leads to fluctuating communication links and rapid power consumption. Therefore, prioritizing the utilization of their data value is crucial, and a certain level of communication latency can be tolerated. Simultaneously, flight time attenuation factors must be considered. The initial drone reward is set based on the following dynamically acquired indicators: the similarity of data distribution between the client and the cluster center is calculated through feature vector comparison; normalized remaining power is reported by the drone in real time; and communication efficiency is predicted in real time using channel state information such as the signal-to-interference-plus-noise ratio output by the digital twin model. Communication resources are not fixedly allocated but dynamically adjusted according to network conditions. The drone reward is expressed as:

[0032] R UAV (i)=α·s data (i,μ k )(1-e -t / 1000 )+β·Energy(i)+γ·η comm (i)

[0033] Among them, R UAV (i) is the reward value for the drone, s data (i,μ k ) represents the cosine similarity of the data distribution between drone i and the cluster center, t is time, e is the base of the natural logarithm, and Energy(i) is the normalized remaining battery power of the drone; η comm(i) represents the communication efficiency between the client and the cluster center, where α, β, and γ are weighting coefficients.

[0034]

[0035] Where B(i) is the real-time bandwidth of client i, B0 is the baseline bandwidth set by the channel twin model, D(i) is the real-time transmission delay of client i, and D0 is the delay tolerance threshold set by the channel twin model.

[0036]

[0037] Here, we assume the model parameter data volume is L (bits) and the effective client transmission rate is R. eff (i), then the transmission rate is expressed as:

[0038]

[0039] in, B(i) is the predicted signal-to-interference-plus-noise ratio output by the twin channel model, and B(i) is the bandwidth of client i.

[0040] T prop (i) is the physical propagation time of the signal from the client to the cluster center, expressed as:

[0041]

[0042] Where d(i) is the physical distance from the client to the cluster center, and c is the speed of light in a vacuum;

[0043]

[0044] Where (x) i ,y i ,z i ), (x c ,y c ,z c These are the real-time GPS coordinates of the client and the cluster center, respectively.

[0045] Since autonomous vehicles focus on high-resolution local ground data acquisition, they have high communication stability but limited computing and power resources. Therefore, prioritizing a balance between communication efficiency and power consumption is crucial. In addition to the aforementioned indicators, the initial reward value for autonomous vehicles also needs to consider the stability factor θ in the direction of movement. stab The magnitude and frequency of directional changes can be calculated by analyzing GPS trajectory points over a short period of time. The reward function is expressed as:

[0046] R UGV (i)=α·s data (i,μ k )+β·Energy(i)+γ·ηcomm (i)+η·θ b

[0047] Among them, R UGV (i) is the reward value of the autonomous vehicle, θ stab It is the directional stability factor;

[0048]

[0049] Where N is the number of GPS trajectory points collected in a short period of time, and N is the change in the orientation angle between the i-th and (i+1)-th trajectory points. It is the direction change frequency weighting function, where k is the adjustment coefficient;

[0050] Simultaneously, a mobility penalty mechanism is introduced to avoid selecting drones that are about to fly out of communication range. Therefore, the faster the drone's speed, the greater the penalty. The mobility penalty is expressed as:

[0051]

[0052] For autonomous vehicles, a penalty is added for directional changes. The lower the directional stability, the larger the penalty, in order to reduce communication interruptions caused by the autonomous vehicle's turning. The penalty function is expressed as:

[0053] Penalty UGV =μ·(1-θ) stab )

[0054] The reward adjustments for drones and autonomous vehicles, derived from the penalty function, are expressed as follows:

[0055] R′ UAV (i)=R UAV (i)-Penalty UAV

[0056] R′ UGV (i)=R UGV (i)-Penalty UGV .

[0057] Next, the UCB value for each client is calculated. The UCB value calculation first calculates the reward value to calculate the average reward value, and then calculates the exploration value based on the average reward value to obtain the UCB value for each client.

[0058] First, the remaining battery power, bandwidth, and latency of the client are collected. Then, the data similarity between the client and the cluster center is calculated to determine the client's reward value. The average reward value is then calculated based on the reward value, and is expressed as follows:

[0059]

[0060] in, It is the reward value that client i receives when it is selected for the sth time;

[0061] The exploration value is calculated based on the average reward value, and then the UCB value of each arm is obtained, expressed as:

[0062]

[0063] Where c is the exploration coefficient and t is the total number of selections;

[0064] For each cluster, no global synchronization is required. The improved UCB algorithm is executed independently within the cluster. The UCB values ​​of the drones and unmanned vehicles are compared. The client with the largest UCB value is selected as the next candidate. Its selection count is incremented by 1. Because increasing the selection count will lead to a decrease in the exploration value, the same client will not be selected indefinitely. The UCB values ​​of the clients within the cluster are recalculated to compare the values ​​and select the next candidate.

[0065] After multiple iterations of calculation, the UCB algorithm balances exploration and utilization, gradually converging to the optimal arm. It selects the clients with the highest UCB values ​​as clients participating in parameter training within the cluster. If a client's battery level is lower than the minimum required battery level or its communication latency is greater than the maximum latency, the client is immediately removed from the candidate pool, and subsequent clients are selected to participate in training.

[0066] In a further embodiment, step three, the process of clustered federated learning between the drone and the unmanned vehicle, further includes:

[0067] Within each cluster, all clients perform local training, and the local model update is represented as:

[0068]

[0069] in, This represents the model parameters updated in round t+1 for the i-th drone or unmanned vehicle in the k-th cluster. Let represent the model parameters updated by the i-th drone in the k-th cluster in the t-th round; η is the learning rate of the model; The model parameters are: The loss function is L k,i Gradient of the model within the time cluster; L k,i Let represent the loss function of the i-th UAV in the k-th cluster;

[0070] After the selected air-to-ground heterogeneous client updates its local parameters multiple times, it uploads its own model parameters to the cluster-center UAV.

[0071] Then, the cluster-centric UAV performs a federated average of the collected model parameters, which is represented as:

[0072]

[0073] in, Let C denote the model parameters of UAV j, the cluster center of the k-th cluster, in round t. k It is the set of clients belonging to cluster center k, |D k | is the total number of samples in the k-th cluster, |D i | is the number of samples for client i within the k-th cluster, and

[0074] The cluster-core drone sends the updated parameters to its affiliated drones and unmanned vehicles. The selected drones and unmanned vehicles perform multiple local updates and then upload the parameters to the cluster-core drone. The drone then aggregates the parameters again and updates them, repeating this process multiple times.

[0075] Finally, the central node aggregates the parameters of each cluster-center UAV, as follows:

[0076]

[0077] Among them, W (t) Indicates cloud model parameters. This represents the total number of samples from all clients participating in the training.

[0078] The central node can send model parameters to the cluster-center drone, which in turn sends them to its client. By repeating the above model training operation, the updated central node model parameters can be obtained. The iteration stops when the difference between the current global model parameters and the previous round model parameters is less than the set threshold or when the maximum number of global communication rounds is reached.

[0079] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: Addressing the technical requirements of collaborative inspection by multiple UAVs and unmanned vehicles in dynamic heterogeneous scenarios of low-altitude intelligent networks, this invention proposes a heterogeneous client autonomous collaboration method based on a twin channel model in an air-to-ground clustered federated learning scenario. This method constructs an air-to-ground collaborative three-dimensional perception and self-organizing decision-making system, using UAVs and unmanned vehicles as heterogeneous terminal nodes, and employs the AP (nearest neighbor propagation) algorithm to achieve dynamic multi-dimensional clustering. Based on data feature similarity and three-dimensional geographical proximity, it effectively solves the instability problems of traditional clustering algorithms in communication fluctuations, equipment energy consumption, and geometrically heterogeneous scenarios, significantly improving clustering efficiency. This paper focuses on improving internal communication efficiency and model convergence speed. Based on this, a multi-dimensional resource-driven client selection strategy is proposed. An improved UCB (Upper Bound Confidence) algorithm is used to design a multi-dimensional reward function, dynamically balancing data distribution similarity, remaining battery power, and communication efficiency. A real-time filtering mechanism is introduced to remove nodes with low battery power or high latency, thereby optimizing global model accuracy and training stability. Furthermore, a clustered federated learning architecture aggregates parameters and pushes them down to edge nodes, reducing reliance on the central server. Besides ensuring the privacy and security of data in sensitive areas, this provides an efficient, robust, and resource-optimized solution for air-ground collaborative tasks, demonstrating significant technical advantages in collaborative inspection scenarios. Attached Figure Description

[0080] Figure 1 This is a scenario diagram of the heterogeneous client autonomous collaboration method based on the twin channel model in the air-ground clustered federated learning scenario of the present invention;

[0081] Figure 2 This is a flowchart of the AP algorithm based on multi-dimensional clustering of the present invention;

[0082] Figure 3 This is a flowchart of the client scheduling based on the improved UCB algorithm of the present invention;

[0083] Figure 4 This is a schematic diagram of the clustered federated learning process of the present invention;

[0084] Figure 5 This is the overall flowchart of the heterogeneous client autonomous collaboration method based on the twin channel model in the air-ground clustered federated learning scenario of the present invention;

[0085] Figure 6 This is a flowchart of the method steps of the present invention. Detailed Implementation

[0086] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0087] Please see Figures 1-6 This invention provides a technical solution: a heterogeneous client autonomous collaboration method based on a twin channel model in an air-to-ground clustered federated learning scenario. To improve the overall model recognition accuracy, convergence rate, and overall resource utilization of UAVs and unmanned vehicles in collaborative inspection tasks, the method specifically includes the following steps:

[0088] During the collaborative inspection task, where UAVs and unmanned vehicles (UAVs) collect information separately, the wide-area image acquisition by UAVs and the local high-resolution image and point cloud data acquisition by UAVs are synchronized by the GPS module. The sensor sampling is triggered according to a preset period based on GPS time, and a timestamp is embedded in the data frame to ensure synchronized acquisition. Both transmit GPS position, environmental perception data and channel status information (CSI) back to the base station in real time. The base station inputs the collected information into the digital twin online channel model to predict the environmental and communication status in real time.

[0089] The central node first distributes the global model parameters to each drone and unmanned vehicle;

[0090] like Figure 2 As shown, this embodiment provides an AP algorithm based on multi-dimensional clustering as a method for clustering drones and unmanned vehicles, including the following steps:

[0091] Step 201: The drone captures environmental images from an aerial overhead view using its onboard high-definition camera, the unmanned vehicle captures ground-level images using its camera, the unmanned vehicle generates high-precision 3D point cloud data using its LiDAR, and the drone and unmanned vehicle obtain their real-time 3D coordinates through the Global Positioning System (GPS).

[0092] Step 202: The autonomous vehicle performs cross-view geometric alignment on the acquired high-resolution images using its own STN-CNN network. First, the STN learns projection transformation parameters, such as the affine matrix, from the autonomous vehicle's head-up view to the drone's top-down view. Then, the tilted view image is reprojected onto the top-down coordinate system according to the transformation parameters to generate the aligned image. The drone and the autonomous vehicle then input their collected images into the CNN for feature extraction, outputting high-level feature vectors.

[0093] The central node calculates the data similarity s between every two clients based on the feature similarity of each client. data(i,j), used to evaluate whether the data distributions among clients are similar, is represented as:

[0094] s data ij = w image · image i j+w pointcloud · pointcloud ij)

[0095]

[0096] Among them, s image (i,j) is the image similarity, f i and f j s are the high-level feature vectors extracted from the image data of clients i and j by the model, where · represents the dot product and ||·|| represents the magnitude of the vector. pointcloud (i,j) represents the point cloud data similarity of the autonomous vehicle, p i p j It is the point cloud data extracted by PointNet from client i and j. The autonomous vehicle extracts point cloud features using PointNet. image and w pointcloud These are the weighting coefficients;

[0097] The cloud calculates the geographical similarity s between any two clients based on the 3D geolocation of each client. geo This is used to assess whether geographical publishing is similar between clients, and is represented as:

[0098]

[0099] Among them, (x i ,y i ,z i ), (x j ,y j ,z j ) represents the three-dimensional geographic coordinates of client i and j, w x w y w z ε is the position weight, and ε is the position correction coefficient.

[0100] Step 203: The cloud platform generates a global similarity matrix S(i,j) through weighted fusion based on data similarity and geographic location similarity, represented as:

[0101] S(i,j)=α·s data (i,j)+β·s geo (i,j)

[0102] Step 204: Set a higher initial preference value for drone nodes and a lower or negative value for unmanned vehicle nodes to ensure that only drones compete for the cluster center;

[0103] Step 205: Update the responsibility value r(i,j), quantifying the superiority of node i choosing j as the cluster center compared to other candidates, expressed as:

[0104] r(i,j)←S(i,j)-max j′≠j [a(i,j′)+S(i,j′)]

[0105] Step 206: Update the attribution value a(i,j), quantify the cumulative support of node j as a cluster center, and the attribution value of node j itself is equal to the positive responsibility value of all other nodes to it, expressed as:

[0106]

[0107] To restrict the affiliation value to non-positive numbers and prevent node i from excessively supporting multiple cluster centers,

[0108]

[0109] Step 207: For each node j, if a(j,j)+r(j,j)>0, then j is the cluster center. If j is an unmanned vehicle (R(j)=0), even if the score is high, it is forcibly excluded. Non-cluster center node i belongs to the cluster center k that maximizes a(j,j)+r(j,j).

[0110] If the central node performs multiple calculations and remains unchanged after several iterations, or if the algorithm executes more than the set number of iterations, then the clustering of drones and unmanned vehicles ends.

[0111] Step 208: If the charge E(j) of the cluster-center UAV is ≤ E min Then, unmanned vehicles and drones are re-clustered to form new clusters and a suitable cluster center is selected.

[0112] like Figure 1 As shown, the UAV performs wide-area aerial image scanning and the unmanned vehicle performs local high-resolution ground data acquisition. Both obtain position and channel status information through GPS and digital twin channel model. After clustering by AP algorithm and scheduling by HT-UCB algorithm, the model parameters are trained locally and aggregated and updated between the cluster and the central node to realize collaborative inspection task scenario with efficient data training and privacy protection in dynamic heterogeneous scenarios.

[0113] Considering that traditional client selection strategies often employ random selection or static mechanisms based on a single metric (such as device battery level), lacking multi-dimensional collaborative optimization of data distribution similarity, resources, and performance, if there are substandard clients, such as drones that are nearly out of power but are forcibly selected due to static screening mechanisms, their unstable communication links may lead to parameter transmission interruptions or even cluster core collapse. Alternatively, camera clients carrying noisy data or adversarial examples may be incorrectly classified into the same cluster due to the non-independent and identically distributed nature of their data distribution and that of drones, causing parameter shifts during the model aggregation stage, resulting in resource waste or training bias.

[0114] Therefore, after the central node completes client clustering based on multidimensional similarity, each cluster initiates the client scheduling process using HT-UCB (Heterogeneous Terminal Perceptual Upper Confidence Bound Algorithm) in the initial stage of federated learning training, such as... Figure 3 As shown, this embodiment also provides a method for client scheduling based on an improved UCB algorithm, the specific steps of which are as follows:

[0115] Step 301: First, for each cluster, before the start of each round of training, the system adopts a batch sequential execution method to select each client in the cluster to participate in a local model training task once to obtain its initial performance data. This process ensures that each client is selected independently once in the initialization phase, avoiding interference between batches, and the selection order only has a short-term impact on the subsequent cumulative calculation of reward value. Then, the initial reward value of each client is calculated.

[0116] Because drones possess high-altitude, wide-area perception capabilities, they can quickly collect large-scale image data. However, their high mobility leads to fluctuating communication links and rapid power consumption. Therefore, prioritizing the utilization of their data value is crucial, and a certain level of communication latency can be tolerated. Simultaneously, flight time attenuation factors must be considered. The initial drone reward is set based on the following dynamically acquired indicators: the similarity of data distribution between the client and the cluster center is calculated through feature vector comparison; normalized remaining power is reported by the drone in real time; and communication efficiency is predicted in real time using channel state information such as the signal-to-interference-plus-noise ratio output by the digital twin model. Communication resources are not fixedly allocated but dynamically adjusted according to network conditions. The drone reward is expressed as:

[0117] R UAV (i)=α·s data (i,μ k )(1-e -t / 1000 )+β·Energy(i)+γ·η comm (i)

[0118] Among them, R UAV (i) is the reward value for the drone, s data (i,μ k) represents the cosine similarity of the data distribution between drone i and the cluster center, t is time, e is the base of the natural logarithm, and Energy(i) is the normalized remaining battery power of the drone; η comm (i) represents the communication efficiency between the client and the cluster center, where α, β, and γ are weighting coefficients.

[0119]

[0120] Where B(i) is the real-time bandwidth of client i, B0 is the baseline bandwidth set by the channel twin model, D(i) is the real-time transmission delay of client i, and D0 is the delay tolerance threshold set by the channel twin model.

[0121]

[0122] Here, we assume the model parameter data volume is L (bits) and the effective client transmission rate is R. eff (i), then the transmission rate is expressed as:

[0123]

[0124] in, B(i) is the predicted signal-to-interference-plus-noise ratio output by the twin channel model, and B(i) is the bandwidth of client i.

[0125] T prop (i) is the physical propagation time of the signal from the client to the cluster center, expressed as:

[0126]

[0127] Where d(i) is the physical distance from the client to the cluster center, and c is the speed of light in a vacuum;

[0128]

[0129] Where (x) i ,y i ,z i ), (x c ,y c ,z c These are the real-time GPS coordinates of the client and the cluster center, respectively.

[0130] Since autonomous vehicles focus on high-resolution local ground data acquisition, they have high communication stability but limited computing and power resources. Therefore, prioritizing a balance between communication efficiency and power consumption is crucial. In addition to the aforementioned indicators, the initial reward value for autonomous vehicles also needs to consider the stability factor θ in the direction of movement. stab The magnitude and frequency of directional changes can be calculated by analyzing GPS trajectory points over a short period of time. The reward function is expressed as:

[0131] RUGV (i)=α·s data (i,μ k )+β·Energy(i)+γ·η comm (i)+η·θ b

[0132] Among them, R UGV (i) is the reward value of the autonomous vehicle, θ stab It is the directional stability factor;

[0133]

[0134] Where N is the number of GPS trajectory points collected in a short period of time, and N is the change in the orientation angle between the i-th and (i+1)-th trajectory points. It is the direction change frequency weighting function, where k is the adjustment coefficient;

[0135] Simultaneously, a mobility penalty mechanism is introduced to avoid selecting drones that are about to fly out of communication range. Therefore, the faster the drone's speed, the greater the penalty. The mobility penalty is expressed as:

[0136]

[0137] For autonomous vehicles, a penalty is added for directional changes. The lower the directional stability, the larger the penalty, in order to reduce communication interruptions caused by the autonomous vehicle's turning. The penalty function is expressed as:

[0138] Penalty UGV =μ·(1-θ) stab )

[0139] The reward adjustments for drones and autonomous vehicles, derived from the penalty function, are expressed as follows:

[0140] R′ UAV (i)=R UAV (i)-Penalty UAV

[0141] R′ UGV (i)=R UGV (i)-Penalty UGV .

[0142] Step 302: Next, calculate the UCB value for each client. The UCB value calculation first calculates the reward value to calculate the average reward value, and then calculates the exploration value based on the average reward value to obtain the UCB value for each client.

[0143] Step 303: Collect the client's remaining battery power, bandwidth, and latency; calculate the data similarity between the client and the cluster center; and thus calculate the client's reward value. The average reward value is then calculated based on the reward value, and is expressed as follows:

[0144]

[0145] in, It is the reward value that client i receives when it is selected for the sth time;

[0146] Step 304: Calculate the exploration value based on the average reward value, and then obtain the UCB value for each arm, expressed as:

[0147]

[0148] Where c is the exploration coefficient and t is the total number of selections;

[0149] Step 305: Within a cluster, compare the UCB values ​​of drones and unmanned vehicles. The client with the largest UCB value is selected as the next target, and its selection count is incremented by 1. Since increasing the selection count will lead to a decrease in the exploration value, the same client will not be selected indefinitely.

[0150] The UCB values ​​of clients within the cluster are recalculated to compare and select the next client to be chosen.

[0151] Step 306: After multiple iterations of calculation, the UCB algorithm will balance exploration and utilization, gradually converge to the optimal arm, and select the clients with the highest UCB values ​​as clients participating in parameter training within the cluster.

[0152] Step 307: If the client's battery level E(j) < E min Or the communication delay D(j) > D max The client is immediately removed from the candidate pool, and the next client is selected to participate in the training.

[0153] After scheduling the client, the drones and unmanned vehicles undergo clustered federated learning, such as... Figure 4 The specific steps include:

[0154] Step 401: The central node first distributes the global model parameters to each drone and unmanned vehicle;

[0155] Step 402: The drones and unmanned vehicles are clustered using the multi-dimensional AP algorithm described above;

[0156] Step 403: Within each cluster, all clients perform local training, and the local model update is represented as:

[0157]

[0158] in, This represents the model parameters updated in round t+1 for the i-th drone or unmanned vehicle in the k-th cluster. Let represent the model parameters updated by the i-th drone in the k-th cluster in the t-th round; η is the learning rate of the model; The model parameters are: The loss function is L k,i Gradient of the model within the time cluster; L k,i Let represent the loss function of the i-th UAV in the k-th cluster;

[0159] After the selected air-to-ground heterogeneous client updates its local parameters multiple times, it uploads its own model parameters to the cluster-center UAV.

[0160] Step 404: Then, the cluster-center UAV performs a federated average of the collected model parameters, expressed as:

[0161]

[0162] in, Let C denote the model parameters of UAV j, the cluster center of the k-th cluster, in round t. k It is the set of clients belonging to cluster center k, |D k | is the total number of samples in the k-th cluster, |D i | is the number of samples for client i within the k-th cluster, and

[0163] The cluster-core drone sends the updated parameters to its affiliated drones and unmanned vehicles. The selected drones and unmanned vehicles perform multiple local updates and then upload the parameters to the cluster-core drone. The drone then aggregates the parameters again and updates them, repeating this process multiple times.

[0164] Step 405: Finally, the central node on the base station side aggregates the parameters of each cluster-center UAV, as follows:

[0165]

[0166] Among them, W (t) Indicates cloud model parameters. This represents the total number of samples from all clients participating in the training.

[0167] Step 406: The central node can send the model parameters to the cluster-center drone, which in turn sends them to its client. By repeating the above model training operation, the updated central node model parameters can be obtained.

[0168] Step 407: Stop iterating when the difference between the current global model parameters and the previous model parameters is less than the set threshold or when the maximum number of global communication rounds is reached.

[0169] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and all such equivalent transformations fall within the protection scope of the present invention.

Claims

1. A heterogeneous client autonomous collaboration method based on a twin channel model in an air-ground clustered federated learning scenario, characterized in that, Includes the following steps: Step 1: The UAV performs wide-area image acquisition, while the unmanned vehicle performs local high-resolution image and point cloud data acquisition. Due to the difference in perspective between the UAV and the unmanned vehicle, the unmanned vehicle image needs to be processed by STN-CNN to change from a head-up view to a top-down view to achieve dimensional unification between the head-up view data of the unmanned vehicle and the top-down view data of the UAV. The central node integrates the similarity of data features of the client-acquired images and the geographical proximity of the client's three-dimensional coordinates. The AP algorithm is used to generate a global similarity matrix, dynamically dividing the UAV and unmanned vehicle into different clusters, and selecting some UAVs as cluster center nodes from all UAVs, with the rest as client members within the cluster. Step 2: Based on the channel state information output by the dynamically constructed and continuously updated digital twin online channel model, each cluster uses the heterogeneous terminal perception upper confidence bound algorithm HT-UCB for dynamic client scheduling during federated learning training. A multi-dimensional reward function is designed based on data distribution similarity, normalized remaining power, communication efficiency, and mobility stability factor to calculate the UCB value of the client. Clients participating in global training are dynamically selected, and low power or high latency nodes are filtered in real time. Step 3: All clients perform local training. The clients with the highest UCB values ​​obtained by the HT-UCB algorithm are selected as heterogeneous air-ground clients. The updated model parameters after local training are uploaded to the cluster core node. The cluster core node performs federated average aggregation of parameters and then distributes them to the clients within the cluster. The base station-side core node finally aggregates all cluster core node parameters to update the global model and iterates in a loop. The iteration stops when the difference between the current global model parameters and the previous round model parameters is less than the set threshold or when the maximum number of global communication rounds is reached.

2. The heterogeneous client autonomous collaboration method based on a twin channel model in a space-ground clustered federated learning scenario according to claim 1, characterized in that: In step one, the central node first sends global model parameters to each drone and unmanned vehicle; The central node calculates the data similarity between each pair of clients based on the high-level feature vectors of the images of unmanned vehicles and drones after model processing, as well as the point cloud data of the unmanned vehicles. It also calculates the geographical similarity based on GPS coordinates and constructs a global similarity matrix using the AP algorithm. The drone and unmanned vehicle groups are then clustered. By updating their responsibility and attribution values, the client with the largest sum of responsibility and attribution values ​​is selected as the cluster center. Non-cluster center clients are assigned to the cluster center that maximizes the sum of their responsibility and attribution values.

3. The heterogeneous client autonomous collaboration method based on a twin channel model in a space-ground clustered federated learning scenario according to claim 2, characterized in that: In step one, the unmanned vehicle image data is processed by STN-CNN and transformed from a head-up view to a top-down view, achieving dimensional unification between the unmanned vehicle head-up view data and the drone top-down view data. Based on this, data feature similarity focuses on mining the heterogeneous information that still exists in the feature distribution of the unmanned vehicle head-up view data and the drone top-down view data, fully considering the heterogeneity in feature distribution between the drone's wide-area image data and the unmanned vehicle's local high-resolution image data; geographical proximity combines the differences between the two in the data acquisition spatial range and scene focus degree; The central node calculates the data similarity s between every two clients based on the feature similarity of each client. data (i,j), used to evaluate whether the data distributions among clients are similar, is represented as: s data i j=w image · image i j+w pointcloud · pointcloud i j) Among them, s image (i,j) is the image similarity, f i and f j s are the high-level feature vectors extracted from the image data of clients i and j by the model, where · represents the dot product and ||·|| represents the magnitude of the vector. pointcloud (i,j) represents the point cloud data similarity of the autonomous vehicle, p i p j It is the point cloud data extracted by PointNet from client i and j. The autonomous vehicle extracts point cloud features using PointNet. image and w pointcloud These are the weighting coefficients; The cloud calculates the geographical similarity s between any two clients based on the 3D geolocation of each client. geo This is used to assess whether geographical publishing is similar between clients, and is represented as: Among them, (x i ,y i ,z i ), (x j ,y j ,z j ) represents the three-dimensional geographic coordinates of client i and j, w x w y w z ε is the position weight, and ε is the position correction coefficient.

4. The heterogeneous client autonomous collaboration method based on a twin channel model in a space-ground clustered federated learning scenario according to claim 3, characterized in that: In step two, after constructing the global similarity matrix, the specific process of the AP algorithm is as follows: The cloud platform generates a global similarity matrix S(i,j) through weighted fusion based on data similarity and geographic location similarity, denoted as: S(i,j)=α·s data (i,j)+β·s geo (i,j) Drone nodes are set with higher initial preference values, while unmanned vehicle nodes are set with low or negative values ​​to ensure that only drones compete for cluster centers. The responsibility value r(i,j) is updated to quantify the superiority of node i in choosing j as the cluster center compared to other candidates, expressed as: r(i,j)←S(i,j)-max j′≠j [a(i,j′)+S(i,j′)] The attribution value a(i,j) is updated, quantifying the cumulative support of node j as a cluster center. The attribution value of node j itself is equal to the positive responsibility value of all other nodes to it, expressed as: To restrict the affiliation value to non-positive numbers and prevent node i from excessively supporting multiple cluster centers, For each node j, if a(j,j)+r(j,j)>0, then j is the cluster center. If j is an unmanned vehicle (R(j)=0), even if the score is high, it is forcibly excluded. Non-cluster center node i belongs to the cluster center k that maximizes a(j,j)+r(j,j).

5. The heterogeneous client autonomous collaboration method based on a twin channel model in a space-ground clustered federated learning scenario according to claim 1, characterized in that: In step two, after the central node completes the client clustering based on multidimensional similarity, each cluster starts the client scheduling process in the initial stage of federated learning training using the heterogeneous terminal perception upper confidence bound algorithm (HT-UCB). First, for each cluster, before the start of each round of training, the system selects each client in the cluster to participate in a local model training task once in batches to obtain its initial performance data. This process ensures that each client is selected independently once in the initialization stage, avoiding interference between batches, and the selection order only has a short-term impact on the cumulative calculation of subsequent reward values. Then, the initial reward value of each client is calculated.

6. The heterogeneous client autonomous collaboration method based on a twin channel model in a space-ground clustered federated learning scenario according to claim 5, characterized in that: The reward values ​​differ for heterogeneous clients. Because drones possess high-altitude, wide-area perception capabilities, they can quickly collect large-scale image data. However, their high mobility leads to fluctuating communication links and rapid power consumption. Therefore, prioritizing the utilization of their data value is crucial, and a certain level of communication latency can be tolerated. Flight time attenuation factors must also be considered. The initial reward value for the drone is set based on the following dynamically acquired indicators: the similarity of data distribution between the client and the cluster center is calculated through feature vector comparison; normalized remaining power is reported by the drone in real time; and communication efficiency is predicted in real time using channel state information such as the signal-to-interference-plus-noise ratio output by the digital twin model. Communication resources are not fixedly allocated but dynamically adjusted according to network conditions. The drone reward value is expressed as: R UAV (i)=α·s data (i,m k (1-e) -t / 1000 )+β·Energy(i)+γ·η comm (i) Among them, R UAV (i) is the reward value for the drone, s data (i,μ k ) represents the cosine similarity of the data distribution between drone i and the cluster center, t is time, e is the base of the natural logarithm, and Energy(i) is the normalized remaining battery power of the drone; η comm (i) represents the communication efficiency between the client and the cluster center, where α, β, and γ are weighting coefficients. Where B(i) is the real-time bandwidth of client i, B0 is the baseline bandwidth set by the channel twin model, D(i) is the real-time transmission delay of client i, and D0 is the delay tolerance threshold set by the channel twin model. Here, we assume the model parameter data volume is L (bits) and the effective client transmission rate is R. eff (i), then the transmission rate is expressed as: in, B(i) is the predicted signal-to-interference-plus-noise ratio output by the twin channel model, and B(i) is the bandwidth of client i. T prop (i) is the physical propagation time of the signal from the client to the cluster center, expressed as: Where d(i) is the physical distance from the client to the cluster center, and c is the speed of light in a vacuum; Where (x) i ,y i ,z i ), (x c ,y c ,z c These are the real-time GPS coordinates of the client and the cluster center, respectively. Since autonomous vehicles focus on high-resolution local ground data acquisition, they have high communication stability but limited computing and power resources. Therefore, prioritizing a balance between communication efficiency and power consumption is crucial. In addition to the aforementioned indicators, the initial reward value for autonomous vehicles also needs to consider the stability factor θ in the direction of movement. stab The magnitude and frequency of directional changes can be calculated by analyzing GPS trajectory points over a short period of time. The reward function is expressed as: R UGV (i)=α·s data (i,m k )+β·Energy(i)+γ·η comm (i)+η·θ b Among them, R UGV (i) is the reward value of the autonomous vehicle, θ stab It is the directional stability factor; Where N is the number of GPS trajectory points collected in a short period of time, and N is the change in the orientation angle between the i-th and (i+1)-th trajectory points. It is the direction change frequency weighting function, where k is the adjustment coefficient; Simultaneously, a mobility penalty mechanism is introduced to avoid selecting drones that are about to fly out of communication range. Therefore, the faster the drone's speed, the greater the penalty. The mobility penalty is expressed as: For autonomous vehicles, a penalty is added for directional changes. The lower the directional stability, the larger the penalty, in order to reduce communication interruptions caused by the autonomous vehicle's turning. The penalty function is expressed as: Penalty UGV =μ·(1-θ stab ) The reward adjustments for drones and autonomous vehicles, derived from the penalty function, are expressed as follows: R′ UAV (i)=R UAV (i)-Penalty UAV R′ UGV (i)=R UGV (i)-Penalty UGV 。 7. A heterogeneous client autonomous collaboration method based on a twin channel model in a space-ground clustered federated learning scenario, as described in claim 6, is characterized in that: After the reward value is set, the overall process of HT-UCB is as follows: The UCB value for each client is calculated separately. The UCB calculation first calculates the reward value to determine the average reward value. Then, the exploration value is calculated based on the average reward value to obtain the UCB value for each arm. The remaining battery power, bandwidth, and latency of each client are collected, and the data similarity between the client and the cluster center is calculated to determine the client's reward value. The average reward value is then calculated based on the reward value, and is expressed as follows: in, It is the reward value that client i receives when it is selected for the sth time; The exploration value is calculated based on the average reward value, and then the UCB value of each arm is obtained, expressed as: Where c is the exploration coefficient and t is the total number of selections; For each cluster, no global synchronization is required. The improved UCB algorithm is executed independently within the cluster. The UCB values ​​of the drones and unmanned vehicles are compared. The client with the largest UCB value is selected as the next candidate. Its selection count is incremented by 1. Because increasing the selection count will lead to a decrease in the exploration value, the same client will not be selected indefinitely. The UCB values ​​of the clients within the cluster are recalculated to compare the values ​​and select the next candidate. After multiple iterations of calculation, the UCB algorithm balances exploration and utilization, gradually converging to the optimal arm. It selects the clients with the highest UCB values ​​as clients participating in parameter training within the cluster. If a client's battery level is lower than the minimum required battery level or its communication latency is greater than the maximum latency, the client is immediately removed from the candidate pool, and subsequent clients are selected to participate in training.

8. The heterogeneous client autonomous collaboration method based on a twin channel model in a space-ground clustered federated learning scenario according to claim 1, characterized in that: In step three, the process of clustered federated learning for drones and unmanned vehicles also includes: Within each cluster, all clients perform local training, and the local model update is represented as: in, This represents the model parameters updated in round t+1 for the i-th drone or unmanned vehicle in the k-th cluster. Let represent the model parameters updated by the i-th drone in the k-th cluster in the t-th round; η is the learning rate of the model; The model parameters are: The loss function is L k,i Gradient of the model within the time cluster; L k,i Let represent the loss function of the i-th UAV in the k-th cluster; After the selected air-to-ground heterogeneous client updates its local parameters multiple times, it uploads its own model parameters to the cluster-center UAV. Then, the cluster-centric UAV performs a federated average of the collected model parameters, which is represented as: in, Let C denote the model parameters of UAV j, the cluster center of the k-th cluster, in round t. k It is the set of clients belonging to cluster center k, |D k | is the total number of samples in the k-th cluster, |D i | is the number of samples for client i within the k-th cluster, and The cluster-core drone sends the updated parameters to its affiliated drones and unmanned vehicles. The selected drones and unmanned vehicles perform multiple local updates and then upload the parameters to the cluster-core drone. The drone then aggregates the parameters again and updates them, repeating this process multiple times. Finally, the central node aggregates the parameters of each cluster-center UAV, as follows: Among them, W (t) Indicates cloud model parameters. This represents the total number of samples from all clients participating in the training. The central node can send model parameters to the cluster-center drone, which in turn sends them to its client. By repeating the above model training operation, the updated central node model parameters can be obtained. The iteration stops when the difference between the current global model parameters and the previous round model parameters is less than the set threshold or when the maximum number of global communication rounds is reached.

Images