Unmanned aerial vehicle cooperative training method and system for vehicle networking split federated learning

Abstract

The application discloses a UAV cooperative training method and system for vehicle networking split federated learning, and the method comprises the following steps: constructing an intersection vehicle networking split federated learning scene, and setting a multi-level energy efficiency business role of a UAV; collecting interval intersection observation data, obtaining a next time system state through a space-time state prediction module; inputting a strategy decision module combined with current observation information to generate a UAV control instruction; matching a corresponding energy efficiency role according to a control action, scheduling a UAV to carry out split federated learning cooperative training and output node submission data; updating information by relying on a node training state single and a gradient debt clearing mechanism, completing a global model update, and realizing UAV cooperative split federated learning training. The application can improve the utilization efficiency of UAV hardware resources under different business bottlenecks. The application can be widely applied to the technical field of split federated learning cooperative training.

Description

Technical Field

[0001] This invention relates to the field of split federated learning collaborative training technology, and in particular to a method and system for collaborative training of unmanned aerial vehicles (UAVs) for split federated learning in the context of vehicle networks. Background Technology

[0002] In the evolution of intelligent transportation systems, real-time prediction of intersection collision risks is a crucial requirement for ensuring the safety of autonomous driving. Traditional centralized cloud-based training models require vehicles to upload large amounts of raw sensor data, which can easily lead to bandwidth congestion and privacy risks. Split Federated Learning (SFL) addresses this by splitting and deploying deep neural networks in slice layers. The vehicle handles front-end feature extraction, while the Remote Unit (RSU) or edge nodes with computing capabilities handle back-end computation for serving the subnetworks. The RSU coordinates model aggregation and parameter synchronization, balancing training efficiency, communication load, and privacy protection without requiring data to leave the local machine.

[0003] Meanwhile, drones, with their high mobility and line-of-sight transmission advantages, have become an important carrier for integrated air-ground collaboration in vehicle-to-everything (V2X) communication. Existing drone-assisted V2X solutions mainly fall into three categories: fixed aerial relays, mobile edge servers, and federated learning clients. Fixed relays focus on coverage enhancement, mobile edge servers on computation offloading, and federated learning clients on aerial perspective data acquisition and model training. While these solutions improve coverage, computing power, or data acquisition capabilities respectively, they generally assume that the drone role is relatively stable during training and fail to model role switching, rotation relay, and SFL slice layer training links in a cohesive manner. Therefore, they cannot handle the training loop breakage problem caused by simultaneous changes in the activation upload path, backend computing nodes, and gradient backhaul path.

[0004] However, most existing technologies limit drones to a single, static role, making it difficult to flexibly address the mixed challenges of alternating communication blind spots and computational surges at urban intersections. Even when some solutions introduce predictive scheduling, split federated learning, or generative agent activation, they typically only focus on resource allocation, missing activation completion, or aggregation waiting time, without further addressing the impact of drones switching between three business roles—independent client, relay collaboration, and computational collaboration server—and the effects on the SFL front-end and back-end training states during formation rotation control. Summary of the Invention

[0005] To address the aforementioned technical problems, the purpose of this invention is to provide a collaborative training method and system for UAVs based on decomposed federated learning for the Internet of Vehicles, which can improve the utilization efficiency of UAV hardware resources under different business bottlenecks.

[0006] The first technical solution adopted in this invention is: a drone collaborative training method for decomposed federated learning in the Internet of Vehicles, comprising the following steps: Construct a vehicle-to-everything (V2X) decomposed federated learning scenario for the target intersection and perform system initialization; determine the networking architecture of roadside units and drone teams, as well as the various energy efficiency classification business roles of drones. The system acquires intersection environmental observation data within a preset time interval, inputs the intersection environmental observation data into the spatiotemporal state prediction module for prediction, and determines the system state prediction result for the next moment. The system state prediction results for the next moment and the current observation state are input into the strategy decision module to generate a control strategy and determine the control actions of the UAV. The control actions of the drone are mapped one-to-one with various energy efficiency classification business roles. The drone is controlled to switch to the energy efficiency classification business role, execute the corresponding mode of split federated learning collaborative training, and output node submission data. Based on the global aggregation phase, the node submission data generated by the split federated learning collaborative training is updated through the node training status sheet and gradient debt settlement mechanism to determine the result of the global model update in this round, thereby realizing the UAV collaborative training of split federated learning.

[0007] Furthermore, the step of acquiring intersection environmental observation data within a preset time interval, inputting the intersection environmental observation data into the spatiotemporal state prediction module for prediction, and determining the system state prediction result for the next moment specifically includes: Obtain multi-dimensional intersection environment observations at the current moment and construct the observation feature vector at the current moment; Input the observation feature vectors of multiple frames within a preset historical window into the temporal prediction network, extract the temporal features, and output the system state prediction vector for the next moment. Calculate the prediction deviation between the observed feature vector at the current moment and the system state prediction vector at the next moment, and input the prediction deviation as a feedback term into the next moment prediction process of the spatiotemporal state prediction module. If the prediction deviation exceeds the preset circuit breaker threshold, the predictive safety circuit breaker mechanism is triggered, switching to a conservative control strategy and adjusting the battery safety threshold. The conservative control strategy and the adjusted battery safety threshold are used as constraint outputs of the system state prediction result to determine the system state prediction result at the next moment.

[0008] Furthermore, the step of inputting the system state prediction result for the next moment and the current observation state into the strategy decision module to generate a control strategy and determine the control actions of the UAV specifically includes: The decision state vector is obtained by concatenating the current observation state, the system state prediction vector in the system state prediction result of the next moment, the prediction deviation, the current UAV business role, and the role transition risk. Based on preset action constraint rules, the candidate action set is masked and filtered to obtain the current set of actionable actions; Input the decision state vector into the reinforcement learning policy network, and output the probability distribution of each action in the current set of possible actions; The control actions of the drone are obtained by sampling based on the probability distribution of each action in the current set of possible actions.

[0009] Furthermore, the training reward value of the reinforcement learning policy network is determined through the following steps: Obtain the model accuracy gain value for this round of split federated learning as a training benefit; Calculate the drone energy consumption value corresponding to the current control action, and normalize the drone energy consumption value as the energy consumption cost item; When the role of the drone service changes in adjacent decision slots, a role switching penalty is applied. Calculate the system service latency and training sample missing ratio corresponding to the current control action, and generate a nonlinear quality of service penalty term; The training reward value of the reinforcement learning policy network is obtained by weighting the training reward, energy consumption cost, role switching penalty, and service quality penalty.

[0010] Furthermore, the control actions of the UAV include rotating control actions. After determining the control actions of the UAV, the system also includes rotating execution steps, which specifically include: When the drone is currently in relay coordination mode, control the standby drone to establish a parallel communication link, complete the link verification, switch the service connection, and then disconnect the original drone link. When the drone is currently in a computational collaboration mode, the standby drone is controlled to preload model parameters. After the current training round ends, the task status and optimizer summary are synchronized, and the drone takes over the next round of computational tasks. With the drone currently in independent client mode, once the current small batch of front-end calculations is completed, the node identity and cache index are synchronized, and the standby drone takes over the subsequent data collection and calculation tasks. After the rotation is completed, the standby drone will take over the current business role and continue to perform split federated learning and collaborative training.

[0011] Furthermore, the step of mapping the control actions of the drone to various energy efficiency classification business roles, controlling the drone to switch to the energy efficiency classification business role, performing corresponding mode-based split federated learning collaborative training, and outputting node submission data specifically includes: When the target business role is in independent client mode, control the drone to shut down the high-power computing and forwarding modules, perform local data acquisition, front-end feature extraction and slice layer feature uploading, and generate client node submission data; When the target business role is in relay collaboration mode, control the drone to fly to the preset relay point, establish a two-way communication forwarding link between the blind spot vehicle and the roadside unit, and ensure the upload of slice layer features and gradient back transmission of the blind spot node. When the target business role is in a computing collaboration mode, the drone is controlled to activate the onboard computing module, take over part of the backend model computing tasks unloaded from the roadside unit, generate the backend computing results of the corresponding node, complete the split federated learning collaborative training, and output the node submission data.

[0012] Furthermore, the step of updating the global model through node training state sheets and gradient debt settlement mechanisms to determine the update result of this round specifically includes: Create a node training status sheet for each node participating in the training, which records the node's business role, training path, and training completion status at each stage. After the preset aggregation time window ends, based on the validity of the node submission data and training path, all nodes are divided into a real submission set, a virtual submission set, and an expired set. Perform a standard split federated learning update process on the nodes in the real submission set to obtain the first part of the model update amount; Proxy activations are generated for nodes in the virtual submission set to maintain training continuity, and corresponding gradient debt slips are generated synchronously to obtain the second part of the model update amount; Nodes in the expired set will not be included in this round of model updates; Based on the first part of the model update and the second part of the model update, the global model update result for this round is obtained by integrating them.

[0013] Furthermore, the step of synchronously generating the corresponding gradient debt statement specifically includes: Calculate the inverse gradient of the slice layer corresponding to the proxy activation; Record the corresponding node number, training round, slice layer information, agent activation credibility, debt validity period and training path information as the corresponding node information; The inverse gradient of the slice layer is integrated with the corresponding node information to generate and store the gradient debt list for this node in this round, which is used for debt settlement when the node is reconnected in the future.

[0014] Furthermore, the tiered debt settlement mechanism specifically includes: When a node reconnects and uploads real activation data, the corresponding historical gradient debt statement for that node is retrieved, and the compatibility score of the corresponding historical debt is calculated. If the compatibility score meets the preset liquidation threshold, the historical debt gradient is added as a compensation term to the front-end model update gradient of the node and participates in the parameter update of the corresponding node. If the compatibility score does not meet the preset liquidation threshold, the corresponding historical debt will be automatically invalidated.

[0015] The second technical solution adopted in this invention is: a drone collaborative training system for decomposed federated learning in the Internet of Vehicles, comprising: The first module is used to construct the vehicle-to-everything (V2X) split federated learning scenario at the target intersection and perform system initialization, determine the networking architecture of roadside units and drone teams, and the various energy efficiency classification business roles of drones. The second module is used to acquire intersection environmental observation data within a preset time interval, input the intersection environmental observation data into the spatiotemporal state prediction module for prediction, and determine the system state prediction result for the next moment. The third module is used to input the system state prediction results for the next moment and the current observation state into the strategy decision module to generate control strategies and determine the control actions of the UAV. The fourth module is used to map the control actions of the drone to various energy efficiency classification business roles, control the drone to switch to the energy efficiency classification business role, perform the corresponding mode of split federated learning collaborative training, and output node submission data. The fifth module is used to update the node submission data generated by the split federated learning collaborative training through the node training status sheet and gradient debt settlement mechanism based on the global aggregation phase, determine the result of the global model update in this round, and realize the UAV collaborative training of split federated learning.

[0016] The beneficial effects of the method and system of this invention are as follows: This invention constructs a vehicle-to-everything (V2X) network split federated learning scenario at the target intersection and initializes the system, determining the networking architecture of roadside units and UAVs, as well as the various energy efficiency-level service roles of the UAVs; further, it acquires intersection environmental observation data within a preset time interval, inputs the intersection environmental observation data into the spatiotemporal state prediction module for prediction, and determines the system state prediction result for the next moment. It can combine real-time observation data to anticipate future traffic flow density and blind spot demand trends, and dynamically calibrate the prediction results using historical deviation feedback, providing a forward-looking and highly confident state input for deep reinforcement learning decision-making. This effectively reduces the ineffective and frequent switching of UAVs between different roles, lowers switching overhead, and improves the smoothness of the control strategy; subsequently, it inputs the system state prediction result for the next moment and the current observation state into the strategy decision module to generate a control strategy and determine the control actions of the UAVs; and it integrates the control actions of the UAVs with various... A one-to-one mapping is established between energy efficiency classification business roles. The drone is controlled to switch to the energy efficiency classification business role, executes the corresponding mode of split federated learning collaborative training, and outputs node submission data. This allows for compensation of the client-side front-end model without interrupting back-end training, reducing model semantic drift caused by role switching, node disconnection, and proxy activation. Finally, based on the global aggregation phase, the node submission data generated by the split federated learning collaborative training is updated through node training status sheets and gradient debt settlement mechanisms to determine the global model update results for this round. This achieves drone collaborative training using split federated learning and proposes a three-state dynamic switching architecture for drones. This allows the same hardware device to flexibly switch between independent client, relay collaboration, and computational collaboration modes according to real-time environmental requirements. This mechanism breaks the resource isolation of traditional solutions, ensuring relay connections for vehicles in communication blind spots and shared computing power for RSUs in computational overload scenarios, improving the utilization efficiency of drone hardware resources under different business bottlenecks. Attached Figure Description

[0017] Figure 1 This is a flowchart of the steps of the UAV collaborative training method for decomposed federated learning in the Internet of Vehicles according to the present invention; Figure 2 This is a structural block diagram of the UAV collaborative training system for decomposed federated learning in the Internet of Vehicles (IoV) of this invention; Figure 3 This is a schematic diagram of the system architecture provided in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the UAV three-state switching logic provided in a specific embodiment of the present invention; Figure 5 This is a schematic diagram of the UAV differentiated rotation principle provided in a specific embodiment of the present invention; Figure 6This is a schematic diagram of the overall implementation process of UAV collaborative training provided in a specific embodiment of the present invention; Figure 7 This is a schematic diagram of the node training status sheet and gradient debt settlement process provided in a specific embodiment of the present invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0019] First, it needs to be explained that, as Figure 6 As shown, this embodiment of the invention constructs a three-state switching architecture for drones based on energy efficiency classification, aiming to address the dual challenges of communication blind spots caused by tall buildings obstructing communication in complex urban intersection scenarios and computational surges caused by sudden RSU overload. In this architecture, drones are assigned three dynamic roles with different energy consumption characteristics: a low-energy-consumption independent client, a medium-energy-consumption relay and coordination node, and a high-energy-consumption computational and coordination server.

[0020] Reference Figure 1 This invention provides a method for collaborative training of unmanned aerial vehicles (UAVs) in a decomposed federated learning approach for vehicle-to-everything (V2X) communication. The method includes the following steps: S100, construct the vehicle-to-everything (V2X) decomposition federated learning scenario for the target intersection and perform system initialization, determine the networking architecture of roadside units and drone teams, and the various energy efficiency classification business roles of drones; In this embodiment, the following was first constructed: Figure 3 The image shows a typical simulation scenario of a crossroads in a core urban area. The scenario's coverage area is set to a length of [missing information]. And the width is Within a rectangular area, an RSU is deployed at its center, with its physical coordinates set to... Simultaneously, the system deploys a formation consisting of one operational drone and M standby drones. The operational drone is restricted to flying in a safe airspace above the intersection. To ensure a stable line-of-sight (LoS) communication link between the drone and ground vehicles at the physical level, the operational drone's flight altitude is [not specified]. It meets the preset safe altitude constraints and candidate hovering point line-of-sight constraints, that is, its altitude is not lower than the minimum safe flight altitude, and the line of sight between it and the target vehicle and RSU is not blocked by obstacles.

[0021] To accurately quantify data transmission performance during the federated learning process, the system sets environmental parameters based on the current mainstream cellular vehicle-to-everything (C-V2X) communication standard. Specifically, the communication carrier frequency... The system uses the 5.9GHz frequency band specifically designed for vehicle-to-everything (V2X) connectivity, resulting in a total system bandwidth of [missing information]. Set to 20MHz, background noise power spectral density The thermal noise standard is set at -174 dBm / Hz. It should be noted that... It needs to be converted to linear power spectral density before being substituted into the noise power calculation. Background noise power is expressed as Based on the above standard parameters, the system will subsequently use Shannon's formula to calculate the real-time data transmission rate between vehicles and network nodes.

[0022] In this scenario, to achieve proactive protection of intersection traffic safety, the system first defines a core computational task – “intersection collision risk prediction.” This task aims to utilize the rich environmental data collected by vehicles and drones to predict potential traffic accident risks in real time. However, directly uploading massive amounts of raw data would cause bandwidth congestion and raise user privacy concerns. Therefore, this invention does not employ centralized training but instead initializes an SFL framework based on a ResNet-18 deep neural network. Simultaneously, three energy-efficient operating modes for the patrol drone are defined: an independent client mode responsible for introducing decomposed meta-learning for personalized data collection and feature extraction; a relay collaborative mode responsible for communication forwarding; and a computational collaborative mode responsible for gradient calculation. To clarify the energy consumption differences between the different modes, the average power consumption of the above three modes is set to a specific value. , and It satisfies a strict increasing relationship, that is Among them, the computing collaboration mode has the highest energy consumption, followed by the relay collaboration mode, and the independent client mode has the lowest. The specific power composition and energy consumption calculation method of each mode will be further given in the subsequent graded energy consumption calculation model.

[0023] In addition, to address the limitations of drone battery life and ensure service continuity, the system has established a formation rotation mechanism driven by intelligent decision-making and set a battery safety threshold. As a mandatory constraint boundary for the decision-making model, during operation, the agent at the RSU end comprehensively considers the current remaining battery power and the future traffic flow prediction results output by the spatiotemporal state prediction module, proactively plans the optimal rotation time, and outputs the rotation command. When the UAV receives the rotation command or the battery power reaches the safety threshold, triggering the mandatory rotation mechanism, and there is a standby UAV that meets the minimum takeover battery power requirement, the system will lock the current task buffer, schedule the standby UAV to reach the takeover position, complete the link verification, and then perform in-flight state synchronization and service takeover; if there is no standby UAV that meets the conditions, or the position and link verification fails, the system enters a low-power conservative mode and triggers a safe return-to-home strategy, thereby achieving a smooth service transition while ensuring flight safety.

[0024] S200: Obtain intersection environment observation data within a preset time interval, input the intersection environment observation data into the spatiotemporal state prediction module for prediction, and determine the system state prediction result for the next moment. In this embodiment, in order to solve the decision lag problem of traditional passive scheduling strategies and the gradient decay and high computational overhead defects faced by traditional long short-term memory networks when processing high-dimensional long-term traffic flows, this invention deploys a time series prediction module based on the selective state-space model (ST-Mamba) at the RSU. This module utilizes its core selective scan mechanism to filter redundant traffic noise in long sequences with linear complexity and dynamically focus on sudden environmental features, thereby making a forward-looking estimate of traffic flow and network demand in future moments.

[0025] First, the system defines the input feature space of the prediction module, and RSU is performed at fixed time intervals. Continuously collect intersection environmental data and at all times. Construct a multidimensional feature vector that includes an error feedback mechanism The feature vector contains observations in seven key dimensions: First, the total number of vehicles in the area. First, it represents the total number of vehicles participating in federated learning within the intersection area at the current moment, directly reflecting the computational load level of the RSU; second, the traffic flow rate at the blind spot boundary. This represents the total number of vehicles entering and exiting the visible boundary of the non-line-of-sight blind zone per unit time. The RSU directly monitors the vehicle entry and exit dynamics at the edge of the visible range, and the activity level of the traffic at this boundary is the most intuitive basis for extrapolating the increase in communication nodes within the blind zone; third, average driving speed It represents the average moving speed of the current traffic flow at the intersection, used to characterize traffic congestion and the rate of topology change; fourth, the channel quality change rate. This represents the first derivative of the current signal-to-noise ratio (SNR) of the communication link, used to reflect the fading trend of the wireless environment; fifth, the current remaining power of the UAV. This represents the available battery energy of the drone at the current moment, and is an important basis for predicting the remaining power at the next moment and determining whether to trigger a rotation; sixth, the UAV's expected energy consumption rate at the current moment. This represents the power consumption of the drone per unit time at the current moment, used to assess endurance pressure; seventh, the prediction deviation feedback from the previous moment. Its calculation method is the weighted Euclidean distance between the current actual observation value and the model prediction value at the previous time, which can be specifically expressed as: ,in, The actual observed subvector at the current moment. This represents the predicted value from the previous time step to the current time step. For the normalized weight matrix (during system initialization) When the value is set to 0 by default to eliminate deviation disturbances when there are no initial prediction results and ensure smooth system startup, this indicator clearly quantifies the prediction confidence of the model. When a prediction error occurs, this feedback value can be used as a correction signal input to the network, enabling the system to have self-correction and anti-interference capabilities.

[0026] Based on the above definition, time The eigenvectors are represented as: To capture long-term periodic patterns while remaining sensitive to current emergencies, the system sets the historical time window length to [value missing]. The prediction module will go past Each time step and the current moment The sequence of feature vectors as input tensors : Next, the sequence is input into the ST-Mamba network for feature extraction and temporal deduction. ST-Mamba is based on a continuous-time state-space model (SSM), which takes a one-dimensional input... Mapping to high-dimensional hidden state The evolution of a continuous system is defined as: To adapt to the discrete time-step sampling of vehicle-to-everything (V2X) networks, this invention introduces a selective step size. The continuous state matrix and input matrix are discretized using a zero-order preservation rule. Unlike traditional linear time-invariant systems, the ST-Mamba of this invention possesses data dependency characteristics, with the input matrix... Output projection matrix and selective step size All are obtained through linear projection from the current input. Dynamically generated. The discretized state transition matrix and input matrix are denoted as follows: and To distinguish it from the continuous-time state-space matrix and input matrix The calculation is as follows: in, This represents the discretized state transition matrix. This represents the discretized input matrix, and both change dynamically with the current input state. It should be noted that the above-mentioned real-time error feedback... As an adjustment selective step size An important input when prediction bias When the size is large, the network adaptively adjusts. This reduces reliance on long-term historical states and increases the weight of current real-time observation inputs in hidden state updates. (Time) The formulas for calculating the discrete hidden layer state update and output prediction are as follows: In the formula, The output projection matrix is ​​dynamically generated from the current input.

[0027] Furthermore, considering that deep learning models may experience severe prediction inaccuracies in extreme scenarios such as sensor failures or sudden traffic accidents, this invention designs a prediction bias safety circuit breaker mechanism. The system responds based on prediction bias feedback. Calculate time Normalized prediction bias rate: In the formula, This represents the actual observation subvector at the current moment. For the normalized weight matrix, A positive number set to prevent the denominator from being zero.

[0028] When prediction deviation rate Exceeding the preset circuit breaker threshold At this time, the system temporarily disables the PPO-based intelligent decision-making network, and the controller executes a preset conservative strategy, adjusting the battery safety threshold during the circuit breaker period to: In the formula, This represents the battery safety threshold under normal operating conditions. This is the temporary safe electrical threshold used during the circuit breaker period. The threshold is increased proportionally. In a preferred embodiment, That is, the temporary safe power threshold is set to 1.15 times the original safe power threshold.

[0029] The system simultaneously suspends and offloads non-essential high-power computing tasks, prioritizing the control link connection between the UAV and the ground. To prevent the system from frequently entering and exiting the circuit breaker state under critical conditions, this invention sets a hysteresis recovery condition: only when consecutive All time steps meet the requirements And satisfy When the circuit breaker is activated, the system releases the circuit breaker and smoothly returns control to the PPO agent. During the circuit breaker protection period, the system suspends writing the current interaction tuple to the reinforcement learning trajectory sampling buffer to avoid non-policy conservative rule actions interfering with the PPO policy gradient estimation.

[0030] Finally, the system maps the hidden state to a predicted output vector for the future time t+1 through a fully connected layer. The output value is processed by the ReLU linear rectifier to ensure its non-negative physical properties. This prediction vector contains three key components that directly correspond to the decision-making requirements of the three roles of the UAV: The above three components serve as the basic prediction output. They are further combined with the current observation state, the link quality change rate, and the UAV role from the previous moment to input into the role risk assessment head, which is used to obtain the role transition risk. The specific physical meaning of the output is as follows: Quantity To predict the total number of vehicles entering the intersection area to participate in federated learning at the next moment, this indicator is used to predict whether the RSU is about to face computational overload and is a key basis for deciding whether to switch to computational cooperative mode; component To predict the number of vehicles entering the non-line-of-sight blind spot in the next moment, the Roadside Unit (RSU) continuously records the last positioning data of vehicles before they enter the blind spot boundary and combines this with historical communication disconnection counts to obtain historical real sample values ​​of this indicator. This data is used for supervised training of the spatiotemporal traffic prediction module. This indicator is used to predict the existence of communication interruption risk and is a key basis for deciding whether to switch to relay cooperative mode. To predict the remaining available power of the drone on duty at the next moment under the current load trend, this indicator is used to determine whether it is necessary to reduce energy consumption in advance or trigger rotation. It is a key basis for deciding whether to maintain independent client mode or start return-to-home rotation.

[0031] Through the above steps, RSU completes the mapping from historical data to future states, transforming the prediction vector... Together with the current observation status, it is sent to the decision-making module, realizing a logical shift from passive response to active perception.

[0032] Furthermore, this invention extends the output of the prediction module to a role transition risk assessment. (Note: The last sentence appears to be incomplete and possibly refers to a separate topic.) The current business role within each decision time slot is Its values ​​correspond to three types of business roles: independent client, relay collaboration node, and computing collaboration server; the rotation control action is inactive. The value of the RSU itself serves as a control event that triggers changes in the role of adjacent decision slots or changes in the training path. RSU is based on the current observation vector. Prediction vector and the current role of drone business Calculate the probability of the character remaining unchanged in the next moment, and define the character's jump risk as: The aforementioned role retention probability can be output by a lightweight role risk assessment head after accessing the prediction vector, or estimated by a classifier trained from historical role switching samples. When the role transition risk exceeds a preset threshold, the system determines that the drone is about to undergo a role transition that may affect the SFL training loop, and initiates a node training state single protection mode in advance during the subsequent aggregation phase. If mapping to SFL training rounds is required, then... Indicates the first The corresponding drone business role during the execution of SFL training.

[0033] S300: Input the system state prediction result for the next moment and the current observation state into the strategy decision module to generate a control strategy and determine the control actions of the UAV; In this embodiment, to achieve optimal scheduling of UAV roles in a highly dynamic environment, the present invention employs the Proximity Policy Optimization (PPO) algorithm as the decision-making core, while the Reliable Utility Unit (RSU), as the intelligent agent, is responsible for collecting environmental information and executing decisions. This step not only constructs a prediction-driven decision-making framework but also establishes a detailed communication and energy consumption calculation model to accurately quantify the benefits and costs of decisions.

[0034] First, define the state space of the agent. To fully understand the system environment and assess the reliability of predictions, the state space is defined by the real-time observation vectors at the current moment. Future prediction vector Prediction error feedback Current role in drone business and the risks of role transition Pieced together, at all times The complete state vector is defined as: in, , and It directly takes the prediction output of the spatiotemporal state prediction module for the next moment, providing the agent with a key forward-looking vision; This indicates the current service role of the UAV within time slot t, with values ​​for independent client, relay collaboration, or computing collaboration server, and is converted into the corresponding numerical code before being input into the policy network. To mitigate role transition risk, this is used to characterize the impact of changes in the current or next time slot's drone service role on the stability of the SFL training path. When the value is large, the agent will increase the triggering priority of subsequent node training state sheet protection, virtual submission, and gradient debt settlement mechanisms while selecting role actions.

[0035] Secondly, the candidate range of basic actions for the intelligent agent is defined. This range includes three types of business role instructions and one rotation control instruction. The first three actions correspond to the three business roles: switching or maintaining independent client mode, switching or maintaining relay collaboration mode, and switching or maintaining computation collaboration mode; the fourth action... Corresponding to the rotation control of drone formations.

[0036] In specific decision-making time slots The current action space after safety constraints and action masking is denoted as ,satisfy: It should be noted that the alternating control action This is only used to initiate the handover process between the on-duty drone and the standby drone, and does not constitute a fourth business role. The current business role of the drone is still recorded as [missing information]. Its values ​​only include three types: independent client, relay collaboration, and computing collaboration server. Once the rotation control action is triggered, the original duty drone will continue to maintain its current operational role until the standby drone completes the takeover. It will complete the necessary communication forwarding, backend computing or frontend data collection tasks; after the handover is completed, the standby drone will inherit the business role.

[0037] When the actions of adjacent decision time slots cause changes in the drone's business role, or when rotation control causes changes in the activated upload path, backend computing node, or gradient backhaul path, the RSU treats this change as an SFL training path state change event and creates or updates the node training state sheet for the affected nodes.

[0038] Next, in order to accurately calculate the various indicators in the reward function, this invention constructs an underlying physical calculation model.

[0039] Establish a communication transmission rate model. Calculate the node's transmission rate using Shannon's formula. With nodes Uplink data transfer rate between : In the formula, For system bandwidth, For transmission power, This refers to the channel gain, which includes path loss and fading. Let be the background noise power under a linear reference, and ,in The linear noise power spectral density is obtained by converting dBm / Hz units. This represents the superimposed interference power from adjacent intersections or the environment. This rate formula is used to calculate the data transmission delay of a specific link. This, in turn, affects the overall system latency.

[0040] A tiered energy consumption calculation model is established. To accurately quantify the energy cost of different task decisions, this invention establishes a tiered energy consumption model based on the activation state of the UAV's hardware modules under different roles. The system calls predefined power tiers and combines them with the current time slot length. Calculate the execution action Energy consumption The specific calculation logic is as follows: Standalone client mode power consumption: when performing an action At this time, the drone needs to maintain flight attitude control, video sensors, and basic communication links to support local data acquisition, front-end feature extraction, and slice layer feature uploading. The energy consumption at this time is: Energy consumption in relay coordination mode: when performing an action At this time, while maintaining flight, the UAV needs to activate the communication receiving, protocol layer forwarding, and data sending modules to support bidirectional relay transmission between the blind-spot vehicle and the RSU. The energy consumption at this time is: Calculate the energy consumption of the collaborative mode: when performing an action At this time, while maintaining flight, the UAV needs to receive the slice layer activation value unloaded by the RSU, activate the onboard GPU to perform server subnet backend computation, and send the slice layer gradient or backend update summary back to the RSU. The energy consumption at this time is: Energy consumption and delay of rotating control actions: when executing actions During the handover period, the original drones will continue to maintain their current operational roles. The system completes the deployment of standby drones, link establishment, status synchronization, and service handover. The total energy consumption corresponding to the rotation control actions is expressed as: The total time delay corresponding to the rotation control action is expressed as follows: In the above formula, Indicates the duration of the current decision slot. The power required to maintain the drone's hovering posture. The sensing power for airborne cameras and radar. The average computational power when the UAV performs meta-learning inner layer updates, front-end subnet computation, and slice layer feature extraction. For data transmission power, For data reception power, For protocol layer relay forwarding power, This refers to the computational power consumed by the embedded GPU during runtime. Indicates the current business role The total energy consumption generated while maintaining existing operations and completing the handover process. This represents the total latency from the triggering of the rotation action to the standby drone completing the business takeover. Both are calculated based on the differentiated rotation process corresponding to the current business role. Using the above formula, the model can quantify the hardware energy consumption differences caused by switching between different roles.

[0041] Design of a multi-objective composite reward function. Based on the above model, this invention designs a multi-objective composite reward function. This function aims to improve the training gains of SFL models while reducing energy consumption costs, switching penalties, and quality of service risks. Its mathematical expression is: In the formula, The model training gain can be estimated by the improvement in global validation accuracy, the decrease in training loss, or their moving average. Indicates the execution of an action Energy consumption; It is the upper bound of the maximum energy consumption that each candidate action in the basic action set may generate within a decision time slot, and is used to normalize the energy consumption of actions including rotation control actions. This is a business role switching indicator function. It takes the value 1 when the UAV's business role changes within two adjacent decision time slots, and 0 otherwise. The rotation control action itself is not directly regarded as a business role switching; its additional cost is covered by rotation energy consumption. and rotation delay reflect; These are the weighting coefficients; This is a non-linear service quality penalty term, defined as follows: In the formula, The system end-to-end SFL closed-loop service latency under the current action mode. This represents the maximum tolerable latency limit for the business. This indicates the proportion of missing training samples due to communication blind spots, path incompatibility, or expired nodes. As the sensitivity scaling factor, when When the timeout is zero, the timeout penalty term is 0; when the timeout exceeds the limit, the penalty value increases exponentially with the normalized timeout ratio, thereby prompting the agent to converge to a scheduling strategy that takes into account real-time performance, energy consumption and training loop stability.

[0042] Finally, the policy update formula for the PPO algorithm is defined. The agent calculates the advantage function. The network parameters of the policy are updated iteratively using the objective function of pruning. : The probability ratio is defined as follows: In the formula, This indicates the action pair between the current policy and the old policy. The probability ratio, Trim hyperparameters for PPO. The function is used to limit the magnitude of policy updates and prevent performance collapse during training. By maximizing this objective function, the agent can learn the optimal control policy that balances SFL training efficiency and drone endurance in complex dynamic environments.

[0043] S400 maps the control actions of the drone to various energy efficiency classification business roles, controls the drone to switch to the energy efficiency classification business role, performs the corresponding mode of split federated learning collaborative training, and outputs node submission data. In this embodiment, upon receiving the control action output by the intelligent agent... Later, when At that time, the duty drone enters one of the execution branches in independent client mode, relay collaboration mode, or computational collaboration mode according to the action type; when At that time, the system will switch to a multi-drone formation rotation process and continue to maintain its current business role until the handover is completed. The corresponding services.

[0044] To intuitively illustrate the state transition relationships between the three business roles and their corresponding environmental triggering conditions, the drone's three-state switching logic is as follows: Figure 4 As shown.

[0045] Execute standalone client mode when the control action instruction is When the network is in a state of good communication and low computational load, the drone switches to a low-power operation mode, shutting down the high-power relay module and server-side GPU computing module, while retaining basic flight control, visual perception, front-end subnet computing, and basic communication modules to complete local data acquisition, personalized front-end updates, slice layer feature extraction, and feature uploading. To overcome the severe data heterogeneity caused by the drone's overhead view and the ground vehicle's head-up view, in a preferred embodiment, the present invention introduces a split meta-learning mechanism into the front-end network in this mode.

[0046] The drone and participating vehicle clients load the basic front-end subnet locally (parameters are...). Instead of directly extracting features, it utilizes its own local data sample set from its own perspective. Perform one or more inner-layer meta updates to generate custom front-end parameters with personalized perspectives. : In the formula, The learning rate is the inner layer learning rate for meta-learning. After obtaining personalized parameters, the drone and vehicle utilize... Perform forward computation to compress the original video data into high-order slice layer feature vectors. The drone then uploads the feature vector to the RSU, where the backend network completes the remaining forward and backward propagation calculations. The gradients of the slice layers calculated by the RSU are then fed back to the drone to update its local frontend parameters. The core value of this mode lies in enriching the sample distribution of the global model using the drone's unique aerial perspective data, and alleviating the convergence bottleneck caused by heterogeneous data through meta-learning. The execution latency of this branch is expressed as: in, The total computational cost includes both inner-layer meta-updates and front-end feature extraction. For the drone's CPU frequency, For the size of the feature data, The gradient magnitude of the slice layer. and These represent the uplink and downlink transmission rates between the drone and the RSU, respectively.

[0047] Execute relay coordination mode when the control action indication is When the prediction module detects a communication requirement in the blind zone, the on-duty UAV activates the communication relay module and maneuvers to a preset relay hovering point. This hovering point must meet geometric constraints, maintaining visual line-of-sight for both the RSU and vehicles in the blind zone. After establishing a connection, the system performs bidirectional protocol layer forwarding: in the uplink, the vehicle in the blind zone sends its calculated slice feature data to the UAV, which then forwards it to the RSU; in the downlink, the RSU sends its calculated slice gradient to the UAV, which then forwards it back to the vehicle. The total closed-loop latency of SFL training in this mode is [not specified]. It consists of the sum of the transmission delay of the two links and the calculation delay at the RSU end: In the formula, , The numbers represent the link rates between the vehicle, the drone, and the RSU, respectively. This represents the computational cost per unit sample for the subsequent subnetting. This branch represents the computing frequency of RSU, reducing the probability of blind zone nodes falling behind due to communication interruptions.

[0048] Execute computational cooperative mode when the control action instruction is When this occurs, it indicates that the RSU is facing a risk of computational resource overload. At this time, the patrol drone activates its high-power onboard GPU module and activates the pre-loaded SFL server subnet in memory as a temporary edge server to share the backend computing pressure of the RSU. This invention pre-splits the ResNet-18 model at specific layers, so the ground vehicle only needs to retain the lightweight client subnet and upload the slice layer activation values ​​calculated by the local front-end network to the RSU. The RSU then unloads the data according to a preset offload ratio. The activation values ​​of some vehicle slices are scheduled and sent to the UAV, which then takes over the forward and backward propagation calculations of the corresponding sample's service terminal network backend; those not unloaded Some backend calculations are still completed by the RSU local service subnet. After the UAV completes backend backpropagation, it sends the corresponding vehicle slice gradient or backend update summary back to the RSU. The RSU then distributes the slice gradient to the vehicle client to complete this round of SFL closed-loop update. Since the intermediate layer activation values ​​are smaller than the original image data and do not directly expose the original sensor data, this mode reduces the privacy exposure risk caused by uploading the original data while sharing the computational burden on the RSU.

[0049] It should be noted that the total service latency in this mode is not the individual UAV computation latency, but rather the system end-to-end latency for one round of SFL closed-loop training for effective vehicles within line of sight in the computational cooperative mode. Since the UAV does not provide relay services in the computational cooperative mode, vehicles in non-line-of-sight blind spots are temporarily unable to participate in this round of training due to communication disruptions. Therefore, the number of effective participating vehicles within line of sight is defined as: in, This indicates the total number of vehicles in the intersection area at the current moment. This represents the number of vehicles that objectively exist in non-line-of-sight blind spots and cannot be connected to the training link under the current computing collaboration mode. This indicates the number of valid vehicles that were able to complete activation uploads and participate in backend calculations through RSU in this round. Vehicles missing from blind spots are included. The service quality penalty is fed back to the PPO agent, causing the agent to tend to choose the relay coordination mode rather than continue to execute the computation coordination mode when the number of vehicles in the blind spot is large.

[0050] Under the above definition, the total system service latency in the computing collaborative mode can be expressed as: Among them, the first item The second term represents the baseline latency for valid vehicles within line of sight to upload slice layer activation values ​​to the RSU; This indicates the dominant backend processing latency when the RSU local backend compute branch and the UAV offload compute branch are executed in parallel; the third item... This represents the backhaul latency of the RSU sending the slice layer gradient to the vehicle client. If the system uses parallel radio resource scheduling, the above upload and download latency can also be equivalently represented by the upper bound of the scheduling window or the maximum single-node transmission latency among the participating vehicles.

[0051] Specifically, the latency of the RSU local backend compute branch is: The delay for calculating the UAV unloading branch is: In the formula, This indicates the proportion of samples where the RSU is offloaded to the drone for backend computation. This indicates the size of the slice layer activation value data uploaded by a single vehicle; This represents the size of the gradient data for a single vehicle slice. This represents the computational cost required for a single sample to complete both forward and backward propagation at the back end of the service terminal network. Indicates the RSU calculation frequency; This indicates the computing frequency of the drone's onboard GPU; This indicates the uplink transmission rate from the vehicle to the RSU; This indicates the offloading transfer rate from the RSU to the drone; This indicates the gradient or update summary return rate from the UAV to the RSU. This represents the downlink gradient backhaul rate from the RSU to the vehicle. Based on the above modeling, Some samples were further calculated by RSU. Some samples are computed by UAVs, and the total back-end processing latency of the system is determined by the longer of the two parallel branches, thus enabling a more accurate characterization of the computational collaboration mode's effect on reducing RSU queuing latency.

[0052] It should be noted that, , and These represent the latency of a single round of SFL closed-loop service in independent client mode, relay collaboration mode, and computing collaboration mode, respectively. and This can represent the closed-loop latency of a single training node; if the same round contains multiple nodes, the total system latency can be taken as the maximum closed-loop latency among the participating nodes, or determined by a unified scheduling window. Apply constraints. In collaborative computing mode, This represents the system end-to-end latency for completing one round of SFL closed-loop training on the effective set of vehicles within line of sight. The RSU local backend computation branch and the UAV offloading computation branch are executed in parallel. The system backend processing latency is determined by... Characterization.

[0053] A multi-drone formation rotation mechanism addresses the limitation of a single drone's onboard battery capacity in sustaining prolonged high-load operations, and the drastically different service continuity requirements for different roles (communication interruptions lead to disconnections, and computational interruptions result in the loss of training gradients). Therefore, this invention designs a task role-adaptive "state-continuous relay" mechanism. This mechanism is triggered by decision commands. When the on-duty drone's battery reaches a threshold or a rotation command is received, the system does not directly execute a physical replacement. Instead, it performs a differentiated process based on the characteristics of the current SFL task flow. Scenario 1: The duty drone is in relay coordination mode, and the system executes the "lateral escort dual-link hot switching" strategy.

[0054] When the system detects that the on-duty drone is providing communication relay services to the blind spot, the standby drone will not directly replace its position. Instead, it will fly to a safe lateral distance from the airspace where the on-duty drone is located and establish a parallel line-of-sight link. At this time, a "dual-link parallel period" is formed between the roadside unit (RSU) and the vehicle in the blind spot, maintained by the two drones. The advantage of this is that the vehicle in the blind spot can switch its logical target address or resource block configuration according to the link redirection command issued by the RSU, thereby reducing the probability of communication interruption during equipment switching and ensuring that the vehicle in the blind spot is always within the effective coverage range of the communication signal during equipment switching, thus achieving a smooth transition of communication services at the physical link level. To avoid signal interference caused by close-range communication between the two drones, the system adopts an orthogonal resource allocation strategy during this parallel period, orthogonally dividing the communication channels of the standby drone and the on-duty drone in the frequency domain or time domain to avoid co-channel interference.

[0055] The system will continuously monitor the link quality of the standby drone. Only when it is confirmed that the signal-to-noise ratio stably exceeds the service quality threshold will the RSU issue a logical link redirection command to the vehicle in the blind spot via the original on-duty drone. After receiving the command, the vehicle will switch its uplink target address to the standby drone, and then the system will disconnect the on-duty drone. This "establish a new connection first, then disconnect the old connection" approach significantly reduces the risk of communication interruption in blind spots due to equipment replacement at the physical level.

[0056] Scenario 2: The duty drone is in computing collaboration mode, and the system executes the "lightweight migration based on training round boundaries" strategy.

[0057] When the patrol drone is performing gradient calculations as an edge server, to avoid data loss due to forced interruption, the system cleverly utilizes the periodic round-by-round characteristics preserved by the "semi-asynchronous aggregation elastic time window" for soft switching. Specifically, to prevent power depletion during the waiting period, the system sets a safety threshold. The system has reserved extra power redundancy to complete a single round of SFL training, ensuring that even if a low power alarm is triggered at the start of a round, the drone will still have enough power to complete the remaining computation tasks of the current round, thus avoiding gradient loss caused by forced interruption.

[0058] With this safety redundancy, the rotation process is as follows: First, while the standby drone is en route to the duty station, the RSU already holds the backend model parameters and static network structure after the most recent global aggregation. Therefore, the standby drone can preload relevant parameters from the RSU via a high-bandwidth backhaul link, without requiring the duty drone to transmit the complete model to the standby drone. Second, the system locks the new task assignment and allows the duty drone to continue completing any unfinished computational tasks from the current training round. Finally, within the handover window after the current training round ends, the duty drone only synchronizes the task progress flag, cache index, and a filtered or compressed summary of the necessary optimizer state to the standby drone. The amount of the above-mentioned continuation data is denoted as [data missing]. Its synchronization delay is expressed as: In the formula, This represents the amount of migration data required to complete the next round of computation tasks, excluding complete model parameters and intermediate activation data during training. This indicates the air-to-air transmission rate between the active-duty drone and the standby drone. For optimizers that require maintaining complete first-order or second-order moment states, the relevant optimizer states are preferably saved uniformly by the RSU and restored when the standby drone completes its takeover, rather than being directly transmitted between the two drones. After receiving the takeover confirmation, the standby drone takes over the next round of computation tasks, thereby avoiding the migration of complete model parameters and a large amount of intermediate training data between drones.

[0059] Scenario 3: When the duty drone is in stand-alone client mode, the system executes a low-overhead replacement strategy at the boundary of the training batch.

[0060] It should be noted that when the duty drone is in independent client mode, since it mainly undertakes local data collection and front-end feature extraction tasks, and does not involve blind zone relay link hot switching or server back-end computing state migration, the system can perform low-overhead replacement after the current small batch of front-end updates is completed. That is, it synchronizes the global front-end initialization parameters, node identity identifiers, local cache indexes and incomplete status sheet numbers, and the standby drone takes over the independent client tasks in the next sampling time slot.

[0061] To intuitively demonstrate the execution flow of the aforementioned differentiated rotation strategy in the time domain, particularly the dual-link overlapping timing in relay mode and the round-interval migration timing in computation mode, this invention constructs a schematic diagram of the UAV adaptive continuous rotation timing principle, as follows: Figure 5 As shown.

[0062] S500, based on the global aggregation stage, updates the node submission data generated by the split federated learning collaborative training through the node training status sheet and gradient debt settlement mechanism to determine the result of the current round of global model update, thereby realizing the UAV collaborative training of split federated learning.

[0063] In this embodiment, as Figure 7 As shown, within each training epoch, RSU acts as the central aggregation server of the system to maintain the node training state and performs a global model update after the aggregation window ends. To address the problems of traditional SFL strict synchronization mechanisms easily leading to lag and blocking, ordinary asynchronous mechanisms easily disrupting epoch boundaries, and generative proxy activation potentially causing missing nodes to have outdated front-end models, this invention implements a role-transition-aware node training state sheet and gradient debt settlement mechanism. This mechanism establishes a node training state sheet for each node in one epoch of SFL training to record whether it has completed forward activation upload, back-end backpropagation, slice layer gradient backpropagation, and front-end parameter updates, instead of only recording one activation upload event.

[0064] First, RSU creates a node training status sheet for participating nodes at the start of each training round: in, Number the nodes. For SFL training rounds, Indicates the first The corresponding drone service role during SFL training execution. This indicates the model slice layer used in the nth round. This indicates the corresponding backend model version number. This indicates that the upload path is activated. Indicates the backend computing execution node. Indicates the gradient backpropagation path. , , , These indicate whether activation upload, backend backpropagation, gradient backpropagation, and frontend update are complete, respectively. If the calculated role transition risk exceeds a preset threshold, or if the output control action causes changes to the drone business role, activation upload path, or gradient backpropagation path, including but not limited to arbitrary switching and rotation actions between independent clients, relay collaboration, and computational collaboration modes, then the RSU enters the node training status sheet protection mode, locking the status sheet number and path information of the current training round.

[0065] Secondly, RSU sets a maximum tolerance aggregation time window, and continuously records the real slice layer activation upload, backend backpropagation, slice layer gradient backpropagation, and frontend update completion status of each node within the window. After the window ends, the system divides the nodes into three categories: real commit set, virtual commit set, and expired set, as detailed below: Furthermore, none of the three overlap. The set of nodes that successfully upload the actual slice layer activation value within the window period and whose paths remain capable of being uploaded back constitutes the actual commit set. Nodes that failed to upload real activations in a timely manner but whose historical activation credibility meets the threshold and can maintain the continuity of backend training through proxy activation constitute a virtual submission set. Nodes with insufficient credibility, excessively long missing periods, or incompatible role paths constitute an expired set. For nodes in the real submission set, the system completes the backend update, slice layer gradient backpropagation, and frontend update according to the regular SFL process; for nodes in the expired set, the system does not use their proxy samples to participate in this round of updates to avoid unreliable features polluting the model.

[0066] For nodes in the virtual submission set, RSU can generate surrogate activations using historical activation caching, lightweight generators, or other distribution estimation methods. These surrogate activations, along with the real activations, are then input into the backend network to maintain the continuity of backend training in this round. It should be noted that the method of generating surrogate activations is not the core limitation of this invention. The core of this invention lies in the fact that once any surrogate activation participates in the backend model update, the system must synchronously record its unpaid gradient responsibility to the client-side frontend model.

[0067] Specifically, when a proxy activation participates in the backend loss calculation, RSU calculates the backward gradient of the slice corresponding to that proxy activation: And generate gradient debt slips: in, For nodes In the Wheel agent activation, Activate the inverse gradient of the corresponding slice layer for the proxy. To activate the agent's credibility, The number of rounds for a node to remain missing. , , These represent the activation upload path, backend computation execution node, and gradient return path corresponding to the debt, respectively. This represents the debt's validity period. This gradient debt statement indicates that the backend model has already been based on nodes. The proxy sample has been updated, but the node... The front-end model has not yet received the slice layer gradient corresponding to the proxy sample.

[0068] To facilitate characterizing the compatibility of the complete training path, nodes are defined. In the The training path triples for the round are: When node When reconnecting and uploading actual activation data in subsequent rounds, the RSU first calculates the debt compatibility score: in, Indicates the credibility of the agent activation. This indicates the statistical distribution difference between proxy activation and reconnection-based genuine activation. This represents the debt decay factor over time. This indicates a training path triplet compatibility indicator function.

[0069] The scoring comprehensively considers the statistical distribution differences between the current real activation and the proxy activation in the debt order, the duration of node missing, and the compatibility of drone role paths. If the compatibility score is lower than a preset threshold, the corresponding debt automatically expires; if the compatibility score meets the threshold, the system performs gradient debt liquidation. In the main embodiment, the system uses a preset fixed slice layer. When the compatibility score meets the threshold, the system does not unconditionally use historical gradients directly, but uses the historical slice layer gradients in the debt order as candidates for front-end compensation gradients, and limits their influence through compatibility score, liquidation weight, and gradient pruning. In other optional embodiments, if the system uses a dynamic slice layer configuration, a cross-layer gradient mapper can be further set. This is used to map historical debt gradients to the current slice layer space.

[0070] During the debt settlement process, the client-side front-end model parameters no longer rely solely on the current round's true gradient updates, but instead incorporate a credible debt compensation term: in, This indicates the learning rate of the client-side front-end model update. The front-end gradient generated for the current real sample. This is the front-end compensation term for the debt slice layer gradient after the current front-end Jacobian mapping; in the main embodiment with a fixed slice layer... The gradient is obtained from the historical slice layer gradient in the debt bill after compatibility gating; in the optional embodiment of dynamic slice layer, For cross-layer mapper Processed debt gradient; This represents the debt liquidation weight. The clip function is used to limit the excessive perturbation of the model by old debts. Through this update, the nodes... The front-end model can compensate for previous back-end updates triggered by proxy activation, thereby reducing the risk of semantic inconsistency between the front-end and back-end models.

[0071] Finally, RSU updates the backend model parameters based on real submission nodes and virtual submission nodes with credibility gating, and can use exponential moving average to smooth the global parameters. It should be noted that debt settlement results in node-level frontend compensation, which is not directly used as a unified frontend gradient shared by all nodes; the compensation update after compatibility gating can first be applied to the corresponding node, and then participate in subsequent frontend aggregation or personalized updates. After completing the global parameter update, RSU distributes the updated global frontend initialization parameters to vehicle and drone clients via downlink broadcast channel, and synchronizes the updated backend subnetwork parameters, model version number, or parameter update summary to the RSU-side backend network and the drone server with computing capabilities; the frontend compensation update formed by node-level debt settlement is applied to the corresponding node and participates in synchronization in subsequent frontend aggregation or personalized updates with weight constraints. If the global model loss converges to a preset threshold or the training epochs reach the maximum limit, training stops and the final intersection collision prediction model is output; otherwise, the system starts a new round of prediction, role scheduling, and node training state single protection loop.

[0072] Regarding the construction and constraints of the overall objective optimization function, given that blind spot non-line-of-sight vehicle data has irreplaceable value in improving the generalization ability of intersection collision prediction models, and that the energy consumption levels and service gains corresponding to different role modes of UAVs vary, the core objective of this invention is defined as follows: under the premise of prioritizing the physical power of the UAV's onboard battery and the maximum tolerable latency constraints of the vehicle-to-everything (V2X) service, the system, through the dynamic optimization of the three-state role of the UAV by the intelligent agent, maximizes the global model accuracy gains and system service stability while controlling the energy consumption cost of the UAV.

[0073] Based on the project objectives, this invention establishes a prediction-driven decision model with physical constraints, transforming the UAV three-state cooperative scheduling problem into a long-term cumulative reward maximization problem under the constraints of power, latency, and SFL training closed loop. The PPO agent outputs role switching or rotation control actions within this constraint framework.

[0074] The overall objective function is to find the optimal policy. This enables the system to operate throughout the entire training cycle. Maximize the expected reward of accumulated discounts within: In the formula, This is a discount factor used to balance the importance of current rewards and future long-term benefits, where instantaneous rewards... It is composed of a weighted average of model training gains, energy costs, handover penalties, and non-linear quality of service penalties: In the formula, The model training gain can be estimated from the improvement in global validation accuracy, the decrease in training loss, or their moving average. The service quality penalty term is: The solution to the above optimization problem is subject to the following three sets of physical constraints and quality of service penalty conditions: C1 represents the role mutual exclusion and discreteness constraint, which specifies the control actions output by the agent. It must belong to the current actionable space. ,Right now: The first three items are actions performed by business roles. This is a rotation control action. Within any decision time slot, the agent outputs only one control action, and the on-duty drone only exists in one of the following business roles: independent client, relay coordination, or computing coordination server. The rotation control action does not constitute a new business role; when the rotation action is triggered, the original on-duty drone continues to maintain its current business role until the handover is completed, after which the standby drone inherits the role.

[0075] C2 represents the federated learning latency constraint. Except for the forced rotation or safe return scenario triggered by the C3 energy safety constraint, the system uses an action mask to retain only those conditions that meet the constraint. The system selects candidate actions; for rotation or conservative control actions that must be performed due to energy security constraints and may exceed the latency threshold, the system uses a service quality penalty item. Severe penalties will be imposed, with energy security constraints as the highest priority.

[0076] C3 represents energy security and mandatory rotation constraints. When the remaining battery power of the operational drone... Falling to or below the safety threshold And there is a requirement to meet the minimum power requirements for takeover. When a standby drone is in operation, the system will determine the current available operational space. Locked for rotating actions The location arrival conditions and link establishment conditions of the standby drone are verified during the rotation action execution phase. If the verification is not met, the system enters a low-power conservative mode and triggers a safe return-to-home strategy.

[0077] Furthermore, to solve the defined complex optimization problem in a continuous state space and a stochastic dynamic environment, this invention employs a proximal policy optimization (PPO) algorithm based on an Actor-Critic architecture for model training and inference. This algorithm ensures the stability and convergence of the training process by limiting the step size of policy updates. The specific implementation process is as follows: Network initialization: The system initializes two deep neural networks: a policy network (Actor). The parameters are This is used to output the probability distribution of actions in a value network (Critic). The parameters are This is used to evaluate the value of the current state and initialize the trajectory sampling buffer to store the sampled trajectory data.

[0078] In each iteration, the Roadside Unit (RSU) acts as an agent, interacting with the environment. The interaction logic follows set action space and energy safety constraints, and the latency priority constraint is reflected through the service quality penalty term in the reward function. First, the agent acquires the current observation vector. And combined with the output prediction vector Current role in drone business and the risk of role transition Construct composite states Next, action masking is performed: Before inputting the policy network, the system performs action masking based on constraint C3. When the battery level of the on-duty drone drops to or below the safety threshold and there is a standby drone with sufficient battery power to take over, the system only retains the rotation action; if there is no standby drone with sufficient battery power to take over, or if the location and link availability of the standby drone fails to be confirmed after it arrives, the system switches to a low-power conservative mode and performs a safe return. Subsequently, the agent samples actions based on the processed probability distribution. After the drone performs this action, the environmental state changes from Transfer to and provide feedback on reward values. Finally, the system checks the circuit breaker status of the current spatiotemporal traffic prediction module: if the system is in the "prediction bias safety circuit breaker" protection period, the RSU will immediately intercept and suspend the input of the current interaction tuple to the trajectory sampling buffer to prevent non-strategic conservative rule actions from being used to update the network; if the system is in a normal intelligent decision-making state, the interaction tuple will be processed normally. Store the data in the trajectory sampling buffer until the buffer is full. Data at each time step.

[0079] Dominance function estimation: The generalized dominance estimation (GAE) algorithm is used to calculate the dominance value at each time step. To reduce the variance of gradient estimation, the time-series difference error is first calculated. : Then recursively calculate the dominance value. : In the formula, This is the GAE smoothing factor, used to control the weights of multi-step guidance.

[0080] The policy network (Actor) is updated by randomly sampling mini-batch data from the buffer to construct the pruning objective function specific to the PPO algorithm. The stochastic gradient ascent method is used to update the policy network parameters. The objective function aims to maximize the policy improvement while preventing excessive updates. Among them, probability ratio , Trim hyperparameters for PPO. The function limits the ratio to Within the range.

[0081] The Critic value network is updated to calculate the predicted value of the state. Compared with the target value of true return The mean squared error (MSE) between the two values ​​is calculated, and the value network parameters are updated using gradient descent. To improve the accuracy of value assessment: in, The final iteration and convergence are repeated, the buffer is cleared, and the next iteration begins. When the total objective function value... Training stops when the improvement no longer becomes significant or when the preset maximum number of training epochs is reached, and the converged optimal policy model is output. Used for online inference control.

[0082] In summary, the embodiments of the present invention have the following technical distinguishing features compared with the prior art: 1) This invention proposes a three-state dynamic collaborative architecture for UAVs oriented towards split federated learning, which enables the same UAV to controllably switch between three business roles: independent client, relay collaborative node, and computing collaborative server. This allows the UAV to no longer be limited to a fixed relay, fixed server, or fixed client, but to dynamically assume different SFL business roles based on communication blind spots, RSU computing overload, and endurance constraints.

[0083] 2) This invention proposes a continuous control mechanism that incorporates UAV rotation control into a three-state collaborative system. The formation rotation action is treated as a control action independent of the three business roles, enabling UAVs to maintain the continuous execution of communication, computing, or front-end data acquisition tasks through standby UAVs when power is limited, without disrupting the definition of the three business roles themselves.

[0084] 3) This invention proposes a predictive constraint-driven three-state scheduling mechanism for unmanned aerial vehicles (UAVs). The spatiotemporal state prediction module outputs the prediction results of future traffic demand, blind spot demand, and remaining battery power of the UAVs. The role risk assessment head outputs the role transition risk by combining the current link quality change rate and the current business role of the UAVs. The strategy decision module selects the three-state business role or rotates control actions under the conditions of action mask, battery threshold, latency constraint, and service quality penalty. This transforms UAV scheduling from a passive response based on instantaneous state to an active scheduling that addresses future communication bottlenecks, computing bottlenecks, and endurance risks.

[0085] 4) This invention proposes an SFL training path protection mechanism for role transition risk, which treats changes in UAV business roles, rotation control, activation upload path, backend computing node, and gradient backhaul path as SFL training path state change events, and triggers a node training state single protection mode when the role transition risk exceeds the threshold or the control action causes a change in the training path.

[0086] 5) This invention proposes a node training state single mechanism for role transition awareness, which records the UAV business role, model slice layer, model version, activation upload path, backend computation execution node, gradient backhaul path, and activation upload, backend backhaul, gradient backhaul and frontend update completion flags for each participating node in each round of SFL training, so that the system can record the complete training closed loop state, rather than just recording whether the node has uploaded and activated.

[0087] 6) This invention proposes a semi-asynchronous aggregation management mechanism based on three types of states: real submission, virtual submission, and expiration. While preserving the training round boundaries, nodes are divided into real submission set, virtual submission set, and expiration set according to the real activation arrival status, proxy activation credibility, missing time, and path compatibility. This reduces strict synchronization blocking while avoiding untrusted proxy samples from polluting the model.

[0088] 7) This invention proposes a gradient debt order generation mechanism that is linked with the virtual submission node. When the agent activates to participate in the training of the service subnet backend, the system synchronously records the agent activation, the reverse gradient of the slice layer corresponding to the agent activation, the confidence level, the number of missing cycles, the training path and the debt validity period, forming a gradient debt order, which is used to characterize the unpaid state of the backend model that has been updated based on the agent sample but the corresponding client frontend has not yet obtained the compensation gradient.

[0089] 8) This invention proposes a gradient debt liquidation mechanism based on statistical differences, time decay, and training path triple compatibility. When a missing node reconnects and uploads a real activation, the system calculates a debt compatibility score based on the statistical distribution differences between the proxy activation and the real activation, the debt duration, and the compatibility of the training path triple composed of the activation upload path, the backend computing node, and the gradient backhaul path. Only when the score meets the threshold is the qualified debt gradient backed to the corresponding client frontend model, thereby compensating for the semantic offset between the frontend and backend models caused by the proxy activation participating in the backend training but the client frontend not being updated synchronously.

[0090] 9) This invention proposes a differentiated continuous rotation mechanism for different UAV business roles. In the relay collaboration mode, a side-flying dual-link hot-switching strategy is adopted. In the computing collaboration mode, a lightweight state transition strategy based on the training round boundary is adopted. In the independent client mode, a low-overhead replacement strategy at the training mini-batch boundary is adopted. In this way, the continuity of communication links, computing services and front-end data acquisition tasks are maintained under the conditions of limited UAV power and formation relay.

[0091] Therefore, the embodiments of the present invention have the following advantages compared with the prior art: 1) This invention breaks through the resource isolation barriers of single-role systems, significantly improving equipment utilization and system flexibility. Existing technologies typically fix drones as single roles such as communication relays or mobile edge servers, leading to resource mismatches in mixed scenarios, such as idle relay drones during computing overload. This invention proposes a three-state dynamic switching architecture for drones, enabling the same hardware device to flexibly switch between independent client, relay collaboration, and computing collaboration modes according to real-time environmental requirements. This mechanism breaks the resource isolation of traditional solutions, ensuring relay connections for vehicles in communication blind spots and computing power sharing for RSUs in computing overload scenarios, thus improving the utilization efficiency of drone hardware resources under different business bottlenecks.

[0092] 2) This invention addresses the issues of SFL training link disruption and synchronization blockage caused by UAV role transitions. During the execution of three-state scheduling, when a UAV undergoes any service role switch, training path change, or formation rotation, the activation upload path, backend computing nodes, and gradient backhaul links in SFL may change synchronously. If a strict synchronization mechanism is used, the entire network will be blocked waiting for disconnected nodes; if nodes are discarded directly, sample bias will occur; if only proxy activation is used to update the backend network, the frontend model of the missing node will not be able to obtain the corresponding gradient. This invention uses a node training state slip protection mechanism to divide node states into real submission, virtual submission, and expiration. During virtual submission, a trusted proxy activation is used to maintain the continuity of backend training and synchronously generate a gradient debt slip, providing a basis for subsequent frontend compensation updates after node reconnection.

[0093] 3) To avoid semantic drift between the front-end and back-end models caused by proxy activation and ensure training consistency under three-state switching, the three-state dynamic architecture introduces a deeper training consistency problem: when RSU uses proxy activation to maintain back-end network updates, the front-end model of missing nodes does not synchronously receive the corresponding backpropagation gradient, and the front-end feature space and the back-end discrimination space will gradually deviate. To address this, this invention proposes a gradient debt liquidation mechanism. When proxy activation participates in back-end training, its slice layer backpropagation gradient, credibility, and validity period are recorded. After the node is reconnected, the liquidation is determined based on the statistical difference between the real activation and the proxy activation, role compatibility, and debt validity. This mechanism can compensate the client front-end model without interrupting back-end training, reducing model semantic drift caused by role switching, node disconnection, and proxy activation.

[0094] 4) This invention achieves a transformation in control mode from "passive response" to "anti-interference active perception," improving system robustness and stability. Addressing the issues of decision lag and system oscillation caused by the lack of environmental prediction in traditional scheduling strategies, as well as the accuracy decline of traditional prediction algorithms under sudden situations, this invention introduces a spatiotemporal traffic prediction module based on the selective state-space model ST-Mamba, which possesses real-time error correction capabilities. This module can combine real-time observation data to anticipate future traffic flow density and blind spot demand trends, and dynamically calibrate prediction results using historical deviation feedback. This provides forward-looking and highly confident state input for deep reinforcement learning decision-making, enabling the system to shift from "processing congestion after it occurs" to "predicting congestion and switching first." Furthermore, it possesses self-adjusting capabilities when predictions are inaccurate, effectively reducing ineffective and frequent switching of UAVs between different roles, lowering switching overhead, and improving the smoothness of the control strategy.

[0095] 5) A seamless relay mechanism based on role adaptation was established, which improved business continuity during long-term operation. Unlike the simple "stop-rotate-restart" strategy in the existing technology, this invention designed "lateral escort hot-switching" and "training boundary state migration" mechanisms for the different business characteristics of UAV relay and computing roles. The mechanism reduces the risk of handover failure in blind zone communication by "establishing a new connection first and then disconnecting the old connection" strategy, and reduces the bandwidth consumption caused by computing task interruption and massive data migration by utilizing the periodic characteristics retained by the semi-asynchronous architecture, thereby improving the business continuity and high reliability of vehicle network safety services during long-term operation.

[0096] Reference Figure 2 A drone collaborative training system for decomposed federated learning in the context of vehicle-to-everything (V2X) includes: The first module 201 is used to construct the vehicle-to-everything (V2X) split federated learning scenario at the target intersection and perform system initialization, determine the networking architecture of roadside units and drone teams, and the various energy efficiency classification business roles of drones. The second module 202 is used to acquire intersection environmental observation data within a preset time interval, input the intersection environmental observation data into the spatiotemporal state prediction module for prediction, and determine the system state prediction result for the next moment. The third module 203 is used to input the system state prediction result for the next moment and the current observation state into the strategy decision module to generate a control strategy and determine the control actions of the UAV. The fourth module 204 is used to map the control actions of the drone to various energy efficiency classification business roles, control the drone to switch to the energy efficiency classification business role, perform the corresponding mode of split federated learning collaborative training, and output node submission data. The fifth module 205 is used to update the node submission data generated by the split federated learning collaborative training through the node training status sheet and gradient debt settlement mechanism based on the global aggregation stage, to determine the result of the global model update in this round, and realize the UAV collaborative training of split federated learning.

[0097] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this is not intended to limit the scope of the embodiments of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the embodiments of the present application.

Claims

1. A collaborative training method for unmanned aerial vehicles (UAVs) based on decomposed federated learning for the Internet of Vehicles (IoV), characterized in that, Includes the following steps: Construct a vehicle-to-everything (V2X) decomposed federated learning scenario for the target intersection and perform system initialization; determine the networking architecture of roadside units and drone teams, as well as the various energy efficiency classification business roles of drones. The system acquires intersection environmental observation data within a preset time interval, inputs the intersection environmental observation data into the spatiotemporal state prediction module for prediction, and determines the system state prediction result for the next moment. The system state prediction results for the next moment and the current observation state are input into the strategy decision module to generate a control strategy and determine the control actions of the UAV. The control actions of the drone are mapped one-to-one with various energy efficiency classification business roles. The drone is controlled to switch to the energy efficiency classification business role, execute the corresponding mode of split federated learning collaborative training, and output node submission data. Based on the global aggregation phase, the node submission data generated by the split federated learning collaborative training is updated through the node training status sheet and gradient debt settlement mechanism to determine the result of the global model update in this round, thereby realizing the UAV collaborative training of split federated learning.

2. The UAV collaborative training method for decomposed federated learning in the Internet of Vehicles according to claim 1, characterized in that, The step of acquiring intersection environmental observation data within a preset time interval, inputting the intersection environmental observation data into the spatiotemporal state prediction module for prediction, and determining the system state prediction result for the next moment specifically includes: Obtain multi-dimensional intersection environment observations at the current moment and construct the observation feature vector at the current moment; Input the observation feature vectors of multiple frames within a preset historical window into the temporal prediction network, extract the temporal features, and output the system state prediction vector for the next moment. Calculate the prediction deviation between the observed feature vector at the current moment and the system state prediction vector at the next moment, and input the prediction deviation as a feedback term into the next moment prediction process of the spatiotemporal state prediction module. If the prediction deviation exceeds the preset circuit breaker threshold, the predictive safety circuit breaker mechanism is triggered, switching to a conservative control strategy and adjusting the battery safety threshold. The conservative control strategy and the adjusted battery safety threshold are used as constraint outputs of the system state prediction result to determine the system state prediction result at the next moment.

3. The UAV collaborative training method for decomposed federated learning in the Internet of Vehicles according to claim 2, characterized in that, The step of inputting the system state prediction result for the next moment and the current observation state into the strategy decision module to generate a control strategy and determine the control actions of the UAV specifically includes: The decision state vector is obtained by concatenating the current observation state, the system state prediction vector in the system state prediction result of the next moment, the prediction deviation, the current UAV business role, and the role transition risk. Based on preset action constraint rules, the candidate action set is masked and filtered to obtain the current set of actionable actions; Input the decision state vector into the reinforcement learning policy network, and output the probability distribution of each action in the current set of possible actions; The control actions of the drone are obtained by sampling based on the probability distribution of each action in the current set of possible actions.

4. The UAV collaborative training method for decomposed federated learning in the Internet of Vehicles according to claim 3, characterized in that, The training reward value of the reinforcement learning policy network is determined through the following steps: Obtain the model accuracy gain value for this round of split federated learning as a training benefit; Calculate the drone energy consumption value corresponding to the current control action, and normalize the drone energy consumption value as the energy consumption cost item; When the role of the drone service changes in adjacent decision slots, a role switching penalty is applied. Calculate the system service latency and training sample missing ratio corresponding to the current control action, and generate a nonlinear quality of service penalty term; The training reward value of the reinforcement learning policy network is obtained by weighting the training reward, energy consumption cost, role switching penalty, and service quality penalty.

5. The UAV collaborative training method for decomposed federated learning in the Internet of Vehicles according to claim 4, characterized in that, The control actions of the UAV include rotating control actions. After determining the control actions of the UAV, the system also includes rotating execution steps, which specifically include: When the drone is currently in relay coordination mode, control the standby drone to establish a parallel communication link, complete the link verification, switch the service connection, and then disconnect the original drone link. When the drone is currently in a computational collaboration mode, the standby drone is controlled to preload model parameters. After the current training round ends, the task status and optimizer summary are synchronized, and the drone takes over the next round of computational tasks. With the drone currently in independent client mode, once the current small batch of front-end calculations is completed, the node identity and cache index are synchronized, and the standby drone takes over the subsequent data collection and calculation tasks. After the rotation is completed, the standby drone will take over the current business role and continue to perform split federated learning and collaborative training.

6. The UAV collaborative training method for decomposed federated learning in the Internet of Vehicles according to claim 5, characterized in that, The step of mapping the control actions of the drone to various energy efficiency classification business roles, controlling the drone to switch to the energy efficiency classification business role, performing split federated learning collaborative training in the corresponding mode, and outputting node submission data specifically includes: When the target business role is in independent client mode, control the drone to shut down the high-power computing and forwarding modules, perform local data acquisition, front-end feature extraction and slice layer feature uploading, and generate client node submission data; When the target business role is in relay collaboration mode, control the drone to fly to the preset relay point, establish a two-way communication forwarding link between the blind spot vehicle and the roadside unit, and ensure the upload of slice layer features and gradient back transmission of the blind spot node. When the target business role is in a computing collaboration mode, the drone is controlled to activate the onboard computing module, take over part of the backend model computing tasks unloaded from the roadside unit, generate the backend computing results of the corresponding node, complete the split federated learning collaborative training, and output the node submission data.

7. The UAV collaborative training method for decomposed federated learning in the Internet of Vehicles according to claim 6, characterized in that, The step of updating the global model in this round by using node training status sheets and gradient debt settlement mechanisms to determine the update result specifically includes: Create a node training status sheet for each node participating in the training, which records the node's business role, training path, and training completion status at each stage. After the preset aggregation time window ends, based on the validity of the node submission data and training path, all nodes are divided into a real submission set, a virtual submission set, and an expired set. Perform a standard split federated learning update process on the nodes in the real submission set to obtain the first part of the model update amount; Proxy activations are generated for nodes in the virtual submission set to maintain training continuity, and corresponding gradient debt slips are generated synchronously to obtain the second part of the model update amount; Nodes in the expired set will not be included in this round of model updates; Based on the first part of the model update and the second part of the model update, the global model update result for this round is obtained by integrating them.

8. The UAV collaborative training method for decomposed federated learning in the Internet of Vehicles according to claim 7, characterized in that, The step of synchronously generating the corresponding gradient debt statement specifically includes: Calculate the inverse gradient of the slice layer corresponding to the proxy activation; Record the corresponding node number, training round, slice layer information, agent activation credibility, debt validity period and training path information as the corresponding node information; The inverse gradient of the slice layer is integrated with the corresponding node information to generate and store the gradient debt list for this node in this round, which is used for debt settlement when the node is reconnected in the future.

9. The UAV collaborative training method for decomposed federated learning in the Internet of Vehicles according to claim 8, characterized in that, The tiered debt settlement mechanism specifically includes: When a node reconnects and uploads real activation data, the corresponding historical gradient debt statement for that node is retrieved, and the compatibility score of the corresponding historical debt is calculated. If the compatibility score meets the preset liquidation threshold, the historical debt gradient is added as a compensation term to the front-end model update gradient of the node and participates in the parameter update of the corresponding node. If the compatibility score does not meet the preset liquidation threshold, the corresponding historical debt will be automatically invalidated.

10. A drone collaborative training system for decomposed federated learning in the Internet of Vehicles, characterized in that, Includes the following modules: The first module is used to construct the vehicle-to-everything (V2X) split federated learning scenario at the target intersection and perform system initialization, determine the networking architecture of roadside units and drone teams, and the various energy efficiency classification business roles of drones. The second module is used to acquire intersection environmental observation data within a preset time interval, input the intersection environmental observation data into the spatiotemporal state prediction module for prediction, and determine the system state prediction result for the next moment. The third module is used to input the system state prediction results for the next moment and the current observation state into the strategy decision module to generate control strategies and determine the control actions of the UAV. The fourth module is used to map the control actions of the drone to various energy efficiency classification business roles, control the drone to switch to the energy efficiency classification business role, perform the corresponding mode of split federated learning collaborative training, and output node submission data. The fifth module is used to update the node submission data generated by the split federated learning collaborative training through the node training status sheet and gradient debt settlement mechanism based on the global aggregation phase, determine the result of the global model update in this round, and realize the UAV collaborative training of split federated learning.

Images