An unmanned aerial vehicle federated learning resource optimization method for hyperspectral data and a system thereof

Abstract

The application relates to a UAV federated learning resource optimization method and system for hyperspectral data, which comprises the following steps: based on a coalition game mechanism, a UAV group is divided into multiple mutually non-overlapping UAV alliances, and in each alliance, one UAV is determined as an alliance cluster head and the rest UAVs are member UAVs; for each alliance, under the condition of meeting the delay constraint, the sum of the calculation energy consumption and the hovering energy consumption is minimized to optimize the CPU frequency of each UAV to obtain an optimal frequency; each UAV uses the optimal CPU frequency to perform local model training on the collected local hyperspectral image data to obtain local model parameters and transmit the local model parameters to the alliance cluster head; based on the aggregated local model parameters of all alliance cluster heads, global model updating is performed; each UAV updates the local model by using the updated global model, and the above steps are iteratively executed until the global model converges. The application realizes effective control of the overall energy consumption of the system and adaptive allocation of resources.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, specifically to a method and system for optimizing UAV federated learning resources for hyperspectral data. Background Technology

[0002] Hyperspectral imaging (HSI) has become a transformative technology in Earth observation, capable of capturing detailed spectral information distributed across hundreds of consecutive narrow bands. This rich imagery data enables precise material identification and quantitative analysis, supporting critical applications such as precision agriculture, environmental monitoring, mineral exploration, and urban planning. With the widespread adoption of lightweight hyperspectral sensors, unmanned aerial vehicles (UAVs), with their high mobility, flexible deployment, and ability to acquire high-resolution data at low altitudes, have become the mainstream platform for remote sensing monitoring. Simultaneously, advancements in artificial intelligence have provided unprecedented opportunities to unlock the potential value of hyperspectral data.

[0003] However, forming stable alliances in dynamic environments involves complex trade-offs between communication benefits, computational power, and physical costs. First, traditional centralized training is impractical for UAV swarms processing massive 3D hyperspectral data cubes. Excessive energy consumption and intolerable latency during long-distance transmission necessitate a distributed, communication-efficient paradigm. Furthermore, in standard star-topology federated learning, the computationally heavy workload of HSI analysis exacerbates synchronization latency. Faster UAVs consume significant power during mechanical hovering while waiting for slower nodes, leading to severe energy imbalances within the swarm. Finally, existing frameworks lack mechanisms to enable UAVs to autonomously form stable alliances. The system needs to balance instantaneous computational demands with collective hovering time in dynamic environments. Summary of the Invention

[0004] This application provides a method and system for optimizing UAV federated learning resources for hyperspectral data, which can solve the problems of high communication overhead, high computational energy consumption, high synchronization delay and unstable federation structure in existing technologies for hyperspectral image classification tasks.

[0005] In a first aspect, embodiments of this application provide a method for optimizing UAV federated learning resources for hyperspectral data, comprising: Based on the alliance game mechanism, the drone swarm is divided into multiple non-overlapping drone alliances, and in each alliance, one drone is designated as the alliance leader and the rest of the drones are member drones. For each alliance, under the condition of meeting latency constraints, the CPU frequency of each UAV is optimized to obtain the optimal frequency by minimizing the sum of computing energy consumption and hovering energy consumption. Each UAV uses the optimal CPU frequency to train a local model on the acquired local hyperspectral image data, obtains local model parameters, and transmits them to the cluster head of its alliance. Global model update based on local model parameters aggregated from all consortium cluster heads; Each UAV updates its local model using the updated global model, and iteratively executes the above steps until the global model converges.

[0006] In conjunction with the first aspect, in one implementation method, the drone swarm is divided into multiple non-overlapping drone alliances based on a coalition game mechanism, specifically including: Initialize the alliance division state of the drone swarm; Iteratively execute the alliance switching operation, control the drones to switch between alliances until the alliance division reaches a stable state, and complete the division of the drone swarm into multiple non-overlapping drone alliances.

[0007] In conjunction with the first aspect, in one implementation, the alliance switching operation is iteratively executed, controlling the drone to switch between alliances until the alliance division reaches a stable state, specifically including: For each drone, determine a candidate alliance for the currently traversed drone from the current alliance partition, and determine the change in utility after the drone joins the candidate alliance; If the change in utility satisfies the preset cooperation preference conditions, then the drone is controlled to switch from its current alliance to the candidate alliance, and the alliance division status is updated. When all drones cease performing switching operations, the alliance division is considered to have reached a stable state.

[0008] In conjunction with the first aspect, in one implementation, determining the change in utility of the drone after joining the candidate coalition specifically includes: The current individual utility of the UAV in the current alliance and the candidate individual utility in the candidate alliance are calculated based on the individual utility function. Based on the difference between the utility of the candidate individual and the utility of the current individual, the change in utility of the drone after joining the candidate alliance is determined; The individual utility function consists of a benefit term and a cost term. The benefit term includes a value reward based on the amount of local data and an incentive subsidy obtained when serving as the cluster head of the consortium. The cost term includes the computational energy consumption for local model training, the transmission energy consumption for model parameter transfer, and the hovering energy consumption during consortium synchronization waiting.

[0009] In conjunction with the first aspect, in one implementation, the cooperative preference conditions are determined based on a trade-off between individual utility and group social welfare; The group social welfare is the sum of the individual utilities of all drones within the alliance; When a switching operation can improve the social welfare of the group, or improve individual utility without reducing the social welfare of the group, it is determined that the preset cooperation preference condition is met.

[0010] In conjunction with the first aspect, in one implementation, for each alliance, under the condition of satisfying latency constraints, the CPU frequency of each UAV is optimized to obtain the optimal frequency with the goal of minimizing the sum of computational energy consumption and hovering energy consumption. Specifically, this includes: Construct an energy consumption optimization objective function with the sum of computational energy consumption and hovering energy consumption as the optimization objective; Initialize the CPU frequency of each drone and use the initialized CPU frequency as the current iteration frequency; Calculate the gradient of the energy consumption optimization objective function with respect to the current iteration frequency; The current iteration frequency is updated in the opposite direction of the gradient to minimize energy consumption and obtain candidate frequencies; The feasibility of the candidate frequencies is verified to ensure that the delay constraints and hardware frequency range constraints are met. Use the verified candidate frequency as the new current iteration frequency to determine whether the convergence condition is met. If the convergence condition is met, the current iteration frequency is output as the optimal frequency. If the convergence condition is not met, return to the step of calculating the gradient.

[0011] In conjunction with the first aspect, in one implementation, a feasibility check is performed on the candidate frequencies to ensure that delay constraints and hardware frequency range constraints are met, specifically including: If the candidate frequency exceeds the frequency range allowed by the hardware, it is corrected to the boundary value of the frequency range to meet the hardware frequency range constraints. If the total latency of the alliance determined based on the candidate frequency exceeds the maximum allowable latency threshold, the candidate frequency is adjusted until the total latency of the alliance meets the maximum allowable latency threshold to ensure that the latency constraint condition is met. If the candidate frequency does not exceed the frequency range allowed by the hardware and the total latency of the alliance determined based on the candidate frequency does not exceed the maximum allowed latency threshold, then the candidate frequency remains unchanged.

[0012] In conjunction with the first aspect, in one implementation, the local hyperspectral image data is a three-dimensional data cube, including spatial dimension and spectral band dimension; The computational load required to process the three-dimensional data cube is proportional to the number of spectral band dimensions, and the computational energy consumption is proportional to the square of the computational load and the CPU frequency.

[0013] In conjunction with the first aspect, in one implementation, a global model update is performed based on the local model parameters aggregated from all consortium cluster heads, specifically including: For the local model parameters aggregated from each consortium cluster head, determine the number of local data samples within the consortium corresponding to each consortium cluster head; Based on the number of local data samples within the alliance, calculate the aggregate weight of the local model parameters of each alliance cluster head in the global model update; The local model parameters uploaded by each consortium cluster head are weighted and fused based on the aggregation weight to obtain the updated global model.

[0014] Secondly, embodiments of this application provide a drone federated learning resource optimization system for hyperspectral data, comprising: a drone swarm containing multiple drones, wherein the drone swarm is configured as follows: the drone swarm is divided into multiple non-overlapping drone alliances based on a federation game mechanism, and within each alliance, one drone is designated as the alliance cluster head, and the remaining drones are member drones; for each alliance, under the condition of satisfying latency constraints, the CPU frequency of each drone is optimized to obtain an optimal frequency with the goal of minimizing the sum of computational energy consumption and hovering energy consumption; the optimal CPU frequency is used to train a local model on the acquired local hyperspectral image data to obtain local model parameters, which are then transmitted to the respective alliance cluster head; the updated global model is received and the local model is updated; The central server is configured to: receive aggregated local model parameters sent by the alliance cluster head; perform global model updates based on the aggregated local model parameters; and distribute the updated global model to the UAV swarm; wherein the UAV swarm and the central server work together to balance energy consumption and hovering energy consumption through alliance game theory and frequency optimization, and iteratively execute until the global model converges.

[0015] The beneficial effects of the technical solutions provided in this application include: This application provides a method and system for optimizing UAV federated learning resources for hyperspectral data. It utilizes a federation game mechanism to divide the UAV swarm, enabling UAVs to autonomously organize into non-overlapping clusters. The cluster head centrally aggregates the model parameters of member UAVs. This hierarchical aggregation method reduces the number of communication nodes directly connected to the central server and replaces some long-distance backhaul links with short-distance links, significantly reducing energy consumption during communication. The introduction of the game mechanism also helps the federation partitioning reach a stable state, avoiding frequent network topology oscillations. Furthermore, a CPU frequency optimization mechanism is introduced within the federation to find the optimal balance between computational and hovering energy consumption while satisfying task latency constraints. Since UAV hovering energy consumption is proportional to task completion time, and computational energy consumption is proportional to the square of the frequency, frequency optimization prevents excessive collective waiting time caused by slow computation by individual UAVs, thereby reducing additional hovering power consumption due to synchronization waiting. This achieves effective control of overall system energy consumption and adaptive resource allocation, ensuring efficient convergence of the global model. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the drone federated learning resource optimization method for hyperspectral data in this application; Figure 2 This diagram illustrates how social welfare changes with the number of iterations under three different sorting methods in this application. Figure 3 This is a schematic diagram of the stable alliance partitioning structure in this application; Figure 4 This is a schematic diagram showing how the number of alliances changes with cluster head subsidies under three different sorting methods in this application; Figure 5 This is a schematic diagram illustrating the variation of total delay with the number of spectral bands in this application; Figure 6 This is a schematic diagram illustrating how the average delay per round changes with the amount of data in this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0018] This application provides a method and system for optimizing UAV federated learning resources for hyperspectral data, which can solve the problems of high communication overhead, high computational energy consumption, high synchronization delay and unstable federation structure in existing technologies for hyperspectral image classification tasks.

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0020] Firstly, embodiments of this application provide a method for optimizing UAV federated learning resources for hyperspectral data, applicable to a system consisting of a central server and a swarm of UAVs. A distributed collaborative system. It includes: 101: Based on a coalition game mechanism, the drone swarm is divided into multiple non-overlapping drone coalitions, and within each coalition, one drone is designated as the coalition leader, and the remaining drones are member drones: First, the system divides the drone swarm based on the coalition formation game mechanism. The drones are divided into multiple non-overlapping drone alliances, forming an alliance division. Each of the drones Belongs to and belongs to only one alliance to satisfy and Constraints. In each alliance Within the system, one drone is selected as the alliance cluster head to bridge communication with the central server, while the remaining drones serve as member drones. Through this hierarchical topology, member drones can use short-distance links to transmit data, thereby replacing the long-distance backhaul link that directly connects to the central server and reducing communication energy consumption.

[0021] 102: Targeting each alliance Under the condition of satisfying latency constraints, the optimal CPU frequency of each UAV is obtained by minimizing the sum of computational energy consumption and hovering energy consumption. Since hyperspectral data processing has a high computational load, and federated learning requires synchronous waiting for the slowest node in the consortium to complete its task, the operating frequency of the UAV directly affects the task completion time and the hovering power consumption during this period. This step calls the CPU frequency optimization algorithm, inputting the consortium partitioning... and constraints Output the optimal frequency vector Optimal frequency vector It includes the optimal frequency configuration for each drone in the alliance. By balancing the power consumption of instantaneous calculations and the energy consumption of collective hovering due to synchronization waiting, it ensures that the frequency value is between the minimum and maximum frequencies allowed by the hardware, thereby achieving the optimal allocation of resources within the alliance and avoiding energy waste or excessive latency caused by improper frequencies.

[0022] 103: Various drones The acquired local hyperspectral image data is trained using the optimal CPU frequency to obtain local model parameters, which are then transmitted to the respective consortium cluster head. The hyperspectral image data is a three-dimensional data cube containing both spatial and spectral band dimensions; the processing load is proportional to the number of spectral bands. After obtaining the local model parameters, member UAVs transmit these parameters to their respective consortium cluster head, which then aggregates them within the consortium and sends the aggregated local model parameters to the central server, completing the uplink communication process and ensuring efficient aggregation of model parameters.

[0023] 104: The central server performs global model updates based on the local model parameters aggregated from all consortium cluster heads; 105: Generate the updated global model and distribute it to the drone swarm. Each UAV updates its local model using the received updated global model, and iteratively executes the above steps until the global model converges.

[0024] This application utilizes a coalition game mechanism to partition the UAV swarm, enabling UAVs to autonomously organize into non-overlapping cluster structures. The cluster head centrally aggregates the model parameters of member UAVs. This hierarchical aggregation method reduces the number of communication nodes directly connected to the central server and replaces some long-distance backhaul links with short-distance links, thereby significantly reducing energy consumption during communication. Furthermore, the introduction of the game mechanism allows the coalition partitioning to tend towards a stable state, avoiding frequent oscillations in the network topology. In addition, a CPU frequency optimization mechanism is introduced within the coalition to find the optimal balance between computational energy consumption and hovering energy consumption while satisfying task latency constraints. Since UAV hovering energy consumption is proportional to task completion time, and computational energy consumption is proportional to the square of the frequency, frequency optimization can prevent excessively long collective waiting times caused by slow computation by individual UAVs, thereby reducing additional hovering power consumption caused by synchronization waiting. This achieves effective control of overall system energy consumption and adaptive resource allocation, ensuring efficient convergence of the global model.

[0025] It should be noted that, in this embodiment, the local hyperspectral image data is a three-dimensional data cube, containing both spatial and spectral band dimensions; specifically, unlike conventional RGB images, each hyperspectral sample is a three-dimensional data cube. ,in The spatial dimension represents the spatial resolution of the image. This indicates the number of spectral bands, typically distributed across hundreds of consecutive narrow bands. (Unmanned aerial vehicles) Collection includes For a local dataset of samples, the computational load required to process the three-dimensional data cube is proportional to the number of spectral band dimensions. This is because hyperspectral imaging technology can capture detailed spectral information distributed across hundreds of consecutive narrow bands. The rich image data makes accurate material identification possible, but also significantly increases the complexity of data processing. This represents the number of CPU cycles required to process one unit of hyperspectral data for a drone. The total number of CPU cycles required to execute one local training round is The formula clarifies the computational load. Directly depends on the number of spectral bands The more bands there are, the more computing resources are required.

[0026] The computational energy consumption is proportional to the square of the computational load and the CPU frequency, specifically modeled as follows: .in For the first The CPU frequency of the drone The effective capacitance coefficient depends on the chip architecture. (Drone) The computation time for performing local training is Its power calculation model is as follows Higher It can reduce computation time and hover time, but it will cause the instantaneous power consumption to increase dramatically by a quadratic factor. Therefore, the cost of computational energy consumption is directly related to the computational load. With frequency The square of this relationship means that a small increase in frequency can lead to a significant increase in energy consumption. The system needs to balance reducing computation time to lower hovering energy consumption with the instantaneous computational power consumption of control. This makes computational energy cost a key variable in resource optimization, ensuring that the model can accurately reflect the physical consumption of the UAV in collaborative tasks.

[0027] Based on the above embodiments, in this embodiment, the drone swarm is divided into multiple non-overlapping drone alliances based on a coalition game mechanism, specifically including steps 1011 to 1012: Step 1011: Initialize the coalition division state of the drone swarm. Specifically, this involves all drones... Randomly select an alliance to form the initial system partition. And set the number of iterations. Since the optimal frequency depends on the alliance size and total data load, it needs to be based on... Execute the CPU frequency optimization algorithm to obtain the initial frequency. The initial individual utility of each drone is calculated. This step provides a feasible starting state for subsequent game iterations, ensuring that each drone has a clear resource allocation configuration in the initial stage, avoiding blind switching that would lead to an unevaluable system state, and laying the foundation for dynamic adjustments to the alliance structure.

[0028] Step 1012: Iteratively execute the alliance switching operation, controlling the drones to switch between alliances until the alliance division reaches a stable state, completing the division of the drone swarm into multiple non-overlapping drone alliances. Specifically, the system enters a loop to determine whether the alliance division has reached a stable state, traversing each drone... To divide the drones currently being traversed from the current alliance A candidate alliance was identified. ,in, ; Let's assume for now that it's a drone Joining the candidate alliances thus forms a potential alliance partition. To accurately assess each potential movement, targeting The affected alliances are running CPU frequency optimization algorithms to obtain temporary frequencies. And determine the change in utility of the drone after it joins the candidate coalition.

[0029] If the change in utility satisfies a preset cooperation preference condition, then the drone is controlled to leave its current alliance. Switch to the candidate alliance and update the alliance partitioning status. At the same time, the optimal frequency and its real-time utility are updated based on the new frequency.

[0030] When all drones cease switching operations in the current iteration, the coalition partitioning is considered to have reached a stable state, the system has converged to Nash equilibrium, the loop is exited, and a stable coalition partitioning is returned. This application theoretically proves that the game is essentially a precise potential game, thus mathematically guaranteeing that the network will converge to a stable alliance partitioning state, avoiding the problem of drastic network partitioning and failure to reach a stable state under selfish preferences. This mechanism ensures the consistency and reliability of network partitioning by making trade-offs between individual computational energy consumption, transmission costs, and group hovering energy consumption caused by synchronization waiting delays, thereby successfully partitioning the drone swarm into multiple non-overlapping drone alliances and providing a stable topology for subsequent resource optimization and model training.

[0031] In this embodiment, determining the change in utility of the drone after it joins the candidate coalition specifically includes: First, the current individual utility of the UAV in the current alliance and the candidate individual utility in the candidate alliance are calculated based on the individual utility function. We then calculate the... Defining individual utility with drones The individual utility function consists of a benefit term and a cost term. Individual utility is defined as reward minus total energy consumption, reflecting the net benefit of the drone participating in collaborative tasks. The benefit term includes a value reward based on the amount of local data and an incentive subsidy received when acting as the leader of the alliance cluster. Specifically, the drone... Data value Its contribution to the accuracy of hyperspectral image classification models is proportional to the amount of data. logarithmic function ,in and The data volume coefficient represents the logarithmic function model, which reflects the diminishing marginal utility of data; that is, as the amount of data increases, the contribution of each unit of data to the model's accuracy gradually decreases. The data reward is a price factor per unit value. With the value of the data product ,in This is used to convert data value into utility units for comparison with energy consumption costs on a unified scale. Additionally, the revenue stream includes subsidies provided by the central server to incentivize drones to act as cluster leaders. via indicator function Judgment, only when drone For cluster head The function value is 1 when the condition is met and the subsidy is received; otherwise, it is 0, thus compensating for the high backhaul communication energy consumption of the cluster head. For indicator functions, The fixed subsidy value provided to the central server is designed to balance the energy consumption differences between the cluster head and other members, and to incentivize drones to assume high-energy-consuming roles.

[0032] Cost items include the computational energy consumption for local model training. Transmission energy consumption of model parameter transmission Backhaul power consumption when acting as cluster head And hovering energy consumption during alliance synchronization waiting period. The energy consumption modeling is as follows: It is positively correlated with the square of the CPU frequency; higher frequencies reduce processing time but drastically increase power consumption. The effective capacitance coefficient, The total number of CPU cycles required to execute a local training round. Transfer energy consumption depends on the role of the member or cluster head; the cluster head is involved in inter-fedition transfer energy consumption. Hovering energy consumption is modeled as follows: Total delay with the alliance Proportional, of which This represents hovering power. This reflects the impact of computational speed on physical costs. To ensure synchronized updates, the consortium must wait for the slowest drone member to complete computation and transmission. During this time, all drones must remain hovering to ensure network connectivity. Faster drones consume hovering power while waiting for slower nodes, leading to uneven energy consumption within the cluster. Therefore, precise calculation of various energy consumption parameters is necessary to determine the true utility. The total individual energy consumption is expressed as... It covers the energy consumption of the entire process from computing to communication to maintaining location, ensuring that the cost item fully reflects the physical consumption.

[0033] Then, based on the difference between the candidate individual utility and the current individual utility, the change in utility of the drone after joining the candidate coalition is determined. The complete expression for individual utility is: .in Price factor per unit value This indicates a data reward. For indicator functions, Subsidies provided to the central server to incentivize drones to act as cluster leaders, therefore Subsidies are provided for cluster heads. This ensures the utility function accurately reflects the net benefit of UAVs in collaborative tasks, providing a quantitative basis for alliance switching decisions. Through this function, the system can weigh individual computational energy consumption, transmission costs, and group hovering energy consumption caused by synchronization delays. This allows UAVs, when autonomously organizing into non-overlapping alliances, to consider not only their own energy consumption but also group social welfare through a cooperation preference mechanism. This balances instantaneous computational demands with collective hovering time in dynamic environments, achieving high-energy-efficiency hyperspectral image classification and avoiding system instability caused by solely pursuing individual gains. This utility calculation process is the core of the alliance formation game. Each potential alliance switch requires recalculating this utility value to determine if the cooperation preference conditions are met, ensuring the network converges to a stable alliance partitioning state and enabling the system to adaptively allocate computational resources under strict constraints.

[0034] Furthermore, the cooperative preference conditions are determined based on a trade-off between individual utility and group social welfare. The main problem addressed in this application is how to maximize social welfare, defined as the sum of the individual utilities of all UAVs. This is equivalent to maximizing the total data value generated by the hyperspectral image classification task minus the total energy consumption of the entire physical network, thereby achieving optimal energy efficiency while ensuring task performance. The group social welfare is the sum of the individual utilities of all UAVs within the alliance, defined as... This is equivalent to maximizing the total data value generated by the hyperspectral mission minus the total energy consumption of the entire physical network. This formula indicates that social welfare is the key factor in alliance partitioning. With frequency allocation The function needs to be maximized by jointly optimizing both. Substituting the costs of each element, the maximization problem can be formulated as follows: This optimization problem is subject to several key constraints, specifically including: constraint C1 is... This ensures that the hyperspectral classification task meets the real-time perception requirements, among which To maximize the total allowable delay and prevent monitoring data from becoming invalid due to task timeout; constraint C2 is... This limits the range of CPU frequencies. and These are the extreme points of the CPU frequency, ensuring safe hardware operation and meeting basic computing requirements; constraint C3 is... This ensures the integrity and mutual exclusivity of the alliance division, meaning each drone must belong to one and only one alliance. System optimization variables include alliance structure. Cluster head selection and CPU frequency allocation By solving this optimization problem, the system can balance instantaneous computational demands with collective hovering time in a dynamic environment, providing clear theoretical objectives and convergence basis for subsequent alliance formation and frequency optimization steps, ensuring that the network will converge to a stable alliance partitioning state.

[0035] A switching operation is considered to meet a preset cooperation preference condition when it improves the social welfare of the group, or improves individual utility without reducing the social welfare of the group. This mechanism improves the social welfare of the alliance by making trade-offs between individual computing energy consumption, transmission costs, and group hovering energy consumption caused by synchronization waiting delays.

[0036] Based on the above embodiments, in this embodiment, for each alliance, under the condition of satisfying latency constraints, the CPU frequency of each UAV is optimized to obtain the optimal frequency with the goal of minimizing the sum of computing energy consumption and hovering energy consumption. Specifically, this includes: First, we construct an energy consumption optimization objective function with the sum of computational energy consumption and hovering energy consumption as the optimization objective: The computational energy consumption model is as follows: ,in The effective capacitance coefficient, For drones The total number of CPU cycles required to execute a local training round. Among them, the drone... Collection includes A local HSI dataset of samples. Unlike RGB images, each HSI sample is a three-dimensional data cube. ,in For spatial dimensions, This indicates the number of spectral bands. Processing HSI data is significantly more complex and related to spectral depth. Proportional. Let This indicates the number of CPU cycles required to process one unit of HSI data. (Drone / Unmanned Aerial Vehicle) The total number of CPU cycles required to execute one local training round is This application introduces CPU frequency. As a key variable in resource allocation, drones... Computation time for performing local training for: ,in For the first The CPU frequency of the drone. Its computational power is modeled as follows: ,in Let be the effective capacitance coefficient. Therefore, the energy cost is calculated as follows: .

[0037] The hovering energy consumption model is as follows: ,in For the overall delay of the alliance, For hovering power, the higher one It can reduce computation time and hover time, but it causes a sharp, quadratic increase in instantaneous power consumption. (Alliance) The federated learning process involves two transport phases: within a federation (members to cluster heads) and between federations (cluster heads to the central server). Indicates alliance The index of the cluster head. For member drones. Its trained model parameters (size is) ) will be transmitted to the cluster head The transmission rate is expressed as: ;in For bandwidth within the alliance, For transmission power, For channel gain, Noise power. Member drone. The transmission time and energy consumption are as follows: .

[0038] It is important to note that, for the cluster head, the time involved in intra-consortium transmissions is... Cluster head The aggregation model forwards global updates to the central server. Its transmission rate is... The transmission time and energy consumption of the cluster head are as follows: ;in For inter-alliance bandwidth, This refers to the transmit power of the cluster head. For channel gain, For noise power. To ensure synchronized updates, the alliance The cluster head can only aggregate and upload data after the slowest drone member has completed its calculations and transmissions. Therefore, the consortium... The total latency in a federated learning round is: ,in Indicates cluster head The calculation time during this time period. Within the alliance, all drones must remain hovering to ensure network connectivity. Any drone The energy cost of hovering is proportional to the total time: ;in The hovering power is used. This application models the hyperspectral image classification task of UAVs as a coalition-forming game. Member UAVs save energy by using short-range transmissions instead of long backhaul links to the central server. The system issues incentive subsidies to the cluster head to compensate for its higher backhaul communication energy consumption.

[0039] Next, the CPU frequency of each drone is initialized, and the initialized CPU frequency is used as the current iteration frequency. Specifically, a lower bound for the frequency is set. This lower bound, together with the maximum frequency threshold, defines the range of CPU frequency during initialization. ,in Set a task-specific lower bound for the frequency and set an iteration counter. This step ensures that the initial frequency meets the minimum hardware requirements and the frequency required by the task, avoiding a frequency that is too low to meet basic computing needs, providing a feasible starting point for subsequent gradient iterations, and ensuring that the optimization process starts from a physically feasible state.

[0040] Then, the gradient of the energy consumption optimization objective function with respect to the current iteration frequency is calculated. Specifically, this includes determining the current delay. And the group of those who fell behind Calculate the gradient of social welfare:

[0041] in, The effective capacitance coefficient; For drones The total number of CPU cycles required to execute a local training round is proportional to the spatial dimension and the number of spectral bands of the hyperspectral data cube. The frequency of the k-th iteration; This is an indicator function that takes a value of 1 when the drone belongs to the set of laggards; The number of drones within the alliance; This refers to the hovering power. The current iteration frequency is updated in the opposite direction of the gradient to minimize energy consumption, resulting in candidate frequencies. Specifically, this involves updating... ,in The step size is defined as follows. This step minimizes the social cost of the alliance by balancing instantaneous power consumption with the collective hovering time caused by lagging members, and the gradient direction guides the frequency adjustment towards reducing total energy consumption. The set of lagging members is defined as follows: The introduction of this feature reflects the impact of synchronous waiting on energy consumption, ensuring that the optimization process takes into account the synergistic effect within the alliance. Faster drones need to weigh the increased computational energy consumption brought about by increasing the frequency against the energy savings brought about by reducing hovering time.

[0042] Next, the candidate frequencies are subjected to feasibility verification to ensure that they meet the delay constraints and hardware frequency range constraints; specifically, this includes projection operations. To ensure physical feasibility. Among them The maximum frequency threshold represents the highest operating frequency allowed by the hardware. The minimum frequency threshold, A task-specific lower bound is established to ensure that basic computational requirements are met. This projection operation limits the updated frequency to the set of feasible hardware constraints, preventing the frequency from exceeding physical limits and causing equipment damage or failure to meet the minimum task requirements.

[0043] The validated candidate frequency is used as the new current iteration frequency. It is then determined whether the convergence condition is met. If the convergence condition is met, the current iteration frequency is output as the optimal frequency; otherwise, the process returns to the step of calculating the gradient. Specifically, this involves meeting the convergence limit requirements. Time return ;in Let be the gradient norm of the social welfare function of the alliance; Assuming a pre-defined convergence tolerance, when the gradient norm falls below this tolerance, it indicates that the frequency update has negligible impact on social welfare, reaching a stable optimization state. The optimal frequency vector is then output. This means that under the alliance structure, the frequency configuration that minimizes the sum of computing energy consumption and hovering energy consumption is minimized, thereby minimizing the social cost of the alliance and ensuring that the resource allocation of each drone within the alliance reaches the optimal balance.

[0044] If the convergence condition is not met, return to the step of calculating the gradient and update the step size. and iterative counter Step size updates are used to adjust the speed of gradient descent, ensuring stable convergence of the iterative process and preventing oscillations. This iterative process continues until the gradient norm falls below the tolerance, allowing the system to converge under strict delay constraints. and frequency range The following implements adaptive allocation of computing resources to obtain the optimal frequency vector. This effectively reduces overall task latency and improves overall social welfare, completing the full iterative process of the CPU frequency optimization algorithm.

[0045] Furthermore, the feasibility of the candidate frequencies is verified to ensure that the delay constraints and hardware frequency range constraints are met, specifically including: If the candidate frequency exceeds the hardware-allowed frequency range, it is corrected to the boundary value of the frequency range to satisfy the hardware frequency range constraint. The hardware-allowed frequency range is defined by a minimum frequency threshold. and maximum frequency threshold These are defined as the minimum and maximum clock frequencies at which the UAV's onboard processor can operate stably, respectively. The specific verification process uses a projection operation, expressed by the formula: When the candidate frequency is lower than Forced upgrade to To prevent tasks from failing due to excessively low frequency, when the candidate frequency is higher than [a certain value]... Forced reduction to This step avoids hardware overload. It ensures that the frequency value after each iteration falls within the feasible set of hardware constraints, thus guaranteeing physical feasibility.

[0046] If the total latency of the alliance determined based on the candidate frequencies exceeds the maximum allowable latency threshold, the candidate frequencies are adjusted until the total latency of the alliance meets the maximum allowable latency threshold to ensure that the latency constraint condition is met. The total latency of the alliance... It depends on the maximum of the computation time and transmission time among all drones in the alliance, where the computation time... Inversely proportional to frequency. Maximum allowable delay threshold. This ensures that the hyperspectral classification task meets the real-time perception requirements. Since higher frequencies require shorter computation times, when a value is detected... At this point, the candidate frequency needs to be increased to reduce computation time, thereby reducing the overall latency of the alliance. This adjustment process needs to be repeatedly verified until... This ensures that the system can adaptively allocate computing resources under strict latency constraints, avoiding a surge in hovering energy consumption due to excessively long synchronization waiting time.

[0047] If the candidate frequency does not exceed the frequency range allowed by the hardware and the total latency of the alliance determined based on the candidate frequency does not exceed the maximum allowed latency threshold, then the candidate frequency remains unchanged. This indicates that the frequency configuration obtained in the current iteration simultaneously satisfies both hardware physical constraints and task timeliness requirements, and is a valid point within the feasible solution space.

[0048] Through the aforementioned verification mechanism, the system can immediately eliminate infeasible solutions after each frequency update, ensuring that the optimization process always proceeds within the constraints. This feasibility verification is a key step in the gradient projection method. It maps the unconstrained gradient update direction back to the feasible region, preventing the algorithm from diverging. Ultimately, this allows the system to balance instantaneous computational power consumption and the collective hovering time caused by lagging solutions, while satisfying real-time performance and hardware safety, to obtain the optimal frequency vector.

[0049] Based on the above embodiments, this embodiment performs a global model update based on the local model parameters aggregated from all consortium cluster heads, specifically including: For the aggregated local model parameters of each consortium cluster head, the number of local data samples within each consortium cluster head is determined: The central server receives aggregated model updates from each consortium cluster head via a global backhaul link. The parameters uploaded by each consortium cluster head hj represent the results of local training by all member UAVs within its consortium Sj. To fairly reflect the contribution of each consortium to the global model, the system counts the total number of hyperspectral image data samples within each consortium. This number is obtained by summing the local dataset Di collected by each UAV i within the consortium. The number of data samples directly reflects the richness of ground feature information possessed by the consortium. The larger the sample size, the higher the reference value of its local model parameters for the global model. Therefore, this number needs to be accurately determined as the basis for subsequent weight allocation, ensuring that consortia with large data volumes occupy a reasonable proportion in the update, avoiding consortia with small data volumes dominating the model update direction, and ensuring that the global model can learn the multi-band spectral features collected by each consortium in a balanced manner.

[0050] Based on the number of local data samples within each alliance, the aggregation weight of the local model parameters of each alliance cluster head in the global model update is calculated. The local model parameters uploaded by each alliance cluster head are then weighted and fused based on this aggregation weight to obtain the updated global model. Specifically, the aggregation weight is proportional to the number of local data samples in each alliance; alliances with larger data volumes receive higher weights. The central server uses this weight to perform a weighted average of the model parameters uploaded by each alliance cluster head, eliminating model bias caused by uneven data distribution across alliances. The weighted fusion process integrates the local knowledge of all alliances, generating a global model that can generalize to the entire monitoring area. The updated global model is then distributed to the UAV swarm for the next round of local model updates. This process iteratively continues until the global model converges, achieving classification and recognition of hyperspectral images. This ensures that the global model accurately reflects the spectral information collected by different alliances, completing the aggregation and redistribution process of the global model, thereby achieving collaborative training while protecting data privacy.

[0051] The embodiments are run on the Matlab simulation platform, and a specific description and verification of the specific implementation of this application is provided: Extensive simulation experiments were conducted to evaluate the performance and robustness of the proposed scheme under different system parameters. The experiments aimed to verify the effectiveness of the federated learning framework based on federated game theory (CFG-FL), particularly its ability to balance energy consumption and latency in hyperspectral image classification tasks. By simulating the dynamic behavior of real UAV swarms, the specific effects of the proposed method on reducing communication overhead and computational energy consumption were quantitatively analyzed, ensuring the feasibility of the theoretical model in practical applications and providing data support for the actual deployment of the algorithm.

[0052] We compared the method of this application with the following two benchmark schemes to highlight the technical advantages of this application. The first is the traditional federated learning scheme, which adopts the standard federated learning architecture with a star topology. In this architecture, all UAVs do not form a federation but communicate directly with the central server. This represents the benchmark case with the highest communication energy consumption and is used to evaluate the communication gain brought by the federation formation mechanism. The second is the fixed frequency scheme, in which the CPU frequency of each UAV is fixed at a constant value instead of being dynamically optimized by the CPU frequency optimization algorithm of this application. This represents the benchmark case of rigid allocation of computing resources. By comparison, we can clearly see the performance gains brought by the federation structure and frequency optimization, and verify the necessity of joint optimization.

[0053] The simulation experiment's spatial setup mirrored a typical low-altitude remote sensing monitoring scenario. UAVs were randomly distributed within a 1000m × 1000m rectangular area. This random distribution reflects the flexibility of UAV deployment and the uncertainty of the environment, avoiding the accidental results caused by specific topologies. The central server, located at coordinates (500, 500), serves as the core node for global model aggregation and distribution. The area size is sufficient to accommodate multiple UAVs working collaboratively, while the central server's location ensures that the maximum communication distance with edge UAVs remains within a reasonable range. This facilitates the assessment of transmission energy consumption differences under different topologies, realistically reflecting the energy consumption comparison between long-distance and short-distance transmission.

[0054] The key system parameters were set based on the hardware characteristics and mission requirements of typical UAVs to ensure that the simulation results have practical reference value. UAV hovering power. Setting it to 100W reflects the average mechanical power consumption of the multi-rotor drone during mission execution. CPU effective capacitance coefficient. for This depends on the onboard processor's chip architecture and is used to accurately calculate dynamic power consumption. Maximum tolerable latency. Setting the time to 150 seconds ensures that the hyperspectral classification task meets the real-time perception requirements, serving as a hard constraint for the frequency optimization algorithm. Model parameter size. The setting of 100 Mbit corresponds to the weight size of the deep learning model and directly affects transmission power consumption. These parameters together constitute the basic constraints of the simulation environment, ensuring that the experimental results can truly reflect the system's performance under physical limitations.

[0055] See Figures 2 to 6 : Figure 2 In this paper, the convergence behavior of social welfare under different preference orders is compared. The selfish ordering exhibits violent oscillations and fails to reach a stable state because drones switch alliances to maximize their individual gains, ignoring group externalities. Although it can achieve high instantaneous peaks, these configurations are transient and impractical. The Pareto ordering produces the lowest social welfare due to its strict switching constraints. In contrast, our proposed scheme guarantees convergence to a stable Nash equilibrium within a finite number of iterations. While its convergent social welfare is slightly lower than the unstable peaks under the selfish preference scheme, it ensures the consistency and reliability of network partitioning, thus validating our proof that this game theory is an exact potential game.

[0056] Figure 3 The paper demonstrates the stable alliance partitioning achieved by the proposed scheme in a network of 20 UAVs. The UAVs autonomously organize into four independent, non-overlapping alliances, each with a designated cluster head responsible for aggregating local updates and communicating with a central server. This cluster topology is a direct result of game theory evolution. In this process, the UAVs weigh the overhead savings from short-range intra-alliance communication against the costs associated with computation and synchronization delays. The formation of spatially adjacent clusters effectively reduces energy consumption for long-distance transmission and mitigates the laggard effect across the entire network.

[0057] Figure 4 This study investigated the impact of cluster head subsidies on network structure. These subsidies are a key factor incentivizing drones to assume the energy-intensive role of cluster head. As subsidies increase, the number of alliances formed under all three preference orders tends to increase. This is because higher rewards make the cluster head role more attractive, encouraging more drones to form their own smaller alliances. However, compared to selfish preference, the proposed scheme can form a greater number of alliances. This indicates its ability to prevent excessive network fragmentation, finding a better balance between the number and size of clusters to maximize overall social welfare. Pareto ordering results in the largest number of alliances, indicating a highly fragmented network. This is because the conditions for Pareto switching operations are too stringent: it prohibits any switching behavior that would cause even a small loss of utility for a single member. This rigidity prevents drones from merging into larger, more efficient groups, causing them to be isolated in small clusters.

[0058] Figure 5 In this study, the performance of the proposed framework was evaluated under various system parameters and compared with benchmark methods. Figure 5 The paper demonstrates the functional relationship between total latency and the number of spectral bands, which directly corresponds to the computational complexity of the HSI classification task. As data complexity increases, the proposed scheme significantly outperforms traditional federated learning and fixed-frequency schemes, achieving the lowest total latency. Traditional federated learning schemes suffer from a severe lag effect, with their total latency entirely dependent on the slowest drone communicating over long-distance links. While fixed-frequency schemes mitigate this issue to some extent, the proposed scheme is superior: it effectively minimizes synchronization wait time by dynamically allocating higher CPU frequencies to each computationally lagging drone in the federation, thereby significantly reducing overall task latency.

[0059] Figure 6 The figure shows how the average communication latency per round changes as the amount of local data allocated to each drone increases. A larger data volume inevitably leads to a heavier computational load. Although the latency of all methods increases with the amount of data, the proposed scheme exhibits the most stable upward trend. Compared to the other two rigid schemes where latency increases sharply, the proposed scheme has the ability to dynamically form alliances and adaptively adjust the CPU frequency, allowing the system to more efficiently absorb the additional computational burden.

[0060] In summary, this application uses Algorithm 1, the Alliance Formation Algorithm, when forming the alliance. The input of this algorithm is the set of drones. Output: Stable alliance division and optimal CPU frequency Algorithm 2, the CPU frequency optimization algorithm, was used for frequency optimization. The input to this algorithm is: alliance partitioning. Constraints Output: Optimal frequency vector .

[0061] In Algorithm 1, switching and merging rules are applied iteratively. Since the optimal frequency depends on the coalition size and total data load, Algorithm 2 is nested within the coalition formation process. Algorithm 1 describes an iterative coalition formation process aimed at achieving Nash-stable partitioning among UAVs. This algorithm begins with an initial random partition, allowing each UAV to systematically explore migrations to randomly selected candidate coalitions. A migration is finalized only if it satisfies the proposed cooperation preferences. To accurately evaluate each potential move, the algorithm invokes Algorithm 2 to optimize resource reallocation for the affected coalitions.

[0062] In this application, we propose a CPU frequency optimization algorithm. Algorithm 2 employs gradient projection to optimize the CPU frequency for a given coalition. Its goal is to maximize the coalition's social welfare by balancing computational energy consumption with latency penalties caused by lagging drones in the coalition. The algorithm initializes the frequency based on a task-specific lower bound and iteratively updates it along the gradient direction of the social welfare function. After each update, it is projected onto a feasible set of hardware constraints to ensure physical feasibility. The iterative process continues until the gradient norm falls below the tolerance.

[0063] This application models a drone network as a federation-forming game model. In this model, multiple drones autonomously organize into non-overlapping clusters to aggregate model update data locally, thereby effectively reducing energy consumption for long-distance transmission. This application introduces a novel cooperative preference mechanism that improves the social welfare of the federation by balancing individual computational energy consumption, transmission costs, and group hovering energy consumption caused by synchronization delays. This application theoretically proves that this game is essentially a precise potential game, mathematically guaranteeing that the network will converge to a stable federation partitioning state. Furthermore, this application proposes an algorithm based on gradient projection to optimize the CPU operating frequency of each drone. By balancing instantaneous computational power consumption and collective hovering time caused by lagging drones, the social cost of the federation is minimized, enabling the system to achieve adaptive allocation of computational resources under strict latency constraints. Extensive simulation experiments demonstrate that, compared to traditional federated learning schemes and fixed-frequency baseline schemes, the proposed method dynamically balances the physical costs of data computation and drone hovering, significantly improving overall social welfare while minimizing system latency.

[0064] Secondly, this application provides a resource optimization system for UAV federated learning based on hyperspectral data, comprising: A drone swarm, comprising multiple drones, is configured as follows: Based on a cooperative game mechanism, the swarm is divided into multiple non-overlapping drone alliances, with one drone designated as the alliance leader and the remaining drones as member drones. For each alliance, under latency constraints, the CPU frequency of each drone is optimized to obtain an optimal frequency, aiming to minimize the sum of computational and hovering energy consumption. The optimal CPU frequency is used to train a local model on the acquired local hyperspectral image data, obtaining local model parameters, which are then transmitted to the alliance leader. The updated global model is received, and the local model is updated accordingly. The central server is configured to: receive aggregated local model parameters sent by the alliance cluster head; perform global model updates based on the aggregated local model parameters; and distribute the updated global model to the UAV swarm; wherein the UAV swarm and the central server work together to balance energy consumption and hovering energy consumption through alliance game theory and frequency optimization, and iteratively execute until the global model converges.

[0065] This application utilizes a coalition game mechanism to partition the UAV swarm, enabling UAVs to autonomously organize into non-overlapping cluster structures. The cluster head centrally aggregates the model parameters of member UAVs. This hierarchical aggregation method reduces the number of communication nodes directly connected to the central server and replaces some long-distance backhaul links with short-distance links, thereby significantly reducing energy consumption during communication. Furthermore, the introduction of the game mechanism allows the coalition partitioning to tend towards a stable state, avoiding frequent oscillations in the network topology. In addition, a CPU frequency optimization mechanism is introduced within the coalition to find the optimal balance between computational energy consumption and hovering energy consumption while satisfying task latency constraints. Since UAV hovering energy consumption is proportional to task completion time, and computational energy consumption is proportional to the square of the frequency, frequency optimization can prevent excessively long collective waiting times caused by slow computation by individual UAVs, thereby reducing additional hovering power consumption caused by synchronization waiting. This achieves effective control of overall system energy consumption and adaptive resource allocation, ensuring efficient convergence of the global model.

[0066] The functions of each module in the above-mentioned UAV federated learning resource optimization system for hyperspectral data correspond to the steps in the above-mentioned UAV federated learning resource optimization method embodiment for hyperspectral data. Their functions and implementation processes will not be described in detail here.

[0067] Thirdly, embodiments of this application provide a drone federated learning resource optimization device for hyperspectral data. The drone federated learning resource optimization device for hyperspectral data can be a personal computer (PC), laptop computer, server, or other device with data processing capabilities.

[0068] In this embodiment, the UAV federated learning resource optimization device for hyperspectral data may include a processor, a memory, a communication interface, and a communication bus.

[0069] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0070] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces for interconnecting devices within the UAV federated learning resource optimization device for hyperspectral data, as well as interfaces for interconnecting the UAV federated learning resource optimization device for hyperspectral data with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0071] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0072] The processor can be a general-purpose processor, which can call the UAV federated learning resource optimization program for hyperspectral data stored in memory and execute the UAV federated learning resource optimization method for hyperspectral data provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the UAV federated learning resource optimization program for hyperspectral data is called can refer to the various embodiments of the UAV federated learning resource optimization method for hyperspectral data in this application, and will not be repeated here.

[0073] Fourthly, embodiments of this application also provide a computer-readable storage medium.

[0074] This application provides a computer-readable storage medium storing a UAV federated learning resource optimization program for hyperspectral data, wherein when the UAV federated learning resource optimization program for hyperspectral data is executed by a processor, it implements the steps of the UAV federated learning resource optimization method for hyperspectral data as described above.

[0075] The method implemented when the UAV federated learning resource optimization procedure for hyperspectral data is executed can be referred to in various embodiments of the UAV federated learning resource optimization method for hyperspectral data in this application, and will not be repeated here.

[0076] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0077] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0078] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0079] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0080] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0081] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0082] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for optimizing UAV federated learning resources for hyperspectral data, characterized in that, It includes: Based on the alliance game mechanism, the drone swarm is divided into multiple non-overlapping drone alliances, and in each alliance, one drone is designated as the alliance leader and the rest of the drones are member drones. For each alliance, under the condition of meeting latency constraints, the CPU frequency of each UAV is optimized to obtain the optimal frequency by minimizing the sum of computing energy consumption and hovering energy consumption. Each UAV uses the optimal CPU frequency to train a local model on the acquired local hyperspectral image data, obtains local model parameters, and transmits them to the cluster head of its alliance. Global model update based on local model parameters aggregated from all consortium cluster heads; Each UAV updates its local model using the updated global model, and iteratively executes the above steps until the global model converges.

2. The UAV federated learning resource optimization method for hyperspectral data as described in claim 1, characterized in that, Based on the coalition game mechanism, the drone swarm is divided into multiple non-overlapping drone coalitions, specifically including: Initialize the alliance division state of the drone swarm; Iteratively execute the alliance switching operation, control the drones to switch between alliances until the alliance division reaches a stable state, and complete the division of the drone swarm into multiple non-overlapping drone alliances.

3. The UAV federated learning resource optimization method for hyperspectral data as described in claim 2, characterized in that, Iteratively execute the alliance switching operation, controlling the drone to switch between alliances until the alliance division reaches a stable state, specifically including: For each drone, determine a candidate alliance for the currently traversed drone from the current alliance partition, and determine the change in utility after the drone joins the candidate alliance; If the change in utility satisfies the preset cooperation preference conditions, then the drone is controlled to switch from its current alliance to the candidate alliance, and the alliance division status is updated. When all drones cease performing switching operations, the alliance division is considered to have reached a stable state.

4. The UAV federated learning resource optimization method for hyperspectral data as described in claim 3, characterized in that, Determine the utility changes of the drone after it joins the candidate coalition, specifically including: The current individual utility of the UAV in the current alliance and the candidate individual utility in the candidate alliance are calculated based on the individual utility function. Based on the difference between the utility of the candidate individual and the utility of the current individual, the change in utility of the drone after joining the candidate alliance is determined; The individual utility function consists of a benefit term and a cost term. The benefit term includes a value reward based on the amount of local data and an incentive subsidy obtained when serving as the cluster head of the consortium. The cost term includes the computational energy consumption for local model training, the transmission energy consumption for model parameter transfer, and the hovering energy consumption during consortium synchronization waiting.

5. The UAV federated learning resource optimization method for hyperspectral data as described in claim 4, characterized in that: The cooperative preference conditions are determined based on a trade-off between individual utility and group social welfare; The group social welfare is the sum of the individual utilities of all drones within the alliance; When a switching operation can improve the social welfare of the group, or improve individual utility without reducing the social welfare of the group, it is determined that the preset cooperation preference condition is met.

6. The UAV federated learning resource optimization method for hyperspectral data as described in claim 1, characterized in that, For each alliance, under the condition of meeting latency constraints, the optimal frequency is obtained by optimizing the CPU frequency of each UAV to minimize the sum of computing energy consumption and hovering energy consumption. Specifically, this includes: Construct an energy consumption optimization objective function with the sum of computational energy consumption and hovering energy consumption as the optimization objective; Initialize the CPU frequency of each drone and use the initialized CPU frequency as the current iteration frequency; Calculate the gradient of the energy consumption optimization objective function with respect to the current iteration frequency; The current iteration frequency is updated in the opposite direction of the gradient to minimize energy consumption and obtain candidate frequencies; The feasibility of the candidate frequencies is verified to ensure that the delay constraints and hardware frequency range constraints are met. Use the verified candidate frequency as the new current iteration frequency to determine whether the convergence condition is met. If the convergence condition is met, the current iteration frequency is output as the optimal frequency. If the convergence condition is not met, return to the step of calculating the gradient.

7. The UAV federated learning resource optimization method for hyperspectral data as described in claim 6, characterized in that, The feasibility of the candidate frequencies is verified to ensure that they meet the delay constraints and hardware frequency range constraints, specifically including: If the candidate frequency exceeds the frequency range allowed by the hardware, it is corrected to the boundary value of the frequency range to meet the hardware frequency range constraints. If the total latency of the alliance determined based on the candidate frequency exceeds the maximum allowable latency threshold, the candidate frequency is adjusted until the total latency of the alliance meets the maximum allowable latency threshold to ensure that the latency constraint condition is met. If the candidate frequency does not exceed the frequency range allowed by the hardware and the total latency of the alliance determined based on the candidate frequency does not exceed the maximum allowed latency threshold, then the candidate frequency remains unchanged.

8. The UAV federated learning resource optimization method for hyperspectral data as described in claim 1, characterized in that: The local hyperspectral image data is a three-dimensional data cube, containing spatial dimension and spectral band dimension; The computational load required to process the three-dimensional data cube is proportional to the number of spectral band dimensions, and the computational energy consumption is proportional to the square of the computational load and the CPU frequency.

9. The UAV federated learning resource optimization method for hyperspectral data as described in claim 1, characterized in that, Global model updates are performed based on the aggregated local model parameters from all consortium cluster heads, specifically including: For the local model parameters aggregated from each consortium cluster head, determine the number of local data samples within the consortium corresponding to each consortium cluster head; Based on the number of local data samples within the alliance, calculate the aggregate weight of the local model parameters of each alliance cluster head in the global model update; The local model parameters uploaded by each consortium cluster head are weighted and fused based on the aggregation weight to obtain the updated global model.

10. A resource optimization system for UAV federated learning of hyperspectral data, characterized in that, It includes: A drone swarm, comprising multiple drones, is configured as follows: Based on the alliance game mechanism, the drone swarm is divided into multiple non-overlapping drone alliances, and in each alliance, one drone is designated as the alliance leader and the rest of the drones are member drones. For each alliance, under the condition of meeting latency constraints, the CPU frequency of each UAV is optimized to obtain the optimal frequency by minimizing the sum of computing energy consumption and hovering energy consumption. The local hyperspectral image data is trained using the optimal CPU frequency to obtain local model parameters, which are then transmitted to the corresponding consortium cluster head. Receive the updated global model and update the local model; The central server is configured as follows: Receive aggregated local model parameters sent by the consortium cluster head; The global model is updated based on the aggregated local model parameters, and the updated global model is then distributed to the drone swarm. The drone swarm works in collaboration with the central server, balancing energy consumption and hovering energy consumption through alliance game theory and frequency optimization, and iteratively executing until the global model converges.

Images