Air-ground cooperative training path planning method and system for target detection task

Abstract

This invention provides a method and system for air-ground cooperative training path planning for target detection tasks, relating to the field of target detection technology. By constructing a globally optimized target model for target detection tasks, this invention addresses the deficiency in existing technologies where UAV trajectory planning and target detection are disconnected. Under the air-ground cooperative federated learning framework, prioritized key data can accurately compensate for the model's learning shortcomings in complex scenarios and low-recognition targets, enabling the model to quickly focus on performance weaknesses for parameter updates. Simultaneously, the association between trajectory planning and the core objective of "minimizing global loss" further ensures the effectiveness and representativeness of parameters uploaded by the ground-based unmanned vehicle. After weighted fusion using a federated averaging algorithm, the newly generated global model possesses more comprehensive feature representation capabilities and stronger generalization abilities.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and specifically to an air-ground collaborative training path planning method and system for target detection tasks. Background Technology

[0002] As target detection technology penetrates into complex scenarios, the contradiction between the data acquisition range and computational efficiency of a single device and the model training requirements is becoming increasingly prominent. Air-ground collaborative technology, with its complementary resource advantages, has become a core path to resolve this contradiction. Air-ground collaboration refers to a dynamic collaborative mechanism between unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGVs) through intelligent algorithms, communication networks, and distributed computing frameworks. This mechanism optimizes the allocation of computing resources, protects data privacy, and facilitates collaborative model training. For target detection tasks, air-ground collaboration can fully leverage the complementary advantages of UAVs and UGVs, such as the high mobility and strong computing power of UAVs and the stable data acquisition capabilities of UGVs. By integrating data from different scenarios and conditions, the target detection model can learn a wider range of feature representations, better adapt to diverse target detection needs, and improve the model's versatility and generalization ability. Simultaneously, it enables the optimized allocation of computing resources, rationally distributing computational tasks to airborne and ground-based equipment, fully utilizing the computing power of both, thereby accelerating model training and inference speed, and improving the real-time performance and accuracy of target detection.

[0003] With technological advancements, air-to-ground collaborative incremental federated learning has become one of the mainstream collaborative training techniques in object detection. This method focuses on the application of UAV position optimization in dynamic data environments. In environments with weak or sudden communication infrastructure failures, UAVs equipped with aggregation servers can act as airborne base stations, providing more stable data transmission and computing services to ground users. Through an incremental federated learning mechanism, UAVs assist ground users in aggregating model parameters and dynamically updating outdated models to adapt to the continuous influx of new data. By jointly optimizing UAV deployment locations, user access frequencies, and sample selection strategies, this method constructs a multi-objective optimization model for energy consumption and training loss while satisfying latency constraints and basic decision variable limitations, achieving a dynamic balance between the two.

[0004] However, existing air-ground collaborative solutions have a core flaw—the UAV trajectory planning is completely disconnected from the target detection task. In other words, the trajectory planning method is designed independently of the target detection task. This problem directly leads to low data acquisition efficiency, making it difficult to match the model training requirements for high-quality and highly targeted data, and seriously restricting the further improvement of the target detection model's performance. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for air-ground cooperative training path planning for target detection tasks, which solves the technical problem that UAV trajectory planning and target detection tasks are completely disconnected in existing air-ground cooperative solutions.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] In a first aspect, the present invention provides an air-ground cooperative training path planning method for target detection tasks, characterized in that a UAV is used as a server in an air-ground cooperative federated learning framework. N The autonomous vehicle acts as a client in the air-ground cooperative federated learning framework; the air-ground cooperative training path planning method includes:

[0010] The server initializes the target detection model as the initial global model and distributes the initial parameters to each autonomous vehicle.

[0011] Each client receives the initial parameters from the server, initializes its local model based on the initial parameters, and constructs and solves the global optimization target model during the local model training process, determining the client selection and drone trajectory during the local model training process; after completing local training, the client uploads the encrypted model parameter increments to the server for aggregation; during the aggregation phase, the federated averaging algorithm is used to weight and fuse the parameter updates according to the proportion of data volume of each client to form a new global model;

[0012] The global optimization objective model includes an objective function and constraints. The objective function is defined as: minimizing global loss by jointly optimizing client selection and UAV trajectory.

[0013] Preferably, the objective function includes:

[0014]

[0015] In the formula:

[0016]

[0017]

[0018] in, This indicates the client's selection; , indicating the drone's trajectory; V For control parameters; This represents the gradient approximation error; and They represent the client respectively. i and nThe local model gradient; It is a client i In the t Initial model parameters for each round; It is a client n In the t Initial model parameters for each round; The length of the virtual energy queue represents the client's... n The extent to which cumulative energy consumption deviates from the budget; Let be a binary choice variable, representing whether a client is selected in round t. n ; For the client n In the round t Total energy consumption; For the client n Maximum budget energy; T This represents the overall round of federal learning and training.

[0019] Preferably, the constraints include:

[0020] (1.1)

[0021] (1.2)

[0022] (1.3)

[0023] (1.4)

[0024] (1.5)

[0025] Constraint 1.1 is a drone concurrency constraint, meaning that the number of drones connecting to unmanned vehicles in each round must not exceed the maximum number of unmanned vehicles that a drone can connect to. ;

[0026] Constraint 1.2 is a coverage constraint, which means that the selected unmanned vehicle must be within the communication coverage area of ​​the drone;

[0027] Constraint 1.3 is a flight distance constraint for the UAV, indicating that the flight distance of the UAV between adjacent rounds shall not exceed [the specified distance]. ;

[0028] Constraint 1.4 is a time delay constraint, meaning that the total time delay of the autonomous vehicle in each round does not exceed a threshold. ;

[0029] Constraint 1.5 is a binary choice variable, indicating whether it is in the first... t Choosing driverless cars in the wheel n ;

[0030] in, Indicates the maximum number of rounds; Indicates drone m and driverless cars n In the t The distance between rounds; Indicates drone m The communication coverage area; This represents the distance between the drone in round t and round t+1. Choose a binary variable. =1 indicates that in the first... t Choosing driverless cars in the wheel n , When =0, it means that in the first... t Wheelchair did not select driverless car n ; Indicates driverless car n Local model computation latency; The communication latency of autonomous vehicle n uploading local model parameters in round t.

[0031] Preferably, solving the global optimization objective model includes:

[0032] The global optimization objective model is transformed into alternating optimization of client selection and UAV trajectory planning, specifically:

[0033] In each iteration, the client selection variable A is fixed first, and the objective function is transformed into a first sub-objective function that is only related to the trajectory. The first sub-objective function is solved using the continuous convex approximation algorithm to optimize the UAV trajectory G and obtain the optimized UAV trajectory. Then, the UAV trajectory G is fixed, and the objective function is transformed into a second sub-objective function that is only related to the client selection A. The constraint conditions are cut using the cutting plane method to obtain a new client selection scheme.

[0034] Preferably, the expression for the first sub-objective function is as follows:

[0035]

[0036] Due to driverless cars n Local model calculation of energy consumption Since the frequency is determined locally by the client and is independent of the trajectory, it can be simplified to the following expression:

[0037]

[0038] in, Indicates the length of the virtual energy queue, and indicates the client. n The extent to which cumulative energy consumption deviates from the budget; Indicates driverless car n Communication energy consumption for uploading local model parameters; driverless car n The local model calculates energy consumption; Indicates the first t The total number of driverless cars selected in each round.

[0039] Preferably, the expression for the second sub-objective function is as follows:

[0040]

[0041] in, This indicates the client's selection; V For control parameters; This represents the gradient approximation error; and They represent the client respectively. i and n The local model gradient; It is a client i In the t Initial model parameters for each round; It is a client n In the t Initial model parameters for each round; The length of the virtual energy queue represents the client's... n The extent to which cumulative energy consumption deviates from the budget; Let be a binary choice variable, representing whether a client is selected in round t. n ; For the client n In the round t Total energy consumption; For the client n Maximum budget energy.

[0042] Preferably, the step of cutting the constraint conditions using the cutting plane method includes:

[0043] Cutting constraints are generated using objective function constraints, specifically including communication capability cutting and energy balance cutting.

[0044] The communication capability segmentation includes:

[0045] When the t The number of clients selected in the round When adding constraints, add the following conditions:

[0046]

[0047] in, For the superselect client set, For relaxation of the bivariate variable, ;

[0048] The energy balance cutting includes: when the client n Cumulative energy consumption Approaching maximum energy consumption When adding constraints, add the following conditions:

[0049]

[0050] in, For the client n The maximum energy reserves available throughout the entire federal learning cycle; For the client n In the t The total energy consumption during participation in federated learning includes local training energy consumption and communication energy consumption; This is the peak energy consumption point, which is the energy consumption generated during the round with the highest energy consumption. This refers to a set of critical time periods, i.e., periods when client energy consumption is significant; This represents the cumulative energy consumption during non-critical periods.

[0051] Secondly, this invention provides an air-ground cooperative training path planning system for target detection tasks, including an edge server and a client within an air-ground cooperative federated learning framework; wherein the edge server is a drone, and the client is... N A driverless car;

[0052] The edge server is used to initialize the target detection model as an initial global model and send the initial parameters to the client.

[0053] The client is used to receive initial parameters from the server, initialize the local model according to the initial parameters, construct and solve the global optimization target model during the local model training process, and determine the client selection and UAV trajectory during the local model training process. After completing the local training, the client uploads the encrypted model parameter increments to the server for aggregation. In the aggregation stage, the federated average algorithm is used to weight and fuse the parameter updates according to the proportion of data volume of each client to form a new global model.

[0054] The global optimization objective model includes an objective function and constraints. The objective function is defined as: minimizing global loss by jointly optimizing client selection and UAV trajectory.

[0055] Thirdly, the present invention provides a computer-readable storage medium, characterized in that it stores a computer program for air-ground cooperative training path planning for a target detection task, wherein the computer program causes a computer to execute the air-ground cooperative training path planning method for a target detection task as described above.

[0056] Thirdly, the present invention provides an electronic device, characterized in that it comprises:

[0057] One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including an air-ground cooperative training path planning method for performing an object detection task as described above.

[0058] (III) Beneficial Effects

[0059] This invention provides a method and system for air-ground cooperative training path planning for target detection tasks. Compared with existing technologies, it has the following advantages:

[0060] This invention addresses the disconnect between UAV trajectory planning and target detection tasks in existing technologies by constructing a globally optimized target model for target detection. Within an air-ground collaborative federated learning framework, prioritized key data accurately compensates for the model's learning shortcomings in complex scenarios and low-resolution targets, enabling the model to quickly focus on performance weaknesses for parameter updates. Furthermore, the association between trajectory planning and the core objective of "minimizing global loss" further ensures the effectiveness and representativeness of parameters uploaded by the ground-based UAV. Through weighted fusion using a federated averaging algorithm, the newly generated global model possesses more comprehensive feature representation capabilities and stronger generalization abilities. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a schematic diagram of a scenario for the air-ground collaborative federated learning framework in an embodiment of the present invention;

[0063] Figure 2 In the ablation experiment during simulation, a fixed step size is used. =0.5, weight V=0.1, UAV trajectory and client selection map;

[0064] Figure 3 In the ablation experiment during simulation, a fixed step size is used. =0.5, weight V=1, UAV trajectory and client selection map;

[0065] Figure 4 In the ablation experiment during simulation, a fixed step size is used. =0.5, weight V=3, UAV trajectory and client selection map;

[0066] Figure 5 In the ablation experiment during simulation, the weight V=1 is fixed, and the step size is... When =0.1, the drone's flight path and client selection map;

[0067] Figure 6 In the ablation experiment during simulation, the weight V=1 is fixed, and the step size is... When the value is 0.5, the drone's flight path and client selection map are shown.

[0068] Figure 7 In the ablation experiment during simulation, the weight V=1 is fixed, and the step size is... When =1, the drone flight path and client selection map;

[0069] Figure 8 This is a schematic diagram of the experimental results under the condition that the client datasets are not independent and identically distributed during the simulation.

[0070] Figure 9 This is a schematic diagram of the experimental results under the condition that the client datasets are independent and identically distributed during the simulation.

[0071] Figure 10 The simulation results show the flight path and client selection graph when using the SCA algorithm + greedy algorithm.

[0072] Figure 11 The simulation uses the SCA algorithm + cutting plane method to determine the flight path and client selection map.

[0073] Figure 12 The simulation uses the SCA algorithm + random selection algorithm to determine the flight path and client selection map.

[0074] Figure 13 This is a comparison chart of the loss values ​​of the cutting plane method and the other two algorithms during simulation.

[0075] Figure 14 This is a comparison chart of the energy consumption of the cutting plane method and the other two algorithms during simulation.

[0076] Figure 15 This is a schematic diagram of the global loss value of the YOLO model during simulation.

[0077] Figure 16 This is a diagram illustrating the global loss value of the YOLO model in a real-world scenario.

[0078] Figure 17 This is a comparison chart of loss values ​​in simulation and real-world scenarios. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] This application provides a method and system for air-ground cooperative training path planning for target detection tasks, which solves the technical problem that UAV trajectory planning and target detection tasks are completely disconnected in existing air-ground cooperative solutions. It tightly integrates UAV trajectory planning with target detection model training to achieve joint optimization of the two.

[0081] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows:

[0082] Existing air-ground coordination solutions mainly suffer from the following shortcomings:

[0083] 1. Most existing UAV trajectory planning methods are performed independently of the target detection task, resulting in low data acquisition efficiency. Existing trajectory planning algorithms are mainly optimized based on coverage, energy efficiency, or obstacle avoidance capabilities, completely ignoring the different contributions of different regions to the target detection task. For example, some regions may contain a large number of targets or complex scenes that are difficult to identify, which are more critical for model training, but existing algorithms assign the same weight to all regions.

[0084] 2. The random client selection strategy employed exhibits significant flaws in the air-to-ground collaborative system. This strategy completely ignores the obvious differences between devices, leading to a substantial decrease in system performance. Specifically, when randomly selecting clients with weak computing power or poor communication conditions, the entire training process is severely slowed down. Furthermore, the data quality collected by UAVs and ground vehicles is affected by various factors, including lighting conditions, shooting angle, and occlusion. The random selection strategy cannot identify and prioritize clients with high data quality, resulting in slow model convergence or even divergence. In some scenarios, introducing clients with low-quality data can degrade model performance. Additionally, the communication environment in the air-to-ground collaborative system is extremely complex, with signal strength and stability fluctuating dramatically with location. Simple random selection cannot avoid areas with poor communication conditions, increasing the risk of communication interruptions and data loss, severely impacting training efficiency and model quality.

[0085] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0086] This invention provides a method for air-ground cooperative training path planning for target detection tasks, using a drone as a mobile edge server in an air-ground cooperative federated learning framework. N The autonomous vehicle serves as a client device in the air-ground collaborative federated learning framework, which is like... Figure 1 As shown, the specific process of this method is as follows:

[0087] The drone initializes the target detection model as the initial global model and distributes the initial parameters to each drone vehicle.

[0088] Each autonomous vehicle receives initial parameters from the server, initializes its local model based on these parameters, and constructs and solves a global optimization target model during the local model training process. It also determines the client selection and drone trajectory during the local model training process. After completing local training, the client uploads encrypted model parameter increments to the server for aggregation. During the aggregation phase, a federated averaging algorithm is used to weight and fuse parameter updates according to the proportion of data volume from each client, forming a new global model.

[0089] The global optimization objective model includes an objective function and constraints. The objective function is defined as: minimizing global loss by jointly optimizing client selection and UAV trajectory.

[0090] The air-ground collaborative training path planning method is described in detail below:

[0091] In this embodiment of the invention, the core challenge in optimizing the training of a target detection model within an air-to-ground collaborative federated learning framework lies in balancing the constraints of UAV communication capabilities with optimal global training performance. Specifically, the communication coverage of UAVs has a physical upper limit, and limited communication bandwidth and signal transmission stability further restrict their data interaction capabilities; while the core objective of model training is to achieve optimal performance. TMinimizing the global loss function within each iteration cycle to achieve the optimal balance between target detection accuracy and generalization ability hinges on two key aspects: First, ensuring the UAV is positioned in the optimal spatial location for communication efficiency is crucial to guarantee data transmission quality and latency stability with ground nodes. Second, accurately selecting the clients (UAVs) with the highest information validity—those nodes holding high-value training data (such as samples from areas of model uncertainty or high-density target data) that can provide the maximum gain for global parameter updates—is essential. These two aspects are interconnected and indispensable: selecting clients outside the optimal communication location will lead to decreased training efficiency due to data transmission failures or excessive latency; neglecting data value in trajectory planning will waste communication resources on low-gain nodes, failing to support effective convergence of the global loss.

[0092] Based on this, the embodiments of the present invention decompose the above-mentioned practical problem into two core optimization dimensions: UAV trajectory planning and client selection, and achieve a collaborative solution through an alternating optimization strategy: dynamically matching communication coverage requirements with the distribution of high-value nodes through trajectory planning, and accurately identifying the training unit that contributes the most to the reduction of global loss through client selection, ultimately achieving a solution in... T Minimize the global loss function within each iteration to achieve the optimal training effect of the object detection model.

[0093] The embodiments of the present invention mainly consider the use of one unmanned aerial vehicle (UAV) and N Within an air-ground collaborative federated learning framework comprised of autonomous vehicles (UEs), specific problems are analyzed, and a globally optimized objective model is solved. The drones act as mobile edge servers, responsible for aggregating and distributing the global model, while the UEs act as client devices, responsible for training and updating the local model. Assume each UE... n ∈{1,2,…,N} stores a local training dataset. D n Its size is ,in and They represent the first j Each input sample and its corresponding label output.

[0094] The air-ground collaborative federated learning framework is as follows:

[0095] This framework implements federated learning based on Flask, a lightweight Python web application framework. Its core functionalities include:

[0096] Routing: Define HTTP endpoints using @app.route();

[0097] Request processing: Retrieve parameters and data using the request object;

[0098] Response handling: Use jsonify() to return JSON formatted data;

[0099] Server deployment: Start the development server using app.run().

[0100] The global model initialization process is as follows:

[0101] The models and datasets are hierarchically stored using a three-dimensional nested dictionary structure called global_models.

[0102] First dimension: Dataset type (MNIST / CIFAR10 / CIFAR100).

[0103] The second dimension: Model architecture (ResNet18 / VGG16 / GoogleNet);

[0104] The third dimension: specific model instances.

[0105] During initialization, a placeholder design is used, and all model pointers are initialized to None. Initialization is completed through the initialize_global_model module. The torch.save(model.state_dict(), path) function is used to generate a binary serialized .pth file with the file name formatted as "dataset_model type_global_model.pth" (e.g., MNIST_ResNet18_global_model.pth) to implement version identification and ensure that the initial download parameters of each client are consistent.

[0106] Meanwhile, the framework employs a dynamic adaptation mechanism, dynamically setting the number of classes based on the dataset name. The `num_classes` parameter is passed during model initialization, automatically adjusting the model's output layer structure. For example, the same ResNet18 class can be adapted to different datasets by passing different `num_classes` parameters, eliminating the need to create independent model classes for each dataset.

[0107] The client-side data partitioning and loading process is as follows:

[0108] During the simulation phase, one server (drone) and four ground clients (unmanned vehicles) were set up. The client data was divided using a linear equal-segmentation strategy: the data partition boundaries were determined by the start_portion and end_portion parameters, and continuous intervals were generated based on the sample order of the original dataset. Each client held 25% of the total training set data.

[0109] After data partitioning, client-side dataset files are generated according to the formats supported by the torchvision.datasets module, and the data is loaded using the get_dataloader module. Loading parameters include batch size (batch_size), dataset path, whether to shuffle the data, and multi-threaded data loading (num_workers), etc. The specific process is as follows:

[0110] 1. The client loads the dataset by specifying the dataset path and type;

[0111] 2. Use the DataLoader module to create a data loader, and configure the batch size, shuffling rules, and multi-threading parameters;

[0112] 3. The iterative data loader acquires batch data and labels for model training and evaluation.

[0113] The process of local model training and parameter update is as follows:

[0114] Local model training and parameter updates involve three steps: "acquisition - training - uploading".

[0115] (1) Acquisition stage:

[0116] The client sends a request to the server's / get_model interface using the HTTP GET method and the requests.get function. It receives a JSON-formatted dictionary of model parameters returned by the server, converts it into a PyTorch tensor, and loads it into the local model.

[0117] (2) Training phase:

[0118] Local training is the core step in client-side model optimization, and the process is as follows:

[0119] Data augmentation: Differentiated preprocessing is performed for different datasets, and all data is normalized to improve the model's generalization ability, accelerate convergence, and enhance numerical stability.

[0120] Forward propagation computation: After loading the global model parameters, the train() function performs forward computation. The input data is transferred to the GPU via inputs.cuda() and propagated layer by layer through network structures such as ResNet18 / VGG16 / GoogleNet (e.g., VGG16 contains 13 convolutional layers and 3 fully connected layers; in the feature extraction stage, spatial downsampling is performed through max pooling, and the final output is the class probability distribution).

[0121] Loss function calculation: The CrossEntropyLoss function is used, and its mathematical expression is:

[0122] In the formula, The total number of categories (e.g., 10 for CIFAR10, 100 for CIFAR100). One-hot encoding of the real label. This is the raw output of the last layer of the model.

[0123] The local loss function is defined as:

[0124] In the formula, Let be the loss function for the object detection task. For model parameters, This is a local dataset.

[0125] Backpropagation optimization: Gradients are automatically calculated using loss.backward(), and parameters (momentum term) are updated using the Adam optimizer. =0.9, adaptive learning rate =0.999, learning rate 0.001); after each batch of training, optimizer.zero_grad() is executed to clear the historical gradients, and the loss and accuracy for each epoch are calculated (output format example: Epoch 1, Loss: 0.1234, Accuracy: 95.67%).

[0126] (3) Upload stage:

[0127] After all clients complete 5 epochs of local training, they upload the model parameters and sample count (for federated aggregation weighting) to the server.

[0128] The server model aggregation and distribution process is as follows:

[0129] The FedAvg algorithm is the core of model aggregation. Its principle is to perform a weighted average of model parameters based on the client sample size. The mathematical expression is:

[0130] In the formula, These are the global model parameters after aggregation in round t+1. The number of clients participating in this round of aggregation, Let k be the local data volume of the kth client. Total data for all clients These are the model parameters for the k-th client after local training in round t.

[0131] The server receives parameters uploaded by clients via the ` / send_model` interface, deserializes the JSON parameters into PyTorch tensors, and stores them in the `client_models` list. After collecting all client parameters, it aggregates the parameters based on the number of local samples using the FedAvg algorithm to generate updated global model parameters. Clients access the ` / get_model` endpoint via an HTTP GET request to complete the parameter update.

[0132] The entire process uses a RESTful API for asynchronous communication. Each cycle includes the steps of "parameter upload → server aggregation → parameter distribution" to achieve air-ground collaborative training.

[0133] The optimization objective of the global optimization objective model is to jointly optimize client selection. and drone trajectory Minimize the global loss function :

[0134] (1)

[0135] In the formula, , For the client n In the T Loss function of the wheel:

[0136] (2)

[0137] in, Let be the loss function for the object detection task. These are global model parameters. For local datasets, N This represents the total number of clients.

[0138] (3)

[0139] in, The total number of target categories, This is the one-hot encoding of the true label class c of the current data sample. The one-hot encoding only contains the true label class c. =1, all others are 0. For the model to the true category c The original predicted score; This indicates that the local model applies to all possible categories. m The original predicted scores (from class 1 to class M);

[0140] The constraints are as follows:

[0141] (4)

[0142] (5)

[0143] (6)

[0144] (7)

[0145] (8)

[0146] (9)

[0147] Equation 4 represents the long-term energy constraint, indicating that each UE in t The total energy consumption of the wheels should not exceed its budget.

[0148] Equation 5 is a constraint on the concurrent capability of UAVs, indicating that the number of UEs connected to by the UAV in each round must not exceed the maximum number of unmanned vehicles that the UAV can connect to.

[0149] Equation 6 is a coverage constraint, which means that the selected UE must be within the communication coverage area of ​​the UAV.

[0150] Equation 7 represents the flight distance constraint for the UAV, indicating that the flight distance between adjacent rounds does not exceed [the specified value]. .

[0151] Equation 8 is a latency constraint, indicating that the total latency of the UE in each round does not exceed the threshold.

[0152] Equation 9 is a binary choice variable, indicating whether it is in the first... t Select from the wheel .

[0153] As can be seen from the above constraints, there are communication and energy constraints in the air-to-ground collaborative training process. These constraints need to take into account energy calculation and communication latency. Specifically:

[0154] Because local training of autonomous vehicles involves training time and energy consumption, establishing autonomous vehicle... n The local latency model and energy consumption model are as follows:

[0155] (10)

[0156] (11)

[0157] in, This is the number of iterations during local training. It is the number of CPU cycles required to process one sample. For local training dataset, It calculates the frequency. kIt is the effective capacitance coefficient.

[0158] After local training, the driverless car n Local model parameters The data is uploaded to the drone; therefore, the communication latency model and energy consumption model for air-to-ground collaboration are established as follows:

[0159] (12)

[0160] (13)

[0161] in, This refers to the size of the local training dataset, i.e., the number of data sets in the local training dataset. It is the uplink transmission rate. It is a drone m And driverless cars in t The distance between rounds. δ It's the signal-to-noise ratio. This is the UE's transmission power.

[0162] From the above model, it can be concluded that in the process of air-ground collaborative training, each epoch The total latency of the UE is the sum of the local calculation latency and the upload latency:

[0163] (14)

[0164] The total power consumption of the UE in each epoch is the sum of the computing power consumption and the upload power consumption:

[0165] (15)

[0166] In the objective function of Equation 1 above, there are two variables: drone position and client selection. Drone position... G Let A be the horizontal coordinates of the drone in each round, a continuous variable. The client selects A as the set of binary decision variables, which are integer variables. This problem involves mixed-integer nonlinear programming with coupled variables, and the energy constraint is a long-term constraint.

[0167] The heterogeneity of client data distribution (Non-IID) causes the local update direction to deviate from the global optimum, which is a core factor affecting convergence speed. According to knowledge of federated learning, convergence analysis in federated learning typically considers the heterogeneity of client data. Heterogeneity leads to increased gradient differences between different clients, thus concluding that the convergence of the global loss is strongly correlated with client gradient differences. Based on this conclusion, this embodiment of the invention introduces gradient approximation error. The objective function in Equation 1 can then be transformed into the form in Equation 16, and the global optimization objective model is as follows:

[0168] Objective function:

[0169] (16)

[0170] Constraints:

[0171] st(4)-(9)

[0172] In the formula:

[0173] (17)

[0174] in, and They represent the client respectively. i and n The local model gradient, It is a client i In the t Initial model parameters for each round.

[0175] The transformed problem is a submodulus maximization problem. This problem refers to the fact that due to resource constraints, it is impossible for all clients to participate in training at the same time. Therefore, this problem can select a subset of clients with the smallest gradient error when all clients participate as the optimal subset to replace all clients.

[0176] The energy constraint in Equation 4 is also a challenge. The client's long-term energy limitations couple client selection and trajectory planning to different learning rounds. Therefore, the Lyapunov optimization framework is adopted, which introduces a virtual energy queue. This approach decomposes long-term constraints into online optimization problems, ultimately yielding the immediate optimization problem to be solved in each round. That is, the objective function represented by Equation 16 can be rewritten in the form of Equation 18.

[0177] (18)

[0178] st(5)-(9)

[0179] (19)

[0180] in, The length of the virtual energy queue represents the client's... n The extent to which cumulative energy consumption deviates from the budget; V These are control parameters used to balance model convergence speed and energy consumption. V The larger the value, the more the algorithm tends to select clients that contribute more to model convergence, rather than energy-efficient clients. Select the approximate error function generated for the client to reflect the model performance loss; Let be a binary choice variable, representing whether a client is selected in round t. n ; For the client n In the round t Total energy consumption; T This represents the overall round of federal learning and training. Represents energy consumption status. It is a virtual energy queue generated by the Lyapunov optimization framework, representing the difference between energy consumption and budget, and its unit is energy. This queue value increases as client energy consumption accumulates, thus "penalizing" high-energy-consuming clients in subsequent rounds, reducing their probability of being selected, and achieving long-term energy balance.

[0181] In Equation 18, each round automatically adjusts client selection by minimizing the objective function that includes queue items. The queue acts as a feedback mechanism; if a client's cumulative energy consumption is too high, queue growth will inhibit the selection of that client in subsequent rounds, thus balancing energy consumption. This ensures that, in the long run, the energy consumption of each client does not exceed the budget.

[0182] The above transformation problem reduces the complexity of solving the objective function by alternately optimizing the client's choice A and the UAV trajectory G. Therefore, the problem will be solved by calculating the client's choice A and the UAV's position G respectively. Specifically:

[0183] When the client selects A as the fixed option =1 client set Determined, therefore For a constant value, the objective function in Equation 18 degenerates into an energy term that is only related to the trajectory:

[0184] (20)

[0185] Due to driverless cars n Local model calculation of energy consumption Since the frequency is determined locally by the client and is independent of the trajectory, it can be simplified to:

[0186] (twenty one)

[0187] Substitute local communication power consumption From the definition, we can obtain:

[0188] (twenty two)

[0189] st(6)-(8)

[0190] Due to the nonlinearity of the function and its constraints, slack variables can be introduced during implementation. The objective function 22 is then transformed into a linear form:

[0191] (twenty three)

[0192] Conditions (6)-(8) in the constraints should be changed to:

[0193] (twenty four)

[0194] (25)

[0195] The problem described above is transformed into a convex quadratic constrained quadratic programming (Convex QCQP) problem, which can therefore be solved using the CVX solver based on the Continuous Convex Approximation (SCA) method. The pseudocode for SCA is shown in Table 1.

[0196] Table 1 Pseudocode for SCA

[0197]

[0198] When the drone trajectory G is fixed, the objective function is transformed into a function that depends only on the client's choice A, i.e., Equation 18 is transformed into Equation 18:

[0199] (26)

[0200] st(4)-(9)

[0201] Under the condition of fixed UAV flight path, the client selection problem can be modeled as a mixed integer programming problem. In this embodiment of the invention, the cutting plane method is used to eliminate non-integer solutions by gradually adding constraints (cutting planes) to approximate the optimal solution.

[0202] First, enter the auxiliary variables. Relaxing the bivariate variables, for the bivariate variables in Equation 9 Perform linearization:

[0203]

[0204] The objective function in Equation 26 is rewritten as Equation 27:

[0205] (27)

[0206] The constraints are modified as follows:

[0207] st(4)-(8)

[0208] In the formula, Represents gradient bias metric:

[0209] (28)

[0210] Set initialization parameters and call Gurobi to solve the relaxation problem to obtain the initial solution. .

[0211] Testing relaxation solutions If the value is an integer, the result is obtained; otherwise, the cutting plane is generated using the objective function constraints to perform the cutting constraint.

[0212] Based on the above formula, the embodiments of the present invention design two types of cutting planes, specifically communication capability cutting and energy balance cutting.

[0213] The communication capabilities are divided as follows:

[0214] From equation 5 in the constraint conditions, we can obtain that when the first... t The number of clients selected in the round When adding constraints, use Equation 29:

[0215] (29)

[0216] in, For the set of superselected clients, i.e., the relaxation solution Given a set of clients with non-integer solutions, constraints are added to force a reduction in the number of selected clients.

[0217] The energy balance is cut as follows:

[0218] From Equation 4 in the constraints, we can obtain that when the client n Cumulative energy consumption Approaching maximum energy consumption Add a constraint – Constraint 30:

[0219] (30)

[0220] in, For the client n The maximum energy reserves available throughout the entire federal learning cycle; For the client n In the t The total energy consumption during participation in federated learning includes local training energy consumption and communication energy consumption; This is the peak energy consumption point, which is the energy consumption generated during the round with the highest energy consumption. This refers to a set of critical time periods, i.e., periods when client energy consumption is significant; This represents the cumulative energy consumption during non-critical periods.

[0221] In Equation 30, the summation on the left is the client's... nThe right side represents the number of times a selection is made during critical periods, while the left side represents the maximum number of peak energy consumption levels a client can tolerate under the remaining energy budget—that is, the maximum allowed number of selections calculated based on the remaining energy budget. This constraint limits the selection frequency during high-energy-consumption periods.

[0222] In air-to-ground collaborative federated learning, clients are distributed across different geographical locations with limited energy budgets. Furthermore, the movement paths of drones can affect communication quality, potentially leading to increased energy consumption for some clients and eventual energy depletion in subsequent rounds. To address long-term energy constraints, the Lyapunov method is introduced in the client selection problem. Each client maintains a virtual queue representing the difference between its energy consumption and budget. The length of the virtual queue determines which client constraints require priority for processing. This achieves a better balance between energy consumption and model accuracy. Equation 31 is used to determine priorities. Prioritization aims to normalize the process, transforming the queue states of each client into relative weights to determine the priority of generating the cutting plane. Then, an iterative solution is performed.

[0223] (31)

[0224] In the formula:

[0225] This is the virtual queue for the current client n; For the current number t The sum of all queues in each round of training.

[0226] The pseudocode for solving it is shown in Table 2:

[0227] Table 2 Pseudocode for the Cutting Plane Method

[0228]

[0229] In the above solution, an optimization model was established with the objective of minimizing the global loss function. However, the objective function contains two completely heterogeneous decision variables: a continuous high-dimensional variable, the UAV trajectory G, and a discrete combined variable, the client selection A. Directly solving them together would be too complex, classifying it as an NP-hard problem. Alternating optimization is needed to reduce complexity, decoupling the original problem into two parallel sub-problems. Furthermore, the UAV motion follows continuous dynamic constraints, making gradient-based optimization suitable, while the client selection is influenced by discrete device states (energy consumption, data volume), making integer programming more appropriate. The separation problem can be addressed using the SCA method and the cutting plane method, and the Lyapunov optimization framework can transform long-term energy constraints into instantaneous penalty terms. Therefore, this embodiment of the invention decomposes the optimization problem into two sub-problems, solving them separately and then performing alternating optimization.

[0230] The basic principle of alternating optimization is to decompose the coupling variables into multiple parts, optimize only one part at a time while keeping the others fixed. The actual steps are as follows:

[0231] 1. Initialization: First, set the initial drone trajectory and client selection variables to provide a starting point for the subsequent iterative optimization process.

[0232] 2. Iterative Process: In each iteration, the client selection variable A is fixed first, and then the continuous convex approximation (SCA) algorithm is used to optimize the UAV trajectory G to obtain the optimized trajectory. Next, the UAV trajectory G is fixed, and the cutting plane method is used to optimize the client selection variable A to obtain a new client selection scheme.

[0233] 3. Variable Update: After completing one round of trajectory and client selection optimization, update the UAV trajectory G, client selection variable A, and Lyapunov virtual queue to reflect the current optimization status and dynamic changes of the system.

[0234] 4. Convergence judgment: Calculate the change in the objective function. If the change is less than the set threshold or the maximum number of iterations is reached, the optimization process is considered to have converged and the iteration is stopped; otherwise, continue to the next round of iterative optimization.

[0235] By using this alternating optimization approach, the complex coupled variable problem is decomposed into two relatively simple subproblems that are solved separately, reducing the complexity of the problem and enabling each subproblem to be optimized more effectively.

[0236] The pseudocode for alternating optimization is shown in Table 3:

[0237] Table 3. Pseudocode for alternating optimization

[0238]

[0239] The effectiveness of the alternating optimization in the embodiments of the present invention is verified by simulation below:

[0240] Within a 500*500 area, 100 fixed clients were selected during the verification process. The datasets of all clients were set to IID (Independent and Identically Distributed). The specific simulation parameters are shown in Table 4.

[0241] Table 4 Simulation Parameter Settings

[0242]

[0243] In the mixed-integer nonlinear programming problem of UAV trajectory planning and client selection, the trajectory update step size is... The objective function weights V are key parameters that significantly influence experimental results, as shown in the pseudocode of the SCA method. = Step length The magnitude of the UAV position update in each iteration is determined. The weight V, as shown in Equation 26, is used to balance the model's variance and energy consumption. It determines the trade-off between model performance and energy consumption during optimization. The following verification process will conduct ablation experiments on these two parameters.

[0244] When the step size is fixed When V = 0.5, the effect of weight V on the experimental results is as follows: Figure 2 , Figure 3 , Figure 4 As shown.

[0245] in, Figure 2 The track length is 342.84, and the total number of selected clients is 100. Figure 3 The track length is 417.19, and the total number of selected clients is 200. Figure 4 The track length is 605.30, and the total number of selected clients is 100.

[0246] When the weight V=1 is fixed, the effect of the step size on the experimental results is as follows: Figure 5 , Figure 6 , Figure 7 As shown.

[0247] Figure 5 The track length is 354.16, and the total number of selected clients is 200. Figure 6 The track length is 417.19, and the total number of selected clients is 200. Figure 7 The track length is 601.96, and the total number of selected clients is 100.

[0248] about Figure 3 , Figure 5 , Figure 6 The explanation for why the total number of selected clients in the isograph is greater than 100: The 100 fixed clients refer to the entire pool of available data resources in the federated learning system. Throughout the simulation, the total number of this group remains fixed at 100. The total number of selected clients, however, is a dynamic scheduling result, representing the frequency with which the algorithm actually mobilizes and uses these resources during the T training rounds. Because a client can be repeatedly selected in multiple rounds, this value is usually greater than 100.

[0249] The ablation experiment results above show that under non-independent and identically distributed dataset conditions, the system focuses more on high-value clients, resulting in uneven distribution of trajectories and choices; while under independent and identically distributed conditions, the system behavior is more balanced. This is highly consistent with the client selection and trajectory optimization mechanisms in the code, indicating that the algorithm can adapt to data distribution characteristics and rationally allocate resources and optimize trajectories.

[0250] The above are the experimental results for the independent and identically distributed (i.i.d.) client datasets. Experimental results for non-independent and identically distributed (i.i.d.) datasets are as follows: Figure 8 , Figure 9 As shown (where, Figure 8 The track length is 285.93, and the total number of selected clients is 100. Figure 9 The track length is 342.84, and the total number of selected clients is 100. (Among them, Figure 8 , Figure 9 Medium weight V=3, step size =0.5. It should be noted that the choice of weight and step size values ​​mainly considers the following: The value of V directly affects the aggressiveness of the drone trajectory and client selection. In a Non-IID environment, due to the large differences in client data distribution, a larger V value is usually needed to promote model convergence, prioritizing clients with high data quality and significant contributions to the global model. In the Non-IID experiment, V was set to 3. This setting makes the drone trajectory more dispersed, and client selection focuses more on its data value rather than simply geographical location, to address data heterogeneity. Simultaneously, for Non-IID data, a broader exploration is needed to find clients that cover more valuable data distributions; therefore, a larger step size is preferred. In the Non-IID experiment, the step size σ was set to 0.5. This setting achieves a good balance between exploration (finding dispersed high-value clients) and utilization (optimizing around the current location), avoiding ineffective coverage of heterogeneous data areas due to a step size that is too small, or trajectory oscillation due to a step size that is too large.

[0251] All the above clients are symmetrically and fixedly distributed. In the next verification, 2-3 categories will be randomly selected from the datasets of each client and sampled. At the same time, the datasets of the clients are also non-independent and identically distributed.

[0252] in Figure 10 It is a combination of SCA algorithm and greedy algorithm. Figure 11 It uses the SCA algorithm combined with the cutting plane method. Figure 12 It consists of the SCA algorithm and a random selection algorithm. Figure 10 The track length is 354.16, and the total number of selected clients is 200. Figure 11 The track length is 692.85, and the total number of selected clients is 100. Figure 12 The flight path length is 464.37, and the total number of selected clients is 200.

[0253] From the different flight paths and client selection graphs of the three algorithms mentioned above, it is easy to see that:

[0254] 1. The goal of the greedy algorithm is to select the client with the greatest marginal benefit. The drone's flight path is short and concentrated, indicating that the greedy algorithm tends to develop in existing good areas rather than explore a wide range, and is prone to getting trapped in local optima.

[0255] 2. Randomized algorithms are unstructured explorations that cover a wide range of clients but are not very efficient.

[0256] 3. The cutting plane method exhibits obvious exploration characteristics, covering a wider area and producing more regular paths. In addition, the algorithm considers global constraints and can cover space more evenly. Therefore, the cutting plane method in this embodiment of the invention has certain advantages over the other two algorithms.

[0257] Using a client containing a random number of MNIST datasets, randomly distributed within a 500*500 area, this embodiment of the invention can alternately optimize these three methods with the SCA algorithm. By comparing the loss values ​​of the model training after selecting the trajectory and client, and the energy consumption results for each game obtained, the following can be observed: Figure 13 , Figure 14 As shown.

[0258] By observing the loss values ​​and energy consumption of each algorithm, the following conclusions can be drawn:

[0259] 1. The greedy algorithm shows a rapid decrease in global loss value in the early stage, but the loss value is relatively high in the later stage. The random selection method shows a slow decrease and fluctuations in the early stage. In comparison, the cutting plane method shows a faster and more stable decrease in global loss value in the early stage, which reflects the stability and effectiveness of the cutting plane method in the optimization process.

[0260] 2. The energy consumption of the greedy algorithm decreases rapidly with the increase of rounds, rises slightly after the fourth round and then stabilizes, indicating that the result of the greedy algorithm tends to be more local optimal. In contrast, the energy consumption of the cutting plane method decreases steadily with the increase of rounds, and the final result is much lower than that of the other two algorithms, which proves the stability and effectiveness of the cutting plane method.

[0261] The YOLOv8 model (i.e., the object detection model) was trained using the aforementioned air-ground collaborative federated learning framework. The dataset was the COCO dataset, with the local datasets of the four clients being independent and identically distributed, each receiving an equal 25% of the training set (approximately 29,572 images). Local training epochs were 3, and the global federated learning rounds were 10. A simulation diagram illustrating the global loss value of the YOLO model is shown below. Figure 15 As shown.

[0262] To further verify the effectiveness of the embodiments of the present invention, this verification process also includes performance verification of target detection in real-world scenarios, as detailed below:

[0263] This experiment used one P450 research drone as the server, three NVIDIA Jetson Orin Nanos, one R550 ROS educational robot, and one R550 PLUS ROS research robot as clients 1-5 respectively. The COCO dataset was used for training. The five clients had identically distributed datasets, each representing 20% ​​of the dataset (approximately 23,657 images). The actual hardware parameters are shown in Table 5.

[0264] Table 5 Actual Hardware Equipment Parameters

[0265]

[0266] In this real-world scenario, the global loss value of the YOLO model is as follows: Figure 16 As shown.

[0267] The comparison chart of loss values ​​between simulation and real-world scenarios is shown below. Figure 17 As shown.

[0268] The following conclusions can be drawn from the comparison of loss values ​​in the above simulation and real-world scenarios:

[0269] 1. Both simulation results and experimental (real-world scenario verification) results show a decreasing trend with the increase of training rounds, indicating that under the federated learning framework designed in this project, the YOLO model can gradually optimize its parameters and learn more effective features.

[0270] 2. The loss value in the experimental environment was consistently higher than that in the simulated environment, and the gap may gradually stabilize or slightly widen with each round. This exposes the performance gap between actual deployment and idealized simulation scenarios.

[0271] 3. The loss curve obtained through experiments may fluctuate due to inaccurate or untimely model updates caused by network latency, packet loss, or communication noise affecting data transmission between the client and server. In practical implementation, custom weights can be used to make the aggregation of model parameters more robust, depending on the specific circumstances.

[0272] In summary, the designed air-ground collaborative distributed machine learning framework can effectively improve the accuracy and robustness of target detection, reduce communication latency, and balance the computational load.

[0273] This invention also provides an air-ground cooperative training path planning system for target detection tasks, including an edge server and a client within an air-ground cooperative federated learning framework; wherein the edge server is a drone, and the client is... N A driverless car;

[0274] The edge server is used to initialize the target detection model as an initial global model and send the initial parameters to the client.

[0275] The client is used to receive initial parameters from the server, initialize the local model according to the initial parameters, construct and solve the global optimization target model during the local model training process, and determine the client selection and UAV trajectory during the local model training process. After completing the local training, the client uploads the encrypted model parameter increments to the server for aggregation. In the aggregation stage, the federated average algorithm is used to weight and fuse the parameter updates according to the proportion of data volume of each client to form a new global model.

[0276] The global optimization objective model includes an objective function and constraints. The objective function is defined as: minimizing global loss by jointly optimizing client selection and UAV trajectory.

[0277] It is understood that the air-ground cooperative training path planning system for target detection tasks provided in this embodiment of the invention corresponds to the air-ground cooperative training path planning method for target detection tasks described above. The explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the air-ground cooperative training path planning method for target detection tasks, and will not be repeated here.

[0278] This invention also provides a computer-readable storage medium storing a computer program for air-ground cooperative training path planning for a target detection task, wherein the computer program causes a computer to execute the air-ground cooperative training path planning method for a target detection task as described above.

[0279] This invention also provides an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including an air-ground cooperative training path planning method for performing an object detection task as described above.

[0280] In summary, compared with existing technologies, it has the following beneficial effects:

[0281] 1. This invention addresses the shortcomings of existing technologies where UAV trajectory planning and target detection tasks are disconnected by constructing a globally optimized target model for target detection tasks. Under the air-ground collaborative federated learning framework, prioritized key data can accurately compensate for the model's learning weaknesses in complex scenarios and low-resolution targets, enabling the model to quickly focus on performance weaknesses for parameter updates. Simultaneously, the association between trajectory planning and the core objective of "minimizing global loss" further ensures the effectiveness and representativeness of parameters uploaded by the ground-based UAV. After weighted fusion using a federated averaging algorithm, the newly generated global model possesses more comprehensive feature representation capabilities and stronger generalization abilities.

[0282] 2. Based on the differences in computing power, data distribution, communication cost, and model contribution of ground nodes, this invention proposes a mixed integer programming strategy based on the cutting plane method. This strategy can dynamically select nodes participating in federated learning based on weight indicators such as data quality, local model accuracy, and energy consumption status, so as to ensure the convergence efficiency and stability of the global model.

[0283] 3. Aiming at minimizing the global loss function, this embodiment of the invention establishes a trajectory planning mathematical model and uses the Continuous Convex Approximation (SCA) method to solve the non-convex optimization problem. This algorithm generates an efficient flight path by comprehensively considering constraints such as UAV flight speed, communication radius, ground node distribution, and dynamic obstacle avoidance.

[0284] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0285] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for air-ground cooperative training path planning for target detection tasks, characterized in that, One drone is used as a server in the air-ground collaborative federated learning framework. N The autonomous vehicle serves as a client in the air-ground collaborative federated learning framework. The air-ground collaborative training path planning method includes: The server initializes the target detection model as the initial global model and distributes the initial parameters to each autonomous vehicle. Each client receives the initial parameters from the server, initializes its local model based on the initial parameters, and constructs and solves the global optimization target model during the local model training process, determining the client selection and drone trajectory during the local model training process; after completing local training, the client uploads the encrypted model parameter increments to the server for aggregation; during the aggregation phase, the federated averaging algorithm is used to weight and fuse the parameter updates according to the proportion of data volume of each client to form a new global model; The global optimization objective model includes an objective function and constraints. The objective function is defined as: minimizing the global loss by jointly optimizing the client selection and the UAV trajectory. The objective function includes: In the formula: in, This indicates the client's selection; , indicating the drone's trajectory; V For control parameters; This represents the gradient approximation error; and They represent the client respectively. i and n The local model gradient; It is a client i In the t Initial model parameters for each round; It is a client n In the t Initial model parameters for each round; The length of the virtual energy queue represents the client's... n The extent to which cumulative energy consumption deviates from the budget; Let be a binary choice variable, representing whether a client is selected in round t. n ; For the client n In the round t Total energy consumption; For the client n Maximum budget energy; T This represents the overall round of federal learning and training.

2. The air-ground cooperative training path planning method for target detection tasks as described in claim 1, characterized in that, The constraints include: （1.1） （1.2） （1.3） （1.4） （1.5） Constraint 1.1 is a drone concurrency constraint, meaning that the number of drones connecting to unmanned vehicles in each round must not exceed the maximum number of unmanned vehicles that a drone can connect to. ; Constraint 1.2 is a coverage constraint, which means that the selected unmanned vehicle must be within the communication coverage area of ​​the drone; Constraint 1.3 is a flight distance constraint for the UAV, indicating that the flight distance of the UAV between adjacent rounds shall not exceed [the specified distance]. ; Constraint 1.4 is a time delay constraint, meaning that the total time delay of the autonomous vehicle in each round does not exceed a threshold. ; Constraint 1.5 is a binary choice variable, indicating whether it is in the first... t Choosing driverless cars in the wheel n ; in, T Represents the overall total number of rounds of federated learning and training; Indicates drone m and driverless cars n In the t The distance between rounds; Indicates drone m The communication coverage area; This represents the distance between the drone in round t and round t+1. Choose a binary variable. =1 indicates that in the first... t Choosing driverless cars in the wheel n , When =0, it means that in the first... t Wheelchair did not select driverless car n ; Indicates driverless car n Local model computation latency; The communication latency of autonomous vehicle n uploading local model parameters in round t.

3. The air-ground cooperative training path planning method for target detection tasks as described in any one of claims 1 to 2, characterized in that, The solution to the global optimization objective model includes: The global optimization objective model is transformed into alternating optimization of client selection and UAV trajectory planning, specifically: In each iteration, the client selection variable A is fixed first, and the objective function is transformed into a first sub-objective function that is only related to the trajectory. The first sub-objective function is solved using the continuous convex approximation algorithm to optimize the UAV trajectory G and obtain the optimized UAV trajectory. Then, the UAV trajectory G is fixed, and the objective function is transformed into a second sub-objective function that is only related to the client selection A. The constraint conditions are cut using the cutting plane method to obtain a new client selection scheme.

4. The air-ground cooperative training path planning method for target detection tasks as described in claim 3, characterized in that, The first sub-objective function is expressed as follows: Due to driverless cars n Local model calculation of energy consumption Since the frequency is determined locally by the client and is independent of the trajectory, it can be simplified to the following expression: in, Indicates the length of the virtual energy queue, and indicates the client. n The extent to which cumulative energy consumption deviates from the budget; Indicates driverless car n Communication energy consumption for uploading local model parameters; driverless car n The local model calculates energy consumption; Indicates the first t The total number of driverless vehicles selected in each round.

5. The air-ground cooperative training path planning method for target detection tasks as described in claim 3, characterized in that, The expression for the second sub-objective function is as follows: in, This indicates the client's selection; V For control parameters; This represents the gradient approximation error; The length of the virtual energy queue represents the client's... n The extent to which cumulative energy consumption deviates from the budget; Let be a binary choice variable, representing whether a client is selected in round t. n ; For the client n In the round t Total energy consumption.

6. The air-ground cooperative training path planning method for target detection tasks as described in claim 3, characterized in that, The method of cutting the constraint conditions using the cutting plane method includes: Cutting constraints are generated using objective function constraints, specifically including communication capability cutting and energy balance cutting. The communication capability segmentation includes: When the t The number of clients selected in the round When adding constraints, add the following conditions: in, For the superselect client set, For relaxation of the bivariate variable, ; The energy balance cutting includes: when the client n Cumulative energy consumption Approaching maximum energy consumption When adding constraints, add the following conditions: in, For the client n The maximum energy reserves available throughout the entire federal learning cycle; For the client n In the t The total energy consumption during participation in federated learning includes local training energy consumption and communication energy consumption; This is the peak energy consumption point, which is the energy consumption generated during the round with the highest energy consumption. This refers to a set of critical time periods, i.e., periods when client energy consumption is significant; This represents the cumulative energy consumption during non-critical periods.

7. A space-ground cooperative training path planning system for target detection tasks, characterized in that, This includes edge servers and clients in an air-to-ground collaborative federated learning framework; the edge server is one drone, and the client is... N A driverless car; The edge server is used to initialize the target detection model as an initial global model and send the initial parameters to the client. The client is used to receive initial parameters from the server, initialize the local model according to the initial parameters, construct and solve the global optimization target model during the local model training process, and determine the client selection and UAV trajectory during the local model training process. After completing the local training, the client uploads the encrypted model parameter increments to the server for aggregation. In the aggregation stage, the federated average algorithm is used to weight and fuse the parameter updates according to the proportion of data volume of each client to form a new global model. The global optimization objective model includes an objective function and constraints. The objective function is defined as: minimizing the global loss by jointly optimizing the client selection and the UAV trajectory. The objective function includes: In the formula: in, This indicates the client's selection; , indicating the drone's trajectory; V For control parameters; This represents the gradient approximation error; and They represent the client respectively. i and n The local model gradient; It is a client i In the t Initial model parameters for each round; It is a client n In the t Initial model parameters for each round; The length of the virtual energy queue represents the client's... n The extent to which cumulative energy consumption deviates from the budget; Let be a binary choice variable, representing whether a client is selected in round t. n ; For the client n In the round t Total energy consumption; For the client n Maximum budget energy; T This represents the overall round of federal learning and training.

8. A computer-readable storage medium, characterized in that, It stores a computer program for air-ground cooperative training path planning for target detection tasks, wherein the computer program causes a computer to execute the air-ground cooperative training path planning method for target detection tasks as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including a method for performing an air-ground cooperative training path planning method for an object detection task as described in any one of claims 1 to 6.

Images