One-shot federated prompting learning method for uav swarm
By employing a one-time communication federated prompting learning method, the collaborative training problem of UAV swarms under resource and privacy protection was solved, enabling efficient personalized model deployment and task execution, and improving the collaborative learning capability and task efficiency of UAV swarms.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-30
- Publication Date
- 2026-06-23
AI Technical Summary
In collaborative learning, drone swarms face challenges such as limited bandwidth and stability of airborne communication links, limited endurance, and limited computing and storage resources. Traditional federated learning methods have failed to effectively adapt to the core characteristics of drone swarms, resulting in increased task latency, high power consumption, poor model adaptability, and low training efficiency, while also lacking privacy protection optimization.
A one-time communication federated prompt learning method is adopted. The pre-trained model and initial prompt parameters are distributed by the central server, and the client performs decentralized processing, extracts the shared semantic main direction and generates personalized components, combines interpolation mixing and random masking to generate enhanced prompt parameters, performs multi-loss optimization, and finally updates the global prompt parameters and deploys personalized models.
It enables efficient collaborative training and task deployment of UAV swarms while ensuring data privacy and resource adaptation, reducing communication costs and improving the model's personalized adaptability and task execution efficiency.
Smart Images
Figure CN122264033A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, and in particular to a one-time communication federated cueing learning method for UAV swarms. Background Technology
[0002] With their core advantages of flexible deployment, wide-area coverage, and collaborative operation, drone swarms have been widely applied in key areas such as forest fire prevention, urban security, agricultural inspection, and disaster relief. Their collaborative learning capabilities directly determine the efficiency and accuracy of visual tasks. However, in actual operational scenarios, drone swarms face significant technical bottlenecks: Firstly, the bandwidth of airborne communication links is limited and their stability is insufficient. At the same time, the endurance of drones is constrained by power. Traditional federated learning requires multiple rounds of client-server communication, which not only increases task latency but also consumes a lot of power due to frequent data transmission, seriously affecting the swarm's operational time. Secondly, different drones have significant differences in their operational scenarios, flight angles, and shooting environments, resulting in highly heterogeneous data distribution. Existing collaborative learning methods have not effectively decoupled globally shared knowledge from local personalized knowledge. In addition, drones have limited computing power and storage capacity, making it difficult to adapt to traditional full-parameter training or complex optimization schemes, resulting in poor local model adaptability and low training efficiency.
[0003] While existing federated prompting learning methods reduce computational overhead through lightweight prompt parameter optimization and adapt to resource-constrained scenarios to some extent, they lack customized design for the core characteristics of drone swarms: either they rely on multi-round communication interactions to balance global consistency and local personalization, failing to meet the limited bandwidth and endurance requirements of drones; or they fail to achieve efficient decoupling between globally shared semantics and locally personalized semantics, making it difficult to cope with the highly heterogeneous nature of drone data; and at the same time, they lack targeted optimization for drone privacy protection scenarios, and the mode of raw data transmission or over-reliance on the global model still poses security risks. Summary of the Invention
[0004] In view of this, the present invention proposes a one-time communication federation prompting learning method for UAV swarms, which enables efficient collaborative training and task deployment of UAV swarms while ensuring data privacy and resource adaptation.
[0005] The technical solution of this invention is implemented as follows: A one-time communication federated prompting learning method for drone swarms includes the following steps: Step S1: The central server distributes the pre-trained visual-language model and initial global prompting parameters to each drone client. Each drone client initializes the initial local prompting parameters and builds a local private dataset. Step S2: The drone client decentralizes the initial global prompting parameters, extracts the shared semantic main direction, projects the initial local prompting parameters onto the orthogonal subspace of the shared semantic main direction, and obtains personalized components. Step S3: Based on the personalized components and the initial local cue parameters, generate enhanced cue parameters through interpolation mixing and random masking; Step S4: Based on the local private dataset, perform multi-loss joint optimization on the enhanced cue parameters and personalized components, and obtain the optimized local cue parameters; Step S5: The central server updates the initial global prompt parameters based on the optimized local prompt parameters and sends the updated global prompt parameters to the drone client.
[0006] Step S6: The drone client builds and deploys a personalized model based on the personalized components and the updated global prompt parameters.
[0007] Preferably, the visual-language model includes an image encoder and a text encoder, and the initial global cue parameters... Let M be a set of learnable context vectors of length M and embedding dimension d, satisfying The initial local prompt parameters .
[0008] Preferably, the specific steps for constructing the local private dataset in step S1 are as follows: The drone client i performs deblurring, denoising, and cleaning processes on the raw image data it collects, and then labels the data according to the categories of downstream tasks to build a local private dataset. Local private dataset ,in For image data, These are the category labels corresponding to the image data.
[0009] Preferably, the specific steps of step S2 are as follows: The drone client calculates the initial global cue parameters. Mean vectors of each dimension The mean vector of each dimension The j-th component is:
[0010] Where M is the initial global prompt parameter. Length, for The j-th element of the m-th vector; The mean vector of each dimension The mean matrix is formed after M copies. , set initial global prompt parameters Subtract the mean matrix row by row Then received a decentralized global hint ; Decentralized global hints Perform a low-rank approximate decomposition to obtain the orthogonal basis directions of the embedding space. Select the first principal component As The main direction is determined and normalized to obtain the shared semantic main direction u. The expression of the shared semantic main direction u is:
[0011] in To prevent division by zero of small constants; Calculate initial local prompt parameters Parallel components along the shared semantic principal direction u The expression for the parallel component is: ,in For the transpose of the shared semantic main direction u; Initial local prompt parameters Subtract parallel components Personalized components are then obtained. .
[0012] Preferably, the specific steps of step S3 are as follows: Using the mixup strategy to personalize components and initial local prompt parameters Generate mixed prompts The hybrid prompt The expression is: ,in Hyperparameters for controlling the shape of the distribution; Generate Bernoulli random variables ,in Let j = 1, 2, ..., M be the probability of dropping, and M be the initial global hint parameter. Length, enhanced prompt parameters The j-th vector satisfies: .
[0013] Preferably, the specific steps of step S4 are as follows: Based on a local private dataset, a class-discriminatory loss is applied to the enhanced cue parameters and personalized components. Two-way enhanced loss and semantic contrast loss Joint optimization of multiple losses; Construct the total loss function ,in and To obtain the loss weights, the optimized local cueing parameters are obtained through gradient descent optimization. ,in N represents the number of drone clients.
[0014] Preferably, the category discrimination loss The cosine similarity calculation based on the image encoder and text encoder is calculated using the following formula:
[0015] in For local private datasets, for Image data in The category labels corresponding to the image data. Represents cosine similarity. Where C is the temperature coefficient and C is the number of downstream task categories. The text hint for category k.
[0016] Preferably, the bidirectional enhancement loss is used to constrain the semantic consistency between the enhanced cue parameters and the initial local cue parameters, and the normalized text features of the enhanced cue parameters are: The normalized text features of the initial local prompt parameters are: B is the feature batch size. and The similarity matrix between them is defined as follows:
[0017] in Temperature coefficient; The bidirectional enhancement loss The expression is:
[0018] Where B is the batch size. .
[0019] Preferably, the specific steps for optimizing the semantic contrast loss are as follows: Constructing text features , as well as ,in , , , For personalized portions, To enhance the prompt parameters, For global reflection hints, the expression is: ,in For initial global hint parameters, To share the main semantic direction; The semantic contrast loss The expression is:
[0020] Where C represents the number of downstream task categories. This is the temperature coefficient.
[0021] Preferably, the updated global prompt parameters in step S5 The expression is:
[0022] Where N is the number of drone clients. For optimized local prompt parameters The elements in The aggregation weight for drone client i is determined by the local private dataset, and its expression is:
[0023] in and These are the local private datasets of drone clients i and j, respectively.
[0024] Compared with the prior art, the beneficial effects of the present invention are: This invention presents a federated cue learning method for drone swarms. After a central server distributes the corresponding visual-language model and initial global cue parameters to all drone clients, each drone client initializes its local cue parameters and constructs a local private dataset. The initial global cue parameters are then decentralized and decomposed into low rank. After extracting the shared semantic principal direction, projection is performed. The resulting personalized components are then structurally mixed and masked to generate semantically consistent enhanced cue parameters. The personalized components and enhanced cue parameters are then used as positive sample pairs, while the enhanced cue parameters and global reflection cue are used as negative sample pairs. Multi-patrol joint optimization separates shared and personalized semantics. Finally, the global cue parameters are updated based on the optimized local cue parameters and sent to the drone clients. The drone clients can then construct and deploy personalized models based on the personalized components and updated global cue parameters, ensuring data privacy and resource adaptation, and enabling efficient collaborative training and task deployment. Attached Figure Description
[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0026] Figure 1 A flowchart of the one-time communication federation prompting learning method for UAV swarms according to the present invention; Figure 2 This is a performance scalability comparison chart of an embodiment of the one-time communication federation prompting learning method for UAV swarms of the present invention under different numbers of clients; Figure 3 This is a performance comparison chart of an embodiment of the one-time communication federation prompt learning method for UAV swarms of the present invention under different learnable prompt lengths. Figure 4 This is a sensitivity analysis diagram of the loss function weight hyperparameters for an embodiment of the federated prompting learning method for unmanned aerial vehicle (UAV) swarms according to the present invention. Detailed Implementation
[0027] To better understand the technical content of this invention, a specific embodiment is provided below, and the invention will be further described in conjunction with the accompanying drawings.
[0028] See Figure 1 The present invention provides a one-time communication federation prompting learning method for UAV swarms, comprising the following steps: Step S1: The central server distributes the pre-trained visual-language model and initial global prompting parameters to each drone client. Each drone client initializes the initial local prompting parameters and builds a local private dataset. In the federated setup of this invention, there are N drone clients and one central server. Each drone client builds a local private dataset D and configures a pre-trained vision-language model, which includes an image encoder. and text encoder The initial global prompt parameters Let M be a set of learnable context vectors of length M and embedding dimension d, satisfying The initial local prompt parameters .
[0029] The specific steps for constructing a local private dataset are as follows: The drone client i performs deblurring, denoising, and cleaning processing on the raw image data it collects, and then labels the data according to the category labels of the downstream tasks to construct the local private dataset. Local private dataset ,in For image data, These are the category labels corresponding to the image data.
[0030] Step S2: The drone client decentralizes the initial global cue parameters, extracts the shared semantic main direction, projects the initial local cue parameters onto the orthogonal subspace of the shared semantic main direction, and obtains personalized components. The specific steps are as follows: The drone client calculates the initial global cue parameters. Mean vectors of each dimension The mean vector of each dimension The j-th component is:
[0031] Where M is the initial global prompt parameter. Length, for The j-th element of the m-th vector can specifically reflect At the center position of each semantic dimension The mean vector of each dimension After being copied M times, a mean matrix with the same shape as the initial global hint parameters is formed. , set initial global prompt parameters Subtract the mean matrix row by row Then received a decentralized global hint This effectively eliminates overall translational deviations and effectively reflects the main direction of internal changes.
[0032] Decentralized global hints Perform a low-rank approximate decomposition to obtain the orthogonal basis directions of the embedding space. Select the first principal component As The main direction is determined and normalized to obtain the shared semantic main direction u. The expression of the shared semantic main direction u is:
[0033] in To prevent division by zero of small constants; Calculate initial local prompt parameters Parallel components along the shared semantic principal direction u The expression for the parallel component is: ,in For the transpose of the shared semantic main direction u; Initial local prompt parameters Subtract parallel components Personalized components are then obtained. This enables drone clients to express independent semantics without conflicting with global trends, achieving clear semantic decoupling. In federated prompt learning, the global model and each drone client model achieve knowledge transfer by sharing prompt parameters, thereby promoting personalized learning in heterogeneous data scenarios. However, existing methods jointly optimize global and local prompts in the same semantic feature space. This directly coupled training mechanism easily leads to the coupling of shared and personalized semantics, making it difficult for local prompts to learn independent representations reflecting the client, thus reducing the model's adaptability in heterogeneous scenarios. To effectively separate shared knowledge from personalized features, this invention introduces orthogonal subspace constraints in the prompt space, geometrically binding the initial global prompt parameters. With initial local prompt parameters To decouple the features, we model shared features along the main semantic direction and learn client-specific features in its orthogonal subspace.
[0034] Step S3: Based on the personalized components and initial local cue parameters, enhanced cue parameters are generated through interpolation mixing and random masking. The specific steps are as follows: Using the mixup strategy to personalize components and initial local prompt parameters Generate mixed prompts The hybrid prompt The expression is: ,in Hyperparameters for controlling the shape of the distribution; Introducing a mixup enhancement mechanism at the cue level improves the model's stability under semantic perturbations and provides the model with diverse data distribution perspectives. Subsequently, some cue semantics are overflowed, making it more robust to noise and distribution shifts, which helps to limit personalization to reliable dimensions.
[0035] Generate Bernoulli random variables ,in Let j = 1, 2, ..., M be the probability of dropping, and M be the initial global hint parameter. Length, enhanced prompt parameters The j-th vector satisfies: .
[0036] Step S4: Based on the local private dataset, perform joint optimization of the enhanced cue parameters and personalized components using multiple losses to obtain the optimized local cue parameters. The specific steps are as follows: Based on a local private dataset, a class-discriminatory loss is applied to the enhanced cue parameters and personalized components. Two-way enhanced loss and semantic contrast loss Joint optimization of multiple losses; Category-based loss The cosine similarity calculation based on the image encoder and text encoder is calculated using the following formula:
[0037] in For local private datasets, for Image data in The category labels corresponding to the image data. Represents cosine similarity. Where C is the temperature coefficient and C is the number of downstream task categories. The text hint for category k.
[0038] The bidirectional enhancement loss is used to constrain the semantic consistency between the enhanced cue parameters and the initial local cue parameters. The normalized text features of the enhanced cue parameters are: The normalized text features of the initial local prompt parameters are: B is the feature batch size. and The similarity matrix between them is defined as follows:
[0039] in Temperature coefficient; The bidirectional enhancement loss The expression is:
[0040] Where B is the batch size. .
[0041] Without adding constraints, the enhanced This may cause the prompt to deviate from the initial local prompt parameters. Therefore, this invention designs a bidirectional enhancement loss, which integrates consistency constraints in two directions to keep the enhanced cue parameters and the initial local cue parameters aligned in the semantic space. Structurally, it connects the enhanced cue parameters with the orthogonal mechanism of the main direction, and in terms of optimization, it ensures the stable convergence of the model and consistent generalization of semantics across clients.
[0042] The specific steps for optimizing semantic contrast loss are as follows: Constructing text features , as well as ,in , , , For personalized portions, To enhance the prompt parameters, For global reflection hints, the expression is: ,in For initial global hint parameters, To share the main semantic direction, excessive concentration of global cues in the main direction can be suppressed, and then... and Form positive sample pairs, and Form negative sample pairs; The semantic contrast loss The expression is:
[0043] Where C represents the number of downstream task categories. As a temperature coefficient, positive sample pairs encourage the model to maintain semantic consistency under enhanced perturbation, while negative sample pairs force personalized prompts and global prompts to separate in the embedding space, thereby enhancing hierarchical expression.
[0044] To further strengthen the difference constraints between global and local semantics, this invention designs semantic contrastive learning, which enhances... With learnable Zooming in maintains high semantic consistency within the positive traffic flow, and removes vectors along the principal direction from global cues. By constructing positive and negative sample pairs in the cue space to enhance the separation between personalized features and global features, the consistency of personalized semantics in orthogonal subspaces can be effectively ensured.
[0045] Then, the total loss function is constructed by combining the aforementioned category discrimination loss constraint and bidirectional enhancement loss constraint. ,in and To obtain the loss weights, the optimized local cueing parameters are obtained through gradient descent optimization. ,in N represents the number of drone clients. This enables effective learning of semantic hierarchy, allowing shared semantics to aggregate along the main direction and personalized features to be effectively unfolded in orthogonal subspaces. This significantly enhances the semantic expressive power of the local model in heterogeneous federated environments.
[0046] Step S5: The central server updates the initial global prompting parameters based on the optimized local prompting parameters, and sends the updated global prompting parameters to the drone client. The expression is:
[0047] Where N is the number of drone clients. For optimized local prompt parameters The elements in The aggregation weight for drone client i is determined by the local private dataset, and its expression is:
[0048] in and These are the local private datasets of drone clients i and j, respectively.
[0049] Federated cue learning enables drone clients to communicate with the server only once, with each drone client i independently updating its local cue parameters based on data distribution. Suppose that in the t-th round of federated iteration, the server can collect optimized local cue parameters from N drone clients and then update the global cue parameters.
[0050] Step S6: The drone client builds and deploys a personalized model based on the personalized components and the updated global cue parameters. After deploying the personalized model, it can be used for inference in downstream vision tasks.
[0051] To verify the effectiveness of the present invention, the following embodiments are provided: Dataset setup: The CLIP dataset, which includes Food101, DTD, Caltech101, Flowers102, and OxfordPets, is provided. The CLIP dataset covers multiple vision domains and can fully reflect the generalization ability of the model.
[0052] Model setup: Two backbone networks are used: ResNet-50 and ResNet-101. Both models are initialized using ImageNet pre-trained weights and adapted to each dataset during federated training.
[0053] Baseline method: To comprehensively evaluate the performance of this embodiment, it is compared with other representative federated prompting learning methods, wherein: FedSCP: The method of this embodiment.
[0054] PromptFL: A unified framework for federated learning based on prompt words.
[0055] PromptFL+Per: Introduces a personalization module on top of PromptFL.
[0056] FedPGP: Balancing Individualization and Generalization Using Low-Rank Decomposition FedOTP: Enhances model personalization by employing optimal transportation mechanisms.
[0057] FedPHA: Using SVD for cue word decomposition preserves the characteristics of the local model.
[0058] GPR-NIAM: A cue learning method for one-shot federation.
[0059] Implementation details: To ensure the fairness of the experimental results, this embodiment was evaluated under the same experimental settings as all baseline methods, with the local training rounds set to (E_l=10) and the global communication rounds set to (E_g=1). The learnable local cue word length (context length) was set to 4, and its embedding dimension was 512. The learning rate was set to 1×10^6. - ³. The training batch size was 32, and the test batch size was 128. Five seeds (0-4) were used, with a discard probability of 0.1. All experiments were performed on an NVIDIA GeForce RTX 3090 GPU.
[0060] Metrics: Two standard classification metrics are used: accuracy (ACC) and F1 score (F1). Accuracy measures the overall correctness of the model's predictions, while the F1 score combines precision and recall, providing a more comprehensive reflection of model performance even in cases of imbalanced sample distribution.
[0061] In comparative experiments, the performance differences between this embodiment and several representative federated cue learning methods were evaluated, with experiments conducted on both ResNet50 and ResNet101 backbone networks. The results show that this embodiment exhibits significant performance advantages across all datasets. Taking the ResNet50 model as an example, in one-shot scenarios, the average accuracy of this embodiment on Caltech101, Flowers102, and OxfordPets reached 80.65%, 47.17%, and 80.35%, respectively, which are 2.06%, 14.75%, and 2.57% higher than FedPHA. This result demonstrates that the method of this embodiment can effectively separate the shared and personalized semantic spaces under the condition of only one round of communication, achieving efficient personalization with low communication costs. Furthermore, compared to the GPR-NIAM method, the average accuracy and F1 score on the DTD dataset are 8.31% and 6.16% higher, respectively, further proving that the proposed cue mask enhancement and contrast constraint scheme further strengthens the personalized representation of the local model.
[0062] To verify the scalability of this embodiment under different client scales, the number of clients was varied to 5, 10, and 15 on the Food101 and Flowers102 datasets, respectively, and the model's ACC and F1-score performance on the test set was evaluated. Figure 2 As shown in the figure, regardless of the number of clients, this embodiment consistently outperforms other advanced methods. As the number of clients increases, the amount of data per client decreases, and dataset heterogeneity increases. The method of this invention consistently maintains its superior performance. This result fully demonstrates that the method of this invention has good scalability and maintains robust personalization capabilities even with increasing distributional differences, thereby ensuring stable model performance in large-scale federated scenarios.
[0063] To evaluate the sensitivity of this embodiment to learnable cue length, cue lengths were set to {2, 4, 8, 16} on the Caltech101 dataset, and the mean and standard deviation of ACC and F1-score were reported. Figure 3 As shown, this embodiment consistently outperforms all comparative methods across different cue lengths, achieving the best performance at a cue length of 4. Performance slightly decreases as the cue length increases further, indicating that excessively long cue lengths may introduce redundancy or noise into a single communication scenario, thereby weakening the model's discriminative ability. This result verifies the effectiveness of the subspace constraint strategy proposed in this embodiment: the model not only significantly reduces communication overhead through single-round communication but also maintains performance stability and excellent generalization ability under different cue length configurations. This characteristic allows the model to better adapt to the computational and storage limitations of different clients, giving it higher practical application value.
[0064] In addition, this embodiment also addresses the loss function. and Sensitivity system analysis was performed on the weights. Specifically, the parameter weights were increased from 0.1 to 0.9 in steps of 0.2 to evaluate the performance of this embodiment. Figure 4 As shown, taking the method on the Flower102 dataset as an example, as the hyperparameters on the x-axis increase, the ACC and F1-score show a trend of "first increasing and then decreasing". Compared with the continuous increase of the y-axis, the performance indicators of the model all show an upward trend.
[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A one-time communication federated prompting learning method for UAV swarms, characterized in that, Includes the following steps: Step S1: The central server distributes the pre-trained visual-language model and initial global prompting parameters to each drone client. Each drone client initializes the initial local prompting parameters and builds a local private dataset. Step S2: The drone client decentralizes the initial global prompting parameters, extracts the shared semantic main direction, projects the initial local prompting parameters onto the orthogonal subspace of the shared semantic main direction, and obtains personalized components. Step S3: Based on the personalized components and the initial local cue parameters, generate enhanced cue parameters through interpolation mixing and random masking; Step S4: Based on the local private dataset, perform multi-loss joint optimization on the enhanced cue parameters and personalized components, and obtain the optimized local cue parameters; Step S5: The central server updates the initial global prompt parameters based on the optimized local prompt parameters and sends the updated global prompt parameters to the drone client. Step S6: The drone client builds and deploys a personalized model based on the personalized components and the updated global prompt parameters.
2. The method for one-time communication federation prompting learning for UAV swarms according to claim 1, characterized in that, The vision-language model includes an image encoder and a text encoder, and the initial global cue parameters... Let M be a set of learnable context vectors of length M and embedding dimension d, satisfying The initial local prompt parameters .
3. The method for one-time communication federation prompting learning for UAV swarms according to claim 1, characterized in that, The specific steps for constructing the local private dataset in step S1 are as follows: The drone client i performs deblurring, denoising, and cleaning processes on the raw image data it collects, and then labels the data according to the categories of downstream tasks to build a local private dataset. Local private dataset ,in For image data, These are the category labels corresponding to the image data.
4. The method for one-time communication federation prompting learning for UAV swarms according to claim 1, characterized in that, The specific steps of step S2 are as follows: The drone client calculates the initial global cue parameters. Mean vectors of each dimension The mean vector of each dimension The j-th component is: Where M is the initial global prompt parameter. Length, for The j-th element of the m-th vector; The mean vector of each dimension The mean matrix is formed after M copies. , set initial global prompt parameters Subtract the mean matrix row by row Then received a decentralized global hint ; Decentralized global hints Perform a low-rank approximate decomposition to obtain the orthogonal basis directions of the embedding space. Select the first principal component As The main direction is determined and normalized to obtain the shared semantic main direction u. The expression of the shared semantic main direction u is: in To prevent division by zero of small constants; Calculate initial local prompt parameters Parallel components along the shared semantic principal direction u The expression for the parallel component is: ,in For the transpose of the shared semantic main direction u; Initial local prompt parameters Subtract parallel components Personalized components are then obtained. .
5. The method for one-time communication federation prompting learning for UAV swarms according to claim 1, characterized in that, The specific steps of step S3 are as follows: Using the mixup strategy to personalize components and initial local prompt parameters Generate mixed prompts The hybrid prompt The expression is: ,in Hyperparameters for controlling the shape of the distribution; Generate Bernoulli random variables ,in Let j = 1, 2, ..., M be the probability of dropping, and M be the initial global hint parameter. Length, enhanced prompt parameters The j-th vector satisfies: .
6. The method for one-time communication federation prompting learning for UAV swarms according to claim 1, characterized in that, The specific steps of step S4 are as follows: Based on a local private dataset, a class-discriminatory loss is applied to the enhanced cue parameters and personalized components. Two-way enhanced loss and semantic contrast loss Joint optimization of multiple losses; Construct the total loss function ,in and To obtain the loss weights, the optimized local cueing parameters are obtained through gradient descent optimization. ,in N represents the number of drone clients.
7. The method for one-time communication federation prompting learning for UAV swarms according to claim 6, characterized in that, The category discrimination loss The cosine similarity calculation based on the image encoder and text encoder is calculated using the following formula: in For local private datasets, for Image data in The category labels corresponding to the image data. Represents cosine similarity. Where C is the temperature coefficient and C is the number of downstream task categories. The text hint for category k.
8. The method for one-time communication federated prompting learning for UAV swarms according to claim 6, characterized in that, The bidirectional enhancement loss is used to constrain the semantic consistency between the enhanced cue parameters and the initial local cue parameters. The normalized text features of the enhanced cue parameters are: The normalized text features of the initial local prompt parameters are: B is the feature batch size. and The similarity matrix between them is defined as follows: in Temperature coefficient; The bidirectional enhancement loss The expression is: Where B is the batch size. .
9. The method for one-time communication federation prompting learning for UAV swarms according to claim 6, characterized in that, The specific steps for optimizing the semantic contrast loss are as follows: Constructing text features , as well as ,in , , , For personalized portions, To enhance the prompt parameters, For global reflection hints, the expression is: ,in For initial global hint parameters, To share the main semantic direction; The semantic contrast loss The expression is: Where C represents the number of downstream task categories. This is the temperature coefficient.
10. The method for one-time communication federation prompting learning for UAV swarms according to claim 1, characterized in that, The updated global prompt parameters in step S5 The expression is: Where N is the number of drone clients. For optimized local prompt parameters The elements in The aggregation weight for drone client i is determined by the local private dataset, and its expression is: in and These are the local private datasets of drone clients i and j, respectively.