Efficient online segmentation learning method and system for unmanned aerial vehicle cluster
By integrating segmentation learning and online learning in UAV swarms, optimizing resource management and sample selection, the problems of poor detection accuracy and high training latency of UAV swarms in unknown scenarios are solved, achieving efficient online training and improved model accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-11
- Publication Date
- 2026-03-13
AI Technical Summary
Existing drone swarms have poor object detection accuracy in unknown scenarios. Traditional offline training methods lead to data drift, and due to the limited computing resources of drones, it is difficult to effectively train complex deep neural networks online.
By deeply integrating segmentation learning with online learning, and through a global model management module, resource management module, dynamic sample processing module, and local model training module, combined with a two-factor importance evaluation mechanism and sub-channel allocation and power control algorithm, efficient online segmentation learning of UAV swarms is achieved, optimizing communication resource allocation and sample selection.
It improves the model detection accuracy and training efficiency of UAV swarms in unknown scenarios, reduces training latency, adapts to the high mobility and limited resources of UAVs, and enhances the model's generalization ability and detection accuracy.
Smart Images
Figure CN121665296A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent self-organizing network technology for unmanned aerial vehicles (UAVs), specifically relating to an efficient online segmentation learning method and system for UAV swarms. Background Technology
[0002] In recent years, unmanned aerial vehicle (UAV) swarm systems have shown enormous development potential and application value. With their significant advantages such as high maneuverability and lightweight design, UAVs have demonstrated unique adaptability in many fields, including fire rescue, agricultural monitoring, and logistics delivery. The cross-integration of UAVs and artificial intelligence technologies has become a cutting-edge research hotspot attracting the attention of both academia and industry. By deploying advanced machine learning models, especially deep neural network architectures, on UAV terminals, UAV systems have acquired autonomous perception, decision-making, and task execution capabilities.
[0003] However, most technical solutions still follow the traditional offline training model, which involves pre-training the model on a fixed training set, then deploying it to the drone after lightweight model processing. For example, federated learning uses drones as clients, training locally and then uploading the model to a server to aggregate it into a global model. When drone swarms perform object detection in unfamiliar scenarios, the accuracy of drone object detection is usually very poor. This is because the use of a limited and fixed training set, lacking training samples from unfamiliar scenarios, causes data drift that severely affects the model's training accuracy.
[0004] Therefore, the concept of online training modeling (FL) was proposed, which improves the adaptability of UAV swarm training models to unfamiliar environments to some extent by continuously incorporating dynamically changing sample data. However, current research and applications of online FL still face significant technical bottlenecks: although existing research has deployed lightweight models (such as YOLOv5) on UAV terminals for online training, when dealing with deep neural networks with large numbers of parameters and complex structures, the training process is difficult to conduct normally due to the limited computing resources of UAVs. Summary of the Invention
[0005] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing an efficient online segmentation learning method and system for UAV swarms. The system includes a global model management module and a resource management module deployed on server UAVs, and a dynamic sample processing module and a local model training module deployed on client UAVs. By deeply integrating segmentation learning (SL) and online learning (OL), an online segmentation learning framework (OnlineSL) is constructed. A two-factor importance evaluation mechanism is used to screen high-value samples to alleviate data drift. Combined with sub-channel allocation and power control algorithms to balance the allocation of communication resources among heterogeneous devices, it can enable UAV swarms to capture dynamic samples in real time in unknown scenarios. Distributed collaborative training improves model detection accuracy while reducing training latency, effectively adapting to the high mobility, limited resources, and dynamic environment characteristics of UAV swarms.
[0006] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:
[0007] A highly efficient online segmentation learning system for UAV swarms, the system comprising a global model management module and a resource management module deployed on server UAVs, a dynamic sample processing module and a local model training module deployed on client UAVs, and a ground station;
[0008] The client drone includes one main client drone and S-1 agent client drones;
[0009] The global model management module is used to store the global DNN model, split the global model into local models, including lower-level sub-models for extracting input abstract representations and upper-level sub-models for prediction, and send the lower-level sub-models to the main client drone and distribute the upper-level sub-models to each client drone for distributed training and updates; at the same time, it receives the updated models uploaded by the client drones, updates the global model by aggregation, calculates the loss and feeds it back to the client drones;
[0010] The resource management module is used to optimize communication resource allocation based on a greedy subchannel allocation strategy and an adaptive power control algorithm.
[0011] The dynamic sample processing module is used to collect online samples in real time and screen them using an importance assessment mechanism and an exponential decay stochastic exploration strategy.
[0012] The local model training module is used to carry the assigned local model, perform forward and backward propagation calculations, send the forward propagation results to the server drone, and perform backward propagation and update the local model after receiving the loss data.
[0013] The ground station is used to store historical training data, assist in sample labeling, and monitor system status. It receives raw collected data transmitted from the client's drone, assists in completing sample preprocessing, and then transmits the data back to the dynamic sample processing module.
[0014] To optimize the above technical solution, the specific measures also include:
[0015] The global model management module described above splits the global model into lower-level sub-models and upper-level sub-models according to the channel dimension.
[0016] The loss calculated by the global model management module mentioned above includes cross-entropy classification loss and collaborative training loss.
[0017] The aforementioned optimized communication resource allocation includes sub-channel allocation and power control. The communication resource configuration of the client UAV is dynamically adjusted through a greedy sub-channel allocation strategy, and the transmission delay is reduced under power constraints by combining an adaptive power control algorithm.
[0018] The greedy sub-channel allocation strategy described above is as follows: devices are sorted in descending order of client drone computing power, and channels are sorted in descending order of channel gain; based on the sorting, at least one channel is first allocated to each device, and the remaining channels are preferentially allocated to the device with the longest total training time, ensuring that each channel is allocated to only one device.
[0019] The adaptive power control algorithm described above is as follows: In the first stage, each device allocates power according to the gain ratio of the assigned channel, ensuring that the power does not exceed the maximum power of a single device; in the second stage, the total power of the channel is calculated. If the power constraint is exceeded Scale the power of each channel proportionally
[0020] The aforementioned dynamic sample processing module employs a dual-factor importance assessment mechanism that combines loss change rate and priority classification to screen high-value samples and eliminate invalid or low-priority samples, ensuring that the training data distribution matches the real-time scene. At the same time, it introduces an exponentially decaying random exploration strategy to improve the accuracy of UAV object detection in unfamiliar scenes.
[0021] The above-mentioned two-factor importance assessment mechanism of fusion loss change rate and priority classification is as follows:
[0022] Based on the sample size of the current training round From historical sample sets Select One sample was used for training, among which These are online samples collected in real time.
[0023] Historical samples If historical loss records If the length is ≥2, then calculate the rate of change of loss. , This represents the training loss for the first two rounds and the previous round; if the length is less than 2, then the rate of change of loss is...
[0024] The importance of samples is calculated using a two-factor importance assessment function: ,in As a dynamic factor;
[0025] Priority classification: by Historical samples were sorted in descending order, and the first few were selected. One sample, compared with online samples Composition of the current training batch Simultaneously update the sample loss record; if at this time... If the length is greater than 2, delete the oldest record.
[0026] An efficient online segmentation learning method for UAV swarms based on the system, comprising:
[0027] S1: The global model management module of the server drone splits the DNN global model into a lower-level sub-model set and an upper-level sub-model set, sends the lower-level sub-model set to a randomly selected main client drone, and sends the upper-level sub-models to the i-th agent client drone respectively;
[0028] S2: The client-side drone's dynamic sample processing module collects online samples in real time, filters them, and then passes them to the local model training module;
[0029] S3: The main client drone runs the lower-level sub-model through the local model training module, and transmits S-1 intermediate outputs to the agent client drone through the uplink. The agent client drone runs the upper-level sub-model and sends the prediction results to the server drone.
[0030] S4: The server drone calculates the loss through the global model management module and broadcasts it to all client drones. After receiving the loss, the client drones perform backpropagation and send the updated upper and lower sub-models to the server drone.
[0031] S5: The server drone's resource optimization module executes resource management algorithms;
[0032] S6: The server drone aggregates the updated upper part of the model uploaded by all client drones and updates the global model; repeat S2-5 until the model converges.
[0033] The present invention has the following beneficial effects:
[0034] This invention employs segmented learning to achieve data sharing among drones. Training a complex model jointly by several drones effectively distributes the computational load, reducing the computational pressure on a single device and improving the model's convergence speed. Online learning enables drones to learn new knowledge in unfamiliar environments, thereby improving the model's accuracy.
[0035] This invention achieves efficient online training of UAV swarms in unknown scenarios through dual-dimensional optimization of dynamic sample selection and resource management. It addresses the problems faced by traditional AI models in UAV swarm applications, such as data drift due to offline training paradigms, difficulties in online training of complex deep neural networks (DNNs) under resource constraints, and poor model generalization ability. Compared to traditional methods, UAV swarm online segmentation learning based on sample selection and resource management deeply integrates segmentation learning and online learning technologies, enabling UAV swarms to capture dynamic sample data in unfamiliar scenarios in real time. Furthermore, it improves the generalization ability and detection accuracy of complex neural network models through a distributed collaborative training mechanism. The resource management optimization algorithm significantly reduces training latency, while the dynamic sample selection algorithm significantly improves model training accuracy, effectively reducing the time it takes for the UAV model to reach the expected accuracy.
[0036] This invention presents an optimized algorithm for subchannel allocation and power control, which effectively avoids the training progress imbalance caused by differences in device performance by dynamically adjusting the communication resource configuration of the client UAV. The dynamic sample selection algorithm employs a dynamic two-factor importance evaluation mechanism, achieving intelligent selection of training data through loss rate of change perception and online sample priority partitioning. Simultaneously, it introduces an exponentially decaying random exploration strategy to improve the accuracy of UAV object detection in unfamiliar scenarios.
[0037] This invention is the first to deeply integrate split learning with online learning, constructing an OnlineSL framework to solve the data drift problem in unknown scenarios caused by offline training in traditional federated learning (FL). Simultaneously, it distributes computational pressure through model splitting, adapting to the limited computing resources of drones. Dynamic sample selection employs a two-factor importance evaluation mechanism (fusing loss value and loss change rate) combined with an exponential decay exploration strategy to effectively screen high-value samples. Compared to traditional fixed sample set training, the model detection accuracy is improved by 1.15-1.25 times. Resource management adopts a joint optimization of "greedy sub-channel allocation + two-stage power control" to balance communication resources of heterogeneous devices. Compared to random allocation and allocation based on received signal strength (RSS), training latency is reduced by 25%-66%, while avoiding training progress imbalances caused by differences in device performance. Attached Figure Description
[0038] Figure 1This is a schematic diagram of the structure of efficient online segmentation learning of UAV swarms used in an embodiment of the present invention, wherein 1 is the global model management module, 2 is the resource optimization module, 3 is the dynamic sample processing module, 4 is the local model training module, and 5 is the ground station.
[0039] Figure 2 This refers to the model accuracy of the present invention example compared to the baseline.
[0040] Figure 3 This is a graph showing the change in the accuracy of the model by different functional modules during the model training process according to an embodiment of the present invention.
[0041] Figure 4 This is a performance comparison of batch latency and total bandwidth between the examples of this invention and the baseline.
[0042] Figure 5 This is the average batch latency of the client drone compared to the baseline in this invention example. Detailed Implementation
[0043] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0044] like Figure 1 As shown, the present invention provides an efficient online segmentation learning system for UAV swarms, comprising a global model management module 1, a resource optimization module 2, a dynamic sample processing module 3, a local model training module 4, and a ground station 5.
[0045] The global model management module 1 and resource management module 2 are deployed on the server drone. The global model management module 1 is responsible for storing the complete DNN model, splitting the model into lower-level sub-models (responsible for input abstract representation extraction) and upper-level sub-models (responsible for prediction) according to the channel dimension, sending the lower-level sub-models to the main client drone, and distributing the upper-level sub-models to each client drone; simultaneously, it receives updated models uploaded by client drones, updates the global model through aggregation, calculates cross-entropy classification loss and co-training loss, and feeds them back to the client drones. The resource management module 2 is responsible for optimizing communication resource allocation, including sub-channel allocation and power control. It prioritizes channel allocation for high-latency devices using a greedy sub-channel allocation strategy, and reduces transmission latency under power constraints using a two-stage adaptive power control algorithm.
[0046] The dynamic sample processing module 3 and the local model training module 4 are deployed on the client drones (including one main client drone and S-1 proxy client drones). The dynamic sample processing module 3 is responsible for collecting online samples in real time, using a two-factor importance assessment mechanism (integrating loss change rate and priority classification) to screen high-value samples and remove invalid or low-priority samples to ensure that the training data distribution matches the real-time scene. The local model training module 4 carries the assigned local model (the main client carries the lower part of the model, and the proxy client carries the upper part of the model), completes forward propagation and backpropagation calculations, sends the forward propagation results to the server drone, and performs backpropagation and updates the local model after receiving the loss data.
[0047] The global model management module 1, resource management module 2, dynamic sample processing module 3, and local model training module 4 constitute the core system of online split learning: the server drone splits the global model and distributes it to the client drones. The client drones filter samples through the dynamic sample processing module, complete training calculations in combination with the local models, and feed the results back to the server drone. The server drones aggregate the models and optimize resource allocation, forming a closed-loop process of "model splitting - sample filtering - distributed training - model aggregation - resource optimization".
[0048] The ground station 5 is used to store historical training data, assist in sample labeling, and monitor system status. It can receive raw collected data transmitted by the client drone, assist in completing sample preprocessing, and then transmit the data back to the dynamic sample processing module 3.
[0049] In this embodiment, a Jetson Nano embedded computer is mounted on a quadcopter drone as an edge computing node, and an online segmentation learning system is formed by a server drone and multiple client drones.
[0050] The efficient online segmentation learning method for UAV swarms based on the sample selection and resource management-based online segmentation learning system includes the following specific steps:
[0051] S1: The global model management module 1 of the server drone splits the complete DNN model into a lower-level sub-model set and an upper-level sub-model set according to the channel dimension. The lower-level sub-model set is sent to the randomly selected master client drone, and the upper-level sub-models are sent to the i-th client drone respectively.
[0052] S2: The client-side drone's dynamic sample processing module 3 collects online samples in real time, adjusts the number of local samples according to the sample capacity of the current training round, and calculates sample importance using a two-factor importance evaluation function. The selected samples are then passed to the local model training module 4.
[0053] In this embodiment, the model is optimized through online segmentation learning, and dynamic sample selection is used to ensure that the training data distribution matches the real-time scene.
[0054] The client-side UAV's dynamic sample processing module 3 collects online samples in real time. Based on the sample size of the current training round It is necessary to start from the historical sample set. Select One sample is used for training:
[0055] Historical samples If historical loss records Length >= 2, calculate the rate of change of loss. If the length is less than 2, ;
[0056] The importance of samples is calculated using a two-factor importance assessment function: (in (This is a dynamic factor that controls the weight of historical changes).
[0057] according to Historical samples were sorted in descending order, and the first few were selected. One sample, compared with online samples Composition of the current training batch Simultaneously update the sample loss record, if If length > 2, delete the oldest record.
[0058] S3: The main client drone runs the lower-level sub-model through the local model training module 4, and transmits S-1 intermediate outputs to the agent client drone through the uplink. The agent client drone runs the upper-level sub-model and sends the prediction results to the server drone.
[0059] In this embodiment, the main client drone runs the lower-level sub-model through the local model training module 4. Calculate the forward propagation delay , ( For batch size, To calculate the strength, This refers to the computational cost of a portion of the model under a single sample. For computing power), S-1 intermediate outputs (size is Transmitted to the agent client drone via uplink, transmission delay ( (uplink speed).
[0060] The agent client drone runs the upper part of the model and calculates the forward propagation delay. ( (This refers to the computational cost of a single sample model) and the prediction results are sent to the server drone.
[0061] S4: The server drone calculates the loss through the global model management module 1 and broadcasts it to all client drones. After receiving the loss, the client drones perform backpropagation and send the updated upper and lower sub-models to the server drone.
[0062] In this embodiment, the server drone calculates the loss through the global model management module 1 and broadcasts it to all client drones, with a broadcast delay. ( (for broadcast rate);
[0063] After the client drone receives the loss, it performs backpropagation; the main client delay... Proxy client latency The updated upper-level sub-model is then sent to the server drone.
[0064] S5: Server drone resource optimization module 2 executes resource management algorithm;
[0065] In this embodiment, by optimizing the allocation of communication resources, including sub-channel allocation and power control, a greedy sub-channel allocation strategy is used to prioritize channel allocation for high-latency devices, and a two-stage adaptive power control algorithm is combined to reduce transmission latency under power constraints.
[0066] The resource management algorithm is as follows:
[0067] Greedy subchannel allocation: based on the client drone's computing power Descending sorting device, sorted by channel gain Sub-channels are sorted in descending order; at least one channel is first allocated to each device, and the remaining channels are allocated preferentially to the total training time. The longest device ensures that each channel is allocated to only one device. The time when the server transmits the activation information to the proxy client. and For the forward and backward propagation time of the proxy client drone, The time lost by the proxy client drone in transmitting data to the server The time it takes for the proxy client drone to transmit gradients to the server;
[0068] Two-stage power control: In the first stage, each device allocates power according to the gain ratio of the assigned channel, ensuring that the power does not exceed the maximum power of a single device. Right now ( (Assign total channel gain to the device); Second stage, calculate total power. If the power constraint is exceeded Scale the power of each channel proportionally
[0069] S6: The server drone aggregates the updated upper part of the model uploaded by all client drones and updates the global model; repeat S2-5 until the model converges (e.g., the detection accuracy reaches the preset threshold).
[0070] This simulation scenario includes one server drone and four client drones (evenly distributed around the server). A wide residual network (WRN-16-8) is used as the DNN model (adapted to split learning to improve target detection accuracy). The local training set is a subset of the CIFAR-10 dataset, with online samples continuously supplemented from CIFAR-10 to simulate dynamic data in unknown scenarios. FedAvg (a classic federated learning algorithm), SFL (split federated learning), and FedProx (resistant to heterogeneous data federated learning) are selected as baselines for model accuracy comparison; random channel allocation, RSS-based channel allocation, and greedy allocation without power control are selected as baselines for resource management comparison.
[0071] Figure 2 The accuracy of the online segmentation learning method relative to the baseline method was tested. It can be seen that after 100 training rounds, the online learning method achieved the highest accuracy, exceeding 80%, while the baseline method did not even reach 80%. The accuracy of the online learning method is 1.15 higher than Fedavg, 1.17 higher than SFL, and 1.25 higher than FedProx. The high performance of the online learning method is attributed to its dynamic sample selection mechanism, which maintains high training efficiency and accuracy for online segmentation learning.
[0072] Figure 3 The importance of online learning and dynamic sample selection was evaluated. Three experimental groups were set up: SL, OnlineSL without dynamic sample selection, and FedAvg as the baseline algorithm. After applying OL, the final training accuracy of the model improved to 90%, which is 1.05 percentage points higher than SL without OL and 1.15 percentage points higher than FedAvg. Further combining a dynamic sample selection mechanism based on ordinary sample updates, the final accuracy exceeded 92%, and compared with the method without dynamic sample selection, it achieved the target accuracy of 85% in fewer training epochs. This is because the dynamic sample selection mechanism improves training efficiency by filtering irrelevant samples, while prioritizing key samples to enhance model accuracy.
[0073] Figure 4The performance of each method in terms of training latency and total bandwidth was compared. The results show that the GSA and PC algorithms significantly reduce training latency compared to the benchmark algorithms. Compared to the random channel allocation scheme, the algorithm of this invention reduces training latency by 66%; compared to the RSS-based channel allocation scheme, the latency is reduced by 54%; and compared to the greedy sub-channel allocation method without power control, the latency is also reduced by 25%. These experimental results demonstrate that the algorithm significantly improves the time efficiency of UAV swarm training by dynamically optimizing sub-channel and power resource allocation, effectively solving the problem of training progress lag caused by uneven distribution of communication resources among heterogeneous devices. Furthermore, as the total bandwidth increases, the training latency also decreases—because the larger total bandwidth expands the sub-channel bandwidth between the client UAV and the server, thereby increasing the data transmission rate.
[0074] Figure 5 The effectiveness of the proposed algorithm in balancing training latency across different clients was evaluated. The average training time for each cluster sample of drones from each client was tested and compared with baseline methods. It can be seen that this invention, through channel allocation and power control, adjusts the training latency of heterogeneous devices in a timely manner, ensuring that all devices have similar training durations, thereby avoiding excessive training latency and imbalance caused by lagging devices. In contrast, other benchmark algorithms do not consider the impact of lagging devices, leading to significant differences in training latency among heterogeneous devices. For random allocation schemes, devices with weaker computing power are assigned to channels with poorer conditions, where lagging devices result in excessively long total training latency.
[0075] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.
Claims
1. A highly efficient online segmentation learning system for unmanned aerial vehicle (UAV) swarms, characterized in that, The system includes a global model management module and a resource management module deployed on the server drone, a dynamic sample processing module and a local model training module deployed on the client drone, and a ground station. The client drone includes one main client drone and S-1 agent client drones; The global model management module is used to store the global DNN model, split the global model into local models, including lower-level sub-models for extracting input abstract representations and upper-level sub-models for prediction, and send the lower-level sub-models to the main client drone and distribute the upper-level sub-models to each client drone for distributed training and updates; at the same time, it receives the updated models uploaded by the client drones, updates the global model by aggregation, calculates the loss and feeds it back to the client drones; The resource management module is used to optimize communication resource allocation based on a greedy subchannel allocation strategy and an adaptive power control algorithm. The dynamic sample processing module is used to collect online samples in real time and screen them using an importance assessment mechanism and an exponential decay stochastic exploration strategy. The local model training module is used to carry the assigned local model, perform forward and backward propagation calculations, send the forward propagation results to the server drone, and perform backward propagation and update the local model after receiving the loss data. The ground station is used to store historical training data, assist in sample labeling, and monitor system status. It receives raw collected data transmitted from the client's drone, assists in completing sample preprocessing, and then transmits the data back to the dynamic sample processing module.
2. The efficient online segmentation learning system for UAV swarms according to claim 1, characterized in that, The global model management module splits the global model into lower-level sub-models and upper-level sub-models according to the channel dimension.
3. The efficient online segmentation learning system for UAV swarms according to claim 1, characterized in that, The loss calculated by the global model management module includes cross-entropy classification loss and collaborative training loss.
4. The efficient online segmentation learning system for UAV swarms according to claim 1, characterized in that, The optimized communication resource allocation includes sub-channel allocation and power control. The communication resource configuration of the client UAV is dynamically adjusted through a greedy sub-channel allocation strategy, and the transmission delay is reduced under power constraints by combining an adaptive power control algorithm.
5. A high-efficiency online segmentation learning system for UAV swarms according to claim 1 or 4, characterized in that, The greedy sub-channel allocation strategy is as follows: devices are sorted in descending order of client UAV computing power, and channels are sorted in descending order of channel gain; according to the sorting, at least one channel is first allocated to each device, and the remaining channels are preferentially allocated to the device with the longest total training time, ensuring that each channel is allocated to only one device.
6. A high-efficiency online segmentation learning system for UAV swarms according to claim 1 or 4, characterized in that, The adaptive power control algorithm is as follows: In the first stage, each device allocates power according to the gain ratio of the assigned channel to ensure that the power does not exceed the maximum power of a single device; The second stage involves calculating the total channel power. If the power constraint is exceeded Scale the power of each channel proportionally 7. The efficient online segmentation learning system for UAV swarms according to claim 1, characterized in that, The dynamic sample processing module employs a dual-factor importance assessment mechanism that combines loss change rate and priority classification to screen high-value samples and remove invalid or low-priority samples, ensuring that the training data distribution matches the real-time scene and improving the accuracy of UAV object detection in unfamiliar scenarios.
8. The efficient online segmentation learning system for UAV swarms according to claim 5, characterized in that, The two-factor importance assessment mechanism of fusion loss change rate and priority classification is as follows: Based on the sample size of the current training round From historical sample sets Select One sample was used for training, among which These are online samples collected in real time. Historical samples If historical loss records If the length is ≥2, then calculate the rate of change of loss. , This is due to training losses from the first two rounds and the previous round; If the length is less than 2, then the rate of change of loss is... ; The importance of samples is calculated using a two-factor importance assessment function: ,in As a dynamic factor; Priority Classification: By Historical samples were sorted in descending order, and the first few were selected. One sample, compared with online samples Composition of the current training batch Simultaneously update the sample loss record; if at this time... If the length is greater than 2, delete the oldest record.
9. A method for efficient online segmentation learning of UAV swarms based on the system described in any one of claims 1-7, characterized in that, include: S1: The global model management module of the server drone splits the DNN global model into a lower-level sub-model set and an upper-level sub-model set, sends the lower-level sub-model set to a randomly selected main client drone, and sends the upper-level sub-models to the i-th agent client drone respectively; S2: The client-side drone's dynamic sample processing module collects online samples in real time, filters them, and then passes them to the local model training module; S3: The main client drone runs the lower-level sub-model through the local model training module, and transmits S-1 intermediate outputs to the agent client drone through the uplink. The agent client drone runs the upper-level sub-model and sends the prediction results to the server drone. S4: The server drone calculates the loss through the global model management module and broadcasts it to all client drones. After receiving the loss, the client drones perform backpropagation and send the updated upper and lower sub-models to the server drone. S5: The server drone's resource optimization module executes resource management algorithms; S6: The server drone aggregates the updated upper part of the model uploaded by all client drones and updates the global model; repeat S2-5 until the model converges.