Edge calculation distribution system and method for cooperative task of unmanned aerial vehicle cluster

Abstract

The invention is suitable for the technical field of unmanned aerial vehicle cluster cooperative control, and provides an edge calculation distribution system and method for unmanned aerial vehicle cluster cooperative tasks, and the system comprises a multi-heterogeneous neural network module, a local fusion module, a global fusion module, a conflict resolution module, a task distribution module, and a self-adaptive optimization module. According to the invention, by constructing a multi-level semantic feature extraction and fusion architecture and combining an intelligent conflict resolution mechanism and a dynamic task allocation algorithm, efficient intelligent collaborative operation of an unmanned aerial vehicle cluster in a complex and changeable environment is realized, all core algorithms can be executed in a distributed manner in the local of the unmanned aerial vehicle through a distributed edge computing architecture, and the distributed edge computing architecture can be applied to the unmanned aerial vehicle cluster. Dependence on a central node is remarkably reduced, communication overhead and delay are greatly reduced, cooperative work can still be maintained even if part of nodes fail, and perfect balance between robustness and performance is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle cluster cooperative control, in particular to an edge computing distribution system and method for unmanned aerial vehicle cluster cooperative tasks. BACKGROUND

[0002] In the prior art, the allocation and execution of unmanned aerial vehicle cluster cooperative tasks mainly rely on centralized or distributed task allocation mechanisms. The centralized scheme usually collects the perception data of each member through a ground control station or a master unmanned aerial vehicle, and performs centralized processing and task allocation at the center node. This way puts high requirements on the computing power and communication bandwidth of the center node. The distributed scheme allows each unmanned aerial vehicle to perform task allocation based on local perception information and simple cooperation protocols (such as market mechanisms, auction algorithms, consensus algorithms, etc.), improving the robustness and flexibility of the system, but it is difficult to obtain the globally optimal allocation effect in complex environments.

[0003] Currently, the mainstream unmanned aerial vehicle cluster system mostly uses a single type of neural network model (such as convolutional neural network CNN, long short-term memory network LSTM, etc.) to extract features and make decisions on perception data, which is difficult to fully exploit multi-modal and multi-granularity information, resulting in limited intelligent level of perception and decision-making. In terms of feature fusion and conflict processing, existing methods mostly use simple weighted averaging or voting mechanisms, lacking deep modeling and resolution means for semantic consistency and conflicts between multiple models or multiple unmanned aerial vehicles, and cannot effectively improve the stability and reliability of the fusion results.

[0004] In summary, the existing technology mostly has the following defects: Firstly, the communication load is too heavy and the scalability is insufficient. With the expansion of the cluster size, the communication load and cooperation complexity increase dramatically, and existing cooperative perception computing technology cannot be directly applied to large-scale unmanned aerial vehicle clusters. The traditional centralized control architecture requires all unmanned aerial vehicles to maintain frequent communication with the center node, which not only consumes a large amount of bandwidth resources but also increases the system response delay, seriously affecting the real-time requirement. When the cluster size exceeds a certain threshold, network congestion and delay problems become the key bottleneck that restricts system performance, making it difficult for the system to maintain efficient communication and cooperative control in large-scale deployment; Secondly, the heterogeneous data fusion capability is severely insufficient. Existing systems generally have problems such as data type heterogenization, inconsistent data quality, and time synchronization difficulties when processing heterogeneous data from different sensors and different formats. Traditional fusion methods often use simple data splicing or weighted averaging, which cannot fully exploit the deep semantic associations between heterogeneous data, resulting in poor fusion results and low information utilization. Especially when dealing with multi-modal sensor data such as vision, radar, and GPS, there is a lack of effective feature assimilation and semantic alignment mechanisms, making it impossible to achieve true intelligent fusion; Furthermore, existing research has obvious limitations in the coordination capability of heterogeneous clusters. Most models do not fully consider the complexity of the real environment and the limitations of the perception capability of unmanned aerial vehicles, mainly designed for homogeneous unmanned aerial vehicle clusters, and are difficult to handle heterogeneous unmanned aerial vehicle coordination scenarios with different models, different loads, and different performance parameters. This limitation leads to a lack of flexibility and adaptability of the system when facing complex tasks, and the centralized, single means cannot meet the fine perception and situation assessment requirements of the unmanned aerial vehicle group; In addition, the lack of real-time and robustness is also an important defect of the existing technology. The traditional neural network fusion method has high computational complexity, which is difficult to meet the strict requirements of real-time of unmanned aerial vehicle clusters. At the same time, the existing system has limited processing capability for abnormal situations such as environmental interference, communication interruption, and equipment failure, and lacks effective fault tolerance mechanism and adaptive adjustment capability.

[0005] Therefore, in view of the above status, it is urgent to provide an edge computing distribution system and method for unmanned aerial vehicle cluster coordination tasks to overcome the deficiencies in current practical applications. SUMMARY

[0006] The purpose of the present application is to provide an edge computing distribution system and method for unmanned aerial vehicle cluster coordination tasks, which effectively solves the problems in the background art.

[0007] The present application is implemented as follows: the edge computing distribution system for unmanned aerial vehicle cluster coordination tasks comprises a multi-heterogeneous neural network module, a local fusion module, a global fusion module, a conflict resolution module, a task allocation module, and an adaptive optimization module. The multi-heterogeneous neural network module is configured on each unmanned aerial vehicle, and is used to extract features from multi-modal raw data collected by the unmanned aerial vehicle through multiple different types of neural network models, and generate corresponding semantic feature vectors; The local fusion module is configured on each unmanned aerial vehicle, and is used to calculate the semantic consistency score between the semantic feature vectors output by different neural network models on the same unmanned aerial vehicle, determine the fusion weight in combination with the confidence of each model, and perform weighted fusion on the semantic feature vectors based on the fusion weight to obtain local fusion features; The global fusion module is used to construct a communication adjacency matrix between unmanned aerial vehicles, calculate the attention weight of the local fusion features of each unmanned aerial vehicle using a graph attention network mechanism, and perform weighted fusion on the local fusion features of multiple unmanned aerial vehicles based on the attention weight to obtain global fusion features; The conflict resolution module is used to detect and resolve semantic discrimination conflicts between multiple models or multiple unmanned aerial vehicles; The task allocation module uses a distributed auction algorithm to allocate tasks based on global semantic representation and task demand vectors; The adaptive optimization module is configured to adjust the fusion weights of the models by a gradient descent optimization algorithm according to the conflict information, so as to reduce the occurrence of future conflicts.

[0008] As a further scheme of the present application, the neural network models in the multi-heterogeneous neural network module include a convolutional neural network, a long short-term memory network, a transformer, and a graph neural network.

[0009] As a further scheme of the present application, in the local fusion module, the specific steps of calculating the semantic consistency score are: For model k1 and model k2 on the i-th UAV, the semantic consistency score is the cosine similarity of the two model output feature vectors; The average consistency score of each model with all other models is calculated. The fusion weight takes into account the confidence and consistency score of the model.

[0010] As a further scheme of the present application, in the local fusion module, the calculation formula of the fusion weight is: ; Wherein, K is the number of models carried by each UAV, α and β is an adjustable weight coefficient, is the model confidence, is the average consistency score of the k-th model on the i-th UAV with all other models.

[0011] As a further scheme of the present application, the process of calculating the attention weight by the global fusion module using a graph attention network mechanism includes: Converting the local fusion features of each UAV into query vectors and key vectors, respectively; Calculating the attention weight between UAVs to form a probability distribution; According to the attention weight, the local fusion features are weighted and summed to generate global fusion features.

[0012] As a further scheme of the present application, the specific working steps of the conflict resolution module are as follows: Conflict resolution is triggered when the similarity of the local features of two UAVs is lower than a preset threshold; A confidence weighted conflict resolution strategy is adopted to perform weighted averaging on the conflict features; Record the conflict information and feed it back to the model training and weight adjustment module.

[0013] As a further scheme of the present application, the specific steps of the task allocation module for task allocation are as follows: Define a semantic requirement vector for each task; The multi-dimensional matching degree between the UAV and the task is calculated, including semantic matching degree, capability coefficient and availability coefficient; The task allocation is performed by using a distributed auction algorithm, and the UAV autonomously bids according to the matching degree.

[0014] As a further scheme of the application, a dynamic task re-allocation mechanism is further included, when it is detected that the task execution is abnormal or a new task arrives, the matching degree is recalculated and the re-allocation process is triggered, and the task switching cost and time loss are considered.

[0015] As a further scheme of the application, the adaptive optimization module includes a performance monitoring unit, which tracks the model confidence, consistency score and task completion rate indicators in real time, and continuously optimizes the fusion weight by using the gradient descent method.

[0016] The edge computing distribution method for UAV cluster cooperative task adopts the system as described above, and the method comprises the following steps: S1. Multi-heterogeneous neural network feature extraction: each UAV extracts features from the collected multi-modal raw data by using multiple different types of neural network models to generate semantic feature vectors; S2. Local fusion: the semantic consistency scores between the semantic feature vectors output by different models on the same UAV are calculated, the fusion weight is determined in combination with the confidence of each model, the semantic feature vectors are weighted and fused to obtain local fusion features; S3. Global fusion: a communication adjacency matrix between the UAVs is constructed, the attention weight of the local fusion features of each UAV is calculated by using a graph attention network mechanism, and the local fusion features are weighted and fused based on the attention weight to obtain global fusion features; S4. Conflict resolution: the similarity between the local fusion features of any two UAVs is monitored, when the similarity is lower than a preset threshold, the conflict resolution mechanism is triggered, the conflict features are weighted and averaged according to the confidence of each UAV, and the conflict information is recorded; S5. Task allocation: a semantic demand vector is defined for each task to be executed, the matching degree between the global fusion features of the UAV and the semantic demand vector of the task is calculated, and the task allocation is performed by using a distributed auction algorithm; S6. Adaptive optimization: according to the conflict information, the fusion weight of each model is adjusted by using a gradient descent optimization algorithm to improve the system performance.

[0017] Compared with the prior art, the application has the following advantages: The prior art mainly adopts a single type of neural network for perception, and it is difficult to fully mine multi-modal information, and the application can fully play the unique advantages of each network through a multi-heterogeneous neural network cooperative mechanism, simultaneously deploy multiple neural network models, and comprehensively capture spatial features, time sequence information and long-distance dependency relationships. In view of the deficiencies of the prior art, such as simple weighted average or voting mechanism and lack of semantic consistency deep modeling, the application establishes a fusion framework based on local and global double-layer consistency discrimination, quantifies model consistency by calculating cosine similarity between feature vectors, and realizes intelligent fusion by combining graph attention network adaptive weight learning. The traditional centralized scheme has a single point of failure risk, and the distributed scheme is difficult to obtain a global optimal effect, and the application can be executed locally and distributedly by all core algorithms of the unmanned aerial vehicle, which significantly reduces the dependence on the center node, greatly reduces the communication overhead and delay, and can still maintain cooperative work even if part of the nodes fail, realizing the perfect balance between robustness and performance. The existing scheme is mainly based on a simple market mechanism and mainly considers physical capability matching, and the application innovatively takes the fused global semantic features as the core basis for task allocation, realizes accurate task-unmanned aerial vehicle matching based on semantic understanding by constructing semantic vector representation of task demand, and fully utilizes the front-end perception fusion results to realize integrated optimization of perception-fusion-decision. DETAILED DESCRIPTION

[0018] The technical scheme of the application will be described below in a clear and complete manner. Obviously, the described embodiments are part of the embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0019] The application will be further explained and described below in combination with specific embodiments.

[0020] The unmanned aerial vehicle cluster cooperative task edge computing distribution system and method provided by the embodiments of the application comprises a multi-heterogeneous neural network module, a local fusion module, a global fusion module, a conflict resolution module, a task allocation module and an adaptive optimization module. The multi-heterogeneous neural network module is configured on each unmanned aerial vehicle, and is used for extracting features from multi-modal raw data collected by the unmanned aerial vehicle through multiple different types of neural network models, and generating corresponding semantic feature vectors. The local fusion module is configured on each UAV and is used to calculate the semantic consistency score between the semantic feature vectors output by different neural network models on the same UAV. It determines the fusion weight by combining the confidence of each model and performs weighted fusion on the semantic feature vectors based on the fusion weight to obtain the local fusion feature. The global fusion module is used to construct a communication adjacency matrix between UAVs, and uses a graph attention network mechanism to calculate the attention weight of the local fusion features of each UAV. Based on the attention weight, the local fusion features of multiple UAVs are weighted and fused to obtain the global fusion features. The conflict resolution module is used to detect and resolve semantic discrimination conflicts between multiple models or multiple UAVs; The task allocation module uses a distributed auction algorithm to allocate tasks based on global semantic representation and task requirement vector. The adaptive optimization module is used to adjust the fusion weights of each model based on the conflict information using a gradient descent optimization algorithm, in order to reduce the occurrence of future conflicts.

[0021] Specifically, the core of the system lies in each drone. (in , (for cluster size) all equipped with Various types of neural network models, including but not limited to Convolutional Neural Networks (CNNs), Long Short-Term Memory Networks (LSTMs), Transformers, and Graph Neural Networks (GNNs), are used for multimodal raw data collected by UAVs. This includes visible light images, thermal imaging data, text information, and data from various sensors, for each neural network model. Perform feature extraction independently to generate corresponding semantic feature vectors: ; in ( (representing a d-dimensional real vector space) represents the first... The first drone The output of each model Multi-dimensional semantic feature vectors. This multi-model parallel processing design enables a single UAV to perform semantic understanding of the same event or target from different angles and granularities, providing a rich and diverse information foundation for subsequent feature fusion. Different models have different strengths; for example, CNN excels at extracting spatial features, LSTM is good at processing temporal information, Transformer can capture long-distance dependencies, and GNN is suitable for processing graph-structured data. Through the collaborative work of multiple models, the system can obtain a more comprehensive and accurate semantic representation.

[0022] Within a single drone, the system needs to effectively fuse the feature outputs of different neural network models for the same event. First, the system calculates the semantic consistency score between any two models. For model k1 and model k2 on the i-th drone, their semantic consistency score is defined as the cosine similarity of the output feature vectors of the two models: ; in, and Let represent the output feature vectors of model k1 and model k2 on the i-th UAV, respectively. This formula measures the similarity between the two feature vectors by calculating the cosine of the angle between them; a value closer to 1 indicates greater semantic consistency between the two models. To comprehensively evaluate the overall consistency level of each model with other models, the system calculates the... The first drone The average consistency score of each model with all other models, using express: ; in For the model With model The semantic consistency score between models is used to determine the fusion weights. The system considers both the model's confidence score and the consistency score. Each drone carries K models, where k,j∈{1,2,…,K,k≠j}. Model confidence score is also considered. This reflects the model's confidence in its output, typically calculated using the maximum value or entropy of the output probability distribution. Fusion weights are used... The calculation formula is as follows: ;

[0023] in α and β These are adjustable weighting coefficients used to balance the importance of confidence and consistency in the weight calculation. Through this weighting mechanism, models with high confidence and consistent judgments with other models will receive greater weights. The final local fusion features are obtained through a weighted average: ;

[0024] This local fusion mechanism can effectively integrate the advantages of multiple models within a single UAV, generating more robust and reliable local semantic representations.

[0025] At the cluster level, the system needs to further integrate the local fusion features of multiple drones to form a globally optimal event semantic representation. The system first constructs a communication adjacency matrix between drones. ,in Indicates drone and The connection strength between the two can be determined based on spatial distance, communication quality, collaboration history, etc. To achieve effective global fusion, the system adopts a graph attention network (GAT) mechanism. This mechanism can adaptively learn the importance of different UAV features and perform weighted fusion accordingly. The calculation process of attention weight is as follows: First, the local fusion features of each UAV are converted into query vectors and key vectors respectively: Query vector: Key vector: where is a learnable parameter matrix that is continuously optimized through the training process. Then calculate the attention weight: ; This formula uses the softmax function to ensure that the sum of all attention weights is 1, forming an effective probability distribution. The attention weight reflects the importance of the features of UAV to the formation of the global features of UAV . The final global fusion feature is: ; Through this global fusion mechanism, each UAV can obtain a global semantic representation that integrates the wisdom of the entire cluster, significantly improving the perception and decision-making ability of individual UAVs.

[0026] In the process of multi-model or multi-UAV collaboration, semantic discrimination conflicts are inevitable. The system detects potential conflicts by monitoring the similarity between local fusion features. When the similarity between the local features of any two UAVs is lower than a pre-set threshold, i.e. , the system will trigger the conflict resolution mechanism. The system's confidence weighted conflict resolution strategy is: According to the confidence of each UAV, the conflict features are weighted and averaged, and the UAV with high confidence occupies a larger weight in the final result: ; where N is the number of UAVs in the cluster, and the confidence weight of the jth UAV is defined as the weighted average of all model confidences: ; The system will record detailed information about each conflict, including the time of conflict, the UAVs involved, the severity of the conflict, etc., forming a conflict information . These information will be fed back to the model training and weight adjustment module, and the model performance will be continuously improved through gradient descent and other optimization algorithms: ; in Let be the fusion weight of the k-th model on the i-th drone at time t+1. For learning rate, Indicates weight Find the partial derivative. Let be the conflict loss function. Let be the set of conflicts at time t. Through this closed-loop feedback mechanism, the system can continuously learn and improve, reducing the occurrence of future conflicts.

[0027] The system provides each task to be executed. ( Define a semantic requirement vector for the total number of tasks. This vector encodes the key characteristics and execution requirements of the task, such as task type, complexity, required sensor type, and real-time requirements. By vectorizing task requirements, the system can achieve precise matching between the task and the UAV's capabilities.

[0028] drones With the task The matching degree between them is calculated using the following formula: ;

[0029] This formula comprehensively considers three key factors: first, semantic matching degree, which measures the degree of fit between the UAV's global semantic features and mission requirements through cosine similarity; second, capability coefficient. This reflects the drone's physical constraints, such as hardware capabilities, battery status, and payload conditions; finally, it includes the availability factor. This multi-dimensional matching calculation takes into account the drone's current workload and schedule. This ensures the scientific rigor and feasibility of task allocation.

[0030] To avoid single points of failure and communication bottlenecks that can arise from centralized allocation, the system employs a distributed auction algorithm for task allocation. In this algorithm, each drone acts as an independent intelligent agent, autonomously bidding based on its own capabilities and task suitability. For the task The bidding strategy is as follows: ;

[0031] This bidding strategy incorporates the concept of opportunity cost, meaning that a drone's bid for a given task equals its match to that task minus its match to other optimal tasks. This strategy ensures that drones prioritize competing for the tasks best suited to them, improving overall allocation efficiency.

[0032] The task allocation problem can be modeled as a constrained optimization problem, with the goal of maximizing the overall matching degree: ;

[0033] where is the task assignment matrix, denotes the task assigned to the drone , otherwise 0. This optimization problem needs to satisfy the following constraints: each task must and only can be assigned to one drone: The task load of each drone cannot exceed its capacity upper limit: where denotes the maximum task carrying capacity of the drone . Through the distributed auction mechanism, each drone can reach a globally near-optimal task assignment solution based on local information.

[0034] Considering various unexpected situations that may occur during actual execution, such as drone failure, task priority change, environmental condition change, etc., the system designs a dynamic task reassignment mechanism. When detecting task execution abnormalities or new tasks arriving, the system will recalculate the matching degree and trigger the reassignment process. During the reassignment process, the system will consider the cost and time loss of task switching to ensure that the benefits brought by reassignment are greater than the switching cost. In addition, the system establishes a comprehensive performance monitoring system to track multiple key indicators in real time to evaluate the system running state. The model confidence is calculated by the maximum value of the output probability distribution: ;

[0035] where is the sigmoid function, which is used to normalize the confidence value to the [0, 1] interval. The consistency score measures the overall consistency level of all model outputs within a single drone: ;

[0036] The task completion rate reflects the actual execution effect of the drone: ;

[0037] These indicators provide important feedback information for the adaptive optimization of the system.

[0038] Based on the performance monitoring results, the system uses the gradient descent method to continuously optimize the fusion weights. The weight update formula considers both the task success rate and the consistency of the two optimization objectives: ;

[0039] where and These are the learning rates for different objectives, and the importance of the two objectives can be adjusted according to actual needs. Through this multi-objective optimization, the system can maintain the stability of the model output while ensuring the performance of the task.

[0040] The entire system's workflow can be described as a continuous closed-loop process. At each time step... The system first performs multimodal data acquisition and feature extraction for each UAV. Several neural network models process the input data in parallel to generate initial feature vectors. The process then proceeds to the local fusion stage, where a local semantic representation is generated by calculating the consistency score and updating the fusion weights. In the global fusion phase, the system constructs a communication graph and calculates attention weights, fusing the local features of multiple drones into global features. Meanwhile, the conflict detection mechanism continuously monitors the system status, and immediately initiates the corresponding resolution procedure upon detecting semantic discrepancies. Based on the fused global semantic features, the task allocation module calculates the matching degree between the UAV and the task, and generates the optimal allocation scheme through a distributed auction algorithm. During task execution, the performance monitoring module continuously collects feedback information, including task completion status, changes in model confidence, and system resource status. Finally, the adaptive optimization module adjusts system parameters based on feedback information, including model fusion weights, task allocation strategies, and conflict resolution rules, forming a complete closed-loop optimization mechanism. This design ensures that the system can maintain efficient and stable operation in a dynamically changing environment and continuously improve performance over time.

[0041] In this embodiment, UAV swarm collaborative perception refers to the sharing of observation information among multiple UAVs. By fusing local and shared observations, the UAVs improve their own perception capabilities and collaboratively decide on flight trajectories accordingly, achieving collaborative detection and tracking of multiple targets. This invention, through the construction of a distributed UAV network architecture, realizes multi-UAV collaborative environmental perception and task execution capabilities. The system supports dynamic task allocation, path planning, and information sharing, enabling efficient swarm collaborative operations in complex environments.

[0042] In the field of multi-source heterogeneous data fusion technology, the semantic fusion mechanism of this invention specifically addresses the problem of unified representation and intelligent fusion of multi-source heterogeneous data. During UAV missions, multi-dimensional information such as positioning data, video streams, and sensor data is generated. Effective fusion of this heterogeneous data is a key technology for achieving intelligent perception. By establishing a local and global two-layer consistency discrimination framework, the system achieves unified representation and intelligent fusion of data from different sources and in different formats, significantly improving the accuracy and efficiency of information processing.

[0043] The technical field of the application is distributed edge computing: the application adopts a distributed edge computing architecture, deploys core algorithms on the edge nodes of the UAVs, and reduces the dependence on the central node. This design fully utilizes the low delay and high real-time advantages of edge computing, and improves the parallel computing capability and fault tolerance of the system through distributed processing. The edge data processing mechanism of the system includes data compression, feature extraction, and task relevance screening functions, effectively reducing the communication load and retaining core information.

[0044] Through the synergistic integration of the above-mentioned multiple technologies, the application constructs a UAV cluster intelligent perception system architecture with the capabilities of "intelligent collaboration + semantic fusion + distributed processing", which can achieve accurate identification, tracking and decision-making of targets in dynamic, complex or high-risk environments, and has significant technical innovation value and broad application prospects.

[0045] The edge computing distribution method for UAV cluster collaborative tasks provided by the embodiment of the application adopts the system as described above, and the method comprises the following steps: S1. Multi-heterogeneous neural network feature extraction: each UAV extracts features from the collected multi-modal raw data through multiple different types of neural network models to generate semantic feature vectors; S2. Local fusion: calculate the semantic consistency scores between the semantic feature vectors output by different models on the same UAV, combine the confidence of each model to determine the fusion weight, and perform weighted fusion on the semantic feature vectors to obtain local fusion features; S3. Global fusion: construct a communication adjacency matrix between the UAVs, calculate the attention weights of the local fusion features of each UAV using a graph attention network mechanism, and perform weighted fusion on the local fusion features based on the attention weights to obtain global fusion features; S4. Conflict resolution: monitor the similarity between the local fusion features of any two UAVs, and when the similarity is lower than a preset threshold, trigger the conflict resolution mechanism, perform weighted averaging on the conflict features according to the confidence of each UAV, and record the conflict information; S5. Task allocation: define a semantic demand vector for each task to be executed, calculate the matching degree between the global fusion features of the UAVs and the semantic demand vector of the task, and perform task allocation using a distributed auction algorithm; S6. Adaptive optimization: adjust the fusion weights of each model through a gradient descent optimization algorithm according to the conflict information to improve the performance of the system.

[0046] In the embodiment, the application constructs an efficient distributed collaborative architecture to solve the communication bottleneck problem, disperses the computing task to each unmanned aerial vehicle node by designing a distributed edge computing architecture, reduces the dependence on the center node, realizes a deep semantic fusion mechanism, constructs a multi-level and multi-dimensional feature extraction and fusion framework by integrating a plurality of heterogeneous neural networks, and realizes deep semantic alignment and intelligent fusion of heterogeneous data. Supporting heterogeneous cluster intelligent collaboration, dynamically allocating tasks and adjusting the collaboration strategy according to the performance characteristics of different unmanned aerial vehicles; By solving these key technical problems, the application can provide more intelligent, efficient and reliable collaborative sensing capability for the unmanned aerial vehicle cluster.

[0047] Finally, it should be pointed out that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the application.

Claims

1. An edge computing distribution system for UAV swarm cooperative task, characterized in that, The application relates to a multi-heterogeneous neural network system for multi-robot task allocation. The multi-heterogeneous neural network module is configured on each unmanned aerial vehicle (UAV) and is used for feature extraction of multi-modal raw data collected by the UAV through multiple different types of neural network models to generate corresponding semantic feature vectors. The local fusion module is configured on each UAV and is used for calculating semantic consistency scores between semantic feature vectors output by different neural network models on the same UAV, determining fusion weights based on confidence of each model, and performing weighted fusion on the semantic feature vectors based on the fusion weights to obtain local fusion features. The global fusion module is used for constructing a communication adjacency matrix between UAVs, calculating attention weights of local fusion features of each UAV by using a graph attention network mechanism, and performing weighted fusion on local fusion features of multiple UAVs based on the attention weights to obtain global fusion features. The conflict resolution module is used for detecting and resolving semantic discrimination conflicts between multiple models or multiple UAVs. The task allocation module performs task allocation based on global semantic representation and task demand vectors by using a distributed auction algorithm. The adaptive optimization module adjusts fusion weights of each model by using a gradient descent optimization algorithm according to the conflict information to reduce future conflicts. The neural network models in the multi-heterogeneous neural network module include convolutional neural networks, long short-term memory networks, transformers and graph neural networks. 2.The edge computing distribution system for UAV swarm cooperative mission according to claim 1, wherein, In the local fusion module, the specific steps for calculating the semantic consistency scores are as follows: 3.The edge computing distribution system for UAV swarm cooperative mission according to claim 1, wherein, For model k1 and model k2 on the ith UAV, the semantic consistency score is the cosine similarity of the feature vectors output by the two models. The average consistency score of each model with all other models is calculated. The fusion weight calculation comprehensively considers the confidence and consistency score of the model. In the local fusion module, the calculation formula of the fusion weight is as follows: 4.The edge computing distribution system for UAV swarm cooperative mission according to claim 1, wherein, The process of calculating the attention weight by using the graph attention network mechanism in the global fusion module includes the following steps: ； wherein K is the number of models carried by each drone, α and β are adjustable weight coefficients, is the model confidence, is the average consistency score of the kth model on the ith drone with all other models. 5.The edge computing distribution system for UAV swarm cooperative mission according to claim 1, wherein, The local fusion features of each UAV are respectively converted into query vectors and key vectors. The attention weight between the UAVs is calculated to form a probability distribution. The local fusion features are weighted and summed according to the attention weight to generate global fusion features. The specific working steps of the conflict resolution module are as follows: 6.The edge computing distribution system for UAV swarm cooperative mission according to claim 1, wherein, Conflict resolution is triggered when the local feature similarity of two UAVs is lower than a preset threshold. A confidence weighted conflict resolution strategy is adopted to perform weighted average on the conflict features. Conflict information is recorded and fed back to the model training and weight adjustment module. The specific steps of the task allocation module for task allocation are as follows:

7. The edge computing distribution system for swarm of UAVs cooperative mission according to claim 1, wherein, A semantic demand vector is defined for each task. A multi-dimensional matching degree between the UAV and the task is calculated, including a semantic matching degree, a capability coefficient and an availability coefficient. A distributed auction algorithm is used for task allocation, and the UAV autonomously bids according to the matching degree. A dynamic task reallocation mechanism is further included, when task execution abnormity or a new task arrives, the matching degree is recalculated and the reallocation process is triggered, and the task switching cost and time loss are considered. 8.The edge computing distribution system for UAV swarm cooperative mission according to claim 1, wherein, ​ 9.The edge computing distribution system for UAV swarm cooperative mission according to claim 1, wherein, The adaptive optimization module includes a performance monitoring unit that tracks model confidence, consistency score, and task completion rate indicators in real time and continuously optimizes the fusion weights using a gradient descent method. 10.Edge computing distribution method for UAV swarm cooperative task, using the system according to any one of claims 1-9, characterized in that, The method comprises the following steps: S1. Multi-heterogeneous neural network feature extraction: Each UAV extracts features from the collected multi-modal raw data using multiple different types of neural network models to generate semantic feature vectors; S2. Local fusion: Calculate the semantic consistency score between the semantic feature vectors output by different models on the same UAV, combine the confidence of each model to determine the fusion weight, and perform weighted fusion on the semantic feature vectors to obtain local fusion features; S3. Global fusion: Construct a communication adjacency matrix between UAVs, calculate the attention weight of the local fusion features of each UAV using a graph attention network mechanism, and perform weighted fusion on the local fusion features based on the attention weight to obtain global fusion features; S4. Conflict resolution: Monitor the similarity between the local fusion features of any two UAVs. When the similarity is lower than the preset threshold, trigger the conflict resolution mechanism, perform weighted averaging on the conflict features according to the confidence of each UAV, and record the conflict information; S5. Task allocation: Define a semantic requirement vector for each task to be executed, calculate the matching degree between the global fusion features of the UAV and the semantic requirement vector of the task, and perform task allocation using a distributed auction algorithm; S6. Adaptive optimization: Adjust the fusion weights of each model using a gradient descent optimization algorithm based on the conflict information to improve system performance.

Images