A dynamic topology-based unmanned aerial vehicle cluster adaptive structure perception cooperation reasoning method

Abstract

The application discloses a kind of unmanned aerial vehicle cluster adaptive structure perception collaborative inference method based on dynamic topology.Aiming at the problems of topological blindness, communication bottleneck and inference instability of industrial unmanned aerial vehicle group in dynamic topology collaborative inference, the method first constructs a multi-modal fusion model based on heterogeneous information network (HIN), and maps the computing, communication and semantic features of the unmanned aerial vehicle into topological perception representation. Secondly, a structure-adaptive knowledge distillation mechanism is established, which dynamically adjusts the compression ratio of intermediate features using the structure similarity weight extracted from HIN, reducing transmission overhead while ensuring inference accuracy. Finally, a topological perception deep reinforcement learning scheduler using integrated graph neural network (GNN) is used to guide the optimization allocation of tasks within the cluster using real-time structure hints, thereby significantly reducing inference delay and energy consumption in highly dynamic mobile environments, and improving the robustness and inference consistency of the system when the topology changes.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and edge computing technology, specifically to an adaptive structure-aware collaborative reasoning method for UAV swarms based on dynamic topology, which is used to realize knowledge collaboration, reasoning acceleration and energy efficiency optimization among heterogeneous devices in a dynamic network environment. Background Technology

[0002] With the widespread application of unmanned aerial vehicle (UAV) swarms in industrial intelligence fields such as power line inspection, real-time disaster monitoring, and smart city management, distributed deep learning inference has become crucial for achieving low-latency and energy-efficient decision-making. Since single-unit UAVs are limited by onboard processing power and battery capacity, current mainstream solutions typically employ a collaborative inference paradigm. This involves partitioning complex deep neural network (DNN) inference tasks and offloading them to other UAV nodes or edge servers within the swarm for execution, thereby alleviating the resource constraints of single-unit UAVs.

[0003] However, in real-world dynamic industrial environments, existing collaborative inference schemes still have significant limitations: First, the high mobility of UAVs leads to non-stationary network topology, making it difficult for existing static partitioning schemes to adapt to time-varying wireless link capacity. This results in bottlenecks during task allocation, making it difficult to guarantee the continuity of inference tasks. Second, existing frameworks often treat computational optimization and communication optimization as independent islands, ignoring the inherent logical connection between the semantic relevance of intermediate features in neural networks and the physical topology. Their fixed-ratio compression methods cannot dynamically adjust the data transmission volume based on the link quality and semantic importance between nodes. Finally, traditional task scheduling mechanisms typically simplify the network state into a flat feature vector. This topology-blind representation cannot capture the complex structured dependencies between nodes, resulting in poor system robustness when facing link degradation or node expansion, easily leading to severe latency fluctuations and handover failures.

[0004] In summary, how to construct a knowledge-based collaborative reasoning framework that can perceive topological structure, achieve deep coupling between semantic features and physical environment, and possess topological elasticity and high-performance scheduling capabilities is a core technical problem that urgently needs to be solved in the field of collaborative operation of industrial-grade UAV swarms. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive structure-aware collaborative reasoning method for UAV swarms based on dynamic topology. By constructing a heterogeneous information network model and a deep reinforcement learning framework for structure awareness, this method solves the problems of high inference latency, high energy consumption, and poor topological robustness of industrial UAV swarms in dynamic environments, and achieves reliable collaborative reasoning with deep coupling between physical environment and semantic features.

[0006] To achieve the above objectives, this invention proposes an adaptive structure-aware cooperative reasoning method for UAV swarms based on dynamic topology, comprising the following key steps:

[0007] 1. Multimodal fusion modeling and feature extraction based on heterogeneous information networks (HIN)

[0008] like Figure 1 As shown, this method first constructs a multimodal fusion model based on heterogeneous information networks (HINs). Specifically, this method models the UAV swarm and its interaction environment as a directed heterogeneous graph. :

[0009]

[0010] A set of nodes mapped from a set of drone devices. It is a heterogeneous edge set. and These are functions that map nodes and edges to their specific set of types. The edge type set is... It includes communication links, task dependencies, and semantic relevance, and is defined as follows:

[0011]

[0012] Indicates the physical layer communication link. This indicates the task dependencies in the logical layer. This represents semantic relevance at the knowledge layer. Based on the above type definition, the global heterogeneous edge set... It is further formalized into the union of three different relational subsets:

[0013]

[0014] Each subset (for Each node captures specific interactions between different devices in physical, logical, or knowledge dimensions. Based on the aforementioned modeling approach, each node... Each is associated with a multimodal feature vector :

[0015]

[0016] in, To calculate the frequency, For communication bandwidth parameters, For semantic embedding vectors, To improve resource utilization, this method uses a graph neural network (GNN) to encode the heterogeneous graph, aggregates multimodal node features through graph convolution operators, and finally generates a topology-aware embedding that can represent the global real-time topology structure through pooling operations. :

[0017]

[0018] This embedding vector serves as a global structural guide, used to guide the intelligent selection of task allocation strategies in subsequent steps.

[0019] 2. Calculate the structural-aware similarity weight index

[0020] To quantify the collaborative potential between nodes in a dynamic topology, this method defines a structure-aware similarity index that integrates logical and physical topology. First, based on communication bandwidth characteristics With transmission threshold Logic for determining the connectivity of physical links:

[0021]

[0022] Among them, I For indicator functions, Used to estimate the effective link capacity between devices. Simultaneously, based on semantic embedding similarity and semantic threshold between nodes. Construct semantic association determination logic:

[0023]

[0024] This represents the semantic cosine similarity between nodes. Based on this, our method integrates physical and logical layer topologies to calculate structure-aware similarity weights. :

[0025]

[0026] To achieve the fusion balance coefficient, This represents the normalized effective bandwidth capacity. This weighted metric is obtained through… The product term achieves strict dimensionality reduction for unavailable physical links, i.e., when the physical link does not meet the transmission threshold ( When the similarity is 0, the similarity weight is forced to zero. This design ensures that the quantified collaboration potential is always based on physical feasibility, thus serving as a structural compass for the scheduling agent in dynamic environments, guiding task flow to reliable paths with high bandwidth and high semantic relevance.

[0027] 3. Perform adaptive feature compression and knowledge distillation.

[0028] Based on the similarity weight This method establishes a structure-adaptive knowledge distillation (SKD) mechanism. First, the structure-adaptive compression factor is defined. It follows a non-linear exponential decay logic:

[0029]

[0030] in, This is the upper limit of the minimum compression ratio. This is the compression sensitivity coefficient. This compression factor is used to reduce the net communication delay during intermediate feature transmission. Net communication energy consumption Redefine:

[0031]

[0032]

[0033] and These represent the basic communication latency and basic energy consumption under uncompressed conditions, respectively. Building upon this, to maintain inference accuracy under significant data compression, this method constructs a joint training architecture where a teacher model guides a student model. The training objective function of this architecture is determined by the classification accuracy loss. Alignment loss with structure-weighted features These are the components. Among them, classification accuracy loss... The calculation is performed using the cross-entropy function:

[0034]

[0035] y represents the ground truth label of the input sample. This represents the logarithm of the student model's output. To achieve topology-aware knowledge transfer, a structure-weighted feature alignment loss is used. Defined as:

[0036]

[0037] This represents the normalized intermediate feature map generated by model M at node v. Finally, by minimizing the total distillation loss... To iteratively update the student model parameters:

[0038]

[0039] in To balance hyperparameters, this mechanism addresses topological correlation. By directly coupling it into the loss function, we can ensure deeper feature compression between highly structurally related node pairs while maintaining key semantic information without loss, thus achieving high-fidelity knowledge collaboration with extremely low communication overhead.

[0040] 4. Calculate the optimization model that integrates hardware constraints and multi-objective costs.

[0041] This method constructs a mathematical model that integrates hardware performance constraints and multi-objective costs based on the aforementioned task allocation decision variables and adaptive compression mechanism. First, considering the local resource limitations of the UAV nodes, memory constraints are introduced to ensure the stability of the inference task. For tasks allocated to nodes... The l-th layer of a deep neural network must satisfy:

[0042]

[0043] in, This represents the weights of layer l and the memory usage of the activation graph. Represents a node The available physical memory capacity. Under this constraint, the single-layer computation latency is defined. Energy consumption of single-layer computing :

[0044]

[0045]

[0046] in, To calculate the load, Calculate the frequency for the node. The effective switching capacitor coefficient is used. For cross-node data transmission, the compression factor from step 3 is utilized. Define intermediate feature transmission delay :

[0047]

[0048] The size of the original output feature. For link bandwidth, and These are the binary decision variables for the computation nodes and the aggregation nodes, respectively. Subsequently, this method constructs the total system delay. Total energy efficiency loss The cumulative calculation logic:

[0049]

[0050]

[0051] Allocate energy consumption for the initial input. This addresses the power loss under corresponding transmission delays. Finally, to achieve Pareto optimality in dynamic environments, this method formalizes the distributed reasoning problem into minimizing the joint cost function. Multi-objective optimization problem:

[0052]

[0053] and For unnormalized weighting factors, This is a constraint penalty term designed to enhance topology robustness. The model provides a quantitative evaluation criterion for the scheduling agent by explicitly coupling logical task mappings with physical heterogeneity, ensuring the global optimality of the inference scheme in resource-constrained and dynamic topology environments.

[0054] 5. Task allocation and scheduling based on deep reinforcement learning

[0055] This method constructs a deep reinforcement learning (DRL) scheduler integrating graph neural networks, utilizing topology-aware embeddings to achieve optimal task allocation. First, this method constructs a comprehensive observation space to generate state vectors. It is composed of dynamic features reflecting instantaneous resource fluctuations, structural embeddings reflecting the global topological background, structural hints reflecting local neighbor relationships, and state vectors reflecting task progress.

[0056]

[0057] It aggregates the remaining computing frequency, available bandwidth, and memory utilization of nodes within the cluster; The global topology-aware embedding generated in step 1; This is the structural hint vector extracted based on the similarity weight matrix in step 2; Used to track the progress of the current inference task within the deep neural network layers. Scheduler's policy network. A multi-head architecture is adopted, and the Softmax function is used to map the internal features to the probability distribution of computation node i and aggregation node j:

[0058]

[0059]

[0060] and These are the feature logarithms generated by the output layer of the policy network. The scheduling agent learns the optimal scheduling policy by maximizing the cumulative reward, with the instantaneous reward function at each decision step... Defined as:

[0061]

[0062] and The normalization constant is The robustness adjustment coefficient is used. This is a constraint penalty term designed to enhance topological robustness. The penalty term... Through Iverson brackets I[ Establish a robust threshold-based system The judgment logic is as follows:

[0063]

[0064] This is a penalty constant. When the similarity weight of the path selected by the scheduling decision in step 2 is lower than the threshold, or when the corresponding physical link is broken, causing the weight to drop to zero, this reward function will generate a high penalty, guiding the scheduling agent to actively avoid unstable links and use structural cues to identify and route to semantic neighbor nodes with high cooperation potential in advance in the dynamic topology.

[0065] 6. Recursive Execution and State Evolution of Distributed Inference Tasks

[0066] To ensure the continuous execution of multi-layer deep neural networks under dynamic topologies, this method employs recursive host transfer logic and a closed-loop feedback mechanism. First, a cooperative coupling function is defined. , used to characterize the transmission quality from the current data-holding node k to the candidate computing node i:

[0067]

[0068] This function explicitly couples physical connectivity with structure-aware compression efficiency. During inference execution, the system follows the hierarchical order of the neural network. Iteratively execute the following closed-loop logic:

[0069] First, extract the structural cue vector from the global similarity matrix based on the index k of the node currently holding the data. and embedding real-time resources Construct the decision state at the current level. Secondly, the optimal computation node i and aggregation node j are selected through the strategy distribution generated by the scheduling agent. If i ≠ k, the intermediate features are compressed and transmitted using the adaptive compression factor in step 3. The resulting transmission delay is:

[0070]

[0071] If i=k, the transmission delay is set to zero. The load calculation is completed at node i. After processing, the inference task enters the state evolution phase, and the data holding node index of layer l+1... The decision probability distribution of the aggregation nodes is sampled and recursively updated:

[0072]

[0073] After each level of inference is completed, the system updates the heterogeneous graph in real time based on the latest positional changes of the drone swarm and the remaining resources. The topological connections and node characteristics are determined, and the similarity matrix calculation in step 2 is retried. Through this recursive execution and real-time dynamic alignment of states, this method can transform a static structural representation into a dynamic navigation map, enabling seamless task switching and knowledge collaboration among UAV nodes in non-stationary industrial environments.

[0074] The inventive principle of this invention:

[0075] The core principle of this invention is to achieve deep fusion of multimodal features of UAV swarms by constructing a Heterogeneous Information Network (HIN). The basic idea is to leverage the nonlinear representation capabilities of graph structures to map communication constraints at the physical layer, task dependencies at the logical layer, and semantic relevance at the knowledge layer to a unified topology-aware space. By defining a structure-aware similarity index, this invention deeply couples the originally isolated computation and communication optimization processes, using structural similarity as prior knowledge to dynamically guide feature compression and knowledge distillation, thereby achieving efficient dimensionality reduction of data at the semantic level. Simultaneously, by combining structural cues extracted from graph neural networks to enhance the perceptual capabilities of the reinforcement learning scheduler, blind path searching is transformed into optimized pathfinding guided by a structured graph, fundamentally solving the problems of blindness and lag in task scheduling under highly dynamic topologies.

[0076] The beneficial effects of this invention are as follows:

[0077] Compared with existing technologies, this invention has significant advantages: First, it significantly improves the energy efficiency of task execution. Through adaptive feature compression, it greatly reduces communication overhead while ensuring inference accuracy, achieving coordinated optimization of inference latency and energy consumption. Second, it has strong topological robustness and decision consistency, enabling it to detect link degradation trends in advance and implement proactive rerouting, effectively avoiding switching failures and latency spikes caused by node movement. Third, it has excellent heterogeneous adaptability and scalability, flexibly responding to the dynamic addition and subtraction of large-scale heterogeneous UAV nodes, providing highly reliable deterministic guarantees for distributed edge intelligent tasks in complex industrial environments. Attached Figure Description

[0078] Figure 1 A block diagram of a structure-aware collaborative reasoning architecture for dynamic topology UAV swarms;

[0079] Figure 2 A comparison chart of system inference latency, energy efficiency, and algorithm decision overhead;

[0080] Figure 3 Figures illustrating system robustness analysis under different topologies, link failures, and mobile speeds;

[0081] Figure 4 A comparison of ablation experiment results and system reliability distribution for the structure-aware mechanism;

[0082] Figure 5 This is a graph evaluating the consistency between the Pareto front distribution and dynamic trajectory performance of the scheduling algorithm.

[0083] Figure 6 A heatmap showing the parametric sensitivity analysis of the impact of core balance coefficient and inference load on inference latency. Detailed Implementation

[0084] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0085] In a preferred embodiment, this solution is applied to industrial power line inspection scenarios. For example... Figure 1 As shown, the system consists of a cluster of 12 heterogeneous drones and edge servers. The computational frequency ratio between drone nodes and edge servers is preset to 1:200, the communication link bandwidth is set to 10Mbps, the physical communication distance threshold is set to 600m, and the average drone speed is set to 10m / s, thus constructing a non-stationary dynamic topology environment. The deep inference task to be processed uses a 5-layer neural network, with a single-layer computational load of approximately 10 GFLOPs, and the initial input feature data size is distributed between 50MB and 200MB.

[0086] During system operation, firstly through, as Figure 1 The architecture shown performs multimodal feature fusion and real-time modeling. During this process, the system acquires the computing frequency (e.g., 0.1 GHz), bandwidth resources, semantic embedding features, and resource utilization of each UAV node in real time. This heterogeneous information is mapped to node feature vectors in a heterogeneous information network, and a graph neural network encoder is used to generate embedding vectors that reflect the global topological context.

[0087] Based on this, the system calculates a structure-aware similarity weight index according to the physical link status and semantic similarity between nodes. During the calculation process, a balancing coefficient is fused. The value is set to 0.3. If two drones become separated by more than 600m due to their flight paths diverging, the system determines that the physical link is broken. At this point, the physical link indicator function automatically forces the similarity weight to zero. This mechanism ensures that collaborative decisions are always based on the premise that the physical link is available, effectively avoiding the risk of handover failure due to high-speed node movement. (Refer to...) Figure 2 and Figure 3 The evaluation results show that this solution demonstrates significant robustness when facing challenges of different topology sizes, link failure rates, and mobility speeds. Especially under extreme conditions with a link failure rate reaching 40%, the system maintains stable latency performance, effectively reducing performance fluctuations caused by drastic topology changes.

[0088] After identifying paths with high collaborative potential, the system performs structure-adaptive feature compression and knowledge distillation. The compression factor is dynamically adjusted based on the calculated similarity weights. For node pairs with high similarity weights, the system uses exponential decay logic to perform deep feature compression, for example, compressing the original feature data to 0.1 to 0.2 times. A structure-weighted feature loss function guides the lightweight student model to capture the key reasoning features of the teacher model. Figure 4 The ablation experiments and reliability distribution results shown demonstrate that, compared to the baseline scheme without topology awareness, this structure-aware mechanism significantly reduces the coefficient of variation of latency while improving decision consistency, achieving deep coupling between semantic importance and physical link quality.

[0089] Task allocation decisions are output in real-time by a deep reinforcement learning scheduler integrating graph neural networks. The scheduling agent perceives the comprehensive state space, including structural cue vectors, and outputs the optimal node allocation and aggregation strategy for each layer of inference tasks by minimizing the joint cost function, which includes a topological penalty term. Figure 5 As shown, the scheduling algorithm in this scheme exhibits excellent convergence characteristics and leads the accuracy-delay Pareto front distribution, maintaining high consistency in inference performance during dynamic trajectory execution. Furthermore, as... Figure 6 The parameter sensitivity analysis shown demonstrates that by adjusting the balance coefficient... The system can flexibly adapt to different task load requirements. For example, under low load conditions, the overhead can be further reduced by increasing the semantic matching weight; while under high load conditions, the system automatically increases its reliance on the quality of the physical link to ensure the real-time completion of inference tasks.

[0090] Throughout the execution process, the system employs recursive host transfer logic. After each layer of neural network inference is completed, the node and link states are updated in real time, and the next decision is re-injected, achieving closed-loop adaptive scheduling for the entire distributed inference process. Through this implementation process, when facing non-stationary topologies caused by high mobility, this solution achieves approximately 49.6% reduction in inference latency and 15.4% improvement in energy efficiency compared to the baseline solution, while also improving inference consistency by approximately 18.7%. This effectively meets the real-time, topology robustness, and reliability requirements of distributed edge intelligence tasks in industrial environments.

Claims

1. A method for adaptive structure-aware cooperative reasoning of unmanned aerial vehicle (UAV) swarms based on dynamic topology, characterized in that, Includes the following steps: 1) Multimodal fusion modeling and feature extraction based on heterogeneous information network (HIN): The UAV swarm and its interaction environment are modeled as a directed heterogeneous graph. The computing resources, communication bandwidth, semantic features and resource utilization of each node are obtained and mapped into multimodal feature vectors. The heterogeneous graph is encoded using a graph neural network, and a topology-aware embedding representing the global real-time topology structure is generated through domain aggregation and feature fusion. 2) Calculate the structure-aware similarity weight index: Based on the heterogeneous graph, calculate the structure-aware similarity weight index to quantify the collaboration potential between nodes by determining the physical link connectivity and semantic relevance; use the physical link determination function to enforce a forced zero constraint on the weight index to exclude collaboration paths that are physically unreachable. 3) Perform structure-adaptive feature compression and knowledge distillation: Based on the similarity weight index, dynamically calculate the structure-adaptive compression factor through nonlinear decay logic; During the knowledge distillation process, the similarity weight index is used to guide the lightweight model to perform feature alignment, thereby completing knowledge collaborative transfer in a compressed state. 4) Optimization model for computing hardware constraints and multi-objective costs: Hardware memory capacity is introduced as a hard constraint for task allocation, and a cumulative cost model including computing latency, energy consumption and transmission latency modulated by the compression factor is constructed to quantitatively evaluate the system inference performance. 5) Task allocation and scheduling based on deep reinforcement learning: Construct a state space including global topology embedding, structural cue vector and task progress; The scheduling agent performs policy iteration based on the reward function including topology constraint penalty term generated by the optimization model, and outputs the node allocation strategy for each inference task. 6) Recursive execution and state evolution of distributed inference tasks: According to the execution hierarchy of the neural network, the data holding node index of the subsequent level is updated by recursively sampling the decision probability distribution generated by the scheduling agent; after each inference task is completed, the heterogeneous graph parameters are updated in real time and the next decision is re-injected to achieve closed-loop adaptive scheduling throughout the process.

2. The method according to claim 1, characterized in that, The multimodal fusion modeling and feature extraction based on heterogeneous information networks (HIN) in step 1) specifically includes: Model the drone swarm and its interactive environment as a directed heterogeneous graph. : A set of nodes mapped from a set of drone devices. It is a heterogeneous edge set. and These are functions that map nodes and edges to their specific set of types; where the edge type set is... It includes communication links, task dependencies, and semantic relevance, and is defined as follows: Indicates the physical layer communication link. This indicates the task dependencies in the logical layer. Represents semantic relevance at the knowledge layer; based on the above type definition, global heterogeneous edge sets It is further formalized into the union of three different relational subsets: Each subset (for Each node captures specific interactions between different devices in physical, logical, or knowledge dimensions; based on the aforementioned modeling approach, each node... Each is associated with a multimodal feature vector : in, To calculate the frequency, For communication bandwidth parameters, For semantic embedding vectors, To improve resource utilization, this method uses a graph neural network (GNN) to encode the heterogeneous graph, aggregates multimodal node features through graph convolution operators, and finally generates a topology-aware embedding that can represent the global real-time topology structure through pooling operations. : This embedding vector serves as a global structural guide, used to guide the intelligent selection of task allocation strategies in subsequent steps.

3. The method according to claim 1, characterized in that, The calculation of the structure-aware similarity weight index in step 2) specifically includes: To quantify the collaborative potential between nodes in a dynamic topology, this method defines a structure-aware similarity index that integrates logical and physical topology. First, based on communication bandwidth characteristics With transmission threshold Logic for determining the connectivity of physical links : Among them, I For indicator functions, Used to estimate the effective link capacity between devices; simultaneously, based on semantic embedding similarity and semantic threshold between nodes. Constructing semantic association determination logic : This method represents the semantic cosine similarity between nodes; based on this, it integrates physical and logical layer topologies to calculate structure-aware similarity weights. : To achieve the fusion balance coefficient, This represents the normalized effective bandwidth capacity; this weighting metric is achieved through... The product term achieves strict dimensionality reduction for unavailable physical links, i.e., when the physical link does not meet the transmission threshold ( When the similarity score is 0, the similarity weight is forced to zero; furthermore, by adjusting the parameters... To prevent the scheduling agent from falling into semantic greed during the decision-making process, i.e. avoiding decision failures caused by focusing only on semantic matching while ignoring physical link quality, this design ensures that the quantified collaboration potential is always based on physical feasibility. Thus, it serves as a structural compass for the scheduling agent in dynamic environments, guiding task flow to reliable paths with high bandwidth and high semantic relevance.

4. The method according to claim 1, characterized in that, Step 3) specifically includes the following: Adaptive feature compression and knowledge distillation of the execution structure. Based on the similarity weight This method establishes a structure-adaptive knowledge distillation (SKD) mechanism; first, it defines a structure-adaptive compression factor. It follows a non-linear exponential decay logic: in, This is the upper limit of the minimum compression ratio. This is the compression sensitivity coefficient; this compression factor is used to account for the net communication delay during intermediate feature transmission. Net communication energy consumption Redefine: and These represent the basic communication latency and basic energy consumption under uncompressed conditions, respectively. Building upon this, to maintain inference accuracy under significant data compression, this method constructs a joint training architecture where a teacher model guides a student model. The training objective function of this architecture is determined by the classification accuracy loss. Alignment loss with structure-weighted features Together they form a whole, including classification accuracy loss. The calculation is performed using the cross-entropy function: y represents the ground truth label of the input sample. The output logarithm of the student model; the structure-weighted feature alignment loss to achieve topology-aware knowledge transfer. Defined as: This represents the normalized intermediate feature map generated by model M at node v; ultimately, it minimizes the total distillation loss. To iteratively update the student model parameters: in To balance hyperparameters, this mechanism addresses topological correlations. By directly coupling it into the loss function, we can ensure deeper feature compression between highly structurally related node pairs while maintaining key semantic information without loss, thus achieving high-fidelity knowledge collaboration with extremely low communication overhead.

5. The method according to claim 1, characterized in that, The optimization model for merging hardware constraints and multi-objective costs in step 4) specifically includes: This method, based on the aforementioned task allocation decision variables and adaptive compression mechanism, constructs a mathematical model that integrates hardware performance constraints and multi-objective costs. First, considering the local resource limitations of UAV nodes, memory constraints are introduced to ensure the stability of the inference task. For tasks allocated to nodes... The l-th layer of a deep neural network must satisfy: in, This represents the weights of layer l and the memory usage of the activation graph. Represents a node Given the available physical memory capacity, and under this constraint, define the single-layer computation latency. Energy consumption of single-layer computing : in, To calculate the load, Calculate the frequency for the node. To optimize the switching capacitor coefficient; for cross-node data transmission, utilize the compression factor from step 3). Define intermediate feature transmission delay : The size of the original output feature. For link bandwidth, and These are the binary decision variables for the computation nodes and aggregation nodes, respectively; subsequently, this method constructs the total system delay. Total energy efficiency loss The cumulative calculation logic: Allocate energy consumption for the initial input. To address the power loss under transmission delay, and ultimately, to achieve Pareto optimality in dynamic environments, this method formalizes the distributed reasoning problem into minimizing the joint cost function. Multi-objective optimization problem: and For unnormalized weighting factors, The constraint penalty term is designed to enhance topological robustness; the model provides a quantitative evaluation criterion for the scheduling agent by explicitly coupling logical task mapping with physical heterogeneous features, ensuring the global optimality of the inference scheme in resource-constrained and dynamic topology environments.

6. The method according to claim 1, characterized in that, Step 5) specifically includes the task allocation and scheduling based on deep reinforcement learning, which includes: This method constructs a deep reinforcement learning (DRL) scheduler integrating graph neural networks, utilizing topology-aware embeddings to achieve optimal task allocation. First, this method constructs a comprehensive observation space to generate state vectors. It is composed of dynamic features reflecting instantaneous resource fluctuations, structural embeddings reflecting the global topological background, structural hints reflecting local neighbor relationships, and state vectors reflecting task progress. It aggregates the remaining computing frequency, available bandwidth, and memory utilization of nodes within the cluster. The global topology-aware embedding generated in step 1), This is the structural cue vector extracted based on the similarity weight matrix in step 2). Used to track the execution progress of the current inference task at the deep neural network level; the scheduler's policy network. A multi-head architecture is adopted, and the Softmax function is used to map the internal features to the probability distribution of computation node i and aggregation node j: and These are the feature logarithms generated by the output layer of the policy network; the scheduling agent learns the optimal scheduling policy by maximizing the cumulative reward, and the instantaneous reward function for each decision step. Defined as: and The normalization constant is The robustness adjustment coefficient is... The penalty term is a constraint penalty term designed to enhance topological robustness. Through Iverson brackets I[ Establish a robust threshold-based system The judgment logic is as follows: This is a penalty constant. When the similarity weight of the path selected by the scheduling decision in step 2) is lower than the threshold, or the corresponding physical link is broken causing the weight to drop to zero, this reward function will generate a high penalty, guiding the scheduling agent to actively avoid unstable links and use structural cues to identify and route to semantic neighbor nodes with high cooperation potential in advance in the dynamic topology; based on the similarity weight in step 2). The construction of the penalty term, and the triggering logic of the penalty term, follows the following segmented conditions: By physical link unavailable ( When =0, force to To maximize this, the mechanism guides the scheduling agent to automatically avoid physically infeasible decisions; furthermore, to prevent the scheduling strategy from falling into semantic greed, i.e., overemphasizing semantic compression while ignoring physical link quality, this method adjusts the balance factor. To optimize the trade-off between semantic compression and physical link quality, the gradient mechanism is expressed as follows: This dual-guidance mechanism ensures that the learning strategy eventually converges to a semantically efficient and topologically robust Pareto-optimal task allocation scheme, thereby enabling reliable reasoning with deterministic guarantees in non-stationary dynamic environments.

7. The method according to claim 1, characterized in that, The recursive execution and state evolution of the distributed reasoning task in step 6) specifically include: To ensure the continuous execution of multilayer deep neural networks under dynamic topologies, this method employs recursive host transfer logic and a closed-loop feedback mechanism. First, a cooperative coupling function is defined. , used to characterize the transmission quality from the current data-holding node k to the candidate computing node i: This function explicitly couples physical connectivity with structure-aware compression efficiency. During inference execution, the system follows the hierarchical order of the neural network. Iteratively execute the following closed-loop logic: First, extract the structural cue vector from the global similarity matrix based on the index k of the currently data-holding node. and embedding real-time resources Construct the decision state at the current level. ; Second, the optimal computation node i and aggregation node j are selected through the strategy distribution generated by the scheduling agent. If i ≠ k, the intermediate features are compressed and transmitted using the adaptive compression factor in step 3. The resulting transmission delay is: If i=k, then the transmission delay is set to zero, and the load calculation is completed at node i. After processing, the inference task enters the state evolution phase, and the data holding node index of layer l+1... The decision probability distribution of the aggregation nodes is sampled and recursively updated: After each level of inference is completed, the system updates the heterogeneous graph in real time based on the latest positional changes of the drone swarm and the remaining resources. The topological connections and node characteristics are determined, and the similarity matrix calculation in step 2) is retried. Through this recursive execution and real-time dynamic alignment of states, this method can transform the static structural representation into a dynamic navigation map, enabling seamless task switching and knowledge collaboration among UAV nodes in non-stationary industrial environments. To ensure the real-time feasibility of the algorithm on resource-constrained UAV vehicle platforms, the overall computational overhead of the distributed reasoning process described in this method is... Represented as: in, The number of layers in a graph neural network. Let be the number of edges. For feature dimension, The hidden layer size; for the number of edges, the following conditions must be met. sparse drone networks, in which For sparse coefficients, the computational complexity is further simplified to This linear scalability ensures that the method has minimal computational overhead and real-time response capability in dynamic environments, supporting vehicle-mounted deployment and online execution in large-scale drone swarms.

Images