Unmanned aerial vehicle lightweight neural network group collaborative reasoning method based on maximum entropy topology

Abstract

The application discloses a UAV lightweight neural network group collaborative reasoning method based on a maximum entropy topology, and belongs to the technical field of UAV intelligent perception and edge artificial intelligence. The method respectively deploys lightweight neural network models obtained through distillation, pruning, quantization and adapter compression on multiple UAVs, abstracts each UAV and its onboard model as a collaborative node, and constructs a candidate collaborative graph according to task correlation, model complementarity, link quality, residual energy, time delay and confidence. A small-world network topology with local high clustering and short average path is generated on the candidate collaborative graph, and then the model collaborative edges are optimized through a maximum entropy objective function, so that the collaborative topology balances between real-time performance, robustness, communication overhead and model diversity. The application can realize efficient collaboration of multiple lightweight models under the condition that the single machine power, load and communication are limited, and can improve the comprehensive reasoning performance of a UAV cluster in a given task scenario.

Description

Technical Field

[0001] This invention relates to the fields of intelligent perception for unmanned aerial vehicles (UAVs), UAV swarm collaboration, edge artificial intelligence, and collaborative reasoning of neural network models, and in particular to a lightweight neural network swarm collaborative reasoning method for UAVs based on maximum entropy topology. Background Technology

[0002] With the development of unmanned aerial vehicle (UAV) platforms, airborne sensors, edge computing chips, and wireless communication technologies, UAVs have gradually evolved from traditional remote-controlled flight platforms into intelligent agents with capabilities such as environmental perception, target recognition, path planning, task decision-making, and swarm collaboration. In scenarios such as agricultural plant protection, forest fire prevention, power line inspection, disaster search and rescue, traffic monitoring, border patrol, and logistics delivery, UAVs need to process data from visible light cameras, infrared cameras, multispectral cameras, LiDAR, inertial navigation units, and positioning units in real time and make rapid decisions in complex environments.

[0003] Existing intelligent inference methods for drones typically employ two technical approaches. The first involves deploying a large-scale deep neural network model on a single drone to achieve high recognition accuracy and decision-making capabilities. The second involves uploading data collected by the drone to ground stations, edge servers, and cloud servers for inference. Both approaches have significant limitations in practical applications: large models require high computing power, memory, power consumption, and heat dissipation, making it difficult to operate stably for extended periods on small, lightweight drones; cloud and ground station inference relies on stable communication links, which are prone to excessive latency, link interruptions, and data transmission failures in highly interference-prone environments such as mountainous areas, disaster sites, and deep farmland.

[0004] In recent years, lightweight neural network models have reduced the number of model parameters and inference overhead through techniques such as pruning, quantization, distillation, low-rank adapters, and neural architecture search, making real-time inference possible on UAVs. However, individual lightweight models are usually limited by capacity, making it difficult to achieve the comprehensive performance of large models in complex environments, cross-view perception, long-tail target recognition, multi-task inference, and uncertainty handling.

[0005] Drone swarm collaboration offers a new approach to solving these problems. Multiple drones can observe the same task area from different spatial locations, perspectives, and sensor modalities. If the lightweight models carried by each drone can collaborate at the model level, rather than simply sharing raw images and geographic locations, then multiple lightweight models can form complementary reasoning capabilities: some models focus on object detection, some on semantic segmentation, and some on trajectory prediction and anomaly recognition. The local advantages of different models are integrated through collaborative fusion to form an overall advantage.

[0006] However, existing UAV swarm collaboration methods focus more on flight formation, task allocation, path planning, and sensor data sharing, and less on the collaborative relationships between onboard neural network models. Even when model collaboration exists, it often employs fixed topology, fully connected communication, and central node aggregation, which can easily lead to problems such as communication redundancy, excessive central node load, single point of failure, and uneven energy consumption. For UAV swarms facing communication constraints, energy limitations, and rapidly changing environments, a model collaborative topology optimization method that can simultaneously balance inference performance, communication efficiency, energy balance, and robustness is needed.

[0007] Small-world networks, characterized by high local clustering and short average path lengths, can improve global information propagation efficiency while maintaining local cooperative stability through a small number of long-range connections. The maximum entropy network approach, under constraints, can maintain topological diversity and uncertainty, preventing cooperative relationships from becoming overly concentrated in a few nodes, thereby enhancing the system's adaptability to node failures, link disturbances, and task changes. Therefore, a model collaborative reasoning method combining small-world topology and maximum entropy topology optimization is needed to improve the inference efficiency and robustness of UAV swarms under conditions of limited communication, energy consumption, and latency. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of the aforementioned background technology by proposing a lightweight neural network swarm collaborative inference method for unmanned aerial vehicles (UAVs) based on maximum entropy topology. This invention no longer interprets UAV collaboration solely as positional and task collaboration between flight platforms, but rather treats the lightweight neural network models carried by the UAVs as collaborative intelligent nodes. Multiple smaller models exchange information, perform complementary inference, and dynamically fuse through an optimized network topology. This achieves overall inference capabilities exceeding those of a single large model under limited computing power, energy, and communication conditions. It solves problems in existing technologies such as the difficulty of long-term real-time deployment of a single large model on UAVs, the limited expressive power of a single lightweight model, high communication overhead due to direct sharing of raw data between UAVs, the difficulty of adapting fixed collaborative topologies to dynamic environments, and the potential for single-point bottlenecks in centralized collaboration.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0010] A lightweight neural network swarm collaborative inference method for UAVs based on maximum entropy topology is characterized by being applied to a UAV swarm consisting of multiple UAVs, and includes the following steps:

[0011] Step S1: Obtain the mission objectives, environmental perception data, drone status data, and communication link data of the drone cluster;

[0012] Step S2: Deploy a lightweight neural network model on each UAV and generate a model state vector for the lightweight neural network model.

[0013] Step S3: Abstract the UAV and its onboard lightweight neural network model into cooperative nodes, and construct a candidate cooperative graph based on the spatial proximity, communication reachability and semantic correlation between cooperative nodes;

[0014] Step S4: Generate a small-world initial topology on the candidate collaboration graph, so that the collaboration nodes maintain local high clustering connection, and reduce the average path length of the collaboration topology through a portion of long-range collaboration edges.

[0015] Step S5: Optimize the cooperative edges in the initial topology of the small world based on the maximum entropy objective function to obtain the cooperative adjacency matrix;

[0016] Step S6: Each cooperating node exchanges compressed model messages according to the cooperative adjacency matrix;

[0017] Step S7: Dynamically weight and fuse the outputs from multiple lightweight neural network models to generate collaborative reasoning results for the drone swarm;

[0018] Step S8: Update the cooperative topology online based on changes in UAV position, link quality, mission stage, remaining energy, and model output uncertainty.

[0019] Furthermore, the lightweight neural network model in step S2 includes one of the following: object detection model, semantic segmentation model, instance segmentation model, object tracking model, scene recognition model, anomaly detection model, path planning model, behavior decision model, and multimodal perception model; the lightweight neural network model is obtained through knowledge distillation, structured pruning, unstructured pruning, low-bit quantization, weight sharing, low-rank adapter, bottleneck adapter, early departure branching, neural architecture search, and combinations thereof; the model state vector includes one of the following: task relevance, model confidence, output uncertainty, remaining energy, link quality, available bandwidth, and inference latency; the model state vector is represented as:

[0020] ,

[0021] in Indicates the first The task relevance of each model to the current task. This indicates the confidence level of the model output. This indicates output uncertainty. Indicates link quality. Indicates available bandwidth. Indicates the remaining energy. Indicates inference delay.

[0022] Furthermore, in step S4, the initial topology of the small world is generated in the following manner:

[0023] First, connect each collaborative node to the UAV based on its spatial proximity and mission area proximity. A local rule network is formed by identifying neighboring nodes; then, the reconnection probability is used to... Some local collaborative edges are replaced with remote collaborative edges, so that the generated collaborative topology satisfies the condition that the clustering coefficient is greater than the preset clustering threshold and the average path length is less than the preset path threshold.

[0024] Furthermore, in step S5, the maximum entropy objective function includes a topological entropy term, a cooperative utility term, a communication overhead term, an energy consumption term, and a delay term, expressed as:

[0025]

[0026]

[0027] in The probability matrix of the cooperative edge. For the first The first collaborative node and the first The probability that there is a collaborative edge between collaborative nodes, H(P) is the topological entropy. For synergistic effect, For communication overhead, For energy consumption expenses, To coordinate delay, , , , and The weighting coefficients; the topological entropy The Bernoulli entropy of the cooperative edge probability is expressed as:

[0028]

[0029] in To prevent the use of strange constants in logarithmic operations.

[0030] Furthermore, the cooperative edge probability Based on edge score The edge score was calculated. The collaborative edge probability is jointly determined by model output similarity, model functional complementarity, task relevance, link quality, historical collaborative gain, node trust, spatial distance, energy overhead, and latency overhead; Calculated using the softmax function: ,in For temperature coefficient; when When the value increases, the distribution of cooperative edges becomes more uniform. When the cooperative topology is reduced, it tends to favor high-scoring cooperative edges; the maximum entropy topology optimization is subject to one of the following constraints: maximum connectivity constraint, minimum connectivity constraint, total communication bandwidth constraint, single node energy consumption constraint, end-to-end latency constraint, link reachability constraint, mission safety area constraint, and UAV flight stability constraint.

[0031] Furthermore, in step S6, the compressed model message includes one of the following: intermediate semantic features, category probability, detection box, segmentation mask summary, trajectory state, confidence, uncertainty, and timestamp; the compressed model message is compressed through feature projection, low-rank mapping, hash encoding, scalar quantization, vector quantization, sparsification, and incremental transmission, and does not directly transmit the original image, original video, and complete sensor data.

[0032] Furthermore, in step S7, the weights of the dynamic weighted fusion are jointly determined by model confidence, output uncertainty, node trust, task relevance, model complementarity, and message timeliness; the dynamic weighted fusion includes feature-level fusion, probability-level fusion, decision-level fusion, and combinations thereof; when multiple model outputs conflict, the conflict results are corrected through consistency verification, confidence suppression, uncertainty penalty, and historical trajectory constraints; the collaborative reasoning results include one of the following: target category, target location, target bounding box, semantic segmentation result, instance segmentation result, target trajectory, anomaly alarm, path planning result, task allocation result, and cluster decision instruction.

[0033] Furthermore, in step S8, when the remaining energy of any UAV is lower than the energy threshold, the link quality is lower than the link threshold, the model uncertainty is higher than the uncertainty threshold, the task area changes, or a UAV node joins or leaves the cluster, a collaborative topology reconstruction is triggered. The UAV cluster includes agricultural inspection UAV clusters, forest fire prevention UAV clusters, disaster search and rescue UAV clusters, border patrol UAV clusters, logistics inspection UAV clusters, power line inspection UAV clusters, and marine monitoring UAV clusters. The UAV cluster equipment includes multiple UAVs, and each UAV includes a flight control unit, a sensor unit, an edge computing unit, a wireless communication unit, and a storage unit.

[0034] Furthermore, a computer device includes a processor and a memory; the processor executes a computer program stored in the memory to implement the steps of the lightweight neural network group collaborative inference method for unmanned aerial vehicles based on maximum entropy topology as described in any one of claims 1-8.

[0035] Furthermore, a computer-readable storage medium is provided for storing a computer program that, when executed by a processor, implements the method of any one of claims 1 to 8.

[0036] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0037] 1. The UAV lightweight neural network swarm collaborative reasoning method based on maximum entropy topology provided by this invention extends UAV collaboration from traditional platform-level collaboration to model-level collaboration, enabling UAVs to not only share location and mission information, but also share semantic reasoning results of onboard lightweight models, thereby improving the level of swarm intelligence.

[0038] 2. The lightweight neural network swarm collaborative inference method for UAVs based on maximum entropy topology provided by this invention adopts the collaborative approach of multiple lightweight models to replace the centralized deployment of a single large model on a single UAV, reducing the computing power, memory, power consumption and heat dissipation pressure of a single machine, and is more suitable for real-time tasks of small and light UAVs.

[0039] 3. The lightweight neural network group collaborative inference method for UAVs based on maximum entropy topology provided by this invention utilizes the local high clustering and short average path characteristics of small-world topology to enable model messages to propagate quickly with fewer communication connections, reducing the bandwidth waste caused by full-connection communication.

[0040] 4. The lightweight neural network group collaborative inference method for UAVs based on maximum entropy topology provided by this invention avoids excessive concentration of collaborative relationships in a few central nodes through maximum entropy topology optimization, which can improve the robustness of the system to node failures, link interruptions, local occlusion and model misjudgments.

[0041] 5. The lightweight neural network swarm collaborative inference method for UAVs based on maximum entropy topology provided by this invention reduces communication bandwidth usage by transmitting compressed model messages such as low-dimensional semantic messages, model confidence, and uncertainties, thereby avoiding direct transmission of original images and complete videos. It can dynamically update the collaborative topology based on the UAV's remaining energy, communication quality, mission stage, and model output uncertainty, thereby improving the adaptability of UAV swarms in complex dynamic scenarios.

[0042] 6. The lightweight neural network swarm collaborative reasoning method for UAVs based on maximum entropy topology provided by this invention enables multiple small models to form a synergistic enhancement effect through model complementarity, dynamic weighted fusion and uncertainty suppression mechanisms, thereby achieving a comprehensive reasoning effect that surpasses that of a single large model in multi-view, multi-task and cross-environment scenarios. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the overall process of collaborative inference for a lightweight UAV model swarm in an embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of the collaborative topology between UAV nodes and lightweight model nodes in an embodiment of the present invention;

[0045] Figure 3 This is a schematic diagram of the small-world network and maximum entropy topology optimization process in an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram illustrating compressed message exchange and collaborative fusion between lightweight models in an embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of topology reconstruction and model collaborative update in a dynamic environment in an embodiment of the present invention. Detailed Implementation

[0048] To make the technical problems to be solved, the technical solutions and the beneficial effects of the present invention clearer, the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0049] In one embodiment, the first The first collaborative node and the first Edge scoring between collaborative nodes Represented as:

[0050]

[0051]

[0052] in, This indicates the similarity between the outputs of two models and their semantic features. This indicates the complementarity of the two models in terms of function, perspective, and modality. Indicates task relevance. Indicates link quality. Indicates historical collaborative gain. Indicates spatial distance and communication distance. Indicates energy consumption expenditure. The coordinate latency is represented by w1 to w8, which are learnable and preset weights.

[0053] Side rating Converted to collaborative edge probability using the softmax function :

[0054]

[0055] in This is the temperature coefficient. When... When the probability is larger, the probability distribution of collaborative edges is more uniform, which is beneficial for topology exploration and robustness; when When the probability is smaller, the probability of collaborative edges is more concentrated on high-scoring edges, which helps to reduce communication overhead.

[0056] The objective function for maximum entropy topology optimization is expressed as:

[0057]

[0058]

[0059] in The topological entropy of the cooperative edge probability matrix. For synergistic effect, For communication overhead, For energy consumption expenses, To coordinate delay, , , , and These are the weighting coefficients.

[0060] Topological entropy It can be represented as:

[0061]

[0062] in To prevent singular constants in logarithmic operations, the topological entropy is maximized to prevent the collaborative relationships from being completely monopolized by a few high-confidence nodes and central nodes, while satisfying task and resource constraints. This enhances the drone swarm's resilience to node failures, link disturbances, and local model misjudgments.

[0063] The optimization process also satisfies the following constraints:

[0064]

[0065]

[0066]

[0067]

[0068]

[0069] in Adjacency matrix The elements in; when the first The first collaborative node and the first When the communication link between the cooperating nodes is unreachable Take 0.

[0070] Collaborative integration mechanism:

[0071] For the A lightweight model, the output of which can be expressed as:

[0072]

[0073] in For intermediate semantic features, The data includes category probabilities, detection results, and segmentation results. For confidence level, The nature of this is uncertain. Compressed messages sent by cooperating nodes to their neighbors. It can be represented as:

[0074]

[0075] in, Represents the compression function. Represents the feature projection function. Represents a timestamp.

[0076] During the fusion phase, the system calculates weights for each model participating in the fusion. :

[0077]

[0078]

[0079] in Indicates the node's trust level. Indicates the complementary contribution of the model. This indicates message latency. The final output can be obtained using probabilistic fusion. Feature-level fusion can also be used. Furthermore, decision-level voting, Bayesian fusion, and attention fusion can also be employed.

[0080] When multiple model outputs conflict, the system performs conflict correction based on output consistency, historical trajectory continuity, spatial geometric constraints, and the magnitude of uncertainty. For example, in a target tracking task, if the target position output by a certain model deviates too much from the historical trajectory and has high uncertainty, the fusion weight of that model at the current time step is reduced.

[0081] Example 1: Collaborative Inference Method for Lightweight UAV Model Swarm

[0082] refer to Figure 1 This embodiment provides a lightweight neural network swarm collaborative inference method for drone swarms. The drone swarm consists of n drones, each equipped with a sensor and an edge computing unit. The edge computing unit can be an embedded GPU, NPU, FPGA, AI acceleration chip, or low-power CPU.

[0083] At the start of a mission, the ground station or the main control module of the UAV swarm issues the mission objectives. For example, in an agricultural inspection mission, the objective is to identify crop diseased areas, count the area affected by pests and diseases, and generate a prescription map; in a disaster search and rescue mission, the objective is to identify personnel, vehicles, fire points, smoke, and obstacles; and in a power line inspection mission, the objective is to identify insulator defects, foreign objects on conductors, and abnormalities in towers.

[0084] Each UAV collects local data based on its own sensors, and its own lightweight model independently completes preliminary inference. Due to differences in viewpoint, distance, altitude, lighting, and sensing modes among the UAVs, the outputs of different lightweight models are complementary. The system calculates the confidence level and uncertainty of each model output and combines this with the UAV's state to form a model state vector.

[0085] refer to Figure 2 The system abstracts each drone and its lightweight model as a collaborative node. If a drone carries multiple models, such as a detection model, a segmentation model, and a tracking model, then each model can be treated as a separate node, or the drone can be considered a composite node containing multiple model child nodes. Candidate collaborative edges are determined by both communication reachability and task relevance.

[0086] refer to Figure 3 The system first constructs an initial small-world topology. For each cooperating node, the system selects several spatially neighboring nodes as local neighbors to ensure stable information exchange within the local area; then, based on model complementarity and link quality, it selects a small number of remote nodes as long-range cooperating objects, enabling information from distant areas to quickly propagate to the entire model group.

[0087] Based on the initial small-world topology, the system further performs maximum entropy topology optimization. Maximum entropy topology optimization does not simply pursue the maximum number of connections, but rather, under constraints of communication, energy consumption, and latency, maintains a moderately dispersed probability of cooperative edges, avoiding all nodes connecting to the same high-performance node and high-power node. This prevents overload of the central node and also maintains cluster inference capabilities in the event of central node failure.

[0088] refer to Figure 4 Each coordinating node only sends compressed model messages. Taking object detection as an example, the compressed model message includes the bounding box coordinates, class probability, detection confidence, low-dimensional feature vector, and timestamp; taking semantic segmentation as an example, the compressed model message includes the mask contour, region statistics, class probability, and low-dimensional semantic features. Compared with directly transmitting the original image, this method significantly reduces communication bandwidth consumption.

[0089] The receiving node receives neighbor messages based on the collaborative adjacency matrix and performs dynamic fusion. If a node has high model confidence, low output uncertainty, high task relevance, and low link latency, then the node's output receives a larger fusion weight; if a node has low battery, unstable link, and high output uncertainty, then the node's output receives a smaller fusion weight.

[0090] refer to Figure 5 The system continuously monitors the drone's status and model output during task execution. When a drone enters an obstructed area, moves far from the task area, experiences a decline in link quality, or has insufficient remaining power, the system reduces its cooperative edge probability and selects new cooperative neighbors. When a new drone joins the cluster, the system incorporates it into the cooperative topology based on its model capabilities and link status.

[0091] Example 2: Agricultural Plant Protection Inspection Scenario

[0092] In one specific embodiment, the present invention is applied to an agricultural plant protection inspection drone swarm. The drone swarm includes multiple drones equipped with visible light cameras and multispectral cameras. Some drones are equipped with lightweight lesion detection models, some with lightweight crop canopy segmentation models, some with lightweight growth assessment models, and some with lightweight path planning models.

[0093] The task area is divided into multiple sub-regions, with each UAV acquiring local images and performing local inference. Since pest and disease areas can be affected by lighting, leaf shading, and shooting angle, individual models are prone to missed detections and false detections. This invention utilizes a small-world cooperative topology, enabling UAVs in adjacent sub-regions to preferentially exchange candidate lesion regions, confidence levels, and uncertainty information. Furthermore, a small number of long-range cooperative edges connect key UAVs in different plots, allowing local detection results to be fused with cross-regional observation information.

[0094] When a suspected high-incidence area of ​​pests and diseases is identified in a certain plot of land, the uncertainty of the drone model output in that area increases. The system automatically increases the cooperative side probability with neighboring drones and multispectral model drones, enabling the detection model and the multispectral growth model to make complementary judgments. The fusion results are used to generate a spatial distribution map of pests and diseases and plant protection prescription suggestions.

[0095] In this embodiment, the present invention can complete cross-UAV model collaboration without transmitting all original aerial images, reducing the communication pressure of large-scale farmland inspections and improving the stability of identifying small lesions, occluded lesions, and boundary lesions.

[0096] Example 3: Disaster Search and Rescue Scenario

[0097] In another specific embodiment, the present invention is applied to a swarm of disaster search and rescue drones. Disaster sites typically present problems such as smoke, obstructions, unstable communication, and complex terrain, making it difficult for a single drone to obtain reliable judgments from a complete perspective.

[0098] In this embodiment, different drones are equipped with lightweight personnel detection models, vehicle detection models, fire point recognition models, smoke segmentation models, and passable area recognition models, respectively. After each model completes inference locally, they only exchange target candidate boxes, region semantic summaries, confidence scores, and uncertainties. The system uses maximum entropy topology optimization to enable drones at different altitudes, from different perspectives, and with different sensing modes to form a complementary and cooperative relationship.

[0099] When the uncertainty of the personnel detection model increases due to smoke obscuring a drone, the system increases the probability of its cooperative edge with infrared drones and other drones from other perspectives; when the link quality of a drone deteriorates, the system temporarily removes it from the critical cooperative path and maintains information propagation efficiency through small-world long-range edges.

[0100] In this embodiment, the present invention can improve the stability and real-time performance of search and rescue target identification under the condition of limited communication at disaster sites, and reduce the risk of missed detection caused by misjudgment by a single model.

[0101] Example 4: Model Training and Deployment Methods

[0102] In one embodiment, a high-precision teacher model is first trained on a cloud or ground server. The teacher model can be a large convolutional neural network, a visual Transformer, a multimodal base model, or a joint model containing multiple task heads. Then, multiple lightweight student models are trained to address the resource limitations of the drone.

[0103] The training loss of the student model can include task loss, distillation loss, and feature alignment loss, expressed as: ,in To detect, segment, and classify task losses, The distillation loss between the output distributions of the teacher model and the student model. For intermediate feature alignment loss, Calibrate loss for uncertainty.

[0104] Different student models can be trained specifically for different task subspaces. For example, model M1 is specifically for recognizing small targets, model M2 is specifically for recognizing occluded targets, model M3 is specifically for processing infrared images, model M4 is specifically for processing multispectral images, and model M5 is specifically for performing path planning. In this way, the outputs of different models are highly complementary during the collaborative inference phase of the UAV swarm.

[0105] During deployment, the system assigns one or more lightweight models to each drone based on its computing power, mission area, sensor type, and remaining battery power. For drones with strong computing power, multiple models can be deployed as composite collaborative nodes; for drones with weak computing power, only a single dedicated model can be deployed.

[0106] Example 5: System Structure

[0107] This invention also provides a lightweight neural network swarm collaborative inference system for unmanned aerial vehicles based on small-world networks and maximum entropy topology optimization, including a task perception module, a model deployment module, a candidate graph construction module, a small-world topology generation module, a maximum entropy topology optimization module, a model message exchange module, a collaborative fusion inference module, and a topology online update module.

[0108] The task awareness module is used to acquire task objectives, task area, environmental awareness data, and UAV status data. The model deployment module is used to load lightweight neural network models on the UAV and record model type, model size, inference speed, task adaptability, and historical performance.

[0109] The candidate graph construction module is used to construct candidate collaborative graphs based on UAV communication reachability, spatial distance, sensor viewpoint, task overlap, and model complementarity. The small-world topology generation module is used to generate an initial topology with locally high clustering and short average paths from the candidate collaborative graphs.

[0110] The maximum entropy topology optimization module determines the probability of cooperative edges based on topology entropy, cooperative utility, communication overhead, energy consumption overhead, and latency overhead, and outputs a cooperative adjacency matrix. The model message exchange module performs compressed message transmission according to the cooperative adjacency matrix. The cooperative fusion inference module dynamically fuses the outputs of multiple models. The online topology update module triggers topology reconstruction based on changes in node state.

[0111] The aforementioned modules can be deployed in the edge computing units of each drone, or partially deployed in temporary coordinating drones and ground stations within the cluster. When the system adopts a decentralized deployment, each drone can independently calculate its own adjacency relationships and coordinate with neighboring nodes; when the system adopts a weakly centralized deployment, the coordination topology can be periodically distributed by temporary coordinating nodes, but other nodes can still take over after the coordinating node fails.

[0112] Example 6: Parameter Setting and Experimental Verification Method

[0113] To verify the effectiveness of this invention, the following experimental method can be used. Set the number of drones... For 4, 8, 16, and 32, each UAV is equipped with a different type of lightweight model, and the following schemes are compared respectively: single UAV large model scheme, single UAV lightweight model scheme, multi-UAV fully connected collaborative scheme, multi-UAV fixed neighborhood collaborative scheme, small-world topology collaborative scheme only, and the small-world network plus maximum entropy topology optimization collaborative scheme of the present invention.

[0114] Evaluation metrics include task accuracy, mean accuracy, segmentation intersection-union ratio, target tracking success rate, end-to-end inference latency, average communication bandwidth, energy consumption per unit task, performance retention rate after node failure, and topology reconstruction time. For agricultural applications, lesion recognition accuracy and prescription accuracy can be further evaluated. Figure 1 Consistency and false negative rate; if applied to disaster search and rescue scenarios, personnel detection rate, false alarm rate and response time can be further evaluated.

[0115] In the experiment, the maximum entropy weight It can be used to control topology diversity and communication overhead weights. It can be used to control bandwidth usage and energy consumption weight. It can be used to extend the flight time of drones, and to reduce latency. This can be used to ensure real-time performance. The above weights can be dynamically adjusted based on the urgency of the task and communication conditions.

[0116] It should be noted that the effect of surpassing a single large model mentioned in this invention refers to the fact that, under a given UAV swarm, multi-view observation and mission scenario, multiple lightweight models can surpass the large model deployed on a single UAV or a single node in terms of comprehensive mission indicators through complementary observation, collaborative fusion and dynamic topology optimization, rather than being limited to necessarily surpassing all large models in any scenario.

[0117] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein, without departing from the essence of the technical solutions of the present invention, and such modifications or substitutions should still fall within the protection scope of the present invention.

Claims

1. A lightweight neural network swarm collaborative inference method for UAVs based on maximum entropy topology, characterized in that, The method is applied to a drone swarm consisting of multiple drones and includes the following steps: Step S1: Obtain the mission objectives, environmental perception data, drone status data, and communication link data of the drone cluster; Step S2: Deploy lightweight neural network models on each UAV and generate model state vectors for the lightweight neural network models. Step S3: Abstract the UAV and its onboard lightweight neural network model into cooperative nodes, and construct a candidate cooperative graph based on the spatial proximity, communication reachability and semantic correlation between cooperative nodes; Step S4: Generate a small-world initial topology on the candidate collaboration graph, so that the collaboration nodes maintain local high clustering connection, and reduce the average path length of the collaboration topology through a portion of long-range collaboration edges. Step S5: Optimize the cooperative edges in the initial topology of the small world based on the maximum entropy objective function to obtain the cooperative adjacency matrix; Step S6: Each cooperating node exchanges compressed model messages according to the cooperative adjacency matrix; Step S7: Dynamically weight and fuse the outputs from multiple lightweight neural network models to generate collaborative reasoning results for the drone swarm; Step S8: Update the cooperative topology online based on changes in UAV position, link quality, mission stage, remaining energy, and model output uncertainty.

2. The lightweight neural network swarm collaborative inference method for UAVs based on maximum entropy topology as described in claim 1, characterized in that, The lightweight neural network model in step S2 includes one of the following: object detection model, semantic segmentation model, instance segmentation model, object tracking model, scene recognition model, anomaly detection model, path planning model, behavior decision model, and multimodal perception model. The lightweight neural network model is obtained through knowledge distillation, structured pruning, unstructured pruning, low-bit quantization, weight sharing, low-rank adapter, bottleneck adapter, early departure branching, neural architecture search, and combinations thereof. The model state vector includes one of the following: task relevance, model confidence, output uncertainty, remaining energy, link quality, available bandwidth, and inference latency. The model state vector is represented as follows: ； in This represents the task relevance of the i-th model to the current task. This indicates the confidence level of the model output. This indicates output uncertainty. Indicates link quality. Indicates available bandwidth. Indicates the remaining energy. Indicates inference delay.

3. The lightweight neural network swarm collaborative inference method for UAVs based on maximum entropy topology as described in claim 1, characterized in that, In step S4, the initial topology of the small world is generated in the following manner: First, connect each collaborative node to the UAV based on its spatial proximity and mission area proximity. A local rule network is formed by identifying neighboring nodes; then, the reconnection probability is used to... Some local collaborative edges are replaced with remote collaborative edges, so that the generated collaborative topology satisfies the condition that the clustering coefficient is greater than the preset clustering threshold and the average path length is less than the preset path threshold.

4. The lightweight neural network swarm collaborative inference method for UAVs based on maximum entropy topology according to claim 1, characterized in that, In step S5, the maximum entropy objective function includes a topological entropy term, a cooperative utility term, a communication overhead term, an energy consumption term, and a delay term, expressed as: ； ； in The probability matrix of the cooperative edge. For the first The first collaborative node and the first The probability that there is a collaborative edge between collaborative nodes, H(P) is the topological entropy. For synergistic effect, For communication overhead, For energy consumption expenses, To coordinate delay, , , , and The weighting coefficients; the topological entropy The Bernoulli entropy of the cooperative edge probability is expressed as: ； in To prevent the use of strange constants in logarithmic operations.

5. The lightweight neural network swarm collaborative inference method for UAVs based on maximum entropy topology according to claim 1, characterized in that, The probability of the cooperative edge Based on edge score The edge score was calculated. The collaborative edge probability is jointly determined by model output similarity, model functional complementarity, task relevance, link quality, historical collaborative gain, node trust, spatial distance, energy overhead, and latency overhead; Calculated using the softmax function: ,in For temperature coefficient; when When the value increases, the distribution of cooperative edges becomes more uniform. When the cooperative topology is reduced, it tends to favor high-scoring cooperative edges; the maximum entropy topology optimization is subject to one of the following constraints: maximum connectivity constraint, minimum connectivity constraint, total communication bandwidth constraint, single node energy consumption constraint, end-to-end latency constraint, link reachability constraint, mission safety area constraint, and UAV flight stability constraint.

6. The lightweight neural network swarm collaborative inference method for UAVs based on maximum entropy topology according to claim 1, characterized in that, In step S6, the compressed model message includes one of the following: intermediate semantic features, category probability, detection box, segmentation mask summary, trajectory state, confidence, uncertainty, and timestamp; the compressed model message is compressed through feature projection, low-rank mapping, hash encoding, scalar quantization, vector quantization, sparsification, and incremental transmission, and does not directly transmit the original image, original video, and complete sensor data.

7. The lightweight neural network swarm collaborative inference method for UAVs based on maximum entropy topology according to claim 1, characterized in that, In step S7, the weights of the dynamic weighted fusion are jointly determined by model confidence, output uncertainty, node trust, task relevance, model complementarity, and message timeliness. The dynamic weighted fusion includes feature-level fusion, probability-level fusion, decision-level fusion, and combinations thereof. When multiple model outputs conflict, the conflict results are corrected through consistency verification, confidence suppression, uncertainty penalty, and historical trajectory constraints. The collaborative reasoning results include one of the following: target category, target location, target bounding box, semantic segmentation result, instance segmentation result, target trajectory, anomaly alarm, path planning result, task allocation result, and cluster decision instruction.

8. The lightweight neural network swarm collaborative inference method for UAVs based on maximum entropy topology according to claim 1, characterized in that, In step S8, a collaborative topology reconstruction is triggered when any UAV's remaining energy is lower than the energy threshold, the link quality is lower than the link threshold, the model uncertainty is higher than the uncertainty threshold, the task area changes, or a UAV node joins or leaves the cluster. The UAV cluster includes agricultural inspection UAV clusters, forest fire prevention UAV clusters, disaster search and rescue UAV clusters, border patrol UAV clusters, logistics inspection UAV clusters, power line inspection UAV clusters, and marine monitoring UAV clusters. The UAV cluster equipment includes multiple UAVs, and each UAV includes a flight control unit, a sensor unit, an edge computing unit, a wireless communication unit, and a storage unit.

9. A computer device, comprising a processor and a memory; characterized in that, When the processor executes the computer program stored in the memory, it implements the steps of the lightweight neural network group collaborative inference method for unmanned aerial vehicles based on maximum entropy topology as described in any one of claims 1 to 8.

10. A computer-readable storage medium for storing a computer program; characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 8.

Images