Combined unmanned aerial vehicle deployment and model division method and system based on collaborative reasoning

Abstract

The embodiment of the invention provides a cooperative reasoning-based joint unmanned aerial vehicle deployment and model division method and system. The method comprises the steps of determining an unmanned aerial vehicle group and a target monitoring area; wherein the unmanned aerial vehicle group comprises a plurality of unmanned aerial vehicles, and the plurality of unmanned aerial vehicles are used for image capture and data classification requests to detect abnormal conditions of a target area; establishing an unmanned aerial vehicle fleet system based on the plurality of unmanned aerial vehicles and the target monitoring area; task allocation is carried out on the multiple unmanned aerial vehicles through the unmanned aerial vehicle cluster system, and target tasks corresponding to the multiple unmanned aerial vehicles are obtained; and based on the target task, performing deployment position optimization and model layer division on the plurality of unmanned aerial vehicles to obtain model layers corresponding to the plurality of unmanned aerial vehicles and optimal deployment positions corresponding to the plurality of unmanned aerial vehicles, so that the plurality of unmanned aerial vehicles execute the target task. According to the scheme, the task completion efficiency of the unmanned aerial vehicle group can be improved.

Description

Technical Field

[0001] This application relates to the fields of UAV cooperative control and artificial intelligence reasoning technology, specifically to a method and system for joint UAV deployment and model partitioning based on cooperative reasoning. Background Technology

[0002] With the widespread application of drone technology in environmental monitoring, disaster relief, and security patrols, drone swarms, with their significant advantages such as high maneuverability, wide field of view coverage, low deployment costs, and the ability to avoid personnel casualties, are being deployed on a large scale in complex and dynamic monitoring areas, requiring real-time execution of intelligent inference tasks such as image recognition and target detection. However, drone equipment is limited by onboard computing power, storage capacity, and energy supply, making it impossible to directly deploy large-scale deep learning models. Existing methods, such as lightweight model design, reduce inference accuracy, while relying on remote edge or cloud servers for task offloading is difficult to meet real-time requirements due to bandwidth limitations and communication latency, failing to fully leverage the inherent rapid response and collaborative operation potential of drone swarms. Furthermore, existing work mainly focuses on inference performance indicators (latency, resource utilization, accuracy), neglecting the fundamental monitoring requirement of maintaining sufficient coverage of the target ground area throughout the process. In practical applications, continuous and stable area coverage is a basic prerequisite for drone swarms to complete their tasks. Even with high inference efficiency, insufficient coverage may lead to the loss of critical information and mission failure. Summary of the Invention

[0003] This application aims to provide a method and system for joint drone deployment and model partitioning based on collaborative reasoning, which can improve the efficiency of drone swarms in completing tasks.

[0004] The technical solution of this application is implemented as follows: In a first aspect, embodiments of this application provide a method for joint UAV deployment and model partitioning based on collaborative reasoning, the method comprising: Identify the drone swarm and target monitoring area; wherein the drone swarm includes multiple drones, which are used for image capture and data classification requests to detect anomalies in the target area; Based on the multiple drones and the target monitoring area, a drone swarm system is established; and the drone swarm system is used to assign tasks to the multiple drones to obtain the target tasks corresponding to each of the multiple drones. Based on the target task, the deployment locations of the multiple drones are optimized and the model layers are divided to obtain the model corresponding to each of the multiple drones and the optimal deployment location corresponding to each of the multiple drones, so that the multiple drones can perform the target task.

[0005] In the above scheme, the step of establishing a drone swarm system based on the multiple drones and the target monitoring area, and allocating tasks to the multiple drones through the drone swarm system to obtain the target tasks corresponding to each of the multiple drones, includes: Based on the multiple drones and the target monitoring area, a communication connection is established to determine the drone swarm system; wherein, any two drones in the drone swarm system can communicate, and their monitoring range is the target monitoring area; The target monitoring area is divided into multiple sub-target monitoring areas; the number of the multiple sub-target monitoring areas is the same as the number of the multiple drones. Based on the multiple sub-target monitoring areas, the drone swarm system allocates tasks to the multiple drones to obtain the target tasks corresponding to each of the multiple drones.

[0006] In the above scheme, the step of optimizing the deployment locations and dividing the model layers of the multiple drones based on the target task to obtain the model layers corresponding to each of the multiple drones and the optimal deployment locations corresponding to each of the multiple drones includes: Assign deep learning models to the multiple drones and determine the model layer corresponding to each of the multiple drones; Based on the target task, the deployment locations of the multiple drones are optimized to obtain the optimal deployment location for each of the multiple drones.

[0007] In the above scheme, assigning deep learning models to the multiple drones and determining the model layer corresponding to each of the multiple drones includes: Determine the constraints; wherein the constraints include a first constraint, a second constraint, a third constraint, a fourth constraint, and a fifth constraint. Based on the first constraint, the second constraint, the third constraint, the fourth constraint, and the fifth constraint, multiple model layers in the deep learning model are respectively assigned to the multiple drones to obtain the model layers corresponding to each of the multiple drones.

[0008] In the above scheme, The first constraint is that the model layer allocated to any one of the plurality of drones does not exceed the maximum memory capacity of the drone. The second constraint is that the sum of the computing energy consumption, data transmission energy consumption, and hovering energy consumption of the plurality of UAVs is less than a preset energy consumption. The third constraint is that each of the multiple model layers in the deep learning model can only be deployed and executed on one drone. The fourth constraint is that the allocation decision variable in the deep learning model is a binary decision variable, which can only take the value 0 or 1. The fifth constraint is that the coverage area of ​​the monitoring area deployed by any one of the plurality of UAVs is greater than the preset coverage area.

[0009] In the above scheme, optimizing the deployment locations of the multiple drones based on the target task to obtain the optimal deployment location for each of the multiple drones includes: Determine the initial position information of the plurality of drones; wherein, the initial position information includes the horizontal coordinates and fixed altitude of the drones; The initial position information is continuously optimized by a successive convex approximation algorithm to obtain the updated position information of the multiple UAVs; The latency is calculated based on the updated location information of the multiple drones; If the latency is lower than a preset threshold, the optimal deployment location for each of the multiple drones is determined.

[0010] In the above scheme, each of the multiple drones simultaneously serves as a data generation node and a computing node; when each drone serves as a data generation node, it captures images and generates inference requests; when each drone serves as a computing node, it performs a portion of the inference task; the specific function of each drone is determined according to the network configuration and task allocation strategy.

[0011] Secondly, embodiments of this application provide a joint UAV deployment and model partitioning system based on collaborative reasoning. The system includes: a determination module, an allocation module, and an optimization partitioning module. The determining module is used to determine the drone swarm and the target monitoring area; wherein, the drone swarm includes multiple drones, which are used for image capture and data classification requests to detect abnormal conditions in the target area; The allocation module is used to establish a drone swarm system based on the multiple drones and the target monitoring area; and to allocate tasks to the multiple drones through the drone swarm system to obtain the target tasks corresponding to each of the multiple drones. The optimization partitioning module is used to optimize the deployment positions and partition the model layers of the multiple drones based on the target task, so as to obtain the model layers corresponding to each of the multiple drones and the optimal deployment positions corresponding to each of the multiple drones, so that the multiple drones can perform the target task.

[0012] Thirdly, embodiments of this application provide a joint UAV deployment and model partitioning device based on collaborative reasoning, comprising: a processor and a memory; wherein, The memory is used to store computer programs; The processor is configured to call and run the computer program from the memory to perform the method as described in the first aspect.

[0013] Fourthly, embodiments of this application provide a computer-readable storage medium storing executable instructions for causing a processor to perform the method described in the first aspect.

[0014] This application provides a method and system for joint UAV deployment and model partitioning based on collaborative reasoning. The method includes: determining a UAV swarm and a target monitoring area; wherein the UAV swarm includes multiple UAVs used for image capture and data classification requests to detect anomalies in the target area; establishing a UAV swarm system based on the multiple UAVs and the target monitoring area; allocating tasks to the multiple UAVs through the UAV swarm system to obtain the target tasks corresponding to each UAV; and optimizing the deployment locations and partitioning the model layers of the multiple UAVs based on the target tasks to obtain the model layers and optimal deployment locations corresponding to each UAV, so that the multiple UAVs can execute the target tasks. In the above scheme, while ensuring the regional coverage, energy constraints, and resource limitations of the distributed UAV swarm, end-to-end inference latency is minimized. The optimization algorithm proposed in this application is highly efficient and converges quickly, enabling the fastest solution for joint UAV deployment and model partitioning, ensuring full coverage of the monitoring area, avoiding blind spots, and improving the task completion efficiency and task response speed of the UAV swarm. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the specification, serve to explain the technical solutions of this application. Obviously, the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0016] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0017] Figure 1This is an optional flowchart illustrating a collaborative reasoning-based joint UAV deployment and model partitioning method provided in an embodiment of this application. Figure 2 A system architecture diagram of a collaborative UAV deployment and model partitioning method based on collaborative reasoning provided in this application embodiment; Figure 3 A schematic diagram of latency optimization curves for a collaborative reasoning-based joint UAV deployment and model partitioning method provided in an embodiment of this application; Figure 4 A schematic diagram of the coverage optimization curve for a collaborative reasoning-based joint UAV deployment and model partitioning method provided in an embodiment of this application; Figure 5 A schematic diagram of the trust domain change curve for a collaborative reasoning-based joint UAV deployment and model partitioning method provided in an embodiment of this application; Figure 6 This application provides an embodiment of an optimized UAV deployment diagram based on a collaborative reasoning-based joint UAV deployment and model partitioning method. Figure 7 A model partitioning result diagram for a collaborative reasoning-based joint UAV deployment and model partitioning method provided in an embodiment of this application; Figure 8 A schematic diagram illustrating the UAV quantity sensitivity analysis of a collaborative reasoning-based joint UAV deployment and model partitioning method provided in this application embodiment; Figure 9 A schematic diagram illustrating the model size sensitivity analysis of a collaborative reasoning-based joint UAV deployment and model partitioning method provided in this application embodiment; Figure 10 A schematic diagram of a joint UAV deployment and model partitioning system based on collaborative reasoning is provided in an embodiment of this application; Figure 11 This is a schematic diagram of a collaborative reasoning-based joint UAV deployment and model partitioning device provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the specific technical solutions of this application will be further described in detail below with reference to the accompanying drawings of the embodiments of this application. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0019] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this application is for the purpose of describing embodiments of this application only and is not intended to be limiting of this application.

[0020] In the following description, references to "some embodiments," "this embodiment," "this application embodiment," and examples, etc., describe a subset of all possible embodiments. However, it is understood that "some embodiments" may be the same subset or different subset of all possible embodiments and may be combined with each other without conflict.

[0021] If the application documents contain similar descriptions such as "first / second", the following explanation shall be added: In the following description, the terms "first / second / third" are used only to distinguish similar objects and do not represent a specific order of objects. It is understood that "first / second / third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0022] This application provides a method for joint UAV deployment and model partitioning based on collaborative reasoning. Figure 1 This is an optional flowchart illustrating a collaborative reasoning-based joint UAV deployment and model partitioning method provided in an embodiment of this application, which will combine... Figure 1 The steps shown are explained.

[0023] S101. Determine the drone swarm and target monitoring area; wherein, the drone swarm includes multiple drones, which are used for image capture and data classification requests to detect anomalies in the target area.

[0024] In some embodiments of this application, each of the multiple drones simultaneously serves as a data generation node and a computing node; when each drone serves as a data generation node, each drone captures images to generate inference requests; when each drone serves as a computing node, each drone performs a portion of the inference task; the specific function of each drone is determined according to the network configuration and task allocation strategy.

[0025] In some embodiments of this application, the joint UAV deployment and model partitioning method based on collaborative reasoning is adapted to UAV monitoring scenarios.

[0026] In some embodiments of this application, the collaborative reasoning-based joint UAV deployment and model partitioning method is adapted to a collaborative reasoning-based joint UAV deployment and model partitioning system.

[0027] In some embodiments of this application, the target monitoring area can be a certain area in a city; multiple drones are selected from drones monitoring the entire city as a drone swarm, which is used to obtain image capture and data classification requests from the target monitoring area.

[0028] S102. Based on multiple drones and target monitoring areas, establish a drone swarm system; and use the drone swarm system to assign tasks to multiple drones to obtain the target tasks corresponding to each drone.

[0029] In some embodiments of this application, a communication connection is established based on multiple drones and a target monitoring area to determine a drone swarm system; wherein any two drones in the drone swarm system can communicate, and the monitoring range is the target monitoring area; the target monitoring area is divided to obtain multiple sub-target monitoring areas; the number of multiple sub-target monitoring areas is consistent with the number of multiple drones; based on the multiple sub-target monitoring areas, the drone swarm system is used to assign tasks to the multiple drones to obtain the target tasks corresponding to each of the multiple drones.

[0030] In some embodiments of this application, there is a one-to-one correspondence between the number of target tasks and multiple sub-target monitoring areas, that is, one drone corresponds to one target task, and one target task corresponds to one sub-target monitoring area.

[0031] S103. Based on the target task, optimize the deployment location and divide the model layer for multiple drones to obtain the model layer corresponding to each drone and the optimal deployment location corresponding to each drone, so that the multiple drones can perform the target task.

[0032] In some embodiments of this application, deep learning models are assigned to multiple drones to determine the model layer corresponding to each drone; based on the target task, the deployment positions of the multiple drones are optimized to obtain the optimal deployment positions corresponding to each drone.

[0033] In some embodiments of this application, initial position information of multiple drones, including the horizontal coordinates and fixed altitude of the drones, is determined; continuous position optimization is performed on the initial position information to obtain updated position information of the multiple drones; based on the updated position information of the multiple drones, the latency is calculated; and if the latency is lower than a preset threshold, the optimal deployment position corresponding to each of the multiple drones is determined.

[0034] Understandably, this approach minimizes end-to-end inference latency while ensuring regional coverage, energy constraints, and resource limitations for the distributed drone swarm. The optimization algorithm proposed in this application is highly efficient and converges quickly, enabling the fastest solution for joint drone deployment and model partitioning schemes. It guarantees full coverage of the monitoring area, avoids blind spots, and improves the drone swarm's task completion efficiency and response speed.

[0035] In some embodiments of this application, S103 can be implemented by S201-S204, as follows: S201. Assign deep learning models to multiple drones and determine the model layer corresponding to each drone.

[0036] In some embodiments of this application, constraints are determined; wherein the constraints include a first constraint, a second constraint, a third constraint, a fourth constraint, and a fifth constraint; based on the first constraint, the second constraint, the third constraint, the fourth constraint, and the fifth constraint, multiple model layers in the deep learning model are assigned to multiple drones respectively, thereby obtaining the model layers corresponding to each of the multiple drones.

[0037] In some embodiments of this application, the first constraint is that the model layer allocated to any one of the multiple drones does not exceed the maximum memory capacity of the drone; the second constraint is that the sum of the computing energy consumption, data transmission energy consumption, and hovering energy consumption of the multiple drones is less than a preset energy consumption; the third constraint is that each model layer in the multiple model layers of the deep learning model can only be deployed and executed on one drone; the fourth constraint is that the allocation decision variable in the deep learning model is a binary decision variable, with values ​​of only 0 or 1; and the fifth constraint is that the coverage area of ​​the monitoring area deployed by any one of the multiple drones is greater than a preset coverage area.

[0038] For example, the first constraint is: ensure that no UAV exceeds its memory capacity when allocating model layers to UAVs.

[0039]

[0040] in, For each UAV n ∈ N Maximum storage capacity; For memory requirements; This indicates the model allocation strategy.

[0041] Second constraint: Limit the energy consumption of UAVs during collaborative reasoning.

[0042]

[0043] in, To calculate energy consumption; Energy consumption for data transmission; Energy consumption during hovering; This refers to available energy.

[0044] Third constraint: Limits each model layer It can only be deployed and executed on a single UAV.

[0045]

[0046] in, This means that each model layer must be assigned to one and only one UAV node.

[0047] Fourth constraint: Ensure that the model's allocation decision variables are binary decision variables, taking only 0 or 1 values.

[0048]

[0049] Among them, if UAV n Execution Model Layer ,but ,otherwise .

[0050] Fifth constraint: Specify the coverage requirements of the UAV deployment for the monitoring area. Indicates the lower limit of the coverage area.

[0051]

[0052] in, To cover the target; This is the coverage threshold.

[0053] S202. Based on the target task, optimize the deployment locations of multiple drones to obtain the optimal deployment locations for each drone.

[0054] In some embodiments of this application, initial position information of multiple drones is determined; wherein, the initial position information includes the horizontal coordinates and fixed altitude of the drones; the initial position information is continuously optimized by a successive convex approximation algorithm to obtain the updated position information of the multiple drones; based on the updated position information of the multiple drones, the time delay is calculated; if the time delay is lower than a preset threshold, the optimal deployment position corresponding to each of the multiple drones is determined.

[0055] For example, this application proposes a two-stage iterative optimization framework, AirInfer, which decomposes the initial problem (joint model partitioning and UAV deployment optimization problem) into two efficiently solvable sub-problems for iterative solving. The steps of the scheme are as follows: Step 1: Initialize parameters, mainly including establishing the system model, setting the objective function, and setting constraints. Establish the system model. Consider a group of unmanned aerial vehicles (UAVs) capable of forming a wireless ad hoc network.

[0056] The network is composed of It consists of several drones, represented as These drones are responsible for detecting anomalies in target areas (e.g., monitoring forest fires or disaster relief sites) through image capture and data classification requests. The drones can function simultaneously as data generation nodes and computing units, meaning a single drone can both capture images to generate inference requests and perform partial inference tasks as a computing node. This dual operational mode allows the drone to dynamically engage in both roles based on network configuration and task allocation strategies. The system architecture is as follows: Figure 2 As shown: Due to limited UAV resources, each UAV is defined... n ∈ N The maximum storage capacity is Available energy is And computing power is For simplicity, assume all UAVs are at a fixed height above the ground. Flight To determine the minimum height required for safety (e.g., avoiding buildings), the UAV's position coordinates are described using Euclidean coordinates. Represents each UAV n ∈ N The horizontal coordinates. Consider a swarm of UAVs collaboratively performing an inference task using a single deep learning model. Where each UAV... n ∈ N Both can generate inference requests, capture images from the monitored area, and transmit them to a designated UAV to perform first-level inference.

[0057] Assume the original model follows the rules set by A chain structure composed of layers, where each model layer Based on predefined memory requirements and computing requirements Features. Using binary decision variables. This indicates the model allocation strategy, which assigns the split model layers to selected UAV nodes. Specifically, if the UAV... n Execution Model Layer ,but ,otherwise Each model layer must be assigned to one and only one UAV node, i.e. .

[0058] set up Indicates a set of key regions The target monitoring area. Definition For UAV equipment The coverage radius is required to maximize the coverage of the entire target monitoring area by the drone swarm, ensuring no blind spots in key areas and guaranteeing the comprehensiveness and real-time nature of data collection. Binary variables are introduced. To indicate key areas Whether it is covered by at least one UAV. To maximize coverage of key areas, the coverage target is defined as... : , ,in, It is an indicator function.

[0059] Latency Model: The latency of UAV swarm collaborative task processing can be divided into computation latency. and communication latency Two parts. UAV n The computational latency for performing a reasoning task can be expressed as: .in, It is the first Layer in UAV n On the processing time, Presentation layer The computational requirements Indicates UAV node The computing power. Computational energy consumption is expressed as... In the formula The effective switching capacitor coefficient.

[0060] Energy Consumption Model: Considering that the UAV-to-ground link is only used to offload classification results, this framework ignores this communication time and emphasizes communication between cooperating UAVs. Under given wireless channel conditions, the UAV device... and The maximum reliable data transfer rate between them is defined as: .in, Indicates allocation to UAV devices bandwidth, This represents the channel power gain at a reference distance of 1 meter. Indicates UAV device The transmission power, Indicates noise power. Indicates UAV device With the UAV device that performs subsequent inference The Euclidean distance between them. Assume the layers... and Assigned to nodes respectively and Then the data transmission delay can be expressed as In the formula Representation Model Layer The amount of intermediate data transmitted. Data transmission energy consumption is approximately the product of transmission power and transmission time, denoted as... Specifically, it is described as follows: In the formula For transmission power. The UAV hovers in a fixed position while collecting data from its coverage area and performing inference tasks for the monitored area. UAV node Hovering energy consumption modeling is In the formula and These are hovering power and hovering duration, respectively. It is assumed that during task execution, all UAVs maintain a fixed hovering time above the monitored area. .

[0061] Initialize the objective function. A target function is defined to evaluate the quality of multi-UAV collaborative inference results. The specific details are as follows: In the monitoring task, each UAV n ∈ N Retrieve pending tasks from its preceding UAV and execute them sequentially. The completion time of all UAVs performing inference tasks is expressed as the sum of computation time and data transfer time, i.e. In the formula It is a UAV n Total time spent on reasoning.

[0062] In UAV collaborative inference scenarios, by rationally partitioning deep learning models and deploying them across resource-constrained UAV clusters, while optimizing UAV deployment locations to ensure maximum coverage of key monitoring areas, the total inference latency is minimized. Therefore, the objective function is set as follows: Ultimately, the goal of solving the multi-UAV cooperative reasoning problem is to find the globally optimal solution. and This refers to the deployment strategy and model partitioning scheme of the UAV cluster, which minimizes the objective function value. In the formula... This represents the time cost of reasoning.

[0063] Step 2: Calculate an efficient collaborative reasoning scheme for the UAV swarm based on a two-stage optimization algorithm.

[0064] Phase 1: Model segmentation based on dynamic programming (DP) algorithm. With the UAV positions fixed, DP is used to distribute the deep learning model layer by layer to each UAV node to minimize cumulative inference latency and meet memory, energy, and computational resource constraints.

[0065] Phase Two: UAV Deployment Optimization Based on Successive Convex Approximation (SCA) Algorithm. Under a fixed model segmentation strategy, a first-order Taylor expansion is performed on the non-convex communication rate to construct a convex approximation subproblem. The deployment location of the UAV is optimized by iteratively solving this subproblem. These two phases are iterated alternately until a local optimum is reached. The specific operations are as follows: 2.1 System Initialization. An initial UAV deployment scheme is generated using region-centric k-means clustering. Set the number of UAVs, coverage radius, memory capacity, computing power, and energy limit; define the set of key reconnaissance points in the monitoring area; and give a minimum coverage threshold. .

[0066] 2.2 In the discrete optimization stage, the UAV position remains fixed, and the dynamic programming (DP) algorithm is used to solve for the optimized model partitioning strategy. The chain-like structure of deep learning models is suitable for sequential decision-making processes. Dynamic programming (DP) can effectively explore the allocation space from model layers to user-defined activity (UAV) while satisfying memory and energy constraints. The specific solution process is as follows: use Represents a state, which means that the first... Layer assigned to UAV n The minimum cumulative delay. In the formula... This is an indicator function for communication delay.

[0067]

[0068] Memory constraints and energy constraints are expressed as follows: , The algorithm flow using dynamic programming (DP) is shown in Table 1. This algorithm can accurately solve the model layer partitioning and allocation problem when the UAV's position is fixed, and efficiently calculate the minimum delay path in polynomial time. The algorithm complexity is O(log n). .

[0069] Table 1

[0070] 2.3 During the continuous optimization phase, while maintaining the current model partitioning strategy, the Successive Convex Approximation Algorithm (SCA) is used to optimize the UAV deployment locations, resulting in... The key strategy of SCA is to iteratively solve nonconvex problems by reformulating them into a sequence of convex subproblems. The objective function contains nonconvex terms related to communication rate. We use a convex approximation to handle the non-convex term.

[0071] In the In the next iteration, for the current iteration point Communication rate Perform convex approximation. Define an auxiliary function. , . It can be rewritten as a... Functions: The gradient can be calculated using this equation:

[0072] Then to Performing a Taylor expansion, we obtain First-order Taylor approximation: .

[0073] based on The optimization target can be Rephrased as decision variables The sub-problem of drone deployment The new optimization problem is a convex optimization problem, which can be solved using efficient convex optimizers such as CVX.

[0074] To simplify the solution, linear relaxation is used instead of the indicator function, and the covering constraint is reformulated as follows: The communication energy consumption is approximately... Therefore, the energy constraint is expressed as .

[0075] The initial problem is transformed into a sequence of convex subproblems using a first-order Taylor expansion, ensuring that the objective function does not increase with each iteration, i.e.: The SCA-based UAV positioning and deployment algorithm is shown in Table 2, with a computational complexity of O(n). O(N 3 ) Within the trust domain, the algorithm converges to a local optimum.

[0076] Table 2

[0077] 2.4 Repeat steps 2.2 and 2.3, generating a monotonically improving solution in each iteration, until the relative time delay improvement falls below a pre-set threshold. The optimal solution obtained at this point and This is the optimal solution under the current state. The details of the two-stage iterative optimization algorithm are shown in Table 3. For practical scenarios, the algorithm typically converges in 12 to 32 iterations.

[0078] Table 3

[0079] The two-stage iterative optimization algorithm avoids the dimensionality explosion problem, reducing the problem complexity from exponential (MINLP) to polynomial time, while ensuring theoretical convergence to a feasible solution through Lyapunov stability improvements. - Local optimum, its computational complexity is .

[0080] To verify the effectiveness and superiority of the collaborative UAV inference framework AirInfer in solving the joint optimization model partitioning and UAV deployment problems, simulation experiments and comparative experiments were set up, with different numbers of UAVs and different sizes of deep learning models.

[0081] To simplify the transmission model and improve the traceability of calculations, the following experimental design was conducted: 100 key areas were set up within a 100m × 100m monitoring range. UAV-related specifications: computing power... Memory capacity Transmission power Coverage radius Communication parameter settings: Bandwidth Channel gain noise power Model parameter settings: Number of model layers 10-24, memory requirements for each layer. Calculate the load The amount of intermediate data output by the layer Energy parameter settings: Hovering power switched capacitor coefficient The default coverage threshold is set to... .

[0082] The baseline algorithms include: a greedy layer allocation method based on SCA deployment (Greedy+SCA); a dynamic programming layer allocation method for random UAV deployment (DP+RW); and a joint optimization method for model layer allocation and UAV deployment using the Particle Swarm Optimization (PSO) algorithm (Joint-PSO).

[0083] Example 1: In the simulation: the number of drones is set to... The model has 10 layers. For example... Figure 3 As shown, the inference latency continuously decreases with increasing iteration count. Within 27 iterations, the latency drops significantly from 1543 ms to 899 ms. This monotonically decreasing latency validates the algorithm's ability to progressively minimize inference latency through model-device co-optimization.

[0084] Figure 4 This demonstrates the trend of coverage improvement during the iteration process. After the initial optimization phase, the coverage curve consistently exceeds... The threshold was determined. By the 11th iteration, the coverage stabilized at 87.0%, indicating that AirInfer effectively complied with the spatial constraints.

[0085] Figure 5The paper demonstrates the improvement trend of the trust domain change curve during the iteration process, with the trust domain radius curve reflecting the trade-off between exploration and exploitation. In the early stages, a larger trust region facilitates important location adjustments, helping the system escape local minima and satisfy coverage constraints. As optimization progresses, the trust region gradually shrinks, supporting finer-grained adjustments, improving communication efficiency while maintaining coverage requirements, and ultimately optimizing inference latency.

[0086] The evolution of a UAV cluster from its initial state to its final deployment formation position is as follows: Figure 6 As shown, the corresponding coverage in key areas improved significantly, rising from 47.0% to 87.0%, confirming that AirInfer achieved comprehensive and effective coverage throughout the monitored area.

[0087] Figure 7 The optimized model layer-to-UAV mapping strategy is demonstrated. Computationally intensive layers are assigned to high-performance UAVs, while data-dependent contiguous layers are placed on the same UAV to eliminate transport hops.

[0088] The results of the above experiments verify AirInfer's capabilities in robustness, effectiveness, and constraint-aware convergence.

[0089] Example 2: The following is a comprehensive sensitivity analysis of the AirInfer comparison baseline algorithm under two key parameters (number of UAVs and model size).

[0090] like Figure 8 As shown, the inference latency of all algorithms decreased as the number of UAVs increased from 5 to 10. This reduction stems from the decrease in the average distance between UAVs, facilitating shorter communication paths and faster data exchange. Among all tested algorithms, the two-stage method achieved the lowest absolute latency across the entire test range. The other three baseline algorithms were more or less affected by the algorithm parameter settings or the randomness of the algorithms themselves, and their latency optimization effects were not as stable and efficient as those of the two-stage algorithm.

[0091] As the model size increases, the latency of all algorithms will increase significantly due to the higher computational demands, such as... Figure 9 As shown, when the number of layers increases to a certain value, the layer optimization of DP can effectively handle chained models, while the Greedy algorithm fails due to the excessive complexity of the model. The Joint-PSO and DP+RW methods failed to converge and did not reach the pre-set coverage threshold, thus rendering their latency optimization results invalid. In contrast, the two-stage algorithm maintained the lowest latency growth rate, indicating that it effectively handled larger-scale model partitioning and cross-chain computation offloading.

[0092] The above results consistently demonstrate the superior performance, robustness, and scalability of the proposed two-stage optimization algorithm. AirInfer exhibits excellent performance under all test parameters and operating conditions.

[0093] Understandably, this application provides the optimal joint model partitioning and drone deployment scheme for multi-drone collaborative inference scenarios, significantly reducing end-to-end inference latency and improving task response speed. It ensures that the algorithm proposed in the monitoring area application reduces the complexity of problem handling from exponential to polynomial level, making it more suitable for practical deployment and supporting dynamic, heterogeneous drone swarms with strong adaptability.

[0094] Based on the collaborative reasoning-based joint UAV deployment and model partitioning method described in the above embodiments, this application also provides a collaborative reasoning-based joint UAV deployment and model partitioning system, such as... Figure 10 As shown, Figure 10 This application provides a schematic diagram of the structure of a collaborative reasoning-based joint UAV deployment and model partitioning system 10. The system includes: a determination module 1001, an allocation module 1002, and an optimization partitioning module 1003. The determining module 1001 is used to determine the drone swarm and the target monitoring area; wherein, the drone swarm includes multiple drones, which are used for image capture and data classification requests to detect abnormal conditions in the target area; The allocation module 1002 is used to establish a drone swarm system based on the multiple drones and the target monitoring area; and to allocate tasks to the multiple drones through the drone swarm system to obtain the target tasks corresponding to each of the multiple drones. The optimization partitioning module 1003 is used to optimize the deployment positions and partition the model layers of the multiple drones based on the target task, so as to obtain the model layers corresponding to each of the multiple drones and the optimal deployment positions corresponding to each of the multiple drones, so that the multiple drones can perform the target task.

[0095] In some embodiments of this application, the determining module 1001 is further configured to establish a communication connection based on the plurality of drones and the target monitoring area to determine the drone swarm system; wherein any two drones in the drone swarm system can communicate, and the monitoring range is the target monitoring area; the allocation module 1002 is further configured to divide the target monitoring area to obtain a plurality of sub-target monitoring areas; the plurality of sub-target monitoring areas are consistent with the number of the plurality of drones; based on the plurality of sub-target monitoring areas, the drone swarm system is used to allocate tasks to the plurality of drones to obtain the target tasks corresponding to each of the plurality of drones.

[0096] In some embodiments of this application, the determining module 1001 is further configured to assign deep learning models to the plurality of drones and determine the model layer corresponding to each of the plurality of drones; The optimization partitioning module 1003 is further configured to optimize the deployment positions of the plurality of UAVs based on the target task, so as to obtain the optimal deployment position for each of the plurality of UAVs.

[0097] In some embodiments of this application, the determining module 1001 is further configured to determine constraints; wherein the constraints include a first constraint, a second constraint, a third constraint, a fourth constraint, and a fifth constraint. The allocation module 1002 is further configured to allocate multiple model layers in the deep learning model to the multiple drones based on the first constraint, the second constraint, the third constraint, the fourth constraint, and the fifth constraint, thereby obtaining the model layers corresponding to each of the multiple drones.

[0098] In some embodiments of this application, the determining module 1001 is further configured to determine the initial position information of the plurality of drones; wherein, the initial position information includes the horizontal coordinates and fixed altitude of the drones; The optimization partitioning module 1003 is further configured to continuously optimize the initial position information using a successive convex approximation algorithm to obtain the updated position information of the multiple drones; calculate the time delay based on the updated position information of the multiple drones; and determine the optimal deployment position for each of the multiple drones if the time delay is lower than a preset threshold.

[0099] Based on the collaborative reasoning-based joint UAV deployment and model partitioning method described in the above embodiments, this application also provides a collaborative reasoning-based joint UAV deployment and model partitioning device, such as... Figure 11 As shown, Figure 11This is a schematic diagram of a collaborative reasoning-based joint UAV deployment and model partitioning device provided in an embodiment of this application. The collaborative reasoning-based joint UAV deployment and model partitioning device 11 includes a processor 1101 and a memory 1102. The memory 1102 is used to store computer programs; the processor 1101 is used to call and run the computer programs from the memory to execute the collaborative reasoning-based joint UAV deployment and model partitioning method as described in the above embodiment.

[0100] In the embodiments of this application, the processor 1101 described above can be at least one of the following: Application-Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), Controller, Microcontroller, and Microprocessor. It is understood that for different devices, the electronic device used to implement the above processor function can also be other types, and the embodiments of this application do not specifically limit it.

[0101] This application provides a computer-readable storage medium storing a computer program for implementing, when executed by a processor, the method for joint UAV deployment and model partitioning based on collaborative reasoning as described in any of the above embodiments.

[0102] For example, the program instructions corresponding to a collaborative reasoning-based joint UAV deployment and model partitioning method in this embodiment can be stored on storage media such as optical discs, hard disks, and USB flash drives. When the program instructions corresponding to the collaborative reasoning-based joint UAV deployment and model partitioning method in the storage media are read or executed by an electronic device, the collaborative reasoning-based joint UAV deployment and model partitioning method described in any of the above embodiments can be implemented.

[0103] Furthermore, the functional modules in this embodiment can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional module. If the integrated unit is implemented as a software functional module and is not sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this embodiment, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the method of this embodiment. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0104] It should be understood that the phrases "one embodiment," "an embodiment," or "some embodiments" mentioned throughout the specification mean that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment," "in one embodiment," or "in some embodiments" appearing throughout the specification do not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the embodiments in this application are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. The descriptions of the various embodiments above tend to emphasize the differences between the various embodiments; their similarities or commonalities can be referred to mutually, and for the sake of brevity, these will not be repeated here.

[0105] The modules described above as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0106] In addition, each functional module in the various embodiments of this application can be integrated into one processing unit, or each module can be a separate unit, or two or more modules can be integrated into one unit; the integrated modules can be implemented in hardware or in the form of hardware plus software functional units.

[0107] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0108] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0109] The features disclosed in the several product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0110] The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0111] The above description is merely an embodiment of this application, but the protection scope of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for joint UAV deployment and model partitioning based on collaborative reasoning, characterized in that, The method includes: Identify the drone swarm and target monitoring area; wherein the drone swarm includes multiple drones, which are used for image capture and data classification requests to detect anomalies in the target area; Based on the multiple drones and the target monitoring area, a drone swarm system is established; and the drone swarm system is used to assign tasks to the multiple drones to obtain the target tasks corresponding to each of the multiple drones. Based on the target task, the deployment locations of the multiple drones are optimized and the model layers are divided to obtain the model layers corresponding to each of the multiple drones and the optimal deployment locations corresponding to each of the multiple drones, so that the multiple drones can perform the target task.

2. The method according to claim 1, characterized in that, The system establishes a drone swarm system based on the multiple drones and the target monitoring area; The system then assigns tasks to the multiple drones through the drone swarm system, obtaining the target tasks corresponding to each drone, including: Based on the multiple drones and the target monitoring area, a communication connection is established to determine the drone swarm system; wherein, any two drones in the drone swarm system can communicate, and their monitoring range is the target monitoring area; The target monitoring area is divided into multiple sub-target monitoring areas; the number of the multiple sub-target monitoring areas is the same as the number of the multiple drones. Based on the multiple sub-target monitoring areas, the drone swarm system allocates tasks to the multiple drones to obtain the target tasks corresponding to each of the multiple drones.

3. The method according to claim 1, characterized in that, Based on the target task, the deployment location optimization and model layer partitioning of the multiple drones are performed to obtain the model layer corresponding to each of the multiple drones and the optimal deployment location corresponding to each of the multiple drones, including: Assign deep learning models to the multiple drones and determine the model layer corresponding to each of the multiple drones; Based on the target task, the deployment locations of the multiple drones are optimized to obtain the optimal deployment location for each of the multiple drones.

4. The method according to claim 3, characterized in that, Assigning deep learning models to the multiple drones and determining the model layer corresponding to each of the multiple drones includes: Determine the constraints; wherein the constraints include a first constraint, a second constraint, a third constraint, a fourth constraint, and a fifth constraint. Based on the first constraint, the second constraint, the third constraint, the fourth constraint, and the fifth constraint, multiple model layers in the deep learning model are respectively assigned to the multiple drones to obtain the model layers corresponding to each of the multiple drones.

5. The method according to claim 4, characterized in that, The first constraint is that the model layer allocated to any one of the plurality of drones does not exceed the maximum memory capacity of the drone. The second constraint is that the sum of the computing energy consumption, data transmission energy consumption, and hovering energy consumption of the plurality of UAVs is less than a preset energy consumption. The third constraint is that each of the multiple model layers in the deep learning model can only be deployed and executed on one drone. The fourth constraint is that the allocation decision variable in the deep learning model is a binary decision variable, which can only take the value 0 or 1. The fifth constraint is that the coverage area of ​​the monitoring area deployed by any one of the plurality of UAVs is greater than the preset coverage area.

6. The method according to claim 3, characterized in that, The step of optimizing the deployment locations of the multiple drones based on the target task to obtain the optimal deployment location for each of the multiple drones includes: Determine the initial position information of the plurality of drones; wherein, the initial position information includes the horizontal coordinates and fixed altitude of the drones; The initial position information is continuously optimized by a successive convex approximation algorithm to obtain the updated position information of the multiple UAVs; The latency is calculated based on the updated location information of the multiple drones; If the latency is lower than a preset threshold, the optimal deployment location for each of the multiple drones is determined.

7. The method according to claim 1, characterized in that, Each of the plurality of drones simultaneously serves as a data generation node and a computing node; when each drone acts as a data generation node, it captures images and generates inference requests; when each drone acts as a computing node, it performs a portion of the inference task; the specific function of each drone is determined according to the network configuration and task allocation strategy.

8. A joint UAV deployment and model partitioning system based on collaborative reasoning, characterized in that, The collaborative reasoning-based joint UAV deployment and model partitioning system includes: a determination module, an allocation module, and an optimization partitioning module, wherein, The determining module is used to determine the drone swarm and the target monitoring area; wherein, the drone swarm includes multiple drones, which are used for image capture and data classification requests to detect abnormal conditions in the target area; The allocation module is used to establish a drone swarm system based on the multiple drones and the target monitoring area; and to allocate tasks to the multiple drones through the drone swarm system to obtain the target tasks corresponding to each of the multiple drones. The optimization partitioning module is used to optimize the deployment positions and partition the model layers of the multiple drones based on the target task, so as to obtain the model layers corresponding to each of the multiple drones and the optimal deployment positions corresponding to each of the multiple drones, so that the multiple drones can perform the target task.

9. A joint UAV deployment and model partitioning device based on collaborative reasoning, characterized in that, include: Processor and memory, of which, The memory is used to store computer programs; The processor is configured to call and run the computer program from the memory to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions for causing a processor to execute, thereby implementing the method of any one of claims 1 to 7.

Images