Unmanned aerial vehicle cluster size model dynamic cooperative reasoning method based on genetic algorithm

Abstract

The application discloses a dynamic cooperative reasoning method for a UAV cluster size model based on a genetic algorithm, which comprises the following steps: acquiring UAV trajectory, memory and computing capacity information and reasoning task distribution information; establishing a dynamic cooperative reasoning model for the UAV cluster size model according to the memory and computing capacity constraints of the UAV; taking the utility value function of reasoning time delay and reasoning accuracy as fitness, and iteratively calculating according to the genetic algorithm to obtain an optimal size model cooperative reasoning scheme with low time delay and high accuracy. The application can effectively improve the reasoning efficiency of the dynamic cooperative reasoning system of the UAV cluster.

Description

Technical Field

[0001] This invention relates to the field of collaborative reasoning technology for unmanned aerial vehicle (UAV) swarms, and specifically to a dynamic collaborative reasoning method for UAV swarm size models based on genetic algorithms. Background Technology

[0002] With the development of artificial intelligence (AI) technology and the maturity of drone technology, deploying AI algorithms to drone swarms to perform important collaborative tasks such as real-time target reconnaissance has become a current research hotspot. However, as the number of model parameters for AI algorithms such as deep learning becomes increasingly large, deploying complex deep learning models to resource-constrained edge devices such as drone swarms presents significant challenges, thus limiting the task collaboration and reasoning capabilities of drone swarms in complex environments.

[0003] To address these issues, the traditional solution is to send data from edge mobile devices to a cloud server, where the cloud server performs inference tasks and returns the results to the edge device. However, due to the need for long-distance data transmission, the traditional cloud computing model suffers from insufficient real-time performance, inadequate bandwidth, high energy consumption, and poor privacy. With the improvement of computing power and storage space in edge devices, as well as the development of model compression and optimization technologies, edge devices can now handle deep neural model inference tasks.

[0004] Patent publication number CN1 16805195A discloses a method and system for collaborative inference in a drone swarm based on model segmentation. The method involves deploying the same pre-trained deep learning model on each drone in the swarm. One drone acts as the source drone, which, upon receiving an inference task request, sends the task to all remaining drones in the swarm. After receiving the inference task from the source drone, each remaining drone performs a statistical evaluation of its own state and available resources, while simultaneously sensing the channel conditions with other drones. The statistical evaluation and sensing results are then fed back to the source drone. Based on the feedback, the inference task is divided into multiple sub-tasks according to the layer structure and number of layers of the deep learning model. These sub-tasks are then assigned to drones that meet the requirements, with each drone executing only one sub-task. This invention improves task execution efficiency by deploying a deep learning model on the drone swarm and through multi-drone collaborative inference. However, this invention requires segmentation and collaborative reasoning for each reasoning task. In practical applications, different drones have different computing capabilities and memory, and different reasoning tasks have different distribution information and computational load. If the same high-parameter deep learning model is deployed on all drones and each reasoning task adopts a collaborative reasoning scheme, it will lead to a large latency in actual operation, thus affecting the real-time performance of the reasoning results. Summary of the Invention

[0005] To address the issues of insufficient real-time performance and poor privacy in traditional cloud computing models, this invention proposes a dynamic collaborative reasoning method for UAV swarm size models based on genetic algorithms. First, it acquires information on UAV trajectory, memory and computing power, as well as the distribution of reasoning tasks. Then, it iterative calculations are performed using a genetic algorithm to derive the optimal size model collaborative reasoning scheme with low latency and high accuracy.

[0006] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0007] A dynamic collaborative reasoning method for a drone swarm size model based on a genetic algorithm, characterized in that the dynamic collaborative reasoning method includes the following steps:

[0008] S1, acquire the trajectory, memory, computing power constraints and inference task distribution information of the drone cluster; the drone cluster consists of N heterogeneous drones with computing power, one of which is defined as the source drone for acquiring input information.

[0009] S2. Based on the constraints of UAV memory and computing power, a dynamic collaborative inference model for UAV swarm size is established, consisting of a first deep neural network model and a second deep neural network model. The dynamic collaborative inference model for UAV swarm size is installed on the source UAV. Each inference task is obtained based on the inference task distribution information. The first deep neural network for multi-UAV collaborative inference or the second deep neural network for single-UAV inference is selected based on the computing power of the UAV swarm and the communication rate derived from the trajectory of the UAV swarm. The accuracy and number of parameters of the first deep neural network model are higher than those of the second deep neural network model.

[0010] S3 uses a genetic algorithm to solve the dynamic collaborative reasoning model of the UAV swarm size model. The solution process includes the following sub-steps:

[0011] S3 1, Initialize the population and encode individuals according to the number of tasks R, with each individual representing R inference schemes; the inference schemes include two types: selecting a first deep neural network model for model segmentation and having N UAVs perform collaborative inference, and selecting a second deep neural network model for local execution by the source UAV. The model segmentation methods are different for different inference schemes.

[0012] S32, calculate the fitness of each individual in the current population based on the utility value function of inference delay and inference accuracy, and select the individual P with the highest fitness;

[0013] S33: Randomly select a group of individuals and select the individual Q with the highest fitness from them through a tournament method;

[0014] S34, crossover mutation of two individuals P and Q to generate new individuals and add them to a new population;

[0015] S35, determine whether the new population size has reached the predetermined population size. If not, proceed to step S33; otherwise, proceed to step S36.

[0016] S36, determine whether the number of iterations of the population has reached the predetermined number of iterations. If not, proceed to step S32; otherwise, proceed to step S37.

[0017] S37: Select the individual with the best fitness among all iterations as the dynamic collaborative reasoning scheme.

[0018] Furthermore, the number of parameters in the first deep neural network model is greater than or equal to 10M, and the number of parameters in the second deep neural network model is less than 10M.

[0019] Furthermore, the communication rate calculation process derived from the trajectory of the UAV swarm in step S2 is as follows:

[0020]

[0021] In the formula, B i，k P represents the bandwidth between drone i and drone k; P represents the transmission power. The distance between the drones at time t is represented by α, which is the path loss exponent; g represents the distance between the drones at time t. 00 These are channel coefficients that follow a complex normal distribution CN(0,1); N o It is additive white Gaussian noise.

[0022] Further, in step S32, the utility function for inference latency and inference accuracy is:

[0023]

[0024] In the formula, R represents the number of inference tasks, Ar represents the accuracy of inference task r, and the accuracy of inference task r is the accuracy of the model on the ImageNet dataset; L r Let λ1 be the inference task r. The inference latency is divided into small model inference latency and large model inference latency according to the inference method of task r. The small model inference latency is the execution latency of the second deep neural network model on the source UAV. The large model inference latency consists of the execution latency of the first deep neural network model of each layer on its assigned UAV and the transmission latency of intermediate data. λ1 and λ2 are the weighting coefficients of accuracy and latency.

[0025] Furthermore, in step S34, the methods for performing crossover mutations on the two individuals P and Q include the following:

[0026] Cross-mutation using the first deep neural network for collaborative inference, cross-mutation using the second neural network for single-machine inference, and cross-mutation of the model segmentation scheme of the first deep neural network.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] The present invention provides a dynamic collaborative reasoning method for UAV swarm size models based on genetic algorithms. This method can obtain an optimal size model collaborative reasoning scheme with low latency and high accuracy under the constraints of UAV memory and computing power, thereby effectively improving the reasoning efficiency of UAV swarm dynamic collaborative reasoning systems. Attached Figure Description

[0029] Figure 1 This is a flowchart of the dynamic collaborative reasoning method for UAV swarm size model based on genetic algorithm according to an embodiment of the present invention;

[0030] Figure 2 This is a scenario diagram for collaborative reasoning with dynamic switching between large and small models in a drone cluster, as described in an embodiment of the present invention. Detailed Implementation

[0031] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0032] This invention provides a dynamic collaborative inference method for UAV swarm size models based on genetic algorithms. First, it acquires UAV trajectory, memory, and computing power information, as well as inference task distribution information. Then, it establishes a UAV execution inference model based on UAV memory and computing power constraints. Using a utility function of inference latency and inference accuracy as the fitness, it iterative calculations are performed using a genetic algorithm to obtain an optimal size model collaborative inference scheme that achieves low latency while ensuring high accuracy. Figure 1 As shown, the dynamic collaborative reasoning method specifically includes the following steps:

[0033] Step S1: Obtain drone trajectory, memory and computing power information, as well as inference task distribution information.

[0034] Step S2: Establish a dynamic collaborative reasoning model for the drone swarm size model based on the constraints of drone memory and computing power. In this embodiment, the drone swarm consists of N heterogeneous drones with computing capabilities, including one source drone that acquires input information. The dynamic collaborative reasoning model for the drone swarm size model is used to obtain each reasoning task based on the distribution information of the reasoning tasks, and selects a first deep neural network for multi-drone collaborative reasoning or a second deep neural network for single-drone reasoning based on the computing power of the drone swarm and the communication rate derived from the trajectory of the drone swarm. The dynamic collaborative reasoning model for the drone swarm size model can be deployed on the source drone. It should be noted that the first deep neural network of this invention can refer to the collaborative reasoning model in the invention with patent publication number CN116805195A, or other similar task processing networks that can be divided into multiple sub-tasks for collaborative reasoning by multiple drones can be selected. The focus of this invention is not on the specific structure of the large model, but on the dynamic selection of the size model. Therefore, the specific structure and working principle of the first deep neural network will not be described in detail here.

[0035] Preferably, the communication rate calculation process derived from the trajectory of the drone swarm is as follows:

[0036]

[0037] In the formula, Bi,k represents the bandwidth between drone i and drone k; P represents the transmission power; denoted by t, where α is the path loss exponent, g0 is the channel coefficient following a complex normal distribution CN(0,1), and N0 is additive white Gaussian noise.

[0038] Step S3: Initialize the population. Encode individuals according to the number of tasks R. Each individual represents R inference schemes. Each inference scheme selects either a large model or a small model for execution. If a large model is selected, it is segmented and N UAVs perform collaborative inference. If a small model is selected, it is executed locally by the source UAV. In this embodiment, the large model is a high-precision complex deep neural network model with more than or equal to 10M parameters (first deep neural network model), and the small model is a relatively accurate lightweight deep neural network model with less than 10M parameters (second deep neural network model).

[0039] Step S4: Calculate the fitness of each individual in the current population based on the utility function of inference delay and inference accuracy, and select the individual P with the highest fitness; preferably, the utility function of inference delay and inference accuracy is:

[0040]

[0041] In the formula, R represents the number of reasoning tasks, and A r For the accuracy of inference task r, the accuracy of inference task r is the accuracy of the model on the ImageNet dataset; L r Let λ1 be the inference task r. The inference latency is divided into small model inference latency and large model inference latency according to the inference method of task r. The small model inference latency is the execution latency of the second deep neural network model on the source UAV. The large model inference latency consists of the execution latency of the first deep neural network model of each layer on its assigned UAV and the transmission latency of intermediate data. λ1 and λ2 are the weighting coefficients of accuracy and latency.

[0042] Step S5: Randomly select a group of individuals and select the individual Q with the highest fitness from them using a tournament method;

[0043] Step S6: Cross over and mutate the two individuals P and Q to generate new individuals and add them to the new population. Preferably, the crossover mutation here includes both crossover mutation of choosing to use the first deep neural network for collaborative inference or choosing to use the second neural network for single-machine inference, and crossover mutation of the specific model segmentation scheme of the first deep neural network. For the former, it refers to the mutation of the choice between the large model and the small model; for the latter, it refers to the mutation of the model segmentation scheme. For example, in the original individuals, a certain inference task chooses the large model for inference and the third layer of the large model is assigned to UAV 1 for processing. After mutation, the inference task still chooses the large model for inference, but the third layer of the large model is assigned to UAV 2 for processing.

[0044] Step S7: Determine whether the new population size has reached the predetermined population size. If not, proceed to step S5; otherwise, proceed to step S8.

[0045] Step S8: Determine whether the number of iterations for the population has reached the predetermined number of iterations. If not, proceed to step S4; otherwise, proceed to step S9.

[0046] Step S9: Select the individual with the best fitness among all generations as the dynamic collaborative reasoning scheme.

[0047] For example, initially, a population might consist of 50 individuals with 7 reasoning tasks. Each reasoning task has a corresponding reasoning scheme, meaning each of the 50 individuals contains 7 reasoning schemes, and each scheme can be either large-model collaborative reasoning or small-model single-machine reasoning. Therefore, the final solution is to select the individual with the highest fitness among these 50 individuals. For instance, the reasoning scheme of the individual with the highest fitness might be to use the large model for three reasoning tasks and the small model for four, while specifying the specific model partitioning method for the large model corresponding to each reasoning task. The aforementioned reasoning process is dynamically updated based on information such as the UAV's status, communication rate, and reasoning task distribution, enabling dynamic selection of the processing method for each reasoning task.

[0048] After obtaining the most fit individual as the dynamic collaborative reasoning scheme for the large and small models, such as Figure 2 As shown, reasoning is performed on each task according to its reasoning scheme. If the task uses small model reasoning, the source UAV performs single-machine reasoning; otherwise, different UAVs execute their assigned layers to complete the reasoning task according to the collaborative reasoning scheme.

[0049] This invention presents a dynamic collaborative inference method for UAV swarm size models based on genetic algorithms. It considers two neural networks: a high-precision, high-parameter complex neural network and a low-precision, low-parameter lightweight neural network. For the high-precision, high-parameter complex neural network, a model segmentation method is used to divide it into multiple sub-networks, with each UAV executing only its own sub-network. For the low-precision, low-parameter lightweight neural network, single-machine inference is performed directly on the source UAV. By dynamically switching between the two networks, highly efficient UAV swarm aerial collaborative inference tasks are achieved. Furthermore, this invention comprehensively considers inference latency and accuracy to construct a fitness function, and uses a genetic algorithm to find the individual with the highest fitness, considering its inference scheme as the optimal inference scheme. Based on this, this invention can obtain an optimal size-model collaborative inference scheme with low latency and high accuracy within the constraints of UAV memory and computing power, thereby effectively improving the inference efficiency of the UAV swarm dynamic collaborative inference system.

[0050] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0051] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0052] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0053] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, causing a series of operational steps to be executed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that run on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0054] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0055] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A dynamic collaborative reasoning method for UAV swarm size model based on genetic algorithm, characterized in that, The dynamic collaborative reasoning method includes the following steps: S1, acquire the trajectory, memory, computing power constraints and inference task distribution information of the drone cluster; the drone cluster consists of N heterogeneous drones with computing power, one of which is defined as the source drone for acquiring input information. S2. Based on the constraints of UAV memory and computing power, a dynamic collaborative inference model for UAV swarm size is established, consisting of a first deep neural network model and a second deep neural network model. The dynamic collaborative inference model for UAV swarm size is deployed on the source UAV. Each inference task is obtained based on the inference task distribution information, and the first deep neural network for multi-UAV collaborative inference or the second deep neural network for single-UAV inference is selected based on the computing power of the UAV swarm and the communication rate derived from the trajectory of the UAV swarm. Among them, the accuracy and number of parameters of the first deep neural network model are higher than those of the second deep neural network model. S3 uses a genetic algorithm to solve the dynamic collaborative reasoning model of the UAV swarm size model. The solution process includes the following sub-steps: S31, initialize the population, encode individuals according to the number of tasks R, each individual represents R inference schemes; the inference schemes include two types, selecting a first deep neural network model for model segmentation and inference collaboratively by N UAVs, and selecting a second deep neural network model for local execution by the source UAV, the model segmentation methods of different inference schemes are different; S32, calculate the fitness of each individual in the current population based on the utility value function of inference delay and inference accuracy, and select the individual P with the highest fitness; S33: Randomly select a group of individuals and select the individual Q with the highest fitness from them through a tournament method; S34, crossover mutation of two individuals P and Q to generate new individuals and add them to a new population; S35, determine whether the new population size has reached the predetermined population size. If not, proceed to step S33; otherwise, proceed to step S36. S36, determine whether the number of iterations of the population has reached the predetermined number of iterations. If not, proceed to step S32; otherwise, proceed to step S37. S37: Select the individual with the best fitness among all iterations as the dynamic collaborative reasoning scheme.

2. The dynamic collaborative reasoning method for UAV swarm size model based on genetic algorithm according to claim 1, characterized in that, The first deep neural network model has a parameter count greater than or equal to 10M, and the second deep neural network model has a parameter count less than 10M.

3. The dynamic collaborative reasoning method for UAV swarm size model based on genetic algorithm according to claim 1, characterized in that, The communication rate calculation process derived from the trajectory of the UAV swarm in step S2 is as follows: In the formula, B i，k P represents the bandwidth between drone i and drone k; P represents the transmission power. denoted by t, where α is the path loss exponent, g0 is the channel coefficient following a complex normal distribution CN(0,1), and N0 is additive white Gaussian noise.

4. The dynamic collaborative reasoning method for UAV swarm size model based on genetic algorithm according to claim 1, characterized in that, In step S32, the utility function for inference latency and inference accuracy is: In the formula, R represents the number of reasoning tasks, and A r For the accuracy of inference task r, the accuracy of inference task r is the accuracy of the model on the ImageNet dataset; L r The inference latency is the latency of inference task r. The inference latency is divided into small model inference latency and large model inference latency according to the inference method of task r. The small model inference latency is the latency of the second deep neural network model executed on the source UAV. The large model inference latency consists of the execution latency of the first deep neural network model of each layer on its assigned UAV and the transmission latency of intermediate data. λ1 and λ2 are the weighting coefficients for accuracy and delay.

5. The dynamic collaborative reasoning method for UAV swarm size model based on genetic algorithm according to claim 1, characterized in that, In step S34, the methods for performing crossover mutation on two individuals P and Q include the following: Cross-mutation using the first deep neural network for collaborative inference, cross-mutation using the second neural network for single-machine inference, and cross-mutation of the model segmentation scheme of the first deep neural network.

Images