Multi-unmanned aerial vehicle cooperative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning

Abstract

The invention provides a multi-unmanned aerial vehicle cooperative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning, and the method comprises the steps: dynamically adjusting the position of an unmanned aerial vehicle based on a potential field method when a task request of a ground vehicle is received, and constructing a topology network in combination with greedy clustering and a DAG topology generation strategy; performing dynamic DNN model partitioning based on a discontinuous layer allocation mechanism and a dynamic programming algorithm DPA, and dynamically allocating DNN discontinuous layers to different unmanned aerial vehicles; performing DNN reasoning according to the topology network and the partitioning scheme; performing cache updating based on the priority and the energy perception strategy; and calculating time delay and overhead. Task allocation and load balancing can be intelligently performed according to the complexity of tasks and the computing power of each unmanned aerial vehicle, so that the overall efficiency of the system is improved, and the processing time delay is reduced; the task and data allocation strategy can be flexibly adjusted according to the network state and the environment change, and the real-time performance and the high efficiency of the task are ensured.

Description

Technical Field

[0001] The present invention relates to the field of UAV collaborative reasoning technology, and in particular to a multi-UAV collaborative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning. Background Art

[0002] With the rapid development of drone technology, multi-drone collaborative reasoning systems are demonstrating significant value in intelligent traffic monitoring. In intelligent driving scenarios, vehicles often face the problem of missing road information due to visual blind spots when cruising long distances on highways, driving on winding roads in mountainous areas, or at the scene of a sudden traffic accident.

[0003] Traditional solutions rely on data fusion between onboard sensors and roadside units, but their sensing range is limited by fixed deployment density and physical obstructions. Drones can rapidly respond to vehicle requests, capturing images of the road ahead from an aerial perspective and forming a dynamic collaborative perception network with vehicles. However, existing cloud-based multi-drone collaborative reasoning systems face significant challenges in real-world road scenarios.

[0004] Traditional multi-UAV collaborative reasoning systems typically rely on cloud computing to process large amounts of data and perform high-load reasoning tasks. In this architecture, the drones primarily perform data collection and preliminary processing, while the cloud handles complex reasoning and analysis. However, this cloud-based processing architecture faces multiple challenges.

[0005] First, cloud computing cannot meet the needs of some application scenarios with extremely high real-time requirements, such as autonomous driving and disaster response. In these scenarios, data collection, processing, and decision-making need to be completed as quickly as possible, and any delay can lead to catastrophic consequences.

[0006] Secondly, cloud computing faces challenges such as communication latency and bandwidth bottlenecks that cannot be ignored. The communication link between drones and the cloud is often limited by bandwidth, transmission distance, and network stability, especially in complex environments. Network instability and data transmission delays can prevent the system from responding quickly to external changes, impacting overall system performance and security.

[0007] Furthermore, multi-UAV collaborative reasoning systems face the challenge of limited computing resources. Each UAV typically has low computing power, relying primarily on lightweight processors and limited memory, suitable for handling lighter computational tasks. However, with the continuous advancement of algorithms such as deep learning and object detection, the complexity and resource requirements of computational tasks are gradually increasing. When faced with computationally demanding tasks such as object detection, object tracking, and large-scale data analysis, the computing power of a single UAV often falls short. In such cases, a single UAV cannot handle all inference tasks, requiring collaboration with other UAVs to complete the task.

[0008] To address these challenges, multi-UAV systems need to improve performance and efficiency by optimizing both the compute and data chains. This requires not only efficient task allocation but also ensuring that each UAV can maximize its capabilities within limited computing resources, avoiding computational bottlenecks. Compute chain optimization involves minimizing system latency and energy consumption through rational task allocation and scheduling, while data chain optimization involves efficiently managing and transmitting data, reducing redundant and ineffective communication. Traditional cloud computing architectures struggle to achieve this efficient task and data optimization. Summary of the Invention

[0009] In response to the shortcomings of the existing technology, the present invention provides a multi-UAV collaborative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning. By optimizing the UAV system topology and dynamically adjusting the DNN model partitioning, the present invention can intelligently perform task allocation and load balancing according to the complexity of the task and the computing power of each UAV, thereby improving the overall efficiency of the system and reducing processing delays.

[0010] The technical solution of the present invention is: a multi-UAV collaborative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning, comprising the following steps:

[0011] S1), initialize the UAV and mission vehicle;

[0012] S2), when receiving the task request of the ground vehicle, determine whether the cache is hit, if not; then execute step S3);

[0013] S3) Dynamically adjust the position of the UAV based on the potential field method, and build a topological network by combining greedy clustering and DAG topology generation strategy;

[0014] S4) Dynamically partition the DNN model based on the discontinuous layer allocation mechanism and the dynamic programming algorithm DPA, and dynamically allocate the DNN discontinuous layers to different drones;

[0015] S5), perform DNN reasoning based on the topological network and partitioning scheme;

[0016] S6), cache update based on priority and energy-aware strategy;

[0017] S7) Calculate delay and overhead.

[0018] Preferably, in step S3), the potential field includes task attraction and repulsion between drones, which drives drones to gather in high-value areas through task attraction, while using repulsion between drones to avoid collisions and interference; and combines the scoring function of drone clustering to screen high-potential nodes, generate a directed acyclic graph that minimizes delay, and unify the task input source through a virtual root node.

[0019] As a preference, in step S3), the task attraction F attract (d) is the gravitational field generated by the vehicle mission request point, the intensity of which is positively correlated with the mission priority, namely:

[0020]

[0021] Where k task is the attraction coefficient; d is the distance between the UAV and the vehicle; R task The mission impact radius;

[0022] The repulsive force F between drones repel (d) Avoid collision and communication interference, the expression is:

[0023]

[0024] Where k uav is the repulsive force coefficient; d safe For a safe distance.

[0025] Preferably, in step S3), under the action of the potential field, the motion trajectory of the UAV is adjusted and moved strictly in accordance with the motion control equation, which is expressed as:

[0026]

[0027] Where, For the mission machine u i Velocity vector at time t+1; It represents the combined force of the UAV's attraction force, obstacle repulsion force, and repulsion force between UAVs; a and b are weight coefficients;

[0028] The best speed in history.

[0029] As a preference, in step S3), each drone u iWith heterogeneous computing and communication capabilities, each drone u i Energy E i (t) Dynamic changes, including motion energy consumption E move (t) and communication energy consumption E comm (t), where the UAV motion energy consumption E move (t) Produced by flight attitude adjustments and displacements, namely:

[0030]

[0031] Where, Indicates the change of kinetic energy; P hov Δt represents the hovering power consumption; m i For drone u i The quality of |v i (t)| is the drone u i The speed of P hov represents the hovering power; Δt represents the time step; when |v i When (t)|<0.1, only hovering power consumption is calculated;

[0032] The communication energy consumption E comm (t) includes the energy consumption of data sending and receiving, namely:

[0033] E comm (t) = P tx ·t tx +P rx ·t rx ; (5)

[0034] Where, P tx represents the transmit power; t tx Indicates the sending time; P rx represents the received power; t rx Indicates the receiving time.

[0035] As a preference, in step S3), the scoring function Score(u i ) is defined as:

[0036]

[0037] In the formula, α, β, and γ are weight coefficients; P tx,i For drone u i The transmission power of R comm Indicates the communication radius; C i For drone u i The computing power of E i For drone u i The remaining energy P max is the maximum transmit power; Cmax is the maximum computing power; E max For maximum energy.

[0038] Preferably, in step S3), the drones are clustered and screened according to their cluster scores. The specific process is as follows:

[0039] First, a preliminary spatial screening is performed to select UAVs within a radius of L meters from the mission point as candidates;

[0040] Next, connectivity verification is carried out to construct a fully connected candidate set and remove isolated nodes to ensure good communication connectivity between candidate drones.

[0041] Finally, the candidate drones are ranked by the greedy algorithm according to the scoring function Score(u i ) Select the first K drones in descending order as the clustering results.

[0042] Preferably, in step S3), a topology network is constructed using a DAG topology generation strategy, specifically comprising the following steps:

[0043] S31) Based on the selected UAV cluster set U = {u1,u2,…,u K}, get the position and transmission power P of the corresponding drone tx,i and computing power C i , and through computing power C i Sort the drones in descending order;

[0044] S32), for each drone u i , select the drone within its communication range and ranked first as the candidate parent node;

[0045] S33) Select the first two parent nodes with the smallest weight according to the link weight to form edge e i,k ;

[0046] S34), add a virtual root node v0, connect it to all nodes without incoming edges, and use the virtual root node v0 as the input source of the task;

[0047] S35), prune redundant edges, delete edges that may cause cycles, and obtain a directed acyclic graph.

[0048] Preferably, in step S4), the dynamic DNN model partitioning includes inter-layer partitioning and intra-layer partitioning. The inter-layer partitioning allocates different layers of the DNN model to devices at different stages, and the intra-layer partitioning further subdivides the computing tasks of the same layer into multiple subtasks, and allocates them to different devices at the same stage for parallel computing according to the computing power ratio of the devices. Specifically, the steps include:

[0049] S41), constructing a heterogeneous graph including drone nodes and DNN layer nodes based on the drone topology network;

[0050] S42) Using the graph attention network GAT to encode the UAV-DNN relationship, the optimal allocation probability between the UAV and the DNN layer is generated through the node features and topological relationships of the heterogeneous graph;

[0051] S43) Dynamically adjust the layer allocation strategy based on the proximal optimization strategy PPO.

[0052] As a preferred embodiment, in step S41), the drone nodes in the heterogeneous graph are represented by Its eigenvector is Among them, P i is the transmission power; loc i is the position coordinate;

[0053] The DNN layer nodes are represented as Its eigenvector Among them, Comp i For layer l i Complexity of Data i For layer l i Output data size;

[0054] The edge types of the heterogeneous graph include drone communication edges and DNN dependency edges Among them, UAV communication edge The link weights in the topological network Determine the link weight Indicates that from the drone u i to droneu k The link communication capability, the link weight The calculation formula is:

[0055]

[0056] Where B represents the channel bandwidth; h i,k Indicates drone u i to u k The channel gain of δ 2 represents the noise power;

[0057] The DNN model dependency edge weight is related to the size of data transmitted between layers. Assuming that layer l j The output of layer l is m The DNN model depends on the edge weights

[0058] Preferably, in step S42), using the graph attention network GAT to encode the drone-DNN model relationship includes the following steps:

[0059] S421), the feature vectors of the drone node and the DNN layer node Project them into high-dimensional space respectively;

[0060] S422) Calculate the association weights between nodes through a multi-head attention mechanism, and each attention head independently learns the interaction relationships of different dimensions;

[0061] S423) Generate the probability matrix P of each drone to the DNN layer i,j ,Right now:

[0062]

[0063] In the formula, Softmax represents the normalization function; and Represent the features of DNN layer nodes and drone nodes respectively.

[0064] Preferably, in step S43), dynamically adjusting the layer allocation strategy based on the proximal end optimization strategy PPO specifically includes the following steps:

[0065] S431), generating a layer allocation plan according to the current strategy;

[0066] S432), calculating rewards and constraint violation penalties;

[0067] S433), optimizing the policy network by gradient descent.

[0068] As a preference, in step S431), according to the distribution probability P i,j , using the maximum probability selection or sampling strategy to generate the allocation result z i,j :

[0069]

[0070] Finally, the complete layer allocation scheme matrix [z i,j ], as the execution decision under the current strategy, and used for subsequent reward calculation and strategy optimization.

[0071] Preferably, in step S432), the calculation formula of the reward R is:

[0072]

[0073] Where, T totalrepresents the total task delay; λ1 and λ2 represent weight coefficients; σ(E) represents the standard deviation of the remaining energy of the UAV; Comm_Cost represents the total communication cost;

[0074] The constraints are:

[0075] Calculation load: ∑ j z i,j Comp i ≤C i , where z i,j Represents the lth j Whether the layer is assigned to drone u i ;

[0076] If the computational load exceeds the threshold, the layer tasks are redistributed using the PPO strategy;

[0077] Energy constraint: E i ≥∑ j z i,j ·(Comp i ·η i ), where η i Indicates drone u i The amount of energy required to perform each unit of computing task;

[0078] If the energy is insufficient, it will be allocated first to the drone with high remaining energy;

[0079] Cache capacity: ∑ j Data j Cache i,j ≤1GB; Cache i,j =1 indicates layer 1 j The output is stored in the UAV u i in the cache;

[0080] If the cache capacity exceeds the threshold, the cache is updated.

[0081] The violation penalty mentioned above refers to a mechanism that provides negative feedback or punishment to behaviors or plans that violate the constraints. The purpose of the punishment is to prevent the algorithm from choosing these infeasible plans.

[0082] Preferably, in step S433), the gradient descent formula is:

[0083]

[0084] Where θ represents the parameters of the policy network; η represents the learning rate; Represents the gradient of the loss function with respect to the parameters;

[0085] Preferably, in step S5), DNN reasoning is performed according to the topological network and partitioning scheme, specifically including the following steps:

[0086] S51) Initialize the inference task environment: initialize the computing state, cache state, and communication link state of each drone, and load and allocate the DNN model layer mapping relationship;

[0087] S52) Execute distributed reasoning in the order of DNN model topology: During the reasoning process, the DNN model is executed layer by layer, and each layer is executed in three stages: reasoning time calculation, cross-node data transmission time calculation, and cache and data management;

[0088] S53) Integrate the final inference results: The output of the last layer of the DNN model is generated by the corresponding drone; if the mission goal requires centralized results, such as returning to the vehicle equipment or control center, the results are uploaded via the 5G link;

[0089] S54), reasoning log and resource status update.

[0090] Preferably, in step S52), the DNN model is executed layer by layer, and the execution of each layer can be divided into three stages: inference time calculation, cross-node data transmission time calculation, and cache and data management, as follows:

[0091] S521) Reasoning calculation time: if the first j Layer is assigned to drone u i , its calculation time T comp (i,j) is:

[0092]

[0093] Among them, Comp j Presentation Layer j The amount of computing resources required, C i For drone u i computing power.

[0094] S522) Cross-node data transmission time: If two adjacent layers l j and l j+1 Assigned to different drones i and u i ′, then the intermediate output needs to be transmitted, and the transmission time is:

[0095]

[0096] Among them, Data j Presentation Layer j The output data size, B i,i′ (t) represents the current moment of UAV u i and ui ′ bandwidth, L prop is a fixed propagation delay; T comm (i,i′,j) represents the communication delay;

[0097] S523) Cache and Data Management: If the cache policy requires it, cache the output locally for subsequent layers to use. If the cache is insufficient, trigger the cache update or cleanup mechanism to ensure the availability and integrity of intermediate data.

[0098] Preferably, in step S53), the calculation formula for recording the overall reasoning completion delay is:

[0099]

[0100] in, Indicates the first j Layer execution drone, Is an indicator function that indicates whether cross-device communication exists.

[0101] Preferably, in step S54), after the inference is completed, the calculation and communication energy consumption of each layer is recorded, including:

[0102] Calculate energy consumption:

[0103] E comp (i,j)=Comp j ·η i ; (15)

[0104] Where η i Calculate energy consumption for the unit.

[0105] Communication energy consumption:

[0106] E comm (t) = P tx ·t tx +P rx ·t rx ; (16)

[0107] Where, P tx Indicates the power consumption of the transmitter; P rx Indicates the power consumption of the receiving end; t rx Indicates the data receiving time; t tx Indicates the transmission time.

[0108] Total energy consumption:

[0109] E total =∑ j E comp (i j ,j)+E comm ; (17)

[0110] The drone’s remaining energy, cache status, communication delay and other information are used for the next round of task scheduling and strategy optimization;

[0111] Preferably, in step S6), each UAV is configured with a limited cache to store intermediate results and task inputs, wherein the cache state is represented by a matrix Cache(t); the cache update based on the priority and energy-aware strategy is specifically as follows:

[0112] S61) Calculate the priority PR of each layer based on the complexity and data size, namely:

[0113] PR j =w1Comp j +w2Data j ; (18)

[0114] Among them, Comp j For layer l j Complexity of Data j For layer l j Output data size; w1 and w2 are weights;

[0115] S62), cache addition and cache elimination, if the cache is not full, directly store the newly allocated layer results; if the cache is full, remove the lowest priority layer;

[0116] j * =arg min PR j ; (19)

[0117] Where, j * The lowest priority layer; PR j is the priority of the jth layer;

[0118] If the remaining energy E i <20%, prioritize retaining high computational load layers;

[0119] S63) Perform status synchronization and broadcast the updated cache status to adjacent drones.

[0120] As a preference, if Cache(t) i,j =1, and requires layer l j , then use cached result Data j , then caching saves time Expressed as:

[0121]

[0122] Where B i,i′ (t) represents the current time t, from the UAV u i Towards the target drone ui′ Communication bandwidth;

[0123] The constraints are:

[0124]

[0125] Where N is the number of layers of the DNN model; Cache(t) max,i (t) is the drone u i The maximum cache capacity at time t.

[0126] As a preference, in step S7), the total delay T for completing task reasoning total (t) includes computational delay and communication delay, namely:

[0127]

[0128] in, Indicates the first j Layer execution drone, Is an indicator function that indicates whether cross-device communication exists.

[0129] In pipeline mode, the calculation delay T compute (t) is determined by the slowest drone and whether there is a cache hit, that is:

[0130]

[0131] Where A i,j (t) represents the lth j Whether the layer is assigned to drone u i ;T cache (t) represents the total computation time saved due to cache hits;

[0132] The communication delay T comm (t) includes uplink, grid and downlink delays, namely:

[0133] T comm (t) = T up (t)+T mesh (t)+T down (t); (24)

[0134] Where, T up (t), T mesh (t), T down (t) are the delays in uplink, grid and downlink stages respectively;

[0135] in,

[0136]

[0137] Where B 5g (t) represents the 5G network bandwidth at time t; D req Indicates the amount of uplink request data; L prop represents the propagation delay;

[0138]

[0139] Where K represents the number of clustered drones; T i,i′ (t) represents the drone u i With drone u i′ The transmission delay factor between i,i′ (t) represents the drone u i to droneu i′ Communication bandwidth between Indicates drone u i to droneu i′ Output data of

[0140]

[0141] Where D result Indicates the amount of downlink result data;

[0142] The time saved by the cache is T cache (t) is expressed as:

[0143]

[0144] The beneficial effects of the present invention are:

[0145] 1. This invention can dynamically allocate computing tasks among multiple drones, avoiding excessive computing load on a single drone. It can flexibly adjust the task and data allocation strategy according to network status and environmental changes, ensuring that even in unstable network conditions, the real-time and efficient performance of tasks can be guaranteed.

[0146] 2. This invention's dynamic DAG topology generation method, based on the potential field method, drives drones to gather in high-value areas through the task attraction field. It also utilizes the repulsive force field between drones to avoid collisions and interference, building stable communication links in real time. It uses a scoring function to screen high-potential nodes, generating a directed acyclic graph that minimizes latency, and unifying task input sources through a virtual root node. This method can improve communication link stability and reduce end-to-end latency.

[0147] 3. This invention introduces a discontinuous layer allocation mechanism and dynamic programming, allowing DNN discontinuous layers to be dynamically allocated to different drones. The load distribution is optimized based on a real-time scoring function related to distance, computing power, and energy status, thereby improving the task completion rate and reducing energy consumption.

[0148] 4. The present invention dynamically sorts cache contents by combining priority indicators of computational complexity and data size, giving priority to retaining high-value layer results; improves the cache hit rate in dynamic task scenarios, and significantly reduces the total system latency and energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0149] Figure 1 Schematic diagram of the process of the present invention;

[0150] Figure 2 It is a flow chart of the method of the present invention;

[0151] Figure 3 A schematic diagram of the topology generated by the present invention;

[0152] Figure 4 Schematic diagram of the partitioning of the DNN model of the present invention. DETAILED DESCRIPTION

[0153] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0154] like Figure 1 and 2 As shown, this embodiment provides a multi-UAV collaborative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning, including the following steps:

[0155] S1), initialize the UAV and mission vehicle;

[0156] S2), when receiving the task request of the ground vehicle, determine whether the cache is hit, if not; then execute step S3);

[0157] S3) Dynamically adjust the position of the drone based on the potential field method, and build a topological network by combining greedy clustering and DAG topology generation strategy;

[0158] In this embodiment, the potential field combines task attraction and inter-drone repulsion. Task attraction drives drones toward high-value areas, while inter-drone repulsion prevents collisions and interference. Drone trajectories strictly follow the motion control equations for adjustment and movement. Meanwhile, drone energy consumption dynamically changes with the progress of communication activities and changes in their motion state. A scoring function for drone clustering is combined to screen high-potential nodes, generating a directed acyclic graph that minimizes latency. Task input sources are unified through a virtual root node.

[0159] In this embodiment, the task attraction F attract (d) is the gravitational field generated by the vehicle mission request point, the intensity of which is positively correlated with the mission priority, namely:

[0160]

[0161] Where k task is the attraction coefficient; d is the distance between the UAV and the vehicle; R task The mission impact radius;

[0162] The repulsive force F between drones repel (d) Avoid collision and communication interference, the expression is:

[0163]

[0164] Where k uav is the repulsive force coefficient; d safe For a safe distance.

[0165] Under the action of the potential field, the motion control equation of the UAV is expressed as:

[0166]

[0167] Where, For the mission machine u i Velocity vector at time t+1; It represents the combined force of the UAV's attraction force, obstacle repulsion force, and repulsion force between UAVs; a and b are weight coefficients;

[0168] The best speed in history.

[0169] In this embodiment, each drone u i With heterogeneous computing and communication capabilities, each drone u i Energy E i (t) Dynamic changes, its energy consumption includes motion energy consumption E move (t) and communication energy consumption E comm (t), where the UAV motion energy consumption E move (t) Produced by flight attitude adjustments and displacements, namely:

[0170]

[0171] Where, Indicates the change of kinetic energy; P hov Δt represents the hovering power consumption; m i UAV i The quality of |v i (t)| is the drone u h The speed of P hov represents the hovering power; Δt represents the time step; when |v i When (t)|<0.1, only hovering power consumption is calculated;

[0172] The communication energy consumption E comm (t) includes the energy consumption of data sending and receiving, namely:

[0173] E comm (t) = P tx ·t tx +P rx ·t rx ; (5)

[0174] Where, P tx Indicates the power consumption of the transmitter; P rx Indicates the power consumption of the receiving end; t rx Indicates the data receiving time; t tx Indicates the transmission time.

[0175] Among them, the transmission time t tx Determined by the data volume D and bandwidth B, that is:

[0176]

[0177] Where N0 represents the noise power spectral density; h represents the channel gain.

[0178] In this embodiment, the scoring function Score(u i ) is defined as:

[0179]

[0180] Wherein, α, β, and γ are weight coefficients; in this embodiment, α = 0.4, β = 0.4, and γ = 0.2;

[0181] P tx,i For drone u i The transmission power of R comm Indicates the communication radius; C i For drone u i The computing power of E i For drone u i The remaining energy P max is the maximum transmission power that can be achieved by all drones; C max The maximum computing power that all drones may have; E max The maximum battery capacity set for all drones.

[0182] Then, the drones are clustered and screened according to their cluster scores. The specific process is as follows:

[0183] First, a preliminary spatial screening was conducted, selecting drones within a 200-meter radius of the mission point as candidates;

[0184] Next, connectivity verification is carried out to construct a fully connected candidate set and remove isolated nodes to ensure good communication connectivity between candidate drones.

[0185] Finally, the candidate drones are ranked by the greedy algorithm according to the scoring function Score(u i ) Select the first K drones in descending order as the clustering results.

[0186] The goal of the DAG topology generation strategy is to build a directed acyclic graph to minimize end-to-end delay, such as Figure 3 As shown in the figure, the DAG topology generation strategy is used to build a topology network, which specifically includes the following steps:

[0187] S31) According to the selected UAV cluster set U={u1,u2,…,u K}, get the position and transmission power P of the corresponding drone tx,i and computing power C i , and through computing power C i Sort the drones in descending order;

[0188] S32), for each drone u i , select the drone within its communication range and ranked first as the candidate parent node;

[0189] S33) Select the first two parent nodes with the smallest weight according to the link weight to form edge e i,k ;

[0190] S34), add a virtual root node v0, connect it to all nodes without incoming edges, and use the virtual root node v0 as the input source of the task;

[0191] S35), prune redundant edges, delete edges that may cause cycles, and obtain a directed acyclic graph.

[0192] S4) Dynamically partition the DNN model based on the discontinuous layer allocation mechanism and the dynamic programming algorithm DPA, and dynamically allocate the DNN discontinuous layers to different drones;

[0193] The DNN model partitioning includes inter-layer partitioning and intra-layer partitioning. Inter-layer partitioning allocates different layers of the DNN model to devices at different stages, while intra-layer partitioning further subdivides the computing tasks of the same layer into multiple subtasks and allocates them to different devices at the same stage for parallel computing according to the computing power ratio of the devices. Figure 4 As shown, the specific steps include:

[0194] S41), constructing a heterogeneous graph including drone nodes and DNN layer nodes based on the drone topology network;

[0195] The UAV nodes in the heterogeneous graph are Its eigenvector is Among them, P i is the transmission power; loc i is the position coordinate;

[0196] The DNN layer nodes are represented as Its eigenvector Among them, Comp i For layer l i Complexity of Data i For layer l i Output data size;

[0197] The edge types of the heterogeneous graph include drone communication edges and DNN dependency edges Among them, UAV communication edge The link weights in the topological network Determine the link weight Indicates that from the drone u i to droneu k The link communication capability, the link weight The calculation formula is:

[0198]

[0199] Where B represents the channel bandwidth; h i,k Indicates drone u i to u k The channel gain of δ 2 represents the noise power;

[0200] The DNN dependency edge weight is related to the size of the data transmitted between layers. Assuming that layer l j The output of layer l is m The input of DNN depends on the edge weight

[0201] S42) Using the graph attention network (GAT) to encode the UAV-DNN relationship, and generating the optimal allocation probability between the UAV and the DNN layer through the node features and topological relationships of the heterogeneous graph; including the following steps:

[0202] S421), the feature vectors of the drone node and the DNN layer node Project them into high-dimensional space respectively;

[0203] S422) Calculate the association weights between nodes through a multi-head attention mechanism, and each attention head independently learns the interaction relationships of different dimensions;

[0204] In this embodiment, the attention weight is calculated based on the query-key value mechanism, and the query-key value calculation is expressed as:

[0205] q i =W q h i ;

[0206] k j =W k h j ;

[0207] Where h i 、h j represents the node feature vector; W q 、W k represents the trainable parameter matrix; q i 、k j Represents the query,key value of the node;

[0208] The LeakyReLU activation function and Softmax normalization are used to ensure that the weights can capture local feature differences while being globally coordinated.

[0209] S423) Generate the probability matrix P of each drone to the DNN layer i,j ,Right now:

[0210]

[0211] In the formula, Softmax represents the normalization function; and Represent the features of DNN layer nodes and drone nodes respectively.

[0212] S43) Dynamically adjust the layer allocation strategy based on the proximal optimization strategy PPO; specifically including the following steps:

[0213] S431), generating a layer allocation plan according to the current strategy;

[0214] According to the distribution probability P i,j , using maximum probability selection or sampling strategy to generate allocation results:

[0215]

[0216] Finally, the complete layer allocation scheme matrix [z i,j ], as the execution decision under the current strategy, and used for subsequent reward calculation and strategy optimization.

[0217] S432), calculating rewards and constraint violation penalties;

[0218] The calculation formula of the reward R is:

[0219]

[0220] Where, T total represents the total delay in completing the task; λ1 and λ2 represent weight coefficients; σ(E) represents the standard deviation of the remaining energy of the UAV; Comm_Cost represents the sum of the communication costs; where, Indicates the main reward; represents energy balance; λ2·∑Comm_Cost is the communication penalty.

[0221] The constraints are:

[0222] Calculation load: ∑ j z i,j Comp i ≤C i , where z i,j Represents the lth j Whether the layer is assigned to drone u i ;

[0223] If the computational load exceeds the threshold, the layer tasks are redistributed using the PPO strategy;

[0224] Energy constraint: E i ≥∑ j z i,j ·(Comp i ·η i ), where η i Indicates drone u i The amount of energy required to perform each unit of computing task;

[0225] If the energy is insufficient, it will be allocated first to the drone with high remaining energy;

[0226] Cache capacity: ∑ j Data j Cache i,j ≤1GB; Cache i,j =1 indicates layer 1 j The output is stored in the UAV u i in the cache;

[0227] If the cache capacity exceeds the threshold, the cache is updated.

[0228] The violation penalty mentioned above refers to a mechanism that provides negative feedback or punishment to behaviors or plans that violate the constraints. The purpose of the punishment is to prevent the algorithm from choosing these infeasible plans.

[0229] S433), optimizing the policy network by gradient descent;

[0230] The gradient descent formula is:

[0231]

[0232] Where θ represents the parameters of the policy network; η represents the learning rate; Represents the gradient of the loss function with respect to the parameters;

[0233] S5) Perform DNN reasoning based on the topological network and DNN model partitioning scheme; specifically, the steps include:

[0234] S51) Initialize the reasoning task environment.

[0235] Initialize the computing state, cache state, and communication link state of each drone, and load and allocate the DNN model layer mapping relationship, that is, which drone is responsible for processing each layer of tasks;

[0236] S52) Execute distributed reasoning in the order of DNN model topology.

[0237] During the inference process, the DNN model is executed layer by layer, and the execution of each layer can be divided into two stages: calculation and data transmission:

[0238] S521) Reasoning calculation time: if the first j Layer is assigned to drone u i , and its calculation time is:

[0239]

[0240] Among them, Comp j Presentation Layer j The amount of computing resources required, C i For drone u i computing power.

[0241] S522) Cross-node data transmission time: If two adjacent layers l j and l j+1 Assigned to different drones i and u i ′, then the intermediate output needs to be transmitted, and the transmission time is:

[0242]

[0243] Among them, Data j Presentation Layer j The output data size, B i,i′ (t) represents the bandwidth at the current moment, L prop is a fixed propagation delay.

[0244] S523) Cache and Data Management: If the cache policy requires it, cache the output locally for subsequent layers to use. If the cache is insufficient, trigger the cache update or cleanup mechanism to ensure the availability and integrity of intermediate data.

[0245] S53) Integrate the final reasoning results.

[0246] The output of the last layer of the DNN model is generated by the corresponding drone. If the mission goal requires centralized results, such as returning them to the vehicle equipment or control center, the results are uploaded via the 5G link. The overall inference completion delay is recorded at this stage:

[0247]

[0248] in, Indicates the first j Layer execution drone, Is an indicator function that indicates whether cross-device communication exists.

[0249] S54), reasoning log and resource status update.

[0250] After inference is completed, the computation and communication energy consumption of each layer is recorded, including:

[0251] Calculate energy consumption:

[0252] E comp (i,j)=Comp j ·η i ; (15)

[0253] Where η i Calculate energy consumption for the unit.

[0254] Communication energy consumption:

[0255] E comm =∑ 传输 (P tx ·t tx +P rx ·t tx ); (16)

[0256] Where, P tx Indicates the power consumption of the transmitter; P rx Indicates the power consumption of the receiving end; t rx Indicates the data receiving time; t tx Indicates the transmission time.

[0257] Total energy consumption:

[0258]

[0259] The drone’s remaining energy, cache status, communication delay and other information are used for the next round of task scheduling and strategy optimization;

[0260] S6), cache update based on priority and energy-aware strategy;

[0261] Each UAV is configured with a limited cache to store intermediate results and task inputs, where the cache state is represented by the matrix Cache(t). The cache update is performed based on the priority and energy-aware strategy, specifically:

[0262] S61) Calculate the priority PR of each layer based on the complexity and data size, namely:

[0263] PR j =w1Comp j +w2Data j ; (18)

[0264] Among them, Comp j For layer l j Complexity of Data j For layer l j Output data size; w1 and w2 are weights;

[0265] S62), cache addition and cache elimination, if the cache is not full, directly store the newly allocated layer results; if the cache is full, remove the lowest priority layer;

[0266] j * =arg min PR j ; (19)

[0267] Where, j * The lowest priority layer; PR j For the first j Priority of the layer;

[0268] If the remaining energy E i <20%, prioritize retaining high computational load layers;

[0269] S63) Perform status synchronization and broadcast the updated cache status to adjacent drones.

[0270] In this embodiment, if Cache(t) i,j =1, and requires layer l j , then use cached result Data j , then caching saves time Expressed as:

[0271]

[0272] Where B i,i′(t) represents the current time t, from the UAV u i Towards the target drone u i′ Communication bandwidth;

[0273] The constraints are:

[0274]

[0275] Where N is the number of layers of the DNN model; Cache(t) max,i (t) is the drone u i The maximum cache capacity at time t.

[0276] S7) Calculate delay and overhead.

[0277] The total delay T to complete task reasoning total (t) includes computational delay and communication delay, namely:

[0278]

[0279] in, Indicates the first j Layer execution drone, Is an indicator function that indicates whether cross-device communication exists.

[0280] In pipeline mode, the calculation delay T compute (t) is determined by the slowest drone and whether there is a cache hit, that is:

[0281]

[0282] Where A i,j (t) represents the lth j Whether the layer is assigned to drone u i ;T cache (t) represents the total computation time saved due to cache hits;

[0283] The communication delay t comm (t) includes uplink, grid and downlink delays, namely:

[0284] T comm (t) = T up (t)+T mesh (t)+T down (t); (24)

[0285] Where, T up (t), T mesh (t), T down (t) are the delays in uplink, grid and downlink stages respectively;

[0286] in,

[0287]

[0288] Where B 5g (t) represents the 5G uplink bandwidth from the vehicle to the UAV at the current time t; D req Indicates the size of the uplink task request data; L prop represents the propagation delay;

[0289]

[0290] Where K represents the number of clustered drones; T i,i′ (t) represents the drone u i With drone u i′ The transmission delay factor between i,i′ (t) represents the drone u i to droneu i′ Communication bandwidth between Indicates drone u i to droneu i′ Output data of

[0291]

[0292] Where D result The amount of data representing the final inference result;

[0293] The time saved by the cache is T cache (t) is expressed as:

[0294]

[0295] The above embodiments and descriptions are only for explaining the principles and best embodiments of the present invention. Without departing from the spirit and scope of the present invention, the present invention may be subject to various changes and improvements, which shall fall within the scope of the invention to be protected.

Claims

1. A multi-UAV collaborative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning, characterized by: The steps include: S1), initialize the UAV and mission vehicle; S2), when receiving the task request of the ground vehicle, determine whether the cache is hit, if not; then execute step S3); S3) Dynamically adjust the position of the drone based on the potential field method, and build a topological network by combining greedy clustering and DAG topology generation strategy; S4) Dynamically partition the DNN model based on the discontinuous layer allocation mechanism and the dynamic programming algorithm DPA, and dynamically allocate the DNN discontinuous layers to different drones; S5), performing DNN reasoning based on the topological network and DNN model partitioning scheme; S6), cache update based on priority and energy-aware strategy; S7) Calculate delay and overhead.

2. The multi-UAV collaborative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning according to claim 1 is characterized by: In step S3), the potential field includes the task attraction and the repulsion between drones. The task attraction drives the drones to gather in high-value areas, while the repulsion between drones is used to avoid collisions and interference. The scoring function of drone clustering is combined to screen high-potential nodes, generate a directed acyclic graph that minimizes delay, and unify the task input source through a virtual root node.

3. The multi-UAV collaborative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning according to claim 2 is characterized by: In step S3), the scoring function Score(u i ) is defined as: In the formula, α, β, and γ are weight coefficients; P tx,i For drone u i The transmission power of R comm Indicates the communication radius; C i For drone u i The computing power of E i For drone u i The remaining energy P max is the maximum transmission power that can be achieved by all drones; C max The maximum computing power that all drones may have; E max The maximum battery capacity set for all drones.

4. The multi-UAV collaborative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning according to claim 3 is characterized by: In step S3), the drones are clustered and screened according to their cluster scores. The specific process is as follows: First, a preliminary spatial screening is performed to select UAVs within a radius of L meters from the mission point as candidates; Next, connectivity verification is carried out to construct a fully connected candidate set and remove isolated nodes to ensure good communication connectivity between candidate drones. Finally, the candidate drones are ranked by the greedy algorithm according to the scoring function Score(u i ) Select the first K drones in descending order as the clustering results.

5. The multi-UAV collaborative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning according to claim 4 is characterized by: In step S3), a topology network is constructed using a DAG topology generation strategy, specifically including the following steps: S31) According to the selected UAV cluster set U={u1,u2,…,u K }, get the corresponding drone u i Position, transmission power P tx,i and computing power C i , and through computing power C i Sort the drones in descending order; S32), for each drone u i , select the drone within its communication range and ranked first as the candidate parent node; S33) Select the first two parent nodes with the smallest weight according to the link weight to form edge e i,k ; S34), add a virtual root node v0, connect it to all nodes without incoming edges, and use the virtual root node v0 as the input source of the task; S35), prune redundant edges, delete edges that may cause cycles, and obtain a directed acyclic graph.

6. The multi-UAV collaborative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning according to claim 1 is characterized by: In step S4), the DNN model partitioning specifically includes the following steps: S41), constructing a heterogeneous graph including drone nodes and DNN layer nodes based on the drone topology network; S42) Using the graph attention network GAT to encode the relationship between the drone and the DNN model, the optimal allocation probability between the drone and the DNN layer is generated through the node features and topological relationships of the heterogeneous graph; S43) Dynamically adjust the layer allocation strategy based on the proximal optimization strategy PPO.

7. The multi-UAV collaborative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning according to claim 6 is characterized by: In step S42), the relationship between the drone and the DNN model is encoded using the graph attention network (GAT), including the following steps: S421), the feature vectors of the drone node and the DNN layer node Project them into high-dimensional space respectively; S422) Calculate the association weights between nodes through a multi-head attention mechanism, and each attention head independently learns the interaction relationships of different dimensions; S423) Generate the probability matrix P of each drone to the DNN layer i,j ,Right now: In the formula, Softmax represents the normalization function; and Represent the features of DNN layer nodes and drone nodes respectively.

8. The multi-UAV collaborative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning according to claim 6 is characterized by: In step S43), the layer allocation strategy is dynamically adjusted based on the proximal end optimization strategy PPO, specifically including the following steps: S431), generating a layer allocation plan according to the current strategy; S432), calculating rewards and constraint violation penalties; S433), optimizing the policy network by gradient descent.

9. The multi-UAV collaborative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning according to claim 1 is characterized by: In step S6), each UAV is configured with a limited cache to store intermediate results and task inputs, wherein the cache state is represented by the matrix Cache(t); the cache update based on the priority and energy-aware strategy is specifically as follows: S61) Calculate the priority PR of each layer based on the complexity and data size, namely: PR j =w1 Comp j +w2Data j ; (18) Among them, Comp j For layer l j Complexity of Data j For layer l j Output data size; w1 and w2 are weights; S62), cache addition and cache elimination, if the cache is not full, directly store the newly allocated layer results; if the cache is full, remove the lowest priority layer; j * =arg min PR j ; (19) Where, j * The lowest priority layer; PR j For the first j Priority of the layer; If the remaining energy E i <20%, prioritize retaining high computational load layers; S63) Perform status synchronization and broadcast the updated cache status to adjacent drones.

10. The multi-UAV collaborative reasoning optimization method based on DAG topology generation and dynamic DNN model partitioning according to claim 1 is characterized by: In step S7), the total delay T for completing task reasoning is total (t) includes computational delay and communication delay, namely: in, Indicates the first j Layer execution drone, is an indicator function, indicating whether there is cross-device communication; In pipeline mode, the calculation delay T compute (t) is determined by the slowest drone and whether there is a cache hit, that is: Where A i,j (t) represents the layer allocation matrix; T cache (t) represents the time saved by cache hit; N represents the number of layers of the DNN model; Comp j Presentation Layer j complexity; The communication delay T comm (t) includes uplink, grid and downlink delays, namely: T comm (t)=T up (t)+T mesh (t)+T down (t); (24) Where, T up (t), T mesh (t), T down (t) are the delays in uplink, grid and downlink stages respectively; in, Where B 5g (t) represents the 5G uplink bandwidth from the vehicle to the UAV at the current time t; D req Indicates the size of the uplink task request data; L prop represents the propagation delay; Where K represents the number of clustered drones; T i,i′ (t) represents the drone u i With drone u i′ The transmission delay factor between i,i′ (t) represents the distance from the drone u i to u i′ Communication bandwidth; Indicates drone u i to droneu i′ Output data of Where D result The amount of data representing the final inference result; The time saved by the cache is T cache (t) is expressed as: Where, Save time for caching; C i (t) is the drone u i Data j For layer l j Output data size.

Images