Air-ground cooperative unmanned aerial vehicle online trajectory planning method for distributed model training

Abstract

The application discloses an air-ground cooperative unmanned aerial vehicle online trajectory planning method for distributed model training, which comprises a distributedly arranged back-end edge server providing computing power support for a large-scale unmanned aerial vehicle cluster, fusing local neural network model parameters of each unmanned aerial vehicle, and generating a global model through online local training and multi-unmanned aerial vehicle neural network parameter fusion; and a reinforcement learning method based on a full connection neural network, which utilizes a distributed multi-target joint trajectory planning method, at the beginning of each time slot, according to a long-term average transmission rate and a long-term average task completion utility, through information interaction with an environment, the unmanned aerial vehicle cluster learns an optimal strategy of joint multi-unmanned aerial vehicle online trajectory planning and local model parameter uploading. The method has the advantages of being capable of performing efficient and reliable distributed model training in a highly dynamic urban environment, and reducing neural network model precision loss under the condition that local data sets of each unmanned aerial vehicle are biased.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically, it relates to an online trajectory planning method for air-ground cooperative unmanned aerial vehicles (UAVs) oriented towards distributed model training. Background Technology

[0002] With the explosive growth in the number of smartphones, wearable devices, and IoT devices, most data is now generated outside the cloud, especially on distributed edge devices. In this distributed network environment, traditional machine learning algorithms deploy neural network training on a single central server. This leads to significant data exchange between data sources and the central server, causing time delays and increasing device power consumption, thus raising communication costs. Training a machine learning model, such as a deep neural network, on various data-generating devices, with model parameters fused by a nearby edge server, not only saves communication costs but also significantly improves information security by preventing the server from handling the raw data. Drones, as aerial mobile devices, offer advantages such as flexibility, mobility, line-of-sight capability, and some computing power, making them suitable for data collection and machine learning model training. Therefore, drone collaboration is a current solution for distributed model training.

[0003] While existing technologies have optimized the training efficiency of distributed model training, they have not conducted extensive research on UAV trajectory optimization and online model training. UAV air-to-ground collaborative scenarios often face complex external environments, making real-time UAV trajectory optimization necessary. Furthermore, given the insufficient initial experience samples for UAVs, it is often necessary for the UAV to cruise within the mission area while simultaneously performing tasks such as target recognition during online training. Summary of the Invention

[0004] In view of this, the present invention proposes an online trajectory planning method for air-ground cooperative UAVs for distributed model training. This method has the advantages of being able to perform efficient and reliable distributed model training in highly dynamic urban environments and reducing the accuracy loss of neural network models under biased local datasets of each UAV.

[0005] To achieve the above technical objectives, the technical solution adopted by this invention is: an online trajectory planning method for air-ground cooperative UAVs oriented towards distributed model training, specifically including the following steps:

[0006] Step 1: Deploy N backend edge nodes and K drones at a fixed flight altitude in the air-ground collaborative edge learning network. Deploy the same neural network model on each drone and initialize the neural network model parameters.

[0007] Step 2: Obtain the status information s of each UAV in the current time slot t.t This includes: the location of the UAV in the current time slot t and the upload status of the UAV neural network model parameters in the previous time slot;

[0008] Step 3: Each drone uses an ∈-greedy strategy to select from a preset set of flight actions. Select a flight action a t,k,1 Fly a fixed distance v; and upload a set of strategies from preset neural network model parameters. The system determines the neural network model parameter upload strategy for each drone. t,k,2 , will [a t,k,1 ,a t,k,2 [This refers to the combined actions of UAV k in the current time slot t;]

[0009] Step 4: Upload strategy a based on the neural network model parameters of each drone. t,k,2 By fusing neural network model parameters, the model parameters of the neural network model can be reconstructed;

[0010] Step 5: Update the position of each drone and obtain the corresponding reward value, and record the status information s of the current time slot t. t The combined actions of UAV k in the current time slot t [a] t,k,1 ,a t,k,2 ], reward value, and the state of the next time slot. t+1 As the experience of the current time slot t, it is stored in the experience pool;

[0011] Step 6: Extract experience from the experience pool and update the parameters of the fully connected neural network model using the Adam optimizer;

[0012] Step 7: Repeat steps 2-6 for each time slot until the fully connected neural network model converges, obtaining the optimal flight maneuver and the optimal neural network model parameter upload strategy for the corresponding time slot.

[0013] Furthermore, step 3 includes the following sub-steps:

[0014] Step 3.1: Construct a set of quantized states Store the quantized state information of the UAV and construct matrix T. k Record the quantized state information of UAV k's access - combined action pairs; the matrix T k The rows represent different state information of drone k, with 1 indicating state information experienced by drone k and 0 indicating state information not experienced by drone k; the matrix T k The column represents the combined actions of drone k, with 1 indicating that drone k has experienced a combined action and 0 indicating that drone k has not experienced a combined action.

[0015] Step 3.2: Using a greedy strategy, the UAV maximizes two objectives with probability 1-1: maximizing the average transmission rate between each UAV and the backend edge node, and the average recognition task completion efficiency, from a preset set of flight actions. Select a flight action a t,k,1 Fly a fixed distance v; and upload a set of strategies from preset neural network model parameters. The system determines the neural network model parameter upload strategy for each drone. t,k,2 , will [a t,k,1 ,a t,k,2 [T is the combined action update matrix of drone k] k The drone randomly selects from a pre-defined set of flight actions with probability ∈ R. Choose an unexplored flight action a t,k,1 Fly a fixed distance v; and upload a set of strategies from preset neural network model parameters. A neural network model parameter upload strategy for selecting an unexplored drone. t,k,2 , will [a t,k,1 ,a t,k,2 [T is the combined action update matrix of drone k] k Select an unexplored action.

[0016] Furthermore, the process of storing the quantified UAV state information in step 3.1 is as follows: if the UAV state information s in the current time slot t is... t With quantized state set If the distance between all elements of a certain quantized state information is less than the corresponding threshold, then s t Do not save; otherwise, save s t Stored to quantized state set middle.

[0017] Furthermore, the process in step 3.2 to maximize the two objectives—average transmission rate and average identification task completion efficiency—is as follows:

[0018] F max =w tr ·F tr,k (s t ,a t,k,1 ,a t,k,2 )+w d ·F d,k (s t ,a t,k,1 )

[0019] Among them, F max F represents the result of maximizing two objectives: the average transmission rate between each drone and the backend edge node, and the average utility of completing the identification task. tr,k (s t ,a t,k,1 ,a t,k,2 This represents the average transmission rate between all drones and the backend edge nodes. a t,k,2 R represents the flight action of the k-th UAV in the current time slot t. k,n Let a represent the transmission rate between the nth backend edge node and the kth drone. t,k,1 w represents the neural network model parameter upload strategy for the k-th UAV in the current time slot t. tr F represents tr,k (s t ,a t,k,1 ,a t,k,2 The weights of F d,k (s t ,a t,k,1 This represents the average identification task completion utility between all drones and backend edge nodes.

[0020] Furthermore, the R k,n for:

[0021]

[0022] Among them, PL k,n PL represents the path loss between the nth backend edge node and the kth drone during transmission. k,n =P k,n (LoS,θ)L LoS +(1-P k,n (LoS,θ))L NLoS P k,n (LoS,θ) represents the probability that the channel between the k-th UAV and the n-th backend edge node is a line-of-sight channel, L LoS This represents the path loss of the line-of-sight channel. η0 represents the additional path loss of the line-of-sight channel, d represents the distance between the UAV and the base station, λ represents the signal wavelength, and L NLoS This represents the path loss of a non-line-of-sight channel. η1 represents the additional path loss of the non-line-of-sight channel, P k Let M represent the signal power of the k-th UAV transmitter, and M represent the channel noise power.

[0023] Furthermore, the F d,k (s t ,a t,k,1 )for:

[0024]

[0025] Where, d k,s This represents the distance between drone k and the corresponding mission point s, (x s ,y s Let (x) be the coordinates of the task point s, and (x) be the coordinates of the point s. t,k ,y t,k Let Δx be the coordinates of UAV k in time slot t. t,k ,Δy t,k These represent the flight distances of the UAV along the x-axis and y-axis, respectively, in time slot t.

[0026] Furthermore, step 4 includes the following sub-steps:

[0027] Step 4.1, if the UAV neural network model parameter upload strategy a t,k,2 To continue training the neural network model, each UAV collects labeled image data in real time while cruising in the air-ground collaborative edge learning network, trains its own neural network model using gradient descent, and sends the model parameters from the trained neural network model to the backend edge nodes within the communication range of the UAV; otherwise, the model parameters from the neural network model are directly sent to the backend edge nodes within the communication range of the UAV.

[0028] Step 4.2: Each backend edge node will fuse the received model parameters and broadcast the fused model parameters to the drones associated with the backend edge nodes;

[0029] Step 4.3: Each UAV reconstructs the model parameters of the neural network model based on the received fused model parameters.

[0030] Furthermore, drones that completed uploading neural network model parameters in the previous time slot in step 4.1 will not upload neural network model parameters again in the current time slot.

[0031] Furthermore, the model parameter fusion process in step 4.2 is as follows:

[0032]

[0033] Where, ω es,n ω' represents the model parameters fused for the current nth backend edge node. es,n N represents the model parameters fused from the nth backend edge node during the previous training round. n N represents the sum of the number of local training samples of the drone participating in training on the nth backend edge node before the current model parameters are fused. n' represents the sum of the number of local training samples of the drone participating in training on the nth backend edge node before the previous round of model parameter fusion. This represents the number of communications between the nth backend edge node and the drone, where k represents... index, ω local,k This indicates that the nth backend edge node receives the model parameters of the kth drone, ω' local,k This indicates that the nth backend edge node received the model parameters from the kth drone during the previous training round, where n... k n represents the current number of training samples for drone k. k ' represents the number of training samples in the previous training round for drone k.

[0034] Furthermore, the model parameters of the reconstructed neural network model in step 4.3 are as follows:

[0035]

[0036] in, The model parameters represent the reconstructed neural network model. This represents the number of backend edge nodes associated with the k-th drone, where n represents... index, ω es,n This represents the model parameters for the current nth backend edge node fusion.

[0037] Compared with existing technologies, this invention has the following advantages: The online trajectory planning method for air-ground cooperative UAVs oriented towards distributed model training utilizes distributed backend edge nodes to provide computing power support for large-scale UAV clusters. This allows for the fusion of local neural network model parameters from each UAV, enabling each UAV to solve the problem of biased datasets in unfamiliar environments. Through online local training and multi-UAV neural network parameter fusion, a global model is generated locally on each UAV, achieving robust and rapid target recognition by UAVs in multiple regions. Parallel to distributed model collaborative training, reinforcement learning algorithms are also used to plan the UAV local model parameter upload strategy and the trajectories of each UAV in real-time online. This invention addresses the communication and data distribution challenges of performing distributed model training in highly dynamic urban environments, specifically targeting multi-UAV collaborative edge learning scenarios. It can be used for online trajectory planning and local model parameter upload strategies in scenarios involving multi-UAV and distributed backend edge node collaborative training of distributed neural network models. Attached Figure Description

[0038] Figure 1 This is a flowchart of the online trajectory planning method for air-ground cooperative UAVs oriented towards distributed model training according to the present invention;

[0039] Figure 2This is a schematic diagram illustrating a scenario for the air-ground cooperative UAV online trajectory planning method for distributed model training according to the present invention.

[0040] Figure 3 This is a flowchart of the operation of the UAV in step 4 of the present invention;

[0041] Figure 4 This is a flowchart of the operation of the backend edge node in step 4 of the present invention. Detailed Implementation

[0042] The technical solution of the present invention will be further explained and described below with reference to the accompanying drawings.

[0043] This invention presents an online trajectory planning method for air-ground cooperative UAVs oriented towards distributed model training. It utilizes distributed backend edge nodes to provide computing power support for large-scale UAV swarms, thereby fusing parameters from the local models of each UAV. Through online local training and multi-UAV neural network parameter fusion, a global model is generated locally on each UAV, enabling robust and rapid target identification by UAVs across multiple regions. Simultaneously, it employs a fully connected neural network reinforcement learning method, utilizing a distributed multi-target joint trajectory planning approach. Through information interaction with the environment, the UAV swarm learns the optimal strategy for joint online trajectory planning and local model parameter uploading among multiple UAVs. Figure 1 The online trajectory planning method for air-to-ground cooperative UAVs specifically includes the following steps:

[0044] Step 1: Deploy N backend edge nodes and K drones at a fixed flight altitude in the air-ground collaborative edge learning network. Deploy the same neural network model on each drone and initialize the neural network model parameters.

[0045] Step 2: Obtain the status information s of each UAV in the current time slot t. t This includes: the location of the UAV in the current time slot t and the upload status of the UAV neural network model parameters in the previous time slot;

[0046] Step 3: Each drone uses an ∈-greedy strategy to select from a preset set of flight actions. Select a flight action a t,k,1 Fly a fixed distance v; and upload a set of strategies from preset neural network model parameters. The system determines the neural network model parameter upload strategy for each drone. t,k,2 , will [a t,k,1 ,a t,k,2As the combined actions of UAV k in the current time slot t; a greedy strategy can maximize the state-action value at each step to achieve local optimum, while adding the probability of ∈ to explore new states and actions can achieve global optimum. Specifically, it includes the following sub-steps:

[0047] Step 3.1: Construct a quantized state set S to store the quantized UAV state information, and construct a matrix T. k Record the quantized state information of UAV k's access - combined action pairs; the matrix T k The rows represent different state information of drone k, with 1 indicating state information experienced by drone k and 0 indicating state information not experienced by drone k; the matrix T k The column represents the combined actions of drone k, with 1 indicating a combined action experienced by drone k and 0 indicating a combined action not experienced by drone k; the process of storing the quantified state information of drone in this invention is as follows: if the state information of drone s in the current time slot t is... t If the distance between any element and a certain quantized state information in the quantized state set S is less than the corresponding threshold, then s t Do not save; otherwise, save s t Store it in the quantized state set S.

[0048] Step 3.2: Using a greedy strategy, the UAV maximizes two objectives with probability 1-1: maximizing the average transmission rate between each UAV and the backend edge node, and the average recognition task completion efficiency, from a preset set of flight actions. Select a flight action a t,k,1 Fly a fixed distance v; and upload a set of strategies from preset neural network model parameters. The system determines the neural network model parameter upload strategy for each drone. t,k,2 , will [a t,k,1 ,a t,k,2 [T is the combined action update matrix of drone k] k The drone randomly selects from a pre-defined set of flight actions with probability ∈ R. Choose an unexplored flight action a t,k,1 Fly a fixed distance v; and upload a set of strategies from preset neural network model parameters. A neural network model parameter upload strategy for selecting an unexplored drone. t,k,2 , will [a t,k,1 ,a t,k,2 [T is the combined action update matrix of drone k] k Select an unexplored action.

[0049] This invention addresses the issues of local neural network model parameter uploading strategy and multi-drone trajectory planning in drone swarms by maximizing two objectives: the average transmission rate between each drone and the backend edge node, and the average recognition task completion utility. Specifically, to determine the flight direction and local model uploading strategy of drone k in time slot t+1, drone k must depend on the observed state (i.e., the positions of each drone in time slot t) and the local model parameter uploading status. Therefore, the local model parameter uploading strategy and multi-drone trajectory planning problem become independent Markov process decision problems. To overcome the curse of dimensionality, a distributed multi-agent model can be used. In this invention, "distributed" means that the action selection of all drones is performed in a distributed and parallel manner. Each drone's decision is only for the current drone's local model uploading strategy and trajectory planning, and the local model uploading strategies and trajectory planning among all drones are independent of each other. The process of maximizing the two objectives of average transmission rate and average recognition task completion utility between each drone and the backend edge node in this invention is as follows:

[0050] F max =w tr ·F tr,k (s t ,a t,k,1 ,a t,k,2 )+w d ·F d,k (s t ,a t,k,1 )

[0051] Among them, F max F represents the result of maximizing two objectives: the average transmission rate between each drone and the backend edge node, and the average utility of completing the identification task. tr,k (s t ,a t,k,1 ,a t,k,2 This represents the average transmission rate between all drones and the backend edge nodes. a t,k,2 R represents the flight action of the k-th UAV in the current time slot t. k,n This represents the transmission rate between the nth backend edge node and the kth drone. PL k,n PL represents the path loss between the nth backend edge node and the kth drone during transmission. k,n =P k,n (LoS,θ)L LoS +(1-P k,n (LoS,θ))L NLoS P k,n(LoS,θ) represents the probability that the channel between the k-th UAV and the n-th backend edge node is a line-of-sight channel, which depends only on the k-coordinate of the UAV and the urban environment; L LoS This represents the path loss of the line-of-sight channel. η0 represents the additional path loss of the line-of-sight channel, d represents the distance between the UAV and the base station, λ represents the signal wavelength, and L NLoS This represents the path loss of a non-line-of-sight channel. η1 represents the additional path loss of the non-line-of-sight channel, P k Let represent the signal power of the k-th UAV transmitter, and M represent the channel noise power. t,k,1 w represents the neural network model parameter upload strategy for the k-th UAV in the current time slot t. tr F represents tr，k (s t a t，k，1 a t，k，2 The weights of F d，k (s t a t，k，1 This represents the average utility of the identification task completion among all drones and backend edge nodes, such as... Figure 2 In each time slot, random target identification task points are generated in the environment. All task points include task point 1, task point 2, ..., task point S. Each UAV k has a corresponding task point s to be identified in each time slot. The UAV has a working limit distance, denoted as d. w When the distance between the drone k and the mission point s is less than the drone's maximum working distance d w When drone k completes the target identification task, it will be assigned a new task point to continue identification. d k，s This represents the distance between drone k and the corresponding mission point s, (x s y s Let (x) be the coordinates of the task point s, and (x) be the coordinates of the point s. t，k y t，k Let Δx be the coordinates of UAV k in time slot t. t，k Δy t，k These represent the flight distances of the UAV along the x-axis and y-axis, respectively, in time slot t.

[0052] Step 4: Upload strategy a based on the neural network model parameters of each drone. t，k，2 By fusing neural network model parameters, the model parameters of the neural network model can be reconstructed; for example... Figure 3-4 Specifically, it includes the following sub-steps:

[0053] Step 4.1, if the UAV neural network model parameter upload strategy a t，k，2To continue training the neural network model, each UAV collects labeled image data in real time while cruising in the air-to-ground collaborative edge learning network. Gradient descent is used to train its own neural network model, and the model parameters from the trained model are sent to the backend edge nodes within the UAV's communication range. In this invention, the model parameters in the neural network model include the weights of each neuron and the deviation data of each activation function. The UAV's communication range is a circular area with a radius equal to a certain distance from the backend edge node. To prevent untrained model parameters from being uploaded to the backend edge nodes, UAVs that have already uploaded their neural network model parameters in the previous time slot will not upload them again in the current time slot; otherwise, the model parameters from the neural network model are directly sent to the backend edge nodes within the UAV's communication range.

[0054] Step 4.2: Each backend edge node fuses the received model parameters and broadcasts the fused model parameters to the drones associated with it. Specifically, since the arrival times of the local model parameters on the drones are inconsistent within a time slot, in order to reserve sufficient time for broadcasting the model parameters, the backend edge node n will broadcast the neural network model parameters ω obtained in time slot t at the beginning of the next time slot. es，n The broadcast is sent to the drone swarm that has established a communication connection with the backend edge node n. After receiving the neural network model parameters from the backend edge node n, the frontend drone k denotes them as ω. es，n The front-end drone k replaces the original local model parameters, denoted as ω. local，k .

[0055] To mitigate the problem of excessive model bias after fusion due to insufficient training samples, local model parameters participating in the fusion process need to be multiplied by weights. The weight value is obtained by dividing the number of local training samples on the current drone k by the sum of the number of local training samples from drones participating in training on the current edge node n. Therefore, the model parameters on the backend edge node need to be subtracted from the weighted local model parameters obtained in the previous round, and then added to the weighted local model parameters received currently. The sum of the number of local training samples from drones participating in training on edge node n before the current and previous rounds of fusion is defined as N. n and N n The above model parameter fusion process can satisfy the following formula: Since an edge node n may receive multiple local model parameters from different drones within a single time slot, using This represents the set of the aforementioned drones, where edge node n will... All local model parameters sent by all UAVs within the system are replaced, replacing the local model parameters of the corresponding UAVs that participated in the previous round of model parameter fusion. Therefore, it is necessary to... The local model parameters on all drones are weighted, accumulated, and replaced, and the model parameter fusion process is updated as follows:

[0056]

[0057] Where, ω es，n ω′ represents the model parameters fused for the current nth backend edge node. es，n N represents the model parameters fused from the nth backend edge node during the previous training round. n N represents the sum of the number of local training samples of the drone participating in training on the nth backend edge node before the current model parameters are fused. n ′ represents the sum of the number of local training samples of the drone participating in training on the nth backend edge node before the previous round of model parameter fusion. This represents the number of communications between the nth backend edge node and the drone, where k represents... index, ω local，k This represents the model parameters received by the nth backend edge node from the kth drone, ω′. local，k This indicates that the nth backend edge node received the model parameters from the kth drone during the previous training round, where n... k n represents the current number of training samples for drone k. k ′ represents the number of training samples in the previous training round for drone k.

[0058] Step 4.3: Each UAV reconstructs the model parameters of the neural network model based on the received fused model parameters;

[0059] Since the front-end drone k may receive multiple neural network model parameters from different back-end edge nodes within a single time slot, it is necessary to average the obtained neural network model parameters. The set of backend edge nodes mentioned above, and the model parameters of the reconstructed neural network model are:

[0060]

[0061] in, This represents the model parameters of the reconstructed neural network model, where n represents... index, ω es，n This represents the model parameters for the current nth backend edge node fusion.

[0062] Step 5: Update the position of each drone and obtain the corresponding reward value, and record the status information s of the current time slot t. t The combined actions of UAV k in the current time slot t [a] t，k，1 a t，k，2 ], reward value, and the state of the next time slot. t+1As the experience of the current time slot t, it is stored in the experience pool;

[0063] Step 6: Extract experience from the experience pool and update the parameters of the fully connected neural network model using the Adam optimizer;

[0064] Step 7: Repeat steps 2-6 for each time slot until the fully connected neural network model converges, obtaining the optimal flight maneuver and the optimal neural network model parameter upload strategy for the corresponding time slot.

[0065] This invention presents an online trajectory planning method for air-ground cooperative UAVs for distributed model training, achieving reliable and efficient distributed model training. In simulation analysis using Alex networks for CIFAR10 dataset recognition, the aforementioned trajectory planning algorithm achieves a neural network model recognition success rate of over 60% within 300 time slots. In contrast, the distributed model training algorithm without trajectory optimization suffers from unstable communication links between the UAV and backend edge nodes, resulting in a dataset deviation recognition success rate that only converges to over 40%, and requires over 400 time slots.

[0066] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for online trajectory planning of air-ground cooperative unmanned aerial vehicles (UAVs) for distributed model training, characterized in that, Specifically, the steps include the following: Step 1: Deploy in the air-ground collaborative edge learning network N Each backend edge node and K A series of drones flying at a fixed altitude, each with an identical neural network model deployed on it, and the neural network model parameters initialized. Step 2: Obtain the current timeslot of each drone. Status information Including: the current time slot The location of the drone and the upload status of the drone neural network model parameters from the previous time slot; Step 3, each drone passes through - Greedy strategy, starting from a pre-set set of flight actions Select a flight action from the options: ={Up, Down, Left, Right, Top Right, Bottom Right, Top Left, Bottom Left} Fly a fixed distance ; and upload a set of strategies from the preset neural network model parameters. ={Neural network model parameters are uploaded to the backend edge nodes, and the neural network model is further trained using the neural network model} to determine the neural network model parameter upload strategy for each drone. ,Will As the current time slot Unmanned aerial vehicle k Combination movements; Step 4: Upload strategy based on the neural network model parameters of each drone. By fusing neural network model parameters, the model parameters of the neural network model can be reconstructed; Step 5: Update the position of each drone and obtain the corresponding reward value, then update the current time slot. Status information Current time slot Unmanned aerial vehicle k Combination movements Reward value, status of the next time slot As the current time slot The experience gained is stored in the experience pool; Step 6: Extract experience from the experience pool and update the parameters of the neural network model using the Adam optimizer; Step 7: Repeat steps 2-6 for each time slot until the neural network model converges, obtaining the optimal flight maneuver and the optimal neural network model parameter upload strategy for the corresponding time slot. .

2. The method for online trajectory planning of air-ground cooperative UAVs for distributed model training according to claim 1, characterized in that, Step 3 includes the following sub-steps: Step 3.1: Construct a set of quantized states It stores quantified drone state information and constructs a matrix. Recording drones Accessed quantized state information - combined action pairs; the matrix The line represents drones Different state information, denoted by 1 for drone. The experienced state information is represented by 0 for the drone. Unexperienced state information; the matrix The column represents drones The combined actions, represented by 1, are the drone's... The combined actions experienced by the drone are represented by 0. Unexperienced combination movements; Step 3.2, using - Greedy strategy, drones use probability... To maximize two objectives—the average transmission rate between each drone and the backend edge node, and the average efficiency in completing the identification task—we need to start from a pre-defined set of flight actions. Select a flight action from the options: ={Up, Down, Left, Right, Top Right, Bottom Right, Top Left, Bottom Left} Fly a fixed distance ; and upload a set of strategies from the preset neural network model parameters. ={Neural network model parameters are uploaded to the backend edge nodes, and the neural network model is further trained using the neural network model} to determine the neural network model parameter upload strategy for each drone. ,Will As a drone k Combination action update matrix Drones are probabilistic Randomly selected from a preset set of flight maneuvers Choose an unexplored flight maneuver from the following options: up, down, left, right, upper right, lower right, upper left, lower left. Fly a fixed distance ; and upload a set of strategies from the preset neural network model parameters. ={Neural network model parameters are uploaded to the backend edge node for further training of the neural network model} Select an unexplored drone for its neural network model parameter upload strategy. ,Will As a drone k Combination action update matrix Select an unexplored action.

3. The online trajectory planning method for air-ground cooperative UAVs oriented towards distributed model training according to claim 2, characterized in that, The process of storing the quantified UAV state information in step 3.1 is as follows: If the current time slot t Status information of the drone With quantized state set If the distance between all elements of a certain quantized state information is less than the corresponding threshold, then Do not save; otherwise, Stored to quantized state set middle.

4. The online trajectory planning method for air-ground cooperative UAVs oriented towards distributed model training according to claim 2, characterized in that, The process in step 3.2 to maximize the two objectives of average transmission rate and average recognition task completion efficiency between each UAV and the backend edge node is as follows: in, This represents the result of maximizing two objectives: the average transmission rate between each drone and the backend edge node, and the average utility of completing the identification task. This represents the average transmission rate between all drones and the backend edge nodes. , Indicates the current time slot t Next k The flight maneuvers of a drone Indicates the first n The backend edge node and the first k Transmission rate between drones Indicates the current time slot t Next k A strategy for uploading neural network model parameters for a single drone. express The weight, This represents the average identification task completion efficiency between all drones and backend edge nodes.

5. The online trajectory planning method for air-ground cooperative UAVs oriented towards distributed model training according to claim 4, characterized in that, The for: in, Indicates the first step in the transmission process n The backend edge node and the first k Path loss between drones , Indicates the first k The drone and the first n The probability that the channel between each backend edge node is a line-of-sight channel. This represents the path loss of the line-of-sight channel. , This represents the additional path loss for the line-of-sight channel. Indicates the distance between the drone and the base station. Indicates the signal wavelength. This represents the path loss of a non-line-of-sight channel. , This represents the additional path loss for non-line-of-sight channels. Indicates the first k The signal power of the drone's transmitter This indicates the channel noise power.

6. The online trajectory planning method for air-ground cooperative UAVs oriented towards distributed model training according to claim 4, characterized in that, The for: in, Indicates drone and corresponding task points The distance between them For the task point coordinates For drones In the time slot coordinates They are time slots drone along x shaft and y Flight distance of the axis K This indicates the number of drones flying at a fixed altitude.

7. The online trajectory planning method for air-ground cooperative UAVs oriented towards distributed model training according to claim 1, characterized in that, Step 4 includes the following sub-steps: Step 4.1, if the UAV neural network model parameter upload strategy... To continue training the neural network model, each UAV collects labeled image data in real time while cruising in the air-ground collaborative edge learning network, trains its own neural network model using gradient descent, and sends the model parameters from the trained neural network model to the backend edge nodes within the communication range of the UAV; otherwise, the model parameters from the neural network model are directly sent to the backend edge nodes within the communication range of the UAV. Step 4.2: Each backend edge node will fuse the received model parameters and broadcast the fused model parameters to the drones associated with the backend edge nodes; Step 4.3: Each UAV reconstructs the model parameters of the neural network model based on the received fused model parameters.

8. The online trajectory planning method for air-ground cooperative UAVs oriented towards distributed model training according to claim 7, characterized in that, In step 4.1, drones that have completed uploading neural network model parameters in the previous time slot will not upload neural network model parameters again in the current time slot.

9. The online trajectory planning method for air-ground cooperative UAVs oriented towards distributed model training according to claim 7, characterized in that, The model parameter fusion process in step 4.2 is as follows: in, Indicates the current number n Model parameters for the fusion of backend edge nodes Indicates the first round of training n Model parameters for the fusion of backend edge nodes Indicates the number of parameters before fusion in the current model. n The sum of the number of local training samples of drones participating in training on each backend edge node. This indicates the number of the previous round of model parameter fusion. n The sum of the number of local training samples of drones participating in training on each backend edge node. Indicates the first n The number of communications between each backend edge node and the drone. k express index, Indicates the current number n The backend edge node received the first k Model parameters of a drone Indicates the first round of training n The backend edge node received the first k Model parameters of a drone Indicates drone Current number of training samples Indicates drone The number of training samples in the previous training round.

10. The online trajectory planning method for air-ground cooperative UAVs oriented towards distributed model training according to claim 7, characterized in that, The model parameters of the reconstructed neural network model in step 4.3 are: in, The model parameters represent the reconstructed neural network model. Indicates the first k The number of backend edge nodes associated with each drone n express index, Indicates the current number n Model parameters for the fusion of backend edge nodes.

Images