Fault-resistant task migration and DNN adaptive segmentation method for low-altitude edge network

Abstract

The invention discloses a low-altitude edge network anti-fault task migration and DNN adaptive segmentation method, which comprises the steps of constructing a multi-unmanned aerial vehicle auxiliary MEC system, and performing DNN task division through a DNN adaptive division strategy; constructing a fault migration model, and migrating an uncompleted DNN task on the failed unmanned aerial vehicle to a normal unmanned aerial vehicle; the weighted energy consumption minimization problem of the multi-unmanned aerial vehicle assisted MEC system is reconstructed into a Markov decision process, and an optimal strategy is learned through interaction with the environment; a DKSAC-PER joint optimization framework is constructed, transmission power is optimized through the DKSAC-PER, calculation resource allocation optimization is performed through a Lagrange multiplier method, a reward value of a current step is calculated through an SAC-PER algorithm, network parameters are updated, and a complex mixed decision space is solved. According to the invention, the flight paths of the remaining unmanned aerial vehicles can be automatically adjusted, the calculation load is redistributed, and the survivability of the system in a severe environment is greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mobile edge computing, and particularly relates to a low-altitude edge network anti-failure task migration and DNN adaptive segmentation method. BACKGROUND

[0002] With the development of low-altitude economy, the improvement of low-altitude intelligent network infrastructure, and the continuous advancement of 6G technology, tasks based on deep neural networks (DNN) are widely used in various industries. However, such tasks are usually computationally intensive and delay-sensitive, and their high computational load and low latency requirements pose a major challenge to the limited computing power and battery life of mobile devices (MD).

[0003] To solve the problem of insufficient computing power of mobile devices, DNN network segmentation technology is proposed. By utilizing the hierarchical structure of DNN models, they are split into multiple parts. The deep network with large computational load is partially offloaded by the deployed edge server, and only a small amount of computing tasks are retained locally. In complex urban or field environments, the channel link between the ground user and the base station is easily blocked by buildings or obstacles, resulting in non-line-of-sight transmission. As a key air node of low-altitude intelligent networks, unmanned aerial vehicles (UAV) can solve this problem by taking advantage of their flexible deployment and line-of-sight transmission.

[0004] However, unmanned aerial vehicles may fail and exit the service network during task execution due to factors such as battery depletion, communication interruption, hardware failure, and adverse weather. This unexpected exit of unmanned aerial vehicles interrupts ongoing communication and computing tasks, leading to incomplete DNN inference and a sharp decline in performance, which in turn affects the overall service continuity and task completion rate of low-altitude intelligent networks. In addition, existing technologies focus more on research in the dimensions of task offloading, trajectory optimization, and resource allocation, and less on robust design in the case of unmanned aerial vehicle exit. There is still a lack of research on dynamic migration and redistribution strategies for DNN tasks after the exit of unmanned aerial vehicles. SUMMARY

[0005] The problem this invention aims to solve is to provide a fault-tolerant task migration and DNN adaptive segmentation method for low-altitude edge networks. This method combines adaptive DNN segmentation and a failed UAV task migration mechanism to increase the system's fault tolerance. It uses the weighted energy consumption in a multi-UAV assisted MEC system as a performance indicator to achieve joint optimization of DNN partitioning decisions, computational task migration, UAV computational resource allocation, and UAV trajectories. Furthermore, it reconstructs the weighted energy consumption minimization problem in a multi-UAV assisted MEC system as a Markov decision process. The method integrates the Tinkelbach transform, Lagrange multiplier method, and a soft actor-commentator algorithm with priority experience replay through the DKSAC-PER joint optimization framework to efficiently solve complex hybrid decision spaces.

[0006] This invention adopts the following technical solution: a method for fault-resistant task transfer and DNN adaptive segmentation in low-altitude edge networks, comprising the following steps:

[0007] Step 1: Construct a multi-UAV assisted MEC system, consisting of a ground mobile device layer and an airborne UAV edge layer. The positions of the UAVs and mobile devices are modeled using a three-dimensional Cartesian coordinate system.

[0008] Step 2: Treat the computing tasks generated by each mobile device in the multi-UAV assisted MEC system as DNN inference tasks, use weighted energy consumption as the performance index, and divide the DNN tasks through a DNN adaptive partitioning strategy.

[0009] Step 3: Construct a fault migration model. By defining UAV state variables and task migration decision variables, the unfinished DNN tasks on the failed UAV are migrated to the normal UAV.

[0010] Step 4: Reconstruct the weighted energy consumption minimization problem of the multi-UAV assisted MEC system into a Markov decision process, map the complex dynamic environment into the observation space of the reinforcement learning agent, and learn the optimal strategy through interaction with the environment.

[0011] Step 5: Construct the DKSAC-PER joint optimization framework to solve the hybrid decision space, including:

[0012] Step 5.1: Optimize the transmission power through the Tinkelbach transform. Introduce auxiliary variables to transform the non-convex problem into a convex optimization problem. Alternately update the transmission power and auxiliary variables until convergence is achieved, and obtain the optimal transmission power.

[0013] Step 5.2: Optimize the allocation of computing resources using the Lagrange multiplier method, introduce time delay constraints and UAV computing capacity constraints, derive the closed-form solution of the optimal computing resource allocation using KKT conditions, and determine the optimal computing resource allocation scheme.

[0014] Step 5.3: Calculate the reward value of the current step and update the network parameters using the SAC-PER algorithm to solve the complex mixed decision space;

[0015] Step 6: Embed the optimization of transmission power into each step of reinforcement learning training. The current association strategy, segmentation decision and trajectory optimization action are given by the Actor network. The value of the action is evaluated by the Critic network. A priority experience replay mechanism is introduced to improve sample efficiency. A weighted energy consumption minimization joint optimization strategy that can adaptively cope with dynamic environment and equipment failure is obtained.

[0016] The results are further applied to mobile edge computing networks assisted by multiple drones, especially in scenarios where drones fail due to power depletion or hardware malfunction. Task migration and adaptive DNN segmentation are used to ensure the continuity of DNN inference tasks and the robustness of the system.

[0017] As a preferred embodiment, in step 1, the multi-UAV assisted MEC system includes A drone equipped with a MEC server and A ground-based mobile device (MD) in this environment will complete the entire flight cycle. Divided into equal intervals There are 3 equal-sized time slots, each with a length of 1. The positions of drones and mobile devices are modeled using a three-dimensional Cartesian coordinate system.

[0018] Each ground mobile device generates a DNN-based computation task in each time slot. A mobile user can only be served by one UAV in the same time slot, while a UAV can serve multiple ground mobile devices in the same time slot. The computation task generated by each mobile device is a DNN inference task.

[0019] As a preferred embodiment, in step 2, the DNN adaptive partitioning strategy is based on There are k types of DNN models, where the k-th type of DNN model contains Layer, user set as Introducing DNN to partition decision variables The task is divided into two parts: Level 1 to Level 2. The layer is calculated locally, the first Layer to the first The layer is unloaded to the drone for computation.

[0020] As a preferred embodiment, in step 3, the fault migration model is defined by defining UAV state variables. and task migration decision variables The unfinished DNN task on the failed UAV is migrated to the normal UAV to ensure the integrity of the DNN task.

[0021] Further, when , it indicates that the UAV is in failure (failure), at this time, task migration is allowed. is in normal operation, at this time, task migration is prohibited; when , it indicates that the UAV is in failure (failure), at this time, task migration is allowed.

[0022] The mobile device unfinished DNN task layer will be migrated to the normal UAV for processing only when and only when .

[0023] Let denote the state of the target UAV , when , it indicates that the target UAV is in failure, and task migration to the failed UAV is not allowed; when , it indicates that the target UAV is normal, and task migration to the target UAV is allowed.

[0024] As a preferred, in step 4, the system needs to obtain the state information of the environment in real time, including the task attributes of the mobile device, the channel state, the three-dimensional position of the UAV, the remaining power and the failure state indication. The problem of minimizing the weighted energy consumption of the multi-UAV assisted MEC system is reconstructed as a Markov decision process for modeling, which is used to map the complex dynamic environment into the observation space of the reinforcement learning agent, so that it can learn the optimal strategy through continuous interaction with the environment.

[0025] Further, the Markov decision process is composed of four tuples , the state space includes: task size, UAV-MD association index, migration variable, UAV acceleration, UAV speed and UAV position information; the action space defines a hybrid action space, including discrete variables (DNN division decision , UAV-MD association , task migration ) and continuous variables (UAV acceleration ); the reward function is designed as the weighted total energy consumption, and penalties including collision, out-of-bound, timeout, speed and acceleration violation are introduced.

[0026] As a preferred, the DKSAC-PER joint optimization framework in step 5 is composed of three coupled steps: transmission power optimization based on Dunkl transform, computational resource allocation optimization based on Lagrange multiplier method, and joint optimization of UAV-MD association factor, task migration, DNN partition decision, and UAV trajectory based on SAC-PER architecture.

[0027] In step 5.1, under the determined Markov decision, the transmission power is optimized by applying Dunkl transform technique to minimize energy consumption, auxiliary variables are introduced to convert the non-convex problem into a convex optimization problem, the objective function is set and the transmission power and auxiliary variables are alternately updated until convergence, the globally optimal transmission power solution is obtained, which greatly reduces the action space dimension of deep reinforcement learning algorithm and improves the convergence accuracy.

[0028] In step 5.2, based on the transmission power determined in step 5.1, the Lagrange function of computational resource allocation is constructed, the delay constraint and the UAV computing capacity constraint are introduced, and the closed-form solution of optimal computational resource allocation is derived using KKT conditions (Karush-Kuhn-Tucker Conditions), to obtain the optimal computational resource allocation scheme that satisfies the delay and capacity constraints.

[0029] In step 5.3, based on the transmission power determined in step 5.1 and the computational resource allocation scheme determined in step 5.2, the optimized power and computational resource frequency are substituted into the environment to calculate the reward value (Reward) of the current step, the remaining variables are optimized using the soft actor-critic network with priority experience replay, and the network parameters are updated. During training, the sample priority is calculated according to the time difference error, and samples with large time difference error are preferentially sampled, and the importance sampling weight is used to correct the bias.

[0030] As a preferred, in step 6, the optimization of transmission power and the optimization of computational resource allocation are embedded in each step of reinforcement learning training.

[0031] After the agent Actor network gives the current association strategy, partition decision, and trajectory optimization, the variables are fixed, and actions such as DNN division, UAV-MD association, task migration, and UAV trajectory control are output; the Critic network evaluates the action value and introduces maximum entropy to enhance the exploration ability and prevent falling into local optimum, and finally outputs the Actor network parameters and the corresponding optimal joint control strategy after iterative training and convergence, which is used to guide the UAV MEC system to perform DNN task partition, mobile device association, task migration, and UAV trajectory planning in each time slot, so as to minimize the weighted energy consumption of the system under the premise of meeting the delay constraint and ensuring fault robustness.

[0032] Compared with the prior art, the present application has the following technical effects:

[0033] 1. The method of the present application introduces a fault-aware task migration mechanism. When the system detects that the UAV has failed due to power consumption or hardware failure, it can adaptively migrate the unfinished DNN task originally unloaded to the failed UAV to other available UAVs. Simulation experiments show that in the dynamic scenario of UAV failure, the present application can autonomously adjust the flight trajectory of the remaining UAVs and redistribute the computing load, ensuring the continuous execution of DNN inference tasks and avoiding complete service interruption due to single-point failure, greatly improving the survival ability of the system in harsh environments.

[0034] 2. The DKSAC-PER joint optimization framework proposed in the present application decouples the complex mixed integer nonlinear programming problem. On the one hand, the optimal solution of transmission power and computing resources is obtained through Dinklage transformation and Lagrange multiplier method respectively, avoiding the blind search of deep reinforcement learning in continuous high-dimensional action space. On the other hand, a priority experience replay mechanism is proposed based on the SAC algorithm, which gives samples with larger time difference errors higher sampling priority, enabling the agent to focus more on high-value experience learning.

[0035] 3. The DKSAC-PER joint optimization framework constructed in the present application effectively overcomes the problems of service interruption, high energy consumption and slow convergence of deep reinforcement learning algorithm in the existing technology under the low-altitude edge network and in the scenario of abnormal exit of UAV.

[0036] 4. The present application uses an adaptive DNN division strategy to dynamically determine the segmentation point of DNN tasks between local processing and edge processing according to channel state and computing capacity, avoiding resource waste caused by traditional "all offloading" or "random division". BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The present application is a multi-UAV assisted MEC system architecture schematic diagram;

[0038] Figure 2 The present application is a SAC-PER algorithm flowchart;

[0039] Figure 3 The present application is a DKSAC-PER algorithm flowchart;

[0040] Figure 4A comparison diagram of convergence performance of the DKSAC-PER algorithm used in the application, a random integrated double Q learning algorithm and a soft actor-critic algorithm under the system model of the application.

[0041] Figure 5 A schematic diagram of three-dimensional flight trajectory optimization of multiple unmanned aerial vehicles in a dynamic failure scenario in the embodiment of the application. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical scheme and advantages of the application more clear, the technical scheme of the application will be further described in detail below with reference to the drawings, and the described embodiments are only a part of the embodiments involved in the application. All non-innovative embodiments of other researchers in the field on the embodiments belong to the protection scope of the application. At the same time, the step numbers in the embodiments are only set for the convenience of description and explanation, and the order between the steps is not limited in any way, and the execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0043] In an embodiment of the application, a low-altitude edge network fault-resistant task migration and DNN adaptive segmentation method, specifically comprising the following steps:

[0044] (I) Construct a multi-unmanned aerial vehicle assisted MEC system model;

[0045] As shown in Figure 1 , the multi-unmanned aerial vehicle assisted MEC system is composed of a ground mobile device layer and an aerial unmanned aerial vehicle edge layer, including a UAV equipped with a MEC server and a ground mobile device (MD), each MD generates a DNN-based computing task in each time slot.

[0046] For ease of processing, the entire flight period is divided into time slots at equal intervals, and each time slot has a length of . A three-dimensional Cartesian coordinate system is used to model the positions of the UAV and the MD. In the th time slot, the horizontal position of the th MD is represented as , and the position of the th UAV is represented as , and the superscript T represents the transpose.

[0047] The movement of the MD follows a Gaussian-Markov model, and the update formulae of its speed and direction are as follows:

[0048] ;

[0049] ;

[0050] in, and These represent the average velocity and direction, respectively. , This represents the coefficient of memory retention. This represents the random fluctuation of the speed update of the i-th mobile device during movement, following a mean and variance of [values ​​to be filled in]. Gaussian distribution; This represents the random fluctuation of the azimuth update of the i-th mobile device during motion engineering, following a mean and variance of [values ​​to be filled in]. The Gaussian distribution.

[0051] The trajectory of a UAV is affected by its flight speed and acceleration The control, and its position update formula is:

[0052] ;

[0053] Meanwhile, to ensure flight safety, the distance between any two drones must meet safety constraints:

[0054] ;

[0055] in, Indicates a collection of drones. Represents a set of time slots. Indicates time slot drones Location, Indicates time slot drones Location, This is a preset safe distance.

[0056] (ii) Dividing DNN tasks

[0057] To achieve flexible allocation of computational load, this embodiment employs a DNN adaptive partitioning strategy. Assume there are... K There are k types of DNN models, where the k-th type of DNN model contains Layer, let the user set of the k-th class DNN model be . ,in .

[0058] Introducing DNN to partition decision variables The task is divided into two parts: Level 1 to Level 2. The layer is calculated locally, the first Layer to the first The computational load of each layer of the DNN is offloaded to the UAV. (FLOPs) and output data volume (bits) are determined by the following formula based on the layer type (convolutional layer or fully connected layer):

[0059] ;

[0060] ;

[0061] in, Indicates the first A collection of convolutional layers in a DNN-like model. Indicates the first The collection of fully connected layers in a DNN-like model. Represents a convolutional layer Input the height of the feature map, Represents a convolutional layer Input the width of the feature map. Represents a convolutional layer Number of input channels, Represents a convolutional layer The number of output channels, Represents a convolutional layer The size of the convolution kernel, Represents a convolutional layer One-dimensional input size, Represents a convolutional layer One-dimensional output size, Memory usage per unit of data.

[0062] (III) Constructing a fault migration model

[0063] This embodiment introduces a binary variable to address potential drone malfunctions (such as battery depletion or hardware failure). This indicates the status of the drone.

[0064] If the first If a UAV malfunctions, the system activates the task migration mechanism:

[0065] ;

[0066] ;

[0067] Among them, when At that time, it indicates that the drone In the time slot Normal operation, task migration is prohibited at this time; when The time indicates drone If a time slot fails (malfunctions), task migration is permitted.

[0068] Denotes the task decision transition variable if and only if At that time, mobile devices Unprocessed DNN task layers will be migrated to normal UAV. Process it.

[0069] Indicates the target drone The state when At that time, it indicates the target drone. The malfunction prevents the mission from being transferred to the failed drone; when At that time, it indicates the target drone. Normal, mission migration to target drone is permitted. .

[0070] (iv) Markov Decision Process

[0071] To address the joint optimization problem in dynamic environments, this embodiment reconstructs the weighted energy consumption minimization problem of a multi-UAV assisted MEC system into a Markov Decision Process (MDP), consisting of quadruples. constitute.

[0072] state space In time slots ,state DNN tasks involving MD UAV flight speed UAV location MD-UAV transmission rate UAV inter-transmission rate and UAV remaining energy .

[0073] Action space :action It includes both discrete and continuous variables, specifically DNN partitioning decisions. UAV-MD correlation index Task migration decision and UAV flight acceleration .

[0074] reward function The objective is to maximize the cumulative reward, which means minimizing the system's weighted total energy consumption while satisfying constraints. The reward function is defined as the negative of the weighted total energy consumption minus multiple penalties.

[0075] ;

[0076] in, This represents the system's weighted total energy consumption. This represents the collision avoidance penalty function for drones. This represents the function that penalizes task timeouts. This represents the function that penalizes drones for crossing boundaries. This represents the penalty function for exceeding the maximum acceleration limit for the drone. This represents the maximum speed exceeding the limit penalty function.

[0077] Let be the state transition probability function, representing the system's state transition probability in the current time slot. In a state And perform the action Then, transition to the next state. The probability of.

[0078] (V) DKSAC-PER Hybrid Optimization Framework

[0079] To address the high dimensionality and convergence difficulties arising from mixed variables (discrete and continuous) in the action space, this embodiment proposes the DKSAC-PER algorithm framework, which decouples the optimization problem into the following three sub-problems:

[0080] 1. Transmission Power Optimization (Dinkelbach Transform): This problem aims to minimize energy consumption. It is a non-convex fractional programming problem, and the objective function is expressed as:

[0081] ;

[0082] in, The optimization objective is to seek the optimal. To minimize the function value, Indicates the UAV-MD correlation factor. Indicates mobile device Transmission power, Indicates the amount of data to be transmitted. Indicates channel bandwidth. Indicates channel gain. Indicates noise power. Indicates the minimum feasible transmission power. This indicates the maximum transmission power.

[0083] To solve the above-mentioned score problem, auxiliary variables are introduced. By using the Tinkelbach transform, the non-convex fractional programming power minimization problem is transformed into a subtractive parametric convex problem. The transformed objective function is expressed as:

[0084] .

[0085] 2. Computational resource allocation optimization (Lagrange multiplier method): Based on the given power and policy, construct the Lagrange function:

[0086] ;

[0087] in, Indicates in time slot drones Assigned to mobile devices The frequency of computational resources is the optimization variable in this formula; These are the Lagrange multipliers corresponding to the delay constraints, used to relax mobile devices. Maximum computational delay constraint; These are Lagrange multipliers corresponding to capacity constraints, used to relax unmanned aerial vehicles (UAVs). Maximum computational capacity constraint; Indicates mobile device In the time slot The latency limit for internal tasks to be processed on the drone; Indicates drone Maximum computing resource capacity; This represents the energy consumption coefficient parameter. ,in, This represents the effective capacitance coefficient of the CPU. This indicates the computational workload of the unloading task:

[0088] ;

[0089] in, Indicates the CPU's processing power. Indicates mobile device The DNN model The computational workload required for each layer Indicates mobile device The total number of layers in the DNN model.

[0090] Further utilize the KKT conditions to obtain the optimal computation frequency that satisfies the time delay constraint. The closed-form solution, based on the resource coordination state With capacity The relationship between them allows for adaptive adjustment of allocation strategies.

[0091] 3. Joint Strategy Optimization (SAC-PER):

[0092] The SAC-PER algorithm flow in this embodiment is as follows: Figure 2 As shown, it consists of three parts: the UAV MEC environment, the priority experience replay buffer, and the SAC agent.

[0093] In complex environments including drone failures, this architecture feeds back the optimized power and resource allocation results to the SAC agent, using the PER mechanism to filter high-value samples and drive the Actor-Critic network to continuously iterate. On one hand, the architecture updates the policy network (Actor) parameters by minimizing KL divergence to balance exploration and exploitation; on the other hand, it optimizes the critic network (Critic) based on Bellman residuals, including the main DNN, target DNN, critic 1, critic 2, etc., and uses a soft update mechanism to synchronize the target network parameters and ensure training stability. Finally, it outputs an optimal joint control strategy that can adaptively cope with drone failures and dynamic environments.

[0094] (vi) DKSAC-PER Algorithm Architecture

[0095] In the DKSAC-PER algorithm architecture described in this embodiment, the processing flow is as follows: Figure 3 As shown, the system first performs system initialization, and then enters a loop iteration process.

[0096] In each iteration: First, the current state is acquired, and actions are generated through the actor network. Next, based on the current state and the generated actions, transmission power and resource allocation are optimized sequentially to determine continuous control variables. Then, actions are executed and rewards are calculated to obtain environmental feedback. Finally, the network parameters are updated using the calculated rewards and experience samples. After the network update, the system performs parameter fusion and determines whether the training termination condition is met. If the termination condition is not met, the system returns to the "acquire current state" step for the next iteration. If the termination condition is met, training ends and the optimal policy is output to minimize the weighted energy consumption of the multi-UAV assisted MEC system.

[0097] Specifically, the temporal difference error for each sample is calculated. :

[0098] ;

[0099] in, Indicates the agent's state and perform actions Then, the feedback reward value obtained from the environment; This represents the discount factor, used to balance the weights of current rewards and future long-term rewards; The estimated future value is based on the target network's prediction of the state at the next time step. and actions Value estimation; The value of the current prediction is based on the current network's understanding of the current state. and Value estimate.

[0100] According to priority Non-uniform sampling is used to make the agent learn from experiences with large prediction errors first.

[0101] in, This indicates the sampling priority of the i-th sample in the experience pool; It represents a very small positive constant, used to prevent the sampling probability from becoming 0 when the time difference error is 0; Indicates the first The larger the absolute value of the temporal difference error of a sample, the higher the contribution of that sample to the model update. This represents the priority adjustment factor, with a value range of [value range missing]. Used to control the strength of priority, when This is considered uniform sampling.

[0102] Furthermore, the loss function of the Actor network incorporates policy entropy to encourage exploration:

[0103] ;

[0104] in, This represents the loss function of the Actor network, and the parameters of the Actor network are updated by minimizing this value. ; This represents the Q-value of the i-th Critic network output. Let be the expectation operator, representing the state sampled from the experience replay buffer. The expected value. Indicates according to the current strategy In state Actions obtained by downsampling The mathematical expectation, This indicates that according to the current Actor network policy Generate Actions . It is an adaptive temperature coefficient.

[0105] Simulation experiments show that the DKSAC-PER algorithm architecture proposed in this invention has significant advantages in handling fault-resistant task transfer and DNN adaptive segmentation tasks in multi-UAV assisted MEC systems. Figure 4 As shown, compared with the random ensemble double Q learning algorithm and the soft actor-critic algorithm, the method of this invention has a faster convergence speed (approximately 20k steps to converge) and can achieve the lowest average weighted energy consumption under different numbers of MDs and different transmission power constraints, proving its robustness and efficiency in dynamic fault environments.

[0106] In this embodiment of the invention, the optimization of three-dimensional flight trajectories of multiple UAVs under dynamic fault scenarios is performed, such as... Figure 5 As shown, in dynamic scenarios where drones malfunction, the method of this invention can autonomously adjust the flight trajectory of the remaining drones and redistribute the computational load, ensuring the continuous execution of DNN inference tasks, avoiding complete service interruption due to single point of failure, and greatly improving the system's survivability in harsh environments.

[0107] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A fault-resistant task transfer and DNN adaptive segmentation method for low-altitude edge networks, characterized in that, Includes the following steps: Step 1: Construct a multi-UAV assisted MEC system, consisting of a ground mobile device layer and an airborne UAV edge layer. The positions of the UAVs and mobile devices are modeled using a three-dimensional Cartesian coordinate system. Step 2: Treat the computing tasks generated by each mobile device in the multi-UAV assisted MEC system as DNN inference tasks, use weighted energy consumption as the performance index, and divide the DNN tasks through a DNN adaptive partitioning strategy. Step 3: Construct a fault migration model. By defining UAV state variables and task migration decision variables, the unfinished DNN tasks on the failed UAV are migrated to the normal UAV. Step 4: Reconstruct the weighted energy consumption minimization problem of the multi-UAV assisted MEC system into a Markov decision process, map the complex dynamic environment into the observation space of the reinforcement learning agent, and learn the optimal strategy through interaction with the environment. Step 5: Construct the DKSAC-PER joint optimization framework to solve the hybrid decision space, including: Step 5.1: Optimize the transmission power through the Tinkelbach transform. Introduce auxiliary variables to transform the non-convex problem into a convex optimization problem. Alternately update the transmission power and auxiliary variables until convergence is achieved, and obtain the optimal transmission power. Step 5.2: Optimize the allocation of computing resources using the Lagrange multiplier method, introduce time delay constraints and UAV computing capacity constraints, derive the closed-form solution of the optimal computing resource allocation using KKT conditions, and determine the optimal computing resource allocation scheme. Step 5.3: Calculate the reward value of the current step and update the network parameters using the SAC-PER algorithm to solve the complex mixed decision space; Step 6: Embed the optimization of transmission power into each step of reinforcement learning training. The current association strategy, segmentation decision and trajectory optimization action are given through the Actor network. The value of the action is evaluated through the Critic network. A priority experience replay mechanism is introduced to improve sample efficiency. A weighted energy consumption minimization joint optimization strategy that can adaptively cope with dynamic environment and equipment failure is obtained.

2. The low-altitude edge network fault-resistant task migration and DNN adaptive segmentation method according to claim 1, characterized in that, The multi-UAV assisted MEC system includes A drone equipped with a MEC server and A ground-based mobile device will cover the entire flight cycle. Divided into equal intervals There are 1 time slot, and the length of each time slot is 1. Each ground mobile device generates a DNN-based computational task in each time slot.

3. The low-altitude edge network fault-resistant task migration and DNN adaptive segmentation method according to claim 2, characterized in that, The positions of the UAV and mobile devices are modeled using a three-dimensional Cartesian coordinate system, as follows: The movement of the mobile device follows a Gauss-Markov model, the first... Mobile device time slots speed and direction The updated formula is: ； ； in, and These represent the average velocity and direction, respectively. , This represents the coefficient of memory retention. This represents the random fluctuation of the speed update of the i-th mobile device during movement, following a mean and variance of [values ​​to be filled in]. Gaussian distribution; This represents the random fluctuation of the azimuth update of the i-th mobile device during motion engineering, following a mean and variance of [values ​​to be inserted here]. Gaussian distribution; No. The trajectory of a drone is affected by its flight speed. and acceleration The control and position update formula is as follows: ； The distance between any two drones satisfies the following safety constraint: ； in, Indicates a collection of drones. Represents a set of time slots. , They represent time slots respectively. drones , Location, This is a preset safe distance.

4. The low-altitude edge network fault-resistant task migration and DNN adaptive segmentation method according to claim 3, characterized in that, The DNN adaptive partitioning strategy is based on There are k types of DNN models, and the k-th type of DNN model contains Layer, user set as Introducing DNN to partition decision variables The task is divided into two parts: Level 1 to Level 2. The layer is calculated locally, the first Layer to the first Layer offload to drone computing; The computational cost of each layer of a DNN and output data volume The layer type is determined by the following formula: ； ； in, Indicates the first A collection of convolutional layers in a DNN-like model. Indicates the first The collection of fully connected layers in a DNN-like model. Represents a convolutional layer Input the height of the feature map, Represents a convolutional layer Input the width of the feature map. , Representing convolutional layers Number of input channels and number of output channels Represents a convolutional layer The size of the convolution kernel, , Representing convolutional layers One-dimensional input size and one-dimensional output size, Memory usage per unit of data.

5. The low-altitude edge network fault-resistant task migration and DNN adaptive segmentation method according to claim 3, characterized in that, The fault migration model introduces binary variables. Indicates the drone's status, when At that time, it indicates that the drone In the time slot In case of a failure, activate task migration: ； ； in, Represents a set of mobile devices. Denotes the task decision transition variable if and only if At that time, mobile devices Unprocessed DNN task layers were migrated to normal drones. Process it; Indicates the target drone The state when At that time, it indicates the target drone. The malfunction prevents the mission from being transferred to the failed drone; when At that time, it indicates the target drone. Normal, mission migration to target drone is permitted. .

6. The low-altitude edge network fault-resistant task migration and DNN adaptive segmentation method according to claim 5, characterized in that, The Markov decision process is composed of quadruples. constitute: For state space: in time slot ,state DNN tasks including mobile devices The flight speed of drones drone location Mobile device-drone transmission rate Inter-UAV transmission rate and the remaining energy of the drone ; For action space: in time slots ,action Includes both discrete and continuous variables, specifically: DNN partitioning decision. Drone-Mobile Device Correlation Indicators Task decision transfer variables and drone flight acceleration ; The reward function is defined as the negative of the weighted total energy consumption minus multiple penalties: ； in, This represents the system's weighted total energy consumption. This represents the collision avoidance penalty function for drones. This represents the function that penalizes task timeouts. This represents the function that penalizes drones for crossing boundaries. This represents the penalty function for exceeding the maximum acceleration limit for the drone. This represents the maximum speed exceeding the limit penalty function; Let be the state transition probability function, representing the system's state transition probability in the current time slot. In a state And perform the action Then, transition to the next state. The probability of.

7. The low-altitude edge network fault-resistant task migration and DNN adaptive segmentation method according to claim 6, characterized in that, The transmission power can be optimized using the Tinkelbach transform, as follows: With the goal of minimizing energy consumption, transmission power optimization is treated as a non-convex fractional programming problem, and the objective function is expressed as: ； in, The optimization objective is to find the optimal transmission power. Minimize the function value; Indicates the drone-mobile device association factor. Indicates mobile device Transmission power, Indicates the amount of data to be transmitted. Indicates channel bandwidth. Indicates channel gain. Indicates noise power. Indicates the minimum feasible transmission power. Indicates the maximum transmission power; Introducing auxiliary variables The power minimization problem in non-convex fractional programming form is transformed into a parameterized convex problem in subtraction form using the Tinkelbach transform. The transformed objective function is: 。 8. The low-altitude edge network fault-resistant task migration and DNN adaptive segmentation method according to claim 7, characterized in that, The Lagrange multiplier method is used to optimize computational resource allocation, as follows: Based on the given power and policy, construct the Lagrange function: ； in, Indicates in time slot drones Assigned to mobile devices The frequency of computing resources; These are the Lagrange multipliers corresponding to the delay constraints, used to relax mobile devices. Maximum computational delay constraint; These are Lagrange multipliers corresponding to capacity constraints, used to relax unmanned aerial vehicles (UAVs). Maximum computational capacity constraint; This represents the energy consumption coefficient parameter; Indicates mobile device In the time slot The latency limit for internal tasks to be processed on the drone; Indicates drone Maximum computing resource capacity; The optimal computational frequency that satisfies the time delay constraint is obtained using the KKT conditions. The closed-form solution, based on the resource coordination state With capacity The relationship between them allows for adaptive adjustment of allocation strategies.

9. The low-altitude edge network fault-resistant task migration and DNN adaptive segmentation method according to claim 8, characterized in that, The optimized power is obtained through the SAC-PER algorithm. and computing resource frequency Substitute the environment, calculate the reward value for the current step, and use a soft actor-critic network with priority experience replay to optimize the remaining variables, including: mobile device association decisions, task migration decisions, DNN segmentation decisions, and drone flight trajectories, to complete the joint optimization.

10. The low-altitude edge network fault-resistant task migration and DNN adaptive segmentation method according to claim 6, characterized in that, In the DKSAC-PER joint optimization framework, the agent outputs the action distribution through the Actor network, the Critic network evaluates the action value, and a priority experience replay mechanism is introduced to improve sample efficiency, as follows: Calculate the temporal difference error for each sample. : ； in, Indicates the agent's state Next action The reward value obtained from feedback in the environment afterwards; This represents the discount factor, used to balance the weights of current rewards and future long-term rewards; The estimated future value is based on the target network's prediction of the state at the next time step. and actions Value estimation; The value of the current prediction is based on the current network's understanding of the current state. and Value estimation; According to priority Non-uniform sampling is used to allow the agent to learn from experiences with large prediction errors first. The loss function of the Actor network encourages exploration by incorporating policy entropy. ； in, Indicates the first in the experience pool Sampling priority of each sample This represents a very small positive constant to prevent the sampling probability from being 0. Indicates the first The temporal difference error of each sample, Indicates the priority adjustment factor; The loss function of the Actor network is defined by minimizing... Update the parameters of the Actor network ; This represents the Q-value of the output of the i-th Critic network; Let be the expectation operator, representing the state sampled from the experience replay buffer. Expected value; Indicates according to the current strategy In state Actions obtained by downsampling The mathematical expectation; It is an adaptive temperature coefficient.

Images