Aging-first multi-unmanned aerial vehicle task unloading and path planning joint optimization method

Abstract

The invention relates to the technical field of multi-unmanned aerial vehicle cooperative edge computing unloading, and discloses a multi-unmanned aerial vehicle task unloading and path planning joint optimization method which comprises the following steps: 1) in a ground equipment-unmanned aerial vehicle cooperative system, dividing tasks into high / low priorities according to time efficiency, calculating local processing time delay, splitting according to needs, and matching initial service unmanned aerial vehicles in a communication range; 2) constructing a communication model, a flight model and a calculation model, and respectively describing equipment transmission, unmanned aerial vehicle movement and task processing characteristics; 3) in combination with constraints such as task priority and computing power load, establishing a dual-agent optimization objective function of ground equipment (unloading decision) and unmanned aerial vehicle (path planning); 4) adopting an improved SAC algorithm fused with an aging priority item, and obtaining a joint optimal scheme of task unloading and path planning through double-agent dynamic interaction iteration; and 5) an unloading instruction is issued according to the scheme, path planning and task processing are executed, and the optimal effect of cooperation of multiple unmanned aerial vehicles is achieved. According to the method, the problems of poor unloading and path adaptability and insufficient timeliness guarantee in a traditional scheme are solved, the high-priority task time delay can be reduced, the multi-unmanned aerial vehicle cooperation efficiency is improved, and the method is suitable for emergency rescue and other timeliness sensitive scenes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of multi-unmanned aerial vehicle cooperative edge computing offloading, and particularly relates to a time-efficiency-priority multi-unmanned aerial vehicle task offloading and path planning joint optimization method. BACKGROUND

[0002] With the rapid iteration of wireless communication technology and unmanned aerial vehicle technology, more and more unmanned aerial vehicles are equipped with embedded units with communication, storage and computing capabilities, which can provide support services in disaster relief, emergency communication, environmental monitoring and other scenarios. The core applications in many scenarios, such as vital sign monitoring data analysis, real-time analysis of rescue site images, and large-scale environmental data collection and processing, are all computing-intensive tasks, which have very high requirements for computing resources and processing time efficiency. In the face of such applications, the computing resources of a single unmanned aerial vehicle and ground equipment are limited, and the battery energy carried by the unmanned aerial vehicle is limited by the load, making it difficult to independently and efficiently complete the task processing requirements.

[0003] The emergence of multi-unmanned aerial vehicle assisted mobile edge computing provides a new way to solve this problem. Compared with traditional fixed edge computing, its core advantage lies in the combination of edge computing nodes and the mobility of unmanned aerial vehicles, through the flexible deployment of unmanned aerial vehicle clusters at the edge of the task scene, to build a "ground equipment-multi-unmanned aerial vehicle" cooperative computing architecture. This architecture can directly offload tasks beyond the processing capacity of the ground equipment to unmanned aerial vehicles, achieve distributed processing of tasks through the cooperative operation of unmanned aerial vehicle clusters, greatly shorten the data transmission distance, reduce the end-to-end delay, and fully utilize the idle computing resources of the cluster, perfectly adapting to the needs of dynamic scenarios such as emergency rescue.

[0004] However, in the previous multi-unmanned aerial vehicle assisted edge computing task offloading scheme, there are the following key problems: 1) the coupling of task offloading and path planning is not fully considered, most schemes handle them independently, only use Euclidean distance to simplify the flight path cost, and do not consider actual constraints such as non-flyable areas, leading to invalidation of the offloading strategy when the path deviates, causing task delay and energy waste; 2) the time efficiency classification requirements of the scene are not fully adapted, a unified optimization strategy is adopted, and emergency tasks such as life rescue cannot be prioritized, and high-priority tasks are prone to time delay due to resource competition.

[0005] Therefore, the present application proposes to solve the defects of the above-mentioned several unmanned aerial vehicle assisted edge computing task offloading schemes. SUMMARY

[0006] Therefore, the present application aims to provide a time-efficiency priority multi-unmanned aerial vehicle task offloading and path planning joint optimization method.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0008] A time-efficiency priority multi-unmanned aerial vehicle task offloading and path planning joint optimization method, characterized in that it comprises the following steps:

[0009] S1. In a cooperative system composed of unmanned aerial vehicles and ground devices, tasks are classified according to priority and tasks that cannot be completed locally are selected for splitting according to a set threshold value, and suitable service unmanned aerial vehicles are preliminarily matched;

[0010] S2. According to the transmission characteristics between devices and the motion characteristics of unmanned aerial vehicles, a communication model and a flight model are respectively constructed; according to the characteristics of tasks and three modes of task processing, including local processing, offloading to a near-end unmanned aerial vehicle for processing and offloading to a far-end unmanned aerial vehicle for processing, a calculation model is constructed;

[0011] S3. Based on the three types of models of communication, flight and calculation in step S2, the core constraint conditions of the system are determined, the optimization objectives of the double intelligent agents are determined and the corresponding objective functions are constructed: the ground devices take the optimization of energy consumption and time delay of task partial offloading as the objective, and the unmanned aerial vehicles take the optimization of energy consumption, time delay and collision avoidance of path planning as the objective;

[0012] S4. An improved SAC algorithm that fuses time-efficiency priority is used to complete the initialization of the interaction of the double intelligent agents, and then the task partial offloading decision of the ground devices and the path planning strategy of the unmanned aerial vehicles are respectively iteratively optimized, and through the dynamic interaction and feedback of the two types of intelligent agents, the joint optimal scheme of task offloading and path planning is obtained;

[0013] Further, the step S1 specifically comprises the following steps:

[0014] S11. Tasks are divided into high-priority tasks and low-priority tasks according to time-efficiency requirements; according to the computing capacity f of the ground device and the computing amount C of the task, the local processing time delay of the task is calculated by T i =C / f, and it is judged whether the local processing time delay of the task exceeds the time delay threshold value set for the corresponding priority task, if yes, the task is split into a plurality of independent sub-tasks of a fixed size according to the size of the task, and a partial offloading decision variable λ i,j,krepresents the allocation proportion of task i; set λ i,j,0 >0 represents partial offloading of task i to UAV j; if not exceeded, set λ i,0,0 =1 represents complete local processing of the task;

[0015] S12, initial service UAV matching: screening the nearest UAV within the communication range (distance ≤ communication radius R) of the ground device as the initial near-end service UAV;

[0016] Further, the step S2 specifically comprises the following steps:

[0017] S21, constructing a communication model: distinguishing between line-of-sight (LoS) channel and non-line-of-sight (NLoS) channel scenarios, combining large-scale fading and channel probability model, obtaining the transmission rate between the ground device and the UAV, and between the UAVs based on the Shannon formula, and then calculating the corresponding transmission delay and transmission energy consumption;

[0018] Further, the step S21 specifically comprises the following steps:

[0019] S211, channel type and probability determination: based on the relative position of the UAV and the communication object, the flight height of the UAV and the environmental parameters (building coverage ratio α, building quantity per unit area β, building height distribution parameter γ), the LoS probability formula of ITU standard is used to calculate the LoS channel probability P(LoS):

[0020]

[0021] Where h TX is the height of the sending end, h RX is the height of the receiving end, and m is the environmental fitting parameter; according to P(LoS), it is determined whether the current channel type is LoS or NLoS;

[0022] S212, large-scale fading calculation: according to the channel type, the large-scale path loss PL(t) is calculated:

[0023]

[0024] Where the free space path loss LFS(t) is:

[0025]

[0026] d(t) is the communication distance at time t, f c is the carrier frequency, c is the speed of light, η LoS , η NLoS are the additional fading losses of LoS and NLoS channels respectively;

[0027] S213, Channel gain calculation: calculate channel gain g(t) based on large-scale fading:

[0028]

[0029] S214, Transmission rate calculation: calculate transmission rate R based on Shannon formula:

[0030]

[0031] Where: P t is the transmit power of the sending end device, G t is the transmit antenna gain of the sending end device, G r is the receive antenna gain of the receiving end device, B is the channel bandwidth, and N0 is the noise power spectral density

[0032] S215, Transmission delay T trans and transmission energy consumption E trans calculation: based on subtask data volume D

[0033]

[0034] E trans = P t · T trans

[0035] S22, Constructing flight model: based on the initial position, flight speed and flight angle of the unmanned aerial vehicle, calculate the real-time position of the unmanned aerial vehicle at any time; calculate the flight delay and flight energy consumption according to the flight distance and flight speed of the unmanned aerial vehicle;

[0036] Further, the step S22 specifically includes the following steps:

[0037] S221, Flight delay calculation:

[0038] Unmanned aerial vehicle real-time position update: assuming that the initial position of the unmanned aerial vehicle n is

[0039]

[0040] The flight angle is ω y (angle with XOY plane), ω z (angle with Z axis), and the real-time flight speed is v n (t), then the three-dimensional position at any time t is:

[0041]

[0042] Flight distance calculation: based on the position displacement, calculate the actual flight distance within t time

[0043]

[0044] Thus, the flight time delay can be obtained by combining the flight distance and the real-time speed

[0045]

[0046] S222, flight energy consumption calculation:

[0047] The rotor thrust force F of the UAV n Calculation:

[0048]

[0049] Where M0 is the mass of the UAV, ||a n is the acceleration module, p is the air density, S n is the rotor area, g n is the gravity vector;

[0050] The thrust size is determined by two scalar parts, M0||a n ||is the thrust component used for accelerating or decelerating the UAV, is the thrust component used to overcome air resistance, and the thrust direction is determined by the vector (v n -M0g n ), and the module ||v n -M0g n ||is the scalar part in the formula;

[0051] The propulsion power of the UAV Calculation: This is the instantaneous power generated by the thrust in the direction of the speed

[0052]

[0053] Thus, the flight energy consumption is obtained

[0054]

[0055] S23, build a calculation model: for the three processing modes of the task, combine the logic of “first split fixed size subtask, then allocate proportion” to calculate the calculation delay and calculation energy consumption of each mode respectively; Where λ i,j,k represents the unloading decision variable; λ i,j,k , λ i,0,0 , λ i,j,0 , λ i,j,k respectively represent the proportion of distributing the split subtask i to the local computing of the ground device, unloading to the computing of the near-end UAV j, and forwarding to the computing of the far-end UAV k through the near-end UAV j, and satisfy ∑ j∈M ∑ k∈M λ i,j,k =1,

[0056] Where N is the set of subtasks and M is the set of drones.

[0057] Furthermore, step S23 specifically includes the following steps:

[0058] S231, Ground Equipment Local Computing Mode: This mode only includes the local computation process of the task. The computation latency and energy consumption are positively correlated with the allocation ratio and the amount of computation, and negatively correlated with the computing power of the ground equipment.

[0059]

[0060] E local =λ i,0,0 ·C i ·P 0,comp

[0061] Where C i Let f0 be the computational load of task i, f0 be the computational capability of the ground equipment, and P be the computational load of task i. 0,comp Calculate energy consumption per unit for ground equipment;

[0062] In addition, the unit computing power consumption P of the computing device comp The "linear power consumption model" is adopted to adapt to the computing energy consumption characteristics of drones and ground equipment. The closer the computing power is to the hardware limit, the higher the unit computing energy consumption.

[0063]

[0064] S232, Offloading to Near-End UAV Computing Mode: Includes "Transmission Process from Ground Equipment to Near-End UAV" and "Near-End UAV Computing Process". The total latency is the sum of transmission latency and computing latency, and the total energy consumption is the sum of transmission energy consumption and computing energy consumption.

[0065]

[0066] E prox =E trans,0-j +λ i,j,0 ·C i ·P j,comp

[0067] Where T trans,0-j E trans,0-j f represents the transmission delay and energy consumption from the ground equipment to the near-end UAV j calculated in S21. j For near-end UAVs, computing power, P j,comp Calculate energy consumption per unit j for near-end UAVs;

[0068] S233, forwarding to remote unmanned aerial vehicle computing mode: including "ground device to near-end unmanned aerial vehicle transmission process", "near-end unmanned aerial vehicle to remote unmanned aerial vehicle forwarding process" and "remote unmanned aerial vehicle computing process", the total delay is the sum of the transmission delay and the computing delay, and the total energy consumption is the sum of the transmission energy consumption and the computing energy consumption:

[0069]

[0070] E dist = E trans,0-j + E trans,j-k + λ i,j,k · C i · P k,comp

[0071] Wherein T trans,j-k , E trans,j-k is the transmission delay and energy consumption of near-end unmanned aerial vehicle j to remote unmanned aerial vehicle k calculated in S21, f k is the computing capacity of remote unmanned aerial vehicle k, P k,comp is the unit computing energy consumption of remote unmanned aerial vehicle k.

[0072] Further, the step S3 specifically comprises the following steps:

[0073] S31, clear system constraints, combined with task offloading, resource allocation, motion safety and timeliness requirements, define five kinds of core constraints to ensure the feasibility and timeliness of system operation:

[0074] Among them, the constraint conditions are as follows:

[0075]

[0076] C4: v min ≤ v n (t) ≤ v max

[0077]

[0078] In the above conditions, wherein N represents the subtask set, M represents the unmanned aerial vehicle set; condition C1) means that the subtask allocation ratio satisfies the normalization. C2) is the limit that the computing power load of ground device and unmanned aerial vehicle is not more than 100%, to avoid overload leading to delay surge. C3) is the limit that the total processing delay of subtask is not more than the corresponding priority threshold. C4) is the constraint that the speed of unmanned aerial vehicle is in the safety interval. C5) is to limit only the unmanned aerial vehicles within the communication range can be used as near-end nodes.

[0079] S32, clear double-agent optimization goal:

[0080] The ground device agent (unloading decision end): taking "time priority and low energy consumption" as the core, minimizing the total energy consumption and maximum transmission delay of sub-task processing under the premise of meeting the above constraints, the objective function is as follows:

[0081] r GD = -(E local + E proxy + E dist + T trans )

[0082] Wherein: E local , E proxy , E dist are the energy consumption of sub-task local computing, unloading to near unmanned aerial vehicle, and forwarding to remote unmanned aerial vehicle through near unmanned aerial vehicle respectively, T trans is the maximum transmission delay of sub-task; the negative sign indicates that the smaller the energy consumption and delay, the better the objective function value.

[0083] S33, unmanned aerial vehicle agent (path planning end): taking "adaptive unloading decision, low energy consumption and no collision" as the core, minimizing flight energy consumption, flight delay and collision risk under the premise of meeting the motion and safety constraints, the objective function is as follows:

[0084]

[0085] Wherein: represents the flight energy consumption of the unmanned aerial vehicle, represents the flight delay of the unmanned aerial vehicle, and Pen represents the collision penalty term of the unmanned aerial vehicle, and the calculation formula is:

[0086]

[0087] Wherein: C penalty is a preset penalty value, is the distance between unmanned aerial vehicles, d safe is a preset safety distance of the unmanned aerial vehicle.

[0088] Further, the step S4 specifically comprises the following steps:

[0089] S41, improve the SAC algorithm by fusing time priority items: in view of the problem that the fixed entropy regular term of the traditional SAC algorithm cannot adapt to the task priority, introduce the time priority item α(t) to replace the fixed entropy regular term, realize the dynamic adjustment of exploration-exploitation balance, for high priority tasks, reduce the exploration range by reducing the time priority item, speed up the convergence to the known efficient strategy; for low priority tasks, expand the exploration boundary by increasing the time priority item, and dig up long-term optimal solution; in each iteration, the double agent updates its own strategy by taking the other party's decision as a reference, and feeds back the optimization direction through a differentiated reward function, realizing the adaptive optimization of the algorithm for different priority tasks;

[0090] S42, double-agent interaction initialization:

[0091] Information sharing: ground device (GD, unloading decision-making agent) and unmanned aerial vehicle (UAV, path planning agent) share global information through edge nodes, including sub-task parameters (computational complexity C i ), device capability (ground device computing capability f0, unmanned aerial vehicle computing capability f j , task priority information (high / low priority identifier);

[0092] Parameter initialization: initialize independent Actor network (output unloading decision λ i,j,k or path action) for GD and each UAV respectively, double Critic network (reduce Q value overestimation); initialize time priority basis parameters: a0=0.5 (basic coefficient), a min =0.1 (minimum exploration coefficient of high priority), a max =0.8 (maximum exploration coefficient of low priority); GD outputs initial sub-task allocation ratio λ i,j,k , UAV preliminarily adjusts flight parameters (speed v n (t), angle ω y / ω z ) according to λ i,j,k ;

[0093] S43, SAC algorithm optimization based on fusion time priority for double-agent decision-making: dynamically adjust exploration-exploitation balance according to task priority, and perform iteration in each round;

[0094] Further, the step S43 specifically comprises the following steps:

[0095] S431, task priority coefficient calculation: for each sub-task i, based on its remaining time T deadline -t elapsed and task type score type score (high priority task=1.0, low priority task=0.3), calculate priority coefficient β(t):

[0096]

[0097] Wherein, ω1=0.7, ω2=0.3 are weights, ∈ is a minimum value to avoid denominator being 0, and β(t) ∈ (0, 1) (after normalization);

[0098] S432, dynamic adjustment of time priority: calculate time priority a(t) according to β(t), and adapt exploration demand of task priority:

[0099] a(t) = a0-(1 - b(t)), a(t) e [a min , a max ]

[0100] High priority task: b(t) large -> a(t) small -> reduce exploration, quickly exploit known efficient policy;

[0101] Low priority task: b(t) small -> a(t) large -> moderate exploration, dig long-term optimal solution;

[0102] S433, Agent action generation: GD (offloading decision), based on its own Actor network, combined with a(t) corresponding to the current b(t), sample sub-task allocation action l i,j,k ; UAV (path planning), based on its own Actor network, combined with a(t) corresponding to the current b(t), sample flight action (speed adjustment, angle steering), if the sub-task is mainly offloaded to the remote UAV k (l i,j,k > 0.5), then control the near-end UAV j and the remote UAV k to keep a cooperative distance d = 2.5d safe ;

[0103] S434, Environment feedback and experience storage: after executing the action, the environment feedbacks the new state (UAV new position, task load change) and the reward; GD reward, r GD = -(E local +E proxy +E dist + T trans )(the smaller the energy consumption and delay, the higher the reward); UAV reward, (Pen is the collision penalty, d ij < d safe , Pen = C penalty , otherwise 0); store the state s t , action a t , reward r t , next state s t+1 to the experience replay pool;

[0104] S435, network update:

[0105] Critic network update, sample samples from the experience pool, calculate the target Q value (merge a(t));

[0106] y j = r j + g · (Q' min + a(t) · logp(a' j+1 | s') j+1 )

[0107] Q' min= min(Q′1(s′ j+1 ,a′ j+1 ),Q′2(s′ j+1 ,a′ j+1 ))

[0108] where y j is the “ideal score” of the jth sample, r j is the immediate reward corresponding to the sample (r GD for GD samples and r UAV for UAV samples), γ = 0.95 is the discount factor, s′ j+1 and a′ j+1 are the next state and action, respectively; the Critic network parameters are updated by gradient descent to minimize the loss function;

[0109] The Actor network is updated, and the Actor loss function (incorporating α(t)) is calculated;

[0110]

[0111] The Actor network parameters are updated by gradient ascent to maximize L π , ensuring that the action policy aligns with the Q-value guidance and the need for timely exploration.

[0112] The timeliness priority parameter is updated, and the timeliness target entropy H timely = H0× β(t) (H0 is the base target entropy, and β(t) is the task priority coefficient) is calculated, and the loss function is constructed:

[0113]

[0114] The loss function L α is minimized by gradient descent to dynamically adjust α0, ensuring that α(t) ∈ [α min , α max ], matching the exploration-exploitation balance for different priority tasks.

[0115] S44, dynamic interaction feedback of the agent: forward feedback, λ i,j,k output by GD is passed to UAV, if λ i,j,k > 0.5 (subtask is mainly forwarded to remote k), UAV adjusts the path to maintain the cooperative distance between j and k, if λ i,j,k < 0.5 (local and near-end processing is the main), j optimizes the trajectory to shorten the distance to the ground device or low-energy hovering, avoiding collision; reverse feedback, the real-time position of UAV is transmitted to GD, if the UAV j is out of the communication range of the ground device (d 0-j (t) > r), GD adjusts λ i,j,0 and λ i,j,k to 0 and reallocates to the UAV within the communication range;

[0116] S45, convergence determination and optimal scheme output: when the cumulative reward fluctuation is less than or equal to the threshold value, it is determined that convergence is reached, and the sub-task offloading distribution matrix Lambda containing the distribution proportion Lambda of all sub-tasks I is output i,j,k ; the UAV trajectory parameter matrix Gamma containing the position of each UAV at any time t speed v n (t), angle omega y / omega z ; both constitute a joint optimal scheme of task offloading and path planning.

[0117] The present application jointly optimizes multi-UAV task offloading and path planning, solves the poor adaptability problem caused by independent decision-making of traditional schemes through a double-agent dynamic interaction feedback collaborative scheduling strategy, and greatly improves the overall operation efficiency of the system; The present application fully considers the time efficiency grading requirements of emergency rescue and other scenes, ensures the controllable time delay of high priority tasks through task priority classification and time efficiency priority dynamic adjustment mechanism, and is more suitable for complex requirements in real scenes with multiple task priorities coexisting; The present application adopts an improved SAC algorithm that fuses time efficiency priorities, and the double agents can learn the optimal decision-making strategy from the interaction process, thereby realizing the joint optimization of task offloading and path planning, which is conducive to more efficient and stable multi-UAV collaborative computing. BRIEF DESCRIPTION OF DRAWINGS

[0118] In order to make the purpose, technical scheme and beneficial effects of the present application clearer, the present application provides the following drawings for illustration:

[0119] Figure 1 The present application is a multi-UAV task offloading and path planning joint optimization overall step flow chart;

[0120] Figure 2 The present application is a joint optimization algorithm flow chart based on SAC that fuses time efficiency priorities; DETAILED DESCRIPTION

[0121] The collaborative computing joint optimization method of the present application will be further described below in conjunction with the drawings:

[0122] The cooperative computing joint optimization strategy provided by the application fully utilizes the cooperative computing resources of the ground equipment and the unmanned aerial vehicle cluster in the multi-unmanned aerial vehicle assisted edge computing scene, constructs a double-agent target function by considering the time efficiency grading demand of the task and the movement safety constraint of the unmanned aerial vehicle, and clearly defines the optimization boundary through defining core constraint conditions such as unloading decision and computing load; then, the joint optimal scheme of sub-task unloading distribution and unmanned aerial vehicle path planning in a dynamic environment is obtained through dynamic interaction feedback iteration of the double-agent of the ground equipment and the unmanned aerial vehicle by combining the improved SAC algorithm fused with the time efficiency priority, so that the optimal strategy of multi-unmanned aerial vehicle cooperative computing and path scheduling is realized.

[0123] Figure 1 The application is a multi-unmanned aerial vehicle task unloading and path planning joint optimization method overall step flow chart, which comprises the following steps: 1) in the cooperative system composed of unmanned aerial vehicles and ground equipment, the task is divided into high / low priority according to the time efficiency demand, the local processing delay is calculated, and it is judged whether to split, and the initial service unmanned aerial vehicle in the communication range is selected after splitting; 2) a communication model, a flight model and a computing model are constructed to respectively depict the device transmission characteristics, the unmanned aerial vehicle movement characteristics and the task processing characteristics, thereby providing a basis for subsequent optimization; 3) based on the three types of models, the core constraints such as unloading decision and computing load are defined by combining the task priority demand, and the double-agent optimization target function of the ground equipment and the unmanned aerial vehicle is established; 4) the improved SAC algorithm fused with the time efficiency priority is used for iterative optimization, and the joint optimal scheme of sub-task unloading distribution and path planning is obtained through the dynamic interaction feedback update strategy of the double-agent; 5) according to the joint optimal scheme of step S4, the ground equipment issues the sub-task unloading instruction, and the unmanned aerial vehicle executes the path planning and task processing operation, so as to realize the optimal effect of multi-unmanned aerial vehicle cooperative task unloading and path scheduling.

[0124] Figure 2 The application is a joint optimization algorithm flow chart based on SAC fused with the time efficiency priority, as shown in Figure 2 The method comprises the following steps:

[0125] Step 1: a ground equipment and unmanned aerial vehicle cooperative system is constructed, which comprises an unloading decision agent (ground equipment), a path planning agent (unmanned aerial vehicle cluster), an edge node and a task processing unit; the sub-task parameters, the equipment capacity and the priority information are synchronized, the Actor / Critic network and the time efficiency priority basic parameter a0 are initialized, and the initial unloading proportion l and the flight parameter are outputted;

[0126] Step 2: a priority coefficient b(t) prediction model based on the residual time and the type score of the task is constructed, and a time efficiency priority a(t) adjustment mechanism is constructed based on b(t) and the basic coefficient a0; the double-agent target function is constructed by combining the communication model, the flight model and the computing model;

[0127] Step 3: The unloading decision agent executes a time priority-based subtask allocation action λ according to α(t) of step 2; the path planning agent executes flight action generation adapted to the unloading demand according to the λ result, and if the subtask is a high proportion of forwarding type, synchronously adjusts the cooperative distance of the unmanned aerial vehicle cluster;

[0128] Step 4: Collect the new state of the system and the double-agent reward, and build an experience pool based on the "state-action-reward-next state" data to provide data support for network updating;

[0129] Step 5: Joint the experience pool data of step 4, update the Critic network, Actor network and time priority α(t) in turn according to the objective function of step 2, and optimize the network parameters by using the gradient descent / ascent algorithm;

[0130] Step 6: The control node determines whether the cumulative reward fluctuation of the continuous preset rounds is ≤ threshold value according to the network updating result of step 5; if yes, outputs the unloading allocation matrix Λ and the trajectory matrix Γ, otherwise returns to step 2 for re-optimization.

[0131] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An aging priority multi-UAV task offloading and path planning joint optimization method, characterized in that, The method comprises the following steps: S1, in a cooperative system composed of a UAV and a ground device, tasks are classified by priority and tasks that cannot be completed locally are selected for splitting according to a set threshold, and an adaptive service UAV is preliminarily matched; S2, a communication model and a flight model are respectively constructed according to the transmission characteristics between devices and the motion characteristics of the UAV; and a calculation model is constructed according to the task characteristics and three modes of task processing, including local processing, unloading to a near-end UAV for processing and unloading to a far-end UAV for processing; S3, based on the three types of models of communication, flight and calculation in step S2, the core constraint conditions of the system are determined, the optimization objectives of the double intelligent agents are determined, and the corresponding objective functions are constructed: the ground device aims to optimize the energy consumption and time delay of task unloading, and the UAV aims to optimize the energy consumption, time delay and collision avoidance of path planning; S4, an improved SAC algorithm that fuses time-sensitive priorities is used to complete the initialization of the interaction of the double intelligent agents, and the task unloading decision of the ground device and the path planning strategy of the UAV are respectively iteratively optimized, and the joint optimal scheme of task unloading and path planning is obtained through the dynamic interaction and feedback of the two types of intelligent agents.

2. The method of claim 1, wherein, The step S1 specifically comprises the following steps: S11, the tasks are divided into high-priority tasks and low-priority tasks according to time requirements, wherein the high-priority tasks are emergency tasks such as disaster area life rescue and emergency communication guarantee, and the low-priority tasks are conventional tasks such as material transportation and environment monitoring; it is judged whether the local processing time delay of the original task exceeds the time delay threshold set for the corresponding priority task, if yes, the two types of tasks are all split into a plurality of independent sub-tasks of a fixed size according to the task size and are allocated; if not, local processing is performed; S12, if the processing time delay threshold is exceeded, the UAVs within the preset range of the ground device are selected as the initial near-end service UAV.

3. The method of claim 1, wherein, The step S2 specifically comprises the following steps: S21, a communication model is constructed: the line-of-sight channel and the non-line-of-sight channel scenarios are distinguished, the channel gain is calculated in combination with the large-scale fading and small-scale fading characteristics, the transmission rate between the ground device and the UAV and between the UAVs is obtained based on the Shannon formula, and then the corresponding transmission time delay and transmission energy consumption are calculated; S22, a flight model is constructed: the real-time position of the UAV at any time is calculated based on the initial position, flight speed and flight angle of the UAV; the flight time delay and flight energy consumption are calculated according to the flight distance and flight speed of the UAV; S23, constructing a calculation model: for the three processing modes of the task, combining the logic of "first splitting fixed-size sub-tasks, and then calculating the computing time delay and computing energy consumption of each mode through λ i,j,k allocation ratio"; wherein λ i,j,k represents the offloading decision variable; λ i,0,0 , λ i,j,0 , λ i,j,k respectively represent the proportion of splitting sub-tasks i allocated to ground device local computing, offloaded to proximal unmanned aerial vehicle j computing, and forwarded to remote unmanned aerial vehicle k computing through proximal unmanned aerial vehicle j, and satisfy Wherein: N is a set of sub-tasks, and M is a set of UAVs.

4. The method of claim 1, wherein, The step S3 specifically comprises the following steps: S31, the system constraint conditions are determined, and five types of core constraint conditions are defined in combination with task unloading, resource allocation, motion safety and time requirement, to ensure the feasibility and timeliness of the system operation; S32, the objective function of the ground device: r GD = -(E local + E proxy + E dist + T trans ) wherein: E local , E proxy , E dist are the computation energy consumptions of the subtask local computation, offloading to the proximal UAV, and forwarding to the distal UAV via the proximal UAV, respectively, T trans is the maximum transmission delay of the subtask; the negative sign indicates that the smaller the energy consumption and delay, the better the objective function value. S33, the objective function of the UAV: wherein: represents the flight energy consumption of the UAV, represents the flight time delay of the UAV, and Pen represents the collision penalty term of the UAV, and the calculation formula is: wherein: C penalty is a preset penalty value, is the distance between the UAVs, d safe is a preset safety distance of the UAV.

5. The method of claim 1, wherein, The step S4 specifically comprises the following steps: S41, improve SAC algorithm by fusing time priority, introduce time priority instead of fixed entropy regular term (high priority reduce exploration accelerate convergence, low priority keep exploration dig optimal solution), double agents update their own strategy according to the other party's decision in each iteration, and feedback optimization direction through differentiated reward function; S42, double agent interaction initialization: ground equipment (unloading decision intelligent agent) and unmanned aerial vehicle (path planning intelligent agent) share sub task parameters, equipment capacity and priority information through edge node; initialize independent actor network, double critic network for both, and initialize time priority basic parameters; ground equipment outputs initial sub task allocation proportion λ, unmanned aerial vehicle adjusts flight parameters (speed, angle) according to the proportion; S43, Iterative optimization of dual-agent decision-making based on SAC algorithm of fusion time priority: taking the remaining time of subtask and the type score of task as input, dynamically calculating the task priority coefficient and time priority; the ground equipment combines the time priority to output the unloading decision action λ based on its own Actor network i,j,k The unmanned aerial vehicle combines the same time priority to output the flight action (speed, angle) based on its own Actor network, and adjusts the cooperation distance; after executing the action, the environment state and reward are collected and stored in the experience replay pool, and finally the network and time priority parameters are updated based on experience iteration; S44, agent dynamic interaction feedback: in positive feedback, unmanned aerial vehicle adjusts path or keeps cooperative distance according to unloading proportion of ground equipment; in reverse feedback, ground equipment dynamically adjusts unloading allocation according to real-time position of unmanned aerial vehicle, ensuring effective cooperation within communication range; S45, output optimal scheme after convergence: when cumulative reward fluctuation ≤ threshold, determine convergence, output sub task unloading allocation matrix Λ and unmanned aerial vehicle trajectory parameter matrix Γ, form joint optimal scheme.

Images