Reliable multi-unmanned aerial vehicle assisted DNN task division and unloading method in MEC system

Abstract

The invention discloses a reliable multi-unmanned-aerial-vehicle-assisted DNN task division and unloading method in an MEC system, and the method comprises the steps: firstly, building a multi-unmanned-aerial-vehicle-assisted MEC architecture, and enabling a plurality of unmanned aerial vehicles (UAVs) to provide a calculation unloading service for a mobile device (MDs); then, in order to ensure reliable communication in a high-dynamic scene, a robust transmission problem is proposed, and a successful transmission probability under a worst channel condition is analyzed by using a conditional risk value method; on this basis, in order to accelerate the reasoning of the DNN, a collaborative DNN division and unloading (CDPO) problem taking the minimization of the total time delay as a target is constructed, and a CPOO algorithm is proposed on the basis of a bat algorithm to carry out optimization solution on task division and unloading based on the DNN. And finally, providing a non-convex optimization problem which simultaneously considers the requirements of reliability assurance and delay sensitivity as a main problem, and solving through a LaCPOO algorithm which combines a subgradient algorithm and a CPOO algorithm.

Description

Technical Field

[0001] The present invention belongs to edge computing tasks, and specifically relates to a reliable multi-UAV assisted DNN task division and offloading method in a MEC system. Background Art

[0002] With the rapid development of the Internet of Things (IoT) and Artificial Intelligence (AI), a large number of novel mobile AI applications, such as image classification, virtual / augmented reality, etc., need to be processed on smart mobile devices (MDs) (e.g., smartphones, laptops, and tablets). However, the rapid development of deep neural networks (DNNs) and the increasing number of parameters have led to significant computational requirements. Due to the limitations of size and computing power, MDs are difficult to meet the requirements of real-time processing, which will seriously affect the quality of service (QoS) requirements. Therefore, edge intelligence has become a promising paradigm for achieving low-latency and reliable DNN inference.

[0003] Due to their advantages of low price, easy deployment and flexible mobility, drones have been applied and developed in fields including but not limited to intelligent sensor data acquisition, target tracking, disaster area monitoring and auxiliary computing. Consider using drones to implement reliable and efficient computing offloading strategies.

[0004] In recent years, the computational complexity of deep neural networks (DNNs) has continued to increase, making it more difficult to run them on resource-constrained edge devices. As a result, DNN partitioning technology has received widespread attention. The core idea of ​​DNN partitioning is to decompose the computational tasks of large neural network models into multiple subtasks and distribute these subtasks among different devices or computing nodes. This partitioning strategy can not only improve the overall computing efficiency, but also effectively reduce the computing pressure of a single device, reduce energy consumption and latency.

[0005] The present invention considers meeting the delay-sensitive requirements of DNN-based tasks while ensuring communication reliability, and proposes a reliable multi-UAV-assisted DNN task division and offloading method in MEC system. Summary of the invention

[0006] Purpose of the invention: The present invention provides a reliable multi-UAV assisted DNN task partitioning and offloading method in a MEC system, so as to realize the reliable offloading of DNN tasks between mobile devices and UAVs in a dynamic network.

[0007] Technical solution: A reliable multi-UAV assisted DNN task partitioning and offloading method in MEC system, including the following steps:

[0008] (1) Construct a multi-UAV-assisted MEC system, dynamically networking a group of mobile devices with offload task requirements, a group of UAVs equipped with MEC servers, and a base station BS with strong computing power;

[0009] (2) Based on the existence of unknown obstacles between MD and UAV, the robustness of transmission reliability in a changing channel fading environment is considered, a reliability assurance communication model is established, and the transmission power optimization problem and constraints are determined. The specific mathematical expression is as follows:

[0010]

[0011] Among them, p mu (t) represents the transmission power allocated by MD m to offload the task to UAV u; (0a) represents the distribution function set with mean μ and variance ∑ for channel fading; (0b) represents the constraint condition to ensure transmission reliability, where γ mu (t) represents the signal-to-noise ratio, γ 0 represents the target signal-to-noise ratio, 1-ε represents the reliability threshold, and ε is the maximum tolerable transmission error rate; (0c) indicates that the transmission power allocated to the mobile device cannot exceed the power threshold;

[0012] (3) A DNN partitioning and offloading model is established. With the goal of minimizing the total inference delay of all MDs for the entire scheduling time T, a collaborative DNN partitioning and offloading problem, namely the CDPO problem, is designed. The specific mathematical expression is as follows:

[0013]

[0014] in like It means that MD m will perform subtask Offload to UAV u, It means that the i-th subtask of MDm in time slot t is arrive The output line of; C1 is the constraint of binary offloading variables; C2 ensures that any MD subtask It can be unloaded to one UAV at most; C3 means that a UAV can only handle Z subtasks at most at a time; C4 ensures effective partitioning; C5 means that the partitioned rows do not intersect each other; C6 indicates that the sum of the number of rows of all subtasks must be equal to the total number of rows C7 Constraint The horizontal distance between MD and UAV cannot exceed the coverage range R of UAV;

[0015] (4) A risk theory approach is proposed to determine the minimum transmission power of each mobile device and drone offloading pair, which is then applied to task offloading. A CPOO algorithm is then proposed to minimize the total inference delay;

[0016] (5) In order to explore the impact of different transmission powers on the reliability probability and total inference delay, a non-convex problem that simultaneously meets the requirements of reliability and delay sensitivity is constructed, and the transmission power and task division offloading decisions are iteratively adjusted and optimized through the LaCPOO algorithm.

[0017] Furthermore, the UAV cluster described in step (1) flies over the area covered by the MEC system, and each UAV is connected to an edge server. The BS, as a cloud server, has powerful computing capabilities. The MD set and the UAV cluster are represented as The entire task scheduling time is divided into T equal-length time slots, each of which is indexed by t; each MD m generates a DNN-based task J in time slot t. m (t), J m (t) can be divided into N m subtasks; let the i-th subtask of MD m be Subtasks can be offloaded to UAVs for collaborative computing. Note that offloading may fail in dynamic networks, and tasks can be sent to BSs for reprocessing. For ease of illustration, a Cartesian coordinate system is constructed, and it is assumed that all MDs and BSs are at zero altitude, while UAVs fly at a fixed altitude H to save energy. Then, at time slot t, the horizontal coordinates of MDs and UAVs are denoted as q m (t) = (x m (t), y m (t)) and q u (t) = (x u (t), y u (t)). The position of BS is denoted as q B =(x B ,y B ).

[0018] Furthermore, step (2) also includes the following calculation process:

[0019] (21) “Transmission reliability” is defined as the probability that the probability of successful transmission exceeds a given reliability threshold. Its mathematical expression is:

[0020] Pr(γ mu (t)≥γ 0 )≥1-ε

[0021] (22) Obtain the signal-to-noise ratio γ of the task offloaded to UAV u at time slot tMDm mu(t), which is calculated as:

[0022]

[0023] where g 0 represents the path loss at a reference distance of 1 meter, represents the fading envelope, N 0 represents the noise power;

[0024] (23) The transmission rate between MD m and UAV u is obtained as:

[0025] R mu (t) = b 0 ·log 2 (1+γ mu (t))

[0026] where b 0 Indicates the bandwidth of transmitted data in orthogonal channels.

[0027] Furthermore, step (3) also includes the following process:

[0028] (31) The division of DNN tasks is defined as follows:

[0029] First, let K be the set of DNN types and denote the number of layers in the feature extraction stage (i.e., convolutional layers and pooling layers) of type-k DNN as L k For convolution / pooling layer l k (1≤l k ≤L k ), whose characteristic is a tuple in It is layer l k The input data size is It is layer l k The number of floating-point operations, where type∈{conv,pool}, can be calculated as:

[0030]

[0031] in Corresponding to layer l k The input height, width and channel values; accordingly, the lth k The output matrix of the layer is associated; where σ is the memory usage of unit data, is the kernel size. Note that MDs do further local processing on the classification / detection stage (i.e., fully connected layers); therefore, the corresponding models are not considered;

[0032] Assume that MD partitions the input data along the height. If it is determined that UAV should partition the last layer L k Which parts of the output lines are sent back to MD, then the AIR algorithm with arbitrary input range is used to trace back all the corresponding input lines of the previous layer, so that each segment of the workload of the previous layer is processed as a separate task.

[0033] On this basis, the partition decision is recorded as That is, the i-th subtask of MDm in time slot t is arrive The output line (i.e. );

[0034] (32) The unloading model between UAV and MD is established as follows:

[0035] When MD m will task When offloading to UAV u, the offloading delay between MD m and UAV u and the computation delay of UAV can be calculated as:

[0036]

[0037] Among them, |, ·| represents the corresponding number of rows, that is, and Represent the input height and data size of the first layer respectively; represents the required computing resources;

[0038] On this basis, once UAV u completes the assigned task, it attempts to send the result to MD m. Since the transmission power of UAV is much greater than that of MD and the result is generally small, the corresponding delay is ignored and the subtask at UAV u is The total inference delay can be expressed as:

[0039]

[0040] (33) Establish an offloading model between BS and MD, as follows:

[0041] Due to the unreliable channel conditions between UAV and MD, it may happen that MD m cannot receive all the returned data from UAVs. In this case, MD will send the entire task to BS for reprocessing. Since the transmission from MD to UAV has taken into account the worst-case channel fading, reliability is guaranteed to a large extent. Therefore, retransmission is a low-probability event. Without considering the transmission power optimization problem of offloading to BS, MD uses maximum power Transmitted to BS. Then, the transmission rate can be calculated as:

[0042]

[0043] where b B and denote the bandwidth and channel gain between MD m and BS respectively. In addition, since BS has sufficient computing resources and strong transmission power, the delay of calculation and return can be ignored. Based on this, the delay calculation formula between UAV u and BS is:

[0044]

[0045] (34) Based on and D mB (t), the total delay to complete the task of MD m in time slot t is:

[0046]

[0047] Among them A m (t) is the set of unloading UAVs of MD m, and the set of UAVs that successfully return the result to MD m is recorded as

[0048] Furthermore, the specific calculation process for determining the minimum transmission power of each mobile device and drone unloading pair using the risk theory method in step (4) is as follows:

[0049] (s41) For a given loss function The probability distribution on and tolerance ε∈(0,1), for a random variable At the level ε, The CVaR is defined as:

[0050]

[0051] in represents the set of all real numbers. For a real number a, (a) + =max(a,0);

[0052] (s42) For For a continuous loss function that is a concave function or a quadratic function, the following equivalence holds:

[0053]

[0054] (s43) Let the loss function as follows:

[0055]

[0056] At the level ε, The feasible set of CVaR can be written as:

[0057]

[0058] where tr(·) represents the matrix trace, and

[0059]

[0060] (s44) To meet the communication requirements of reliability, the minimum transmission power required for MD m to successfully transmit the task to UAV u under the worst channel conditions is derived, and the calculation formula is:

[0061]

[0062] in β={β mu (t)}.

[0063] In addition, the solution to the problem of minimizing the total inference delay in step (4) includes the following six stages:

[0064] (s41) Initialize position: use P, v p denote the size of the bat population and the speed of the pth bat respectively; for bat p, (Par p , x p )=(H, X) represents the position, and Represent the initial volume and initial pulse occurrence rate respectively. The total number of iterations is denoted by W; In this stage, the initial position of each bat is generated by a greedy method; First, for each task, the partition of the last convolutional layer is randomly initialized, and the transmission data size of each subtask and the computational workload of each layer are obtained by using the partition decision of the AIR algorithm; Then, each subtask is sorted in descending order according to the computational workload of each MD; For each subtask of each MD, the offloading delay of all reachable UAVs is first calculated, and then the subtask with only one reachable UAV is first offloaded; Next, the UAVs are assigned to the subtasks of the remaining MDs in descending order of the minimum delay, and the assignment order is random; At the same time, if a UAV reaches its service limit, the task is offloaded to the UAV with the second lowest service delay until all the first subtasks of the MDs are offloaded; Then, start offloading the second subtasks of all MDs until all subtasks of the MD are offloaded; Finally, the partitioning and offloading decisions are obtained;

[0065] (s42) Update speed and position: Update the speed v of bat p p Set as the number of different partitioning decisions of the current global best bat; update the position of bat p by exchanging the two ranges of partitioning decisions, and the number of exchanges is equal to the speed; then, obtain the unloading decision through the above greedy strategy;

[0066] (s43) Update the local optimal bat: compare the pulse emission rate r and the random value rand. If rand is greater than r, try to generate a new solution using the same method as above; if rand is less than A and its value is less than the optimal solution, update it to the latest optimal solution;

[0067] (s44) Update volume A and pulse emission rate r: Update Where γ and α represent the pulse emission rate increment coefficient and volume attenuation coefficient respectively; where w is the current iteration number;

[0068] (s45) Update the global optimal bat stage: sort all bats, update the global optimal bat p and its position (Par p , x p );

[0069] (s46) Iterative search phase: Repeat the above steps (s42) to (s46) until the number of iterations reaches W.

[0070] Furthermore, the specific process of constructing a non-convex problem that satisfies both reliability and delay sensitivity requirements in step (5) is as follows:

[0071] The transmit power allocation subproblem calculates the minimum transmit power to ensure the reliability of communication. The CDPO problem minimizes the total delay by partitioning the input data of the task and unloading it to the drone according to the minimum transmit power. If a higher transmit power is used to unload the subtask, the partitioning and unloading decisions and the total delay will change accordingly. In order to explore the relationship between different transmission powers on the reliability probability and the total reasoning delay, a weight coefficient is assigned to each objective. Then, the constraints and objective functions of the two subproblems are combined into the following main problem:

[0072]

[0073] where ψ 1 and ψ 2 are the weight parameters of total delay and transmission power, respectively, and ψ 1 +ψ 2 =1.

[0074] In addition, the specific process of iteratively adjusting and optimizing the transmission power and task partitioning offloading decision by the LaCPOO algorithm in step (5) is as follows:

[0075] (s51) Obtain the Lagrangian relaxation of the main problem, which is expressed as:

[0076]

[0077] where λ = {λ 1 ,λ2 ,λ 3 ,λ 4 ,λ 5 ,λ 6}, And for 1≤t≤T, there is

[0078]

[0079] (s52) Define the Lagrangian dual function as:

[0080]

[0081] (s53) Solve the Lagrangian dual problem by maximizing g(λ) subject to the domain constraints of the dual variables:

[0082]

[0083] (s54) The LaCPOO algorithm is proposed to solve the dual problem. The algorithm finds a feasible solution by iteratively combining subgradient projection with the CPOO method. In each iteration, the Lagrange multiplier is updated with a fixed step size, and the transmission power is adjusted according to the KKT condition. After updating the transmission power, the CPOO algorithm solves the CDPO problem, and the solution is used to update the multiplier in the next iteration. Specifically, it includes the following two processes:

[0084] Internal minimization solution: By about Taking the derivative of β,χ and setting them equal to zero, we get:

[0085]

[0086] The transmission power can be calculated from the above equation: Once the transmission power is updated, the CDPO problem can be solved using the proposed CPOO algorithm, which enables the LaCPOO algorithm to obtain a feasible solution to the main problem;

[0087] External maximization solution: Given the subgradient of g(λ) is:

[0088]

[0089] where p mu (t), M mu (t), β mu (t) is the solution to the inner minimization problem in the current iteration; for real variables π, [π] + =max(π,0); for real matrix π, [π] +=max(π,0).

[0090] The algorithm first initializes the power allocation Auxiliary variables β and Lagrange multipliers; then the CPOO algorithm is used to calculate the unloading decision x and the partitioning variable In each iteration, the Lagrange multipliers are updated by the subgradient method and then recalculated using the CPOO algorithm β, and adjust x and This process is repeated until it converges to a predefined threshold. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 It is a schematic diagram of the structure of the system constructed by the present invention;

[0092] Figure 2 It is a schematic diagram of the DNN partitioning method;

[0093] Figure 3 This is a comparison chart of the average latency of different algorithms under different DNN model types;

[0094] Figure 4 This is a comparison chart of the average transmission power of different algorithms under different DNN model types;

[0095] Figure 5 This is a comparison chart of average delays of different algorithms at different bandwidths;

[0096] Figure 6 is the effect of different algorithms on average delay at different offloading reliability thresholds;

[0097] Figure 7 is the effect of different algorithms on the average transmit power at different offloading reliability thresholds;

[0098] Figure 8 is the weight parameter ψ 1 Impact diagram on average delay and average transmit power. DETAILED DESCRIPTION

[0099] In order to explain the technical solution disclosed in the present invention in detail, further elaboration is given below in conjunction with the embodiments and the accompanying drawings of the specification.

[0100] In recent years, artificial intelligence (AI) applications based on deep neural networks (DNNs) have received extensive attention in mobile edge computing (MEC). However, ensuring reliable transmission and meeting delay-sensitive requirements in dynamic networks remain challenging. In this paper, a multi-UAV-assisted MEC architecture is designed, in which multiple UAVs provide computational offloading services for mobile devices (MDs). In order to ensure reliable communication in highly dynamic scenarios, a robust transmission problem is proposed, and the probability of successful transmission under the worst-case channel conditions is analyzed using the conditional value-at-risk method. On this basis, in order to accelerate the reasoning of DNNs, a collaborative DNN partitioning and offloading (CDPO) problem is constructed by proposing a collaborative DNN partitioning and offloading scenario, which is NP-hard. Furthermore, an efficient collaborative partitioning offloading optimization (CPOO) algorithm is proposed based on the bat algorithm to solve the CDPO problem. In addition, considering the requirements of reliability guarantee and delay sensitivity, a non-convex optimization problem is proposed as the main problem, and the LaCPOO algorithm, which combines the subgradient algorithm and the CPOO algorithm, is proposed. Simulation results show that compared with other algorithms, the proposed method can accelerate DNN reasoning while ensuring reliability.

[0101] Furthermore, the reliability-oriented DNN collaborative reasoning and offloading architecture proposed in the present invention allows multiple UAVs to assist MDs that do not have sufficient computing power to complete complex DNN model reasoning tasks. In order to ensure reliable communication in dynamic offloading networks, a robust transmission problem is established, and the conditional risk value method is applied to calculate the minimum transmission power of MD. In order to accelerate DNN reasoning, an NP-hard collaborative DNN partitioning and offloading CDPO problem is constructed, which aims to minimize the total task delay by jointly optimizing offloading decisions and DNN partitioning in dynamic networks. Furthermore, in order to explore the impact of transmission power on the total reasoning delay of the task, a master problem that considers both transmission power and delay is constructed.

[0102] (1) Build a multi-UAV-assisted MEC system, which dynamically networks a group of mobile devices MDs with offload task requirements, a group of UAVs equipped with MEC servers, and a base station BS with strong computing power.

[0103] like Figure 1 As shown in the figure, the proposed multi-UAV MEC system consists of a group of MDs A group of drones equipped with MEC servers are used as edge servers, denoted as And a cloud server with strong computing power as a base station BS. In addition, the entire task scheduling time is divided into T time slots of equal length, each time slot is indexed by t. Each MD m generates a DNN-based task J in time slot t m (t), Jm (t) can be divided into N m Subtasks. The i-th subtask of MD m is Subtasks can be offloaded to UAVs for collaborative computing. Note that offloading may fail in dynamic networks, and tasks can be sent to BSs for reprocessing. For ease of illustration, a Cartesian coordinate system is constructed and it is assumed that all MDs and BSs are at zero altitude, while UAVs fly at a fixed altitude H to save energy. Then, at time slot t, the horizontal coordinates of MDs and UAVs are denoted as q m (t) = (x m (t), y m (t)) and q u (t) = (x u (t), y u (t)). The position of BS is denoted as q B =(x B ,y B ).

[0104] (2) Based on the existence of unknown obstacles between MD and UAV, the robustness of transmission reliability in a changing channel fading environment is considered, a reliability-guaranteed communication model is established, and the transmission power optimization problem and constraints are determined.

[0105] At time slot t, if MD m decides to offload any subtask to UAV u. Assume that data is transmitted in orthogonal channels and denote the bandwidth as b 0 The transmission rate between MD m and UAV u can be calculated by the following formula:

[0106] R mu (t) = b 0 ·log 2 (1+γ mu (t))

[0107] where γ mu (t) is the received signal-to-noise ratio (SNR) when UAV u performs MD m task in time slot t, which can be obtained by the following calculation:

[0108]

[0109] where g 0 is the path loss at a reference distance of 1 meter, p mu (t) is the allocated transmission power between MD m and UAV u, The fading envelope.

[0110] Since there are unknown obstacles between MD and UAV, the robustness of transmission reliability in a changing channel fading environment is considered. It is assumed that the channel fading obeys a distribution with a mean of μ and a variance of ∑. The distribution function set is expressed as:

[0111]

[0112] A successful transmission requires that the received signal-to-noise ratio is higher than a certain threshold γ 0 Given Channel fading distribution The reliability of transmission is measured by the probability that the probability of successful transmission exceeds the reliability threshold, so the reliability of successful transmission between MD m and UAV u is:

[0113]

[0114] Where 1-ε represents the reliability threshold, and ε is the maximum tolerable transmission error rate.

[0115] To ensure reliable communication under the worst channel fading conditions, the allocated transmit power should be sufficient to offload the task to the corresponding UAV u with rapidly changing channel gain. For task offloading from MD m to UAV u, the maximum transmit power of MD m is assumed to be The minimum transmission power p will be calculated mu The problem statement for (t) is as follows:

[0116]

[0117] (3) A DNN partitioning and offloading model is established with the goal of minimizing the total inference delay of all MDs for the entire scheduling time T. A collaborative DNN partitioning and offloading problem, namely the CDPO problem, is designed.

[0118] First, let K be the set of DNN types and denote the number of layers in the feature extraction stage (i.e., convolutional layers and pooling layers) of type-k DNN as L k For convolution / pooling layer l k (1≤l k ≤L k ), whose characteristic is a tuple in It is layer l k The input data size is It is layer l k The number of floating-point operations, where type∈{conv,pool}, can be calculated as:

[0119]

[0120] in Corresponding to layer l k The input height, width and channel values ​​of . Accordingly, the lth k The output matrix of the layer is The port is the memory usage of the unit data. is the kernel size. Note that MDs do further local processing on the classification / detection stage (i.e., fully connected layers). Therefore, the corresponding models are not considered.

[0121] Assume that MD partitions the input data along the height. If it is determined that the drone should k Which parts of the output lines are sent back to MD, then the AIR algorithm with arbitrary input range is used to trace back all the corresponding input lines of the previous layer, so that each segment of the workload of the previous layer is processed as a separate task, such as Figure 2 shown.

[0122] On this basis, consider the i-th subtask of MD m in time slot t as Let the partition decision be That is, the i-th subtask of MD m in time slot t is arrive The output line (i.e. Record the uninstall variable as This means that MD m will perform subtask Uninstall to Then, the unloading model is expressed as follows:

[0123] 1) Offloading to UAV: ​​When MD m transfers the task When offloading to UAV u, the offloading delay between MD m and UAV u can be calculated as:

[0124]

[0125] Among them, |, ·| represents the corresponding number of rows, that is, and Represent the input height and data size of the first layer respectively.

[0126] After receiving the input data of MD m, UAV u starts local calculation. The delay of UAV u calculating the MD m task in time slot t is:

[0127]

[0128] in represents the required computing resources. Once UAV u completes the assigned task, it attempts to send the result to MD m. Since the transmission power of the UAV is much greater than that of the MD and the result is generally small, the corresponding delay is ignored. On this basis, UAV u performs the subtask The total inference delay can be expressed as:

[0129]

[0130] 2) Offloading to BS: Due to the unreliable channel conditions between the UAV and MD, it may happen that MD m cannot receive all the returned data from the UAV. In this case, MD will send the entire task to BS for reprocessing. Since the transmission from mobile device to UAV has taken into account the worst-case channel fading, reliability is guaranteed to a large extent. Therefore, retransmission is a low-probability event. Without considering the transmission power optimization problem to BS, MD uses maximum power. Transmitted to BS. Then, the transmission rate can be calculated as:

[0131]

[0132] where b B and denote the bandwidth and channel gain between MD m and BS, respectively. In addition, since BS has sufficient computing resources and strong transmission power, the delay of calculation and return can be ignored. Then, the delay between UAV u and BS can be calculated as:

[0133]

[0134] Let Λ m (t) is the set of unloading drones of MD m, and the set of drones that successfully return the results to MD m is recorded as Therefore, the total delay to complete the task of MD m in time slot t is:

[0135]

[0136] Based on the above analysis, given a set of MDs and UAVs collection For 1≤t≤T, make and Assume that all tasks are offloaded with the minimum transmission power calculated in P0. With the goal of minimizing the total inference delay of all MDs for the entire scheduling time T, the collaborative DNN partitioning and offloading (CDPO) problem can be expressed as follows, with the following constraints: C1 is the constraint on the binary offloading variable; C2 ensures that any subtask It can be unloaded to one UAV at most; C3 means that a UAV can only handle Z subtasks at most at a time; C4 ensures effective partitioning; C5 means that the partitioned rows do not intersect each other; C6 indicates that the sum of the number of rows of all subtasks must be equal to the total number of rows C7 Constraint The horizontal distance between MD and UAV cannot exceed the coverage range R of UAV.

[0137]

[0138] (4) A risk theory method is proposed to determine the minimum transmission power of each mobile device and drone offloading pair, which is then applied to task offloading; a CPOO algorithm is then proposed to minimize the total inference delay.

[0139] (41) Risk Theory Method

[0140] For the above problem P0, the goal is to minimize the transmission power between MD m and UAV u while ensuring reliability. However, since the chance constraint is usually non-convex, P0 is difficult to solve without a specific assumption on the fading distribution. Then, a tractable reformulation of P0 is derived using the conditional value at risk (CVaR). First, for a given loss function The probability distribution on and tolerance ε∈(0,1), for a random variable At the level ε, The CVaR is defined as:

[0141]

[0142] in represents the set of all real numbers. For a real number a, (a) + =max(a,0). In addition, for For a continuous loss function that is a concave function or a quadratic function, the following equivalence holds:

[0143]

[0144] Assuming the loss function as follows:

[0145]

[0146] At the level ε, The feasible set of CVaR can be written as:

[0147]

[0148] where tr(·) represents the matrix trace, and

[0149]

[0150] So, suppose β={β mu (t)}, the problem P0 can be transformed into the following cone optimization problem:

[0151]

[0152] (42) Solution to the CDPO problem

[0153] Aiming at the CDPO problem (i.e. P1), inspired by the bat algorithm, the CPOO algorithm is proposed. The bat algorithm simulates the behavior of micro-bats in nature using echolocation technology to hunt and avoid obstacles, and is suitable for solving combinatorial optimization problems. The main steps of the algorithm are as follows:

[0154] 1) Initialize position: use P, v p denote the size of the bat population and the speed of the pth bat, respectively. For bat p, (Par p , x p )=(H, X) represents the position, and Denote the initial volume and initial pulse occurrence rate, respectively. The total number of iterations is denoted by W. In this stage, the initial position of each bat is generated by a greedy method. First, for each task, the partition of the last convolutional layer is randomly initialized, and the transmission data size of each subtask and the computational workload of each layer are obtained by using the partition decision of the AIR algorithm. Then, each subtask is sorted in descending order according to the computational workload of each MD. For each subtask of each MD, the offloading delay of all reachable UAVs is first calculated, and then the subtask with only one reachable UAV is offloaded first. Next, the UAVs are assigned to the subtasks of the remaining MDs in descending order of the minimum delay, and the assignment order is random. At the same time, if a UAV reaches its service limit, the task is offloaded to the UAV with the second lowest service delay until all the first subtasks of the MDs are offloaded. Then, start offloading the second subtasks of all MDs until all subtasks of the MD are offloaded. Finally, the partitioning and offloading decisions are obtained.

[0155] 2) Update speed and position: Update the speed v of bat p p Let be the number of different partitioning decisions of the current global best bat.,The position of bat p is updated by exchanging the two ranges of partitioning,decisions with the number of exchanges being equal to the speed.,Then, the unloading decision is obtained by the above,greedy strategy.

[0156] 3) Update the local optimal bat: Compare the pulse emission rate r and the random value rand. If rand is greater than r, try to generate a new solution using the same method as above. If rand is less than A and its value is less than the optimal solution, update it to the latest optimal solution.

[0157] 4) Update volume A and pulse emission rate r: Update Where γ and α represent the pulse emission rate increment coefficient and volume attenuation coefficient respectively. Where w is the current iteration number.

[0158] 5) Update the global optimal bat: Sort all bats and update the global optimal bat p and its position (Par p , x p ).

[0159] 6) Iterative search: Repeat steps 2) to 5) above until the number of iterations reaches W.

[0160] (5) In order to explore the impact of different transmission powers on the reliability probability and total inference delay, a non-convex problem that simultaneously meets the requirements of reliability and delay sensitivity is constructed, and the transmission power and task division offloading decisions are iteratively adjusted and optimized through the LaCPOO algorithm.

[0161] As mentioned before, the first sub-problem calculates the minimum transmit power to ensure the reliability of communication, and the second sub-problem minimizes the total delay by partitioning the input data of the task and offloading it to the UAV with the minimum transmit power. If the sub-task is offloaded with a higher transmit power, the partitioning and offloading decisions as well as the total delay will change accordingly. In order to explore the relationship between different transmission powers on the reliability probability and the total inference delay, a weight coefficient is assigned to each objective. Then, the constraints and objective functions of the two sub-problems are combined into the following main problem:

[0162]

[0163] where ψ 1 and ψ 2 are the weight parameters of total delay and transmission power, respectively, and ψ 1 +ψ 2 =1.

[0164] Since the objective function is a coupling of different variables, this is a non-convex optimization problem. The sub-gradient descent algorithm is used to update the transmission power. The Lagrangian relaxation of this problem is:

[0165]

[0166] where λ = {λ 1 ,λ 2 ,λ 3 ,λ4 ,λ 5 ,λ 6}, And for 1≤t≤T, there is

[0167]

[0168] The Lagrange dual function is:

[0169]

[0170] The Lagrangian dual problem is solved by maximizing g(λ) subject to the domain constraints of the dual variables:

[0171]

[0172] Then, the LaCPOO algorithm is proposed to solve the dual problem. The algorithm finds a feasible solution by iteratively combining subgradient projection with the CPOO method. During each iteration, the Lagrange multipliers are updated with a fixed step size, while the transmission power is adjusted according to the KKT condition. After updating the transmission power, the CPOO algorithm solves the secondary problem and the solution is used to update the multipliers in the next iteration.

[0173] 1) Internal minimization solution: By about Taking the derivative of β,χ and setting them equal to zero, we get:

[0174]

[0175] Transmission power It can be calculated by the above formula. Once the transmission power is updated, the CDPO problem can be solved using the proposed CPOO algorithm, which enables the LaCPOO algorithm to obtain a feasible solution to the main problem.

[0176] 2) External maximization solution: Given the subgradient of g(λ) is:

[0177]

[0178] where p mu (t), M mu (t), β mu (t) is the solution to the inner minimization problem in the current iteration. For real variables π, [π] + =max(π,0); for real matrix π, [π] + =max(π,0).

[0179] The algorithm first initializes the power allocation Auxiliary variables β and Lagrange multipliers. Then the CPOO algorithm is used to calculate the unloading decision x and the partitioning variable In each iteration, the Lagrange multipliers are updated by the subgradient method and then recalculated using the CPOO algorithm β, and adjust x and This process is repeated until it converges to a predefined threshold.

[0180] Finally, this embodiment establishes a simulation model and implements the above algorithm based on the constructed scenario. The default settings are introduced as follows. MDs are randomly distributed in the network area of ​​400m×400m, and a drone cluster consisting of 4 drones is used as an edge server. The drone flies at a fixed altitude H of 100m. The flight speed of the drone in the clockwise direction is 10m / s. The initial position is set to q 1 =(150,150)m,q 2 =(150,250)m,q 3 =(250,150)m,q 4 =(250, 250)m. The maximum coverage radius R of the drone is set to 200 meters. The base station is deployed at (200, 200)m. Bandwidth b 0 and b B Set to 10MHz and 15MHz respectively. Background noise N 0 The maximum transmission power of MD is 2mW, and the maximum computing power of MD and UAV is set to 1, 4~8GHz respectively. The processing density is set to 8 double-precision FLOPS / cycle. In addition, the population size and flight frequency of bats are set to 20 and 50 respectively. The reliability threshold is set to 0.9, assuming that the channel fading follows a distribution with a mean of 2 and a variance of 0.4. By default, the classic DNN model, AlexNet, is used as the DNN-based task.

[0181] For performance comparison, the following 3 baselines were implemented:

[0182] 1. Local: Process all DNN inferences locally;

[0183] 2. RPOM: randomly divide the last convolutional layer and randomly unload subtasks;

[0184] 3.Edgent-P: Each MD selects the nearest UAV and then divides the DNN-based task into two parts according to the computing resources. The former is offloaded to the UAV and the latter is processed locally.

[0185] In order to evaluate the performance under different DNN types, famous DNN models, TinyYOLOv2, NiN and VGG16 are also selected for comparison. Figure 3 The average latency for different DNN model types is shown. As shown in the figure, Local leads to the highest inference latency due to limited computing power. For most classic DNN models, the proposed LaCPOO algorithm provides the lowest average latency even in uncertain environments, demonstrating the versatility of LaCPOO.

[0186] The average transmit power under different DNN model types is as follows: Figure 4 As shown, it can be seen that in Edgent-P and PROM, the average transmission power remains almost unchanged, indicating that the adaptability of these two algorithms is poor. Compared with CPOO, LaCPOO greatly reduces the transmission power, which shows that the proposed algorithm can ensure the minimum energy consumption while maintaining reliability compared with other algorithms in different models.

[0187] Figure 5 The average delay under different bandwidths is depicted. As shown in the figure, all algorithms gradually reduce the average delay as the transmission bandwidth increases. It is observed that the performance of PROM is greatly affected by the bandwidth. This is because the randomness of partitioning and offloading decisions causes the transmission delay to account for a large proportion of the total delay. Specifically, the LaCPOO algorithm always produces the lowest average delay through cooperative partitioning and offloading.

[0188] Figure 6 and 7 The performance under different offloading reliability thresholds is evaluated. Figure 6 As shown in , as the reliability threshold increases, the minimum transmission power increases to ensure reliable offloading. And among all algorithms, the proposed LaCPOO has a lower average transmission power, which shows that it can reduce the energy consumption of MD while ensuring reliable communication. Figure 7 As shown in Figure 2, as the reliability threshold increases, the average delay decreases. This is mainly because a larger reliability threshold requires a larger transmit power, which increases the transmission rate, so the average delay also decreases accordingly.

[0189] Figure 8 Given different ψ 1 The average delay and transmission power under the setting. 1 A larger value of represents a relatively larger weight of the delay metric in the objective function. 1 As ψ increases gradually to 1, the delay metric gradually dominates the value of the objective function, resulting in a gradual decrease in the average delay. In addition, the average transmit power increases as its proportion in the objective function decreases. 1When set to 0.8, the LaCPOO algorithm is suitable for balancing delay and transmission power.

Claims

1. A reliable multi-UAV assisted DNN task partitioning and offloading method in MEC system, characterized in that: The steps include: (1) Construct a multi-UAV-assisted MEC system, dynamically networking a group of mobile devices with offload task requirements, a group of UAVs equipped with MEC servers, and a base station BS with strong computing power; (2) Based on the existence of unknown obstacles between MD and UAV, the robustness of transmission reliability in a changing channel fading environment is considered, a reliability-guaranteed communication model is established, and the transmission power optimization problem and constraints are determined. The specific mathematical expression is as follows: Among them, p mu (t) represents the transmission power allocated by MD m to offload the task to UAV u; (0a) represents the distribution function set with mean μ and variance ∑ for channel fading; (0b) represents the constraint condition to ensure transmission reliability, where γ mu (t) represents the signal-to-noise ratio, γ0 represents the target signal-to-noise ratio, 1-ε represents the reliability threshold, and ε is the maximum tolerable transmission error rate; (0c) indicates that the transmission power allocated to the mobile device cannot exceed the power threshold; (3) A DNN partitioning and offloading model is established. With the goal of minimizing the total inference delay of all MDs for the entire scheduling time T, a collaborative DNN partitioning and offloading problem, namely the CDPO problem, is designed. The specific mathematical expression is as follows: in like It means that MD m will perform subtask Offload to UAV u, It means that the i-th subtask of MDm in time slot t is arrive The output line of; C1 is the constraint of binary offloading variables; C2 ensures that any MD subtask It can be unloaded to at most one UAV: ​​C3 means that a UAV can only process at most Z subtasks at a time; C4 ensures effective partitioning; C5 means that the partitioned rows do not intersect each other; C6 indicates that the sum of the number of rows of all subtasks must be equal to the total number of rows C7 Constraint The horizontal distance between MD and UAV cannot exceed the coverage range R of UAV; (4) A risk theory approach is proposed to determine the minimum transmission power of each mobile device and drone offloading pair, which is then applied to task offloading; a CPOO algorithm is then proposed to minimize the total inference delay; (5) In order to explore the impact of different transmission powers on the reliability probability and total inference delay, a non-convex problem that simultaneously meets the requirements of reliability and delay sensitivity is constructed, and the transmission power and task division offloading decisions are iteratively adjusted and optimized through the LaCPOO algorithm.

2. The reliable multi-UAV assisted DNN task partitioning and offloading method in the MEC system according to claim 1 is characterized in that: In step (1), the UAV cluster flies over the area covered by the MEC system, and each UAV is connected to an edge server. The BS, as a cloud server, has powerful computing capabilities. The MD set and the UAV cluster are represented as The entire task scheduling time is divided into T equal-length time slots, each of which is indexed by t; each MD m generates a DNN-based task J in time slot t. m (t), J m (t) can be divided into N m subtasks; let the i-th subtask of MD m be Subtasks can be offloaded to UAVs for collaborative computing. Note that offloading may fail in dynamic networks, and tasks can be sent to BSs for reprocessing. For ease of illustration, a Cartesian coordinate system is constructed, and it is assumed that all MDs and BSs are at zero altitude, while UAVs fly at a fixed altitude H to save energy. Then, at time slot t, the horizontal coordinates of MDs and UAVs are denoted as q m (t) = (x m (t), y m (t)) and q u (t) = (x u (t), y u (t)); the position of BS is denoted as q B =(x B ,y B ).

3. The reliable multi-UAV assisted DNN task partitioning and offloading method in the MEC system according to claim 1 is characterized in that: Step (2) also includes the following calculation process: (21) “Transmission reliability” is defined as the probability that the probability of successful transmission exceeds a given reliability threshold. Its mathematical expression is: Pr(γ mu (t)≥γ0)≥1-ε (22) Obtain the signal-to-noise ratio γ of the task offloaded to UAV u at time slot tMDm mu (t), which is calculated as: Where g0 represents the path loss at a reference distance of 1 meter, represents the fading envelope, and N0 represents the noise power; (23) The transmission rate between MD m and UAV u is obtained as: R mu (t)=b0·log2(1+γ mu (t)) Where b0 represents the bandwidth of the transmitted data in the orthogonal channel.

4. The reliable multi-UAV assisted DNN task partitioning and offloading method in the MEC system according to claim 1 is characterized in that: Step (3) also includes the following process: (31) The division of DNN tasks is defined as follows: First, let K be the set of DNN types, and denote the total number of convolutional and pooling layers of type-k DNN as L k ; For convolution / pooling layer l k (1≤l k ≤L k ), whose characteristic is a tuple in It is layer l k The input data size is It is layer l k The number of floating-point operations, where type∈{conv,pool}, can be calculated: in Corresponding to layer l k The input height, width and channel values; accordingly, the lth k The output matrix of the layer is associated; where σ is the memory usage of unit data, is the kernel size; note that MDs will do further local processing on the fully connected layer; therefore, the corresponding model is not considered; Assuming that MD partitions the input data along the height, if it is determined that UAV should partition the last layer L k Which parts of the output rows are sent back to MD, then the arbitrary input range algorithm AIR algorithm is used to backtrack all the corresponding input rows of the previous layer, so that each segment of the previous layer's regional workload is processed as a separate task. On this basis, the partition decision is recorded as That is, the i-th subtask of MDm in time slot t is arrive The output line of (32) The unloading model between UAV and MD is established as follows: When MD m will task When offloading to UAV u, the offloading delay between MD m and UAV u and the computation delay of UAV can be calculated as: where |·,·| represents the corresponding number of rows, i.e. and Represent the input height and data size of the first layer respectively; represents the required computing resources; On this basis, UAV u sub-task The total inference delay can be expressed as: (33) The offloading model between BS and MD is established, and the delay calculation formula between UAV u and BS is: where b B and denote the bandwidth and channel gain between MD m and BS, respectively, and R mB (t) is the transmission rate, which is calculated as: (34) Based on and D mB (t), the total delay to complete the task of MD m in time slot t is: where Λ m (t) is the set of unloading UAVs of MD m, and the set of UAVs that successfully return the result to MD m is recorded as 5. The reliable multi-UAV assisted DNN task partitioning and offloading method in the MEC system according to claim 1 is characterized in that: The specific calculation process for determining the minimum transmission power of each mobile device and drone unloading pair using the risk theory method in step (4) is as follows: (s41) For a given loss function The probability distribution on and tolerance ε∈(0,1), for a random variable At the level ε, The CVaR is defined as: in represents the set of all real numbers. For a real number a, (a) + =max(a,0); (s42) For For a continuous loss function that is a concave function or a quadratic function, the following equivalence holds: (s43) Let the loss function as follows: At the level ε, The feasible set of CVaR can be written as: where tr(·) represents the matrix trace, and (s44) To meet the communication requirements of reliability, the minimum transmission power required for MD m to successfully transmit the task to UAV u under the worst channel conditions is derived, and the calculation formula is: among them β={β mu (t)}。 6. The reliable multi-UAV assisted DNN task partitioning and offloading method in the MEC system according to claim 1 is characterized in that: The solution to the problem of minimizing the total inference delay in step (4) includes the following six stages: (s41) Initialization position: user, v p denote the size of the bat population and the speed of the pth bat respectively; for bat p, (Par p , x p )=(H, X) represents the position, and denote the initial volume and initial pulse occurrence rate respectively, and the total number of iterations is denoted by W; in this stage, the initial position of each bat is generated by a greedy method; first, for each task, the partition of the last convolutional layer is randomly initialized, and the transmission data size of each subtask and the computational workload of each layer are obtained by using the partition decision of the AIR algorithm; then, each subtask is sorted in descending order according to the computational workload of each MD; for each subtask of each MD, the offloading delay of all reachable UAVs is first calculated, and then the subtask with only one reachable UAV is offloaded first; next, the UAVs are assigned to the subtasks of the remaining MDs in descending order of the minimum delay, and the assignment order is random; at the same time, if a UAV reaches its service limit, the task is offloaded to the UAV with the second lowest service delay until all the first subtasks of the MDs are offloaded; then, the second subtasks of all MDs are started to be offloaded until all the subtasks of the MD are offloaded; finally, the partitioning and offloading decisions are obtained; (s42) Update speed and position: Update the speed v of bat p p Set as the number of different partitioning decisions of the current global best bat; update the position of bat p by exchanging the two ranges of partitioning decisions, and the number of exchanges is equal to the speed; then, obtain the unloading decision through the above greedy strategy; (s43) Update the local optimal bat: compare the pulse emission rate r and the random value rand. If rand is greater than r, try to generate a new solution using the same method as above; if rand is less than A and its value is less than the optimal solution, update it to the latest optimal solution; (s44) Update volume A and pulse emission rate r: Update Where γ and α represent the pulse emission rate increment coefficient and volume attenuation coefficient respectively; where w is the current iteration number; (s45) Update the global optimal bat stage: sort all bats, update the global optimal bat p and its position (Par p , x p ); (s46) Iterative search phase: Repeat the above steps (s42) to (s45) until the number of iterations reaches W.

7. The reliable multi-UAV assisted DNN task partitioning and offloading method in the MEC system according to claim 1 is characterized in that: The specific process of constructing a non-convex problem that satisfies both reliability and delay sensitivity requirements in step (5) is as follows: The transmit power allocation subproblem calculates the minimum transmit power to ensure the reliability of communication. The CDPO problem minimizes the total delay by partitioning the input data of the task and unloading it to the drone according to the minimum transmit power. If a higher transmit power is used to unload the subtask, the partitioning and unloading decisions and the total delay will change accordingly. In order to explore the relationship between different transmission powers on the reliability probability and the total reasoning delay, a weight coefficient is assigned to each objective. Then, the constraints and objective functions of the two subproblems are combined into the following main problem: Where ψ1 and ψ2 are weight parameters of total delay and transmission power respectively, and ψ1+ψ2=1.

8. The reliable multi-UAV assisted DNN task partitioning and offloading method in the MEC system according to claim 1 is characterized in that: The specific process of iteratively adjusting and optimizing the transmission power and task partitioning offloading decision by the LaCPOO algorithm in step (5) is as follows: (s51) Obtain the Lagrangian relaxation of the master problem, which is expressed as: where λ = {λ 1 ,λ 2 ,λ 3 ,λ 4 ,λ 5 ,λ 6 }, And for 1≤t≤T, there is (s52) Define the Lagrangian dual function as: (s53) Solve the Lagrangian dual problem by maximizing g(λ) subject to the domain constraints of the dual variables: (s54) The LaCPOO algorithm is proposed to solve the dual problem; the algorithm finds a feasible solution by iteratively combining subgradient projection with the CPOO method; In each iteration, the Lagrange multiplier is updated with a fixed step size, and the transmission power is adjusted according to the KKT condition. After updating the transmission power, the CPOO algorithm solves the CDPO problem, and the solution is used to update the multiplier in the next iteration. Specifically, it includes the following two processes: Internal minimization solution: By about Taking the derivative of β, x and setting it equal to zero, we get The transmission power can be calculated from the above equation: Once the transmission power is updated, the CDPO problem can be solved using the proposed CPOO algorithm, which enables the LaCPOO algorithm to obtain a feasible solution to the main problem; External maximization solution: Given the subgradient of g(λ) is: where p mu (t), M mu (t), β mu (t) is the solution to the inner minimization problem in the current iteration; For a real variable π, [π] + =max(π,0); for real matrix π, [π] + =max(π,0); The algorithm first initializes the power allocation Auxiliary variables β and Lagrange multipliers; then the CPOO algorithm is used to calculate the unloading decision x and the partitioning variable In each iteration, the Lagrange multipliers are updated by the subgradient method and then recalculated using the CPOO algorithm β, and adjust x and This process is repeated until it converges to a predefined threshold.

Images